diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/README.md b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/README.md new file mode 100644 index 0000000000000000000000000000000000000000..9f95018b886ef13a991a20d2fbe17b5b89774f6b --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/README.md @@ -0,0 +1,130 @@ +## 一、准备运行环境 + + **表 1** 版本配套表 + + | 配套 | 版本 | 环境准备指导 | + | ----- | ----- |-----| + | Python | 3.10.2 | - | + | torch | 2.1.0 | - | + +### 1.1 获取CANN&MindIE安装包&环境准备 +- [800I A2](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=4&model=32) +- [Duo卡](https://www.hiascend.com/developer/download/community/result?module=pt+ie+cann&product=2&model=17) +- [环境准备指导](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/80RC2alpha002/softwareinst/instg/instg_0001.html) + +### 1.2 CANN安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构，{soc}表示昇腾AI处理器的版本。 +chmod +x ./Ascend-cann-toolkit_{version}_linux-{arch}.run +chmod +x ./Ascend-cann-kernels-{soc}_{version}_linux.run +# 校验软件包安装文件的一致性和完整性 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --check +./Ascend-cann-kernels-{soc}_{version}_linux.run --check +# 安装 +./Ascend-cann-toolkit_{version}_linux-{arch}.run --install +./Ascend-cann-kernels-{soc}_{version}_linux.run --install + +# 设置环境变量 +source /usr/local/Ascend/ascend-toolkit/set_env.sh +``` + + +### 1.3 MindIE安装 +```shell +# 增加软件包可执行权限，{version}表示软件版本号，{arch}表示CPU架构。 +chmod +x ./Ascend-mindie_${version}_linux-${arch}.run +./Ascend-mindie_${version}_linux-${arch}.run --check + +# 方式一：默认路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install +# 设置环境变量 +cd /usr/local/Ascend/mindie && source set_env.sh + +# 方式二：指定路径安装 +./Ascend-mindie_${version}_linux-${arch}.run --install-path=${AieInstallPath} +# 设置环境变量 +cd ${AieInstallPath}/mindie && source set_env.sh +``` + +### 1.4 Torch_npu安装 +安装pytorch框架 版本2.1.0 +[安装包下载](https://download.pytorch.org/whl/cpu/torch/) + +使用pip安装 +```shell +# {version}表示软件版本号，{arch}表示CPU架构。 +pip install torch-${version}-cp310-cp310-linux_${arch}.whl +``` +下载 pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +```shell +tar -xzvf pytorch_v{pytorchversion}_py{pythonversion}.tar.gz +# 解压后，会有whl包 +pip install torch_npu-{pytorchversion}.xxxx.{arch}.whl +``` +## 二、下载本仓库 + +### 2.1 下载到本地 +```shell +git clone https://gitee.com/ascend/ModelZoo-PyTorch.git +``` + +### 2.2 环境依赖安装 +```bash +pip3 install -r requirements.txt +``` + +## 三、Stable-Diffusion-v1.5 使用 + +### 3.1 权重及配置文件说明 +下载权重和配置文件 +```shell +# stable-diffusion-v1.5: +git clone https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5 + +# stable-diffusion-2.1: +git clone https://huggingface.co/stabilityai/stable-diffusion-2-1-base +``` + +### 3.2 修改配置文件 +将model_index.json中所有的`diffusers`字段修改为`stablediffusion` + +### 3.3 单卡功能测试 +设置权重路径 +```shell +# stable-diffusion-v1.5 (使用已下载的权重) +model_base='./stable-diffusion-v1-5' + +# stable-diffusion-v2.1 (使用已下载的权重) +model_base='./stable-diffusion-2-1-base' +``` +执行命令： +```shell +python3 inference_stablediffusion.py \ + --model ${model_base} \ + --prompt_file ./prompts/prompts.txt \ + --steps 50 \ + --save_dir ./results \ + --device 0 +``` +参数说明： +- --model：模型权重路径。 +- --prompt_file：提示词文件。 +- --steps: 图片生成迭代次数。 +- --save_dir：生成图片的存放目录。 +- --device：推理设备ID。 + + +### 3.4 模型推理性能 + +性能参考下列数据。 + +### StableDiffusion v1.5 +| 硬件形态 | 迭代次数 | 平均耗时| +| :------: |:----:|:----:| +| Atlas 800I A2 (32G) | 50 | 2.821s | + + +### StableDiffusion v2.1 +| 硬件形态 | 迭代次数 | 平均耗时| +| :------: |:----:|:----:| +| Atlas 800I A2 (32G) | 50 | 2.619s | \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/inference_stablediffusion.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/inference_stablediffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..6dbc957391cc02ce148d1bbf449ef4d071fca53b --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/inference_stablediffusion.py @@ -0,0 +1,93 @@ +import torch +import torch_npu +import time +import os +import argparse +from stablediffusion import StableDiffusionPipeline + + +def parse_arguments(): + parser = argparse.ArgumentParser() + parser.add_argument( + "--prompt_file", + type=str, + default="./prompts/prompts.txt", + help="The prompts file to guide images generation.", + ) + parser.add_argument( + "--negative_prompt", + type=str, + default="", + help="The prompt or prompts to guide what to not include in image generation.", + ) + parser.add_argument( + "--steps", + type=int, + default=50, + help="The number of denoising steps.", + ) + parser.add_argument( + "--model", + type=str, + default="./stable-diffusion-v1.5", + help="The path of stable-diffusion.", + ) + parser.add_argument( + "--device", + type=int, + default=0, + help="NPU device id.", + ) + parser.add_argument( + "--save_dir", + type=str, + default="./results", + help="Path to save result audio files.", + ) + return parser.parse_args() + + +def main(): + args = parse_arguments() + save_dir = args.save_dir + if not os.path.exists(save_dir): + os.makedirs(save_dir) + + torch_npu.npu.set_device(args.device) + torch.manual_seed(1) + npu_stream = torch_npu.npu.Stream() + + pipe = StableDiffusionPipeline.from_pretrained(args.model, torch_dtype=torch.float16) + pipe.to("npu") + + total_time = 0 + prompts_num = 0 + average_time = 0 + skip = 2 + with os.fdopen(os.open(args.prompt_file, os.O_RDONLY), "r") as f: + for i, prompt in enumerate(f): + with torch.no_grad(): + npu_stream.synchronize() + begin = time.time() + image = pipe( + prompt=prompt, + negative_prompt=args.negative_prompt, + num_inference_steps=args.steps, + ).images[0] + npu_stream.synchronize() + end = time.time() + if i > skip - 1: + total_time += end - begin + prompts_num = i+1 + image_save_path = os.path.join(save_dir, f"images_{i}.png") + image.save(image_save_path) + if prompts_num > skip: + average_time = total_time / (prompts_num-skip) + else: + raise ValueError("Infer average time skip first two prompts, ensure that prompts.txt \ + contains more than three prompts") + print(f"Infer average time: {average_time:.3f}s\n") + + +if __name__ == "__main__": + main() \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/prompts/prompts.txt b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/prompts/prompts.txt new file mode 100644 index 0000000000000000000000000000000000000000..a375a0bb63931d0d5da6c6d91df1e14f870f47d0 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/prompts/prompts.txt @@ -0,0 +1,16 @@ +Beautiful illustration of The ocean. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Islands in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Seaports in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The waves. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Grassland. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Wheat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Hut Tong. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The boat. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Pine trees. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Bamboo. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of The temple. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Cloud in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Sun in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Spring. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Lotus. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper +Beautiful illustration of Snow piles. in a serene landscape, magic realism, narrative realism, beautiful matte painting, heavenly lighting, retrowave, 4 k hd wallpaper \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/requirements.txt b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..6d39028a97778341c84978d9643b966f25895618 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/requirements.txt @@ -0,0 +1,5 @@ +setuptools==57.5.0 +torch==2.1.0 +diffusers==0.26.3 +transformers==4.46.0 +open_clip_torch==2.20.0 \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..529145f57160863a45c35608f5ffcb6fbbc23e99 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/__init__.py @@ -0,0 +1,5 @@ +from .pipeline import StableDiffusionPipeline, DiffusionPipeline +from .models import UNet2DConditionModel +from .models.model_utils import ModelMixin +from .schedulers import PNDMScheduler, SchedulerMixin +from .vae import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/activations.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/activations.py new file mode 100644 index 0000000000000000000000000000000000000000..c045238004962da70fad9a62b8b4a25f200a1e95 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/activations.py @@ -0,0 +1,123 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.models.lora import LoRACompatibleLinear + + +ACTIVATION_FUNCTIONS = { + "swish": nn.SiLU(), + "silu": nn.SiLU(), + "mish": nn.Mish(), + "gelu": nn.GELU(), + "relu": nn.ReLU(), +} + + +def get_activation(act_fn: str) -> nn.Module: + """Helper function to get activation function from string. + + Args: + act_fn (str): Name of activation function. + + Returns: + nn.Module: Activation function. + """ + + act_fn = act_fn.lower() + if act_fn in ACTIVATION_FUNCTIONS: + return ACTIVATION_FUNCTIONS[act_fn] + else: + raise ValueError(f"Unsupported activation function: {act_fn}") + + +class GELU(nn.Module): + r""" + GELU activation function with tanh approximation support with `approximate="tanh"`. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + approximate (`str`, *optional*, defaults to `"none"`): If `"tanh"`, use tanh approximation. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, approximate: str = "none", bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + self.approximate = approximate + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate, approximate=self.approximate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32), approximate=self.approximate).to(dtype=gate.dtype) + + def forward(self, hidden_states): + hidden_states = self.proj(hidden_states) + hidden_states = self.gelu(hidden_states) + return hidden_states + + +class GEGLU(nn.Module): + r""" + A [variant](https://arxiv.org/abs/2002.05202) of the gated linear unit activation function. + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + self.proj = linear_cls(dim_in, dim_out * 2, bias=bias) + + def gelu(self, gate: torch.Tensor) -> torch.Tensor: + if gate.device.type != "mps": + return F.gelu(gate) + # mps: gelu is not implemented for float16 + return F.gelu(gate.to(dtype=torch.float32)).to(dtype=gate.dtype) + + def forward(self, hidden_states, scale: float = 1.0): + args = () if USE_PEFT_BACKEND else (scale,) + hidden_states, gate = self.proj(hidden_states, *args).chunk(2, dim=-1) + return hidden_states * self.gelu(gate) + + +class ApproximateGELU(nn.Module): + r""" + The approximate form of the Gaussian Error Linear Unit (GELU). For more details, see section 2 of this + [paper](https://arxiv.org/abs/1606.08415). + + Parameters: + dim_in (`int`): The number of channels in the input. + dim_out (`int`): The number of channels in the output. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__(self, dim_in: int, dim_out: int, bias: bool = True): + super().__init__() + self.proj = nn.Linear(dim_in, dim_out, bias=bias) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.proj(x) + return x * torch.sigmoid(1.702 * x) diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention.py new file mode 100644 index 0000000000000000000000000000000000000000..9b2d570658f9af3e08ca74ffe65f9add1065f1a3 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention.py @@ -0,0 +1,676 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND +from diffusers.utils.torch_utils import maybe_allow_in_graph +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import GEGLU, GELU, ApproximateGELU +from .attention_processor import Attention +from .embeddings import SinusoidalPositionalEmbedding +from .normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm + + +def _chunked_feed_forward( + ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None +): + # "feed_forward_chunk_size" can be used to save memory + if hidden_states.shape[chunk_dim] % chunk_size != 0: + raise ValueError( + f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`." + ) + + num_chunks = hidden_states.shape[chunk_dim] // chunk_size + if lora_scale is None: + ff_output = torch.cat( + [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + else: + # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete + ff_output = torch.cat( + [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)], + dim=chunk_dim, + ) + + return ff_output + + +@maybe_allow_in_graph +class GatedSelfAttentionDense(nn.Module): + r""" + A gated self-attention dense layer that combines visual features and object features. + + Parameters: + query_dim (`int`): The number of channels in the query. + context_dim (`int`): The number of channels in the context. + n_heads (`int`): The number of heads to use for attention. + d_head (`int`): The number of channels in each head. + """ + + def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int): + super().__init__() + + # we need a linear projection since we need cat visual feature and obj feature + self.linear = nn.Linear(context_dim, query_dim) + + self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head) + self.ff = FeedForward(query_dim, activation_fn="geglu") + + self.norm1 = nn.LayerNorm(query_dim) + self.norm2 = nn.LayerNorm(query_dim) + + self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0))) + self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0))) + + self.enabled = True + + def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor: + if not self.enabled: + return x + + n_visual = x.shape[1] + objs = self.linear(objs) + + x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :] + x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x)) + + return x + + +@maybe_allow_in_graph +class BasicTransformerBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm (: + obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`. + attention_bias (: + obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter. + only_cross_attention (`bool`, *optional*): + Whether to use only cross-attention layers. In this case two cross attention layers are used. + double_self_attention (`bool`, *optional*): + Whether to use two self-attention layers. In this case no cross attention layers are used. + upcast_attention (`bool`, *optional*): + Whether to upcast the attention computation to float32. This is useful for mixed precision training. + norm_elementwise_affine (`bool`, *optional*, defaults to `True`): + Whether to use learnable elementwise affine parameters for normalization. + norm_type (`str`, *optional*, defaults to `"layer_norm"`): + The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`. + final_dropout (`bool` *optional*, defaults to False): + Whether to apply a final dropout after the last feed-forward layer. + attention_type (`str`, *optional*, defaults to `"default"`): + The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`. + positional_embeddings (`str`, *optional*, defaults to `None`): + The type of positional embeddings to apply to. + num_positional_embeddings (`int`, *optional*, defaults to `None`): + The maximum number of positional embeddings to apply. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + attention_bias: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_elementwise_affine: bool = True, + norm_type: str = "layer_norm", # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'layer_norm_i2vgen' + norm_eps: float = 1e-5, + final_dropout: bool = False, + attention_type: str = "default", + positional_embeddings: Optional[str] = None, + num_positional_embeddings: Optional[int] = None, + ada_norm_continous_conditioning_embedding_dim: Optional[int] = None, + ada_norm_bias: Optional[int] = None, + ff_inner_dim: Optional[int] = None, + ff_bias: bool = True, + attention_out_bias: bool = True, + ): + super().__init__() + self.only_cross_attention = only_cross_attention + + self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero" + self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm" + self.use_ada_layer_norm_single = norm_type == "ada_norm_single" + self.use_layer_norm = norm_type == "layer_norm" + self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous" + + if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None: + raise ValueError( + f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to" + f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}." + ) + + self.norm_type = norm_type + self.num_embeds_ada_norm = num_embeds_ada_norm + + if positional_embeddings and (num_positional_embeddings is None): + raise ValueError( + "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined." + ) + + if positional_embeddings == "sinusoidal": + self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings) + else: + self.pos_embed = None + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + if norm_type == "ada_norm": + self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_zero": + self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm1 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim if only_cross_attention else None, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None or double_self_attention: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + if norm_type == "ada_norm": + self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm) + elif norm_type == "ada_norm_continuous": + self.norm2 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "rms_norm", + ) + else: + self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim if not double_self_attention else None, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + out_bias=attention_out_bias, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + if norm_type == "ada_norm_continuous": + self.norm3 = AdaLayerNormContinuous( + dim, + ada_norm_continous_conditioning_embedding_dim, + norm_elementwise_affine, + norm_eps, + ada_norm_bias, + "layer_norm", + ) + + elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]: + self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine) + elif norm_type == "layer_norm_i2vgen": + self.norm3 = None + + self.ff = FeedForward( + dim, + dropout=dropout, + activation_fn=activation_fn, + final_dropout=final_dropout, + inner_dim=ff_inner_dim, + bias=ff_bias, + ) + + # 4. Fuser + if attention_type == "gated" or attention_type == "gated-text-image": + self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim) + + # 5. Scale-shift for PixArt-Alpha. + if norm_type == "ada_norm_single": + self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5) + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = 0 + + def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0): + # Sets chunk feed-forward + self._chunk_size = chunk_size + self._chunk_dim = dim + + def forward( + self, + hidden_states: torch.FloatTensor, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + timestep: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + class_labels: Optional[torch.LongTensor] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm1(hidden_states, timestep) + elif self.norm_type == "ada_norm_zero": + norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1( + hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm1(hidden_states) + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif self.norm_type == "ada_norm_single": + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = ( + self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1) + ).chunk(6, dim=1) + norm_hidden_states = self.norm1(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa + norm_hidden_states = norm_hidden_states.squeeze(1) + else: + raise ValueError("Incorrect norm used") + + if self.pos_embed is not None: + norm_hidden_states = self.pos_embed(norm_hidden_states) + + # 1. Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 2. Prepare GLIGEN inputs + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + gligen_kwargs = cross_attention_kwargs.pop("gligen", None) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + if self.norm_type == "ada_norm_zero": + attn_output = gate_msa.unsqueeze(1) * attn_output + elif self.norm_type == "ada_norm_single": + attn_output = gate_msa * attn_output + + hidden_states = attn_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + # 2.5 GLIGEN Control + if gligen_kwargs is not None: + hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"]) + + # 3. Cross-Attention + if self.attn2 is not None: + if self.norm_type == "ada_norm": + norm_hidden_states = self.norm2(hidden_states, timestep) + elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]: + norm_hidden_states = self.norm2(hidden_states) + elif self.norm_type == "ada_norm_single": + # For PixArt norm2 isn't applied here: + # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103 + norm_hidden_states = hidden_states + elif self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"]) + else: + raise ValueError("Incorrect norm") + + if self.pos_embed is not None and self.norm_type != "ada_norm_single": + norm_hidden_states = self.pos_embed(norm_hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=encoder_attention_mask, + **cross_attention_kwargs, + ) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + # i2vgen doesn't have this norm 🤷‍♂️ + if self.norm_type == "ada_norm_continuous": + norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"]) + elif not self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm3(hidden_states) + + if self.norm_type == "ada_norm_zero": + norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None] + + if self.norm_type == "ada_norm_single": + norm_hidden_states = self.norm2(hidden_states) + norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp + + if self._chunk_size is not None: + # "feed_forward_chunk_size" can be used to save memory + ff_output = _chunked_feed_forward( + self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale + ) + else: + ff_output = self.ff(norm_hidden_states, scale=lora_scale) + + if self.norm_type == "ada_norm_zero": + ff_output = gate_mlp.unsqueeze(1) * ff_output + elif self.norm_type == "ada_norm_single": + ff_output = gate_mlp * ff_output + + hidden_states = ff_output + hidden_states + if hidden_states.ndim == 4: + hidden_states = hidden_states.squeeze(1) + + return hidden_states + + +@maybe_allow_in_graph +class TemporalBasicTransformerBlock(nn.Module): + r""" + A basic Transformer block for video like data. + + Parameters: + dim (`int`): The number of channels in the input and output. + time_mix_inner_dim (`int`): The number of channels for temporal attention. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + """ + + def __init__( + self, + dim: int, + time_mix_inner_dim: int, + num_attention_heads: int, + attention_head_dim: int, + cross_attention_dim: Optional[int] = None, + ): + super().__init__() + self.is_res = dim == time_mix_inner_dim + + self.norm_in = nn.LayerNorm(dim) + + # Define 3 blocks. Each block has its own normalization layer. + # 1. Self-Attn + self.norm_in = nn.LayerNorm(dim) + self.ff_in = FeedForward( + dim, + dim_out=time_mix_inner_dim, + activation_fn="geglu", + ) + + self.norm1 = nn.LayerNorm(time_mix_inner_dim) + self.attn1 = Attention( + query_dim=time_mix_inner_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + cross_attention_dim=None, + ) + + # 2. Cross-Attn + if cross_attention_dim is not None: + # We currently only use AdaLayerNormZero for self attention where there will only be one attention block. + # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during + # the second cross attention block. + self.norm2 = nn.LayerNorm(time_mix_inner_dim) + self.attn2 = Attention( + query_dim=time_mix_inner_dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + ) # is self-attn if encoder_hidden_states is none + else: + self.norm2 = None + self.attn2 = None + + # 3. Feed-forward + self.norm3 = nn.LayerNorm(time_mix_inner_dim) + self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu") + + # let chunk size default to None + self._chunk_size = None + self._chunk_dim = None + + def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs): + # Sets chunk feed-forward + self._chunk_size = chunk_size + # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off + self._chunk_dim = 1 + + def forward( + self, + hidden_states: torch.FloatTensor, + num_frames: int, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + # Notice that normalization is always applied before the real computation in the following blocks. + # 0. Self-Attention + batch_size = hidden_states.shape[0] + + batch_frames, seq_length, channels = hidden_states.shape + batch_size = batch_frames // num_frames + + hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels) + + residual = hidden_states + hidden_states = self.norm_in(hidden_states) + + if self._chunk_size is not None: + hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size) + else: + hidden_states = self.ff_in(hidden_states) + + if self.is_res: + hidden_states = hidden_states + residual + + norm_hidden_states = self.norm1(hidden_states) + attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None) + hidden_states = attn_output + hidden_states + + # 3. Cross-Attention + if self.attn2 is not None: + norm_hidden_states = self.norm2(hidden_states) + attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states) + hidden_states = attn_output + hidden_states + + # 4. Feed-forward + norm_hidden_states = self.norm3(hidden_states) + + if self._chunk_size is not None: + ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size) + else: + ff_output = self.ff(norm_hidden_states) + + if self.is_res: + hidden_states = ff_output + hidden_states + else: + hidden_states = ff_output + + hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels) + hidden_states = hidden_states.permute(0, 2, 1, 3) + hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels) + + return hidden_states + + +class SkipFFTransformerBlock(nn.Module): + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + kv_input_dim: int, + kv_input_dim_proj_use_bias: bool, + dropout=0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + attention_out_bias: bool = True, + ): + super().__init__() + if kv_input_dim != dim: + self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias) + else: + self.kv_mapper = None + + self.norm1 = RMSNorm(dim, 1e-06) + + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=cross_attention_dim, + out_bias=attention_out_bias, + ) + + self.norm2 = RMSNorm(dim, 1e-06) + + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + out_bias=attention_out_bias, + ) + + def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs): + cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {} + + if self.kv_mapper is not None: + encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states)) + + norm_hidden_states = self.norm1(hidden_states) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + norm_hidden_states = self.norm2(hidden_states) + + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + **cross_attention_kwargs, + ) + + hidden_states = attn_output + hidden_states + + return hidden_states + + +class FeedForward(nn.Module): + r""" + A feed-forward layer. + + Parameters: + dim (`int`): The number of channels in the input. + dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`. + mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + final_dropout (`bool` *optional*, defaults to False): Apply a final dropout. + bias (`bool`, defaults to True): Whether to use a bias in the linear layer. + """ + + def __init__( + self, + dim: int, + dim_out: Optional[int] = None, + mult: int = 4, + dropout: float = 0.0, + activation_fn: str = "geglu", + final_dropout: bool = False, + inner_dim=None, + bias: bool = True, + ): + super().__init__() + if inner_dim is None: + inner_dim = int(dim * mult) + dim_out = dim_out if dim_out is not None else dim + linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear + + if activation_fn == "gelu": + act_fn = GELU(dim, inner_dim, bias=bias) + if activation_fn == "gelu-approximate": + act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias) + elif activation_fn == "geglu": + act_fn = GEGLU(dim, inner_dim, bias=bias) + elif activation_fn == "geglu-approximate": + act_fn = ApproximateGELU(dim, inner_dim, bias=bias) + + self.net = nn.ModuleList([]) + # project in + self.net.append(act_fn) + # project dropout + self.net.append(nn.Dropout(dropout)) + # project out + self.net.append(linear_cls(inner_dim, dim_out, bias=bias)) + # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout + if final_dropout: + self.net.append(nn.Dropout(dropout)) + + def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor: + compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear) + for module in self.net: + if isinstance(module, compatible_cls): + hidden_states = module(hidden_states, scale) + else: + hidden_states = module(hidden_states) + return hidden_states diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention_processor.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention_processor.py new file mode 100644 index 0000000000000000000000000000000000000000..ffc4fbfb232b83b85ae946ce2544eab02efb94b2 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/attention_processor.py @@ -0,0 +1,1019 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from importlib import import_module +from typing import Callable, Optional, Union + +import torch +import torch_npu +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate, logging +from diffusers.utils.import_utils import is_xformers_available +from diffusers.models.attention_processor import ( + CustomDiffusionAttnProcessor, FusedAttnProcessor2_0, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0, + SpatialNorm, SlicedAttnProcessor, SlicedAttnAddedKVProcessor, LoRAAttnProcessor, LoRAAttnProcessor2_0, + LoRAXFormersAttnProcessor, LoRAAttnAddedKVProcessor, IPAdapterAttnProcessor, IPAdapterAttnProcessor2_0, + CustomDiffusionAttnProcessor, XFormersAttnAddedKVProcessor, XFormersAttnProcessor, +) +from diffusers.models.lora import LoRACompatibleLinear, LoRALinearLayer + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +class Attention(nn.Module): + r""" + A cross attention layer. + + Parameters: + query_dim (`int`): + The number of channels in the query. + cross_attention_dim (`int`, *optional*): + The number of channels in the encoder_hidden_states. If not given, defaults to `query_dim`. + heads (`int`, *optional*, defaults to 8): + The number of heads to use for multi-head attention. + dim_head (`int`, *optional*, defaults to 64): + The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): + The dropout probability to use. + bias (`bool`, *optional*, defaults to False): + Set to `True` for the query, key, and value linear layers to contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to False): + Set to `True` to upcast the attention computation to `float32`. + upcast_softmax (`bool`, *optional*, defaults to False): + Set to `True` to upcast the softmax computation to `float32`. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + cross_attention_norm_num_groups (`int`, *optional*, defaults to 32): + The number of groups to use for the group norm in the cross attention. + added_kv_proj_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the added key and value projections. If `None`, no projection is used. + norm_num_groups (`int`, *optional*, defaults to `None`): + The number of groups to use for the group norm in the attention. + spatial_norm_dim (`int`, *optional*, defaults to `None`): + The number of channels to use for the spatial normalization. + out_bias (`bool`, *optional*, defaults to `True`): + Set to `True` to use a bias in the output linear layer. + scale_qk (`bool`, *optional*, defaults to `True`): + Set to `True` to scale the query and key by `1 / sqrt(dim_head)`. + only_cross_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to only use cross attention and not added_kv_proj_dim. Can only be set to `True` if + `added_kv_proj_dim` is not `None`. + eps (`float`, *optional*, defaults to 1e-5): + An additional value added to the denominator in group normalization that is used for numerical stability. + rescale_output_factor (`float`, *optional*, defaults to 1.0): + A factor to rescale the output by dividing it with this value. + residual_connection (`bool`, *optional*, defaults to `False`): + Set to `True` to add the residual connection to the output. + _from_deprecated_attn_block (`bool`, *optional*, defaults to `False`): + Set to `True` if the attention block is loaded from a deprecated state dict. + processor (`AttnProcessor`, *optional*, defaults to `None`): + The attention processor to use. If `None`, defaults to `AttnProcessor2_0` if `torch 2.x` is used and + `AttnProcessor` otherwise. + """ + + def __init__( + self, + query_dim: int, + cross_attention_dim: Optional[int] = None, + heads: int = 8, + dim_head: int = 64, + dropout: float = 0.0, + bias: bool = False, + upcast_attention: bool = False, + upcast_softmax: bool = False, + cross_attention_norm: Optional[str] = None, + cross_attention_norm_num_groups: int = 32, + added_kv_proj_dim: Optional[int] = None, + norm_num_groups: Optional[int] = None, + spatial_norm_dim: Optional[int] = None, + out_bias: bool = True, + scale_qk: bool = True, + only_cross_attention: bool = False, + eps: float = 1e-5, + rescale_output_factor: float = 1.0, + residual_connection: bool = False, + _from_deprecated_attn_block: bool = False, + processor: Optional["AttnProcessor"] = None, + out_dim: int = None, + ): + super().__init__() + self.inner_dim = out_dim if out_dim is not None else dim_head * heads + self.query_dim = query_dim + self.cross_attention_dim = cross_attention_dim if cross_attention_dim is not None else query_dim + self.upcast_attention = upcast_attention + self.upcast_softmax = upcast_softmax + self.rescale_output_factor = rescale_output_factor + self.residual_connection = residual_connection + self.dropout = dropout + self.fused_projections = False + self.out_dim = out_dim if out_dim is not None else query_dim + + # we make use of this private variable to know whether this class is loaded + # with an deprecated state dict so that we can convert it on the fly + self._from_deprecated_attn_block = _from_deprecated_attn_block + + self.scale_qk = scale_qk + self.scale = dim_head**-0.5 if self.scale_qk else 1.0 + + self.heads = out_dim // dim_head if out_dim is not None else heads + # for slice_size > 0 the attention score computation + # is split across the batch axis to save memory + # You can set slice_size with `set_attention_slice` + self.sliceable_head_dim = heads + + self.added_kv_proj_dim = added_kv_proj_dim + self.only_cross_attention = only_cross_attention + + if self.added_kv_proj_dim is None and self.only_cross_attention: + raise ValueError( + "`only_cross_attention` can only be set to True if `added_kv_proj_dim` is not None. Make sure to set either `only_cross_attention=False` or define `added_kv_proj_dim`." + ) + + if norm_num_groups is not None: + self.group_norm = nn.GroupNorm(num_channels=query_dim, num_groups=norm_num_groups, eps=eps, affine=True) + else: + self.group_norm = None + + if spatial_norm_dim is not None: + self.spatial_norm = SpatialNorm(f_channels=query_dim, zq_channels=spatial_norm_dim) + else: + self.spatial_norm = None + + if cross_attention_norm is None: + self.norm_cross = None + elif cross_attention_norm == "layer_norm": + self.norm_cross = nn.LayerNorm(self.cross_attention_dim) + elif cross_attention_norm == "group_norm": + if self.added_kv_proj_dim is not None: + # The given `encoder_hidden_states` are initially of shape + # (batch_size, seq_len, added_kv_proj_dim) before being projected + # to (batch_size, seq_len, cross_attention_dim). The norm is applied + # before the projection, so we need to use `added_kv_proj_dim` as + # the number of channels for the group norm. + norm_cross_num_channels = added_kv_proj_dim + else: + norm_cross_num_channels = self.cross_attention_dim + + self.norm_cross = nn.GroupNorm( + num_channels=norm_cross_num_channels, num_groups=cross_attention_norm_num_groups, eps=1e-5, affine=True + ) + else: + raise ValueError( + f"unknown cross_attention_norm: {cross_attention_norm}. Should be None, 'layer_norm' or 'group_norm'" + ) + + if USE_PEFT_BACKEND: + linear_cls = nn.Linear + else: + linear_cls = LoRACompatibleLinear + + self.linear_cls = linear_cls + self.to_q = linear_cls(query_dim, self.inner_dim, bias=bias) + + if not self.only_cross_attention: + # only relevant for the `AddedKVProcessor` classes + self.to_k = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + self.to_v = linear_cls(self.cross_attention_dim, self.inner_dim, bias=bias) + else: + self.to_k = None + self.to_v = None + + if self.added_kv_proj_dim is not None: + self.add_k_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + self.add_v_proj = linear_cls(added_kv_proj_dim, self.inner_dim) + + self.to_out = nn.ModuleList([]) + self.to_out.append(linear_cls(self.inner_dim, self.out_dim, bias=out_bias)) + self.to_out.append(nn.Dropout(dropout)) + + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + if processor is None: + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + self.set_processor(processor) + + def set_use_memory_efficient_attention_xformers( + self, use_memory_efficient_attention_xformers: bool, attention_op: Optional[Callable] = None + ) -> None: + r""" + Set whether to use memory efficient attention from `xformers` or not. + + Args: + use_memory_efficient_attention_xformers (`bool`): + Whether to use memory efficient attention from `xformers` or not. + attention_op (`Callable`, *optional*): + The attention operation to use. Defaults to `None` which uses the default attention operation from + `xformers`. + """ + is_lora = hasattr(self, "processor") and isinstance( + self.processor, + LORA_ATTENTION_PROCESSORS, + ) + is_custom_diffusion = hasattr(self, "processor") and isinstance( + self.processor, + (CustomDiffusionAttnProcessor, CustomDiffusionXFormersAttnProcessor, CustomDiffusionAttnProcessor2_0), + ) + is_added_kv_processor = hasattr(self, "processor") and isinstance( + self.processor, + ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, + SlicedAttnAddedKVProcessor, + XFormersAttnAddedKVProcessor, + LoRAAttnAddedKVProcessor, + ), + ) + + if use_memory_efficient_attention_xformers: + if is_added_kv_processor and (is_lora or is_custom_diffusion): + raise NotImplementedError( + f"Memory efficient attention is currently not supported for LoRA or custom diffusion for attention processor type {self.processor}" + ) + if not is_xformers_available(): + raise ModuleNotFoundError( + ( + "Refer to https://github.com/facebookresearch/xformers for more information on how to install" + " xformers" + ), + name="xformers", + ) + elif not torch.cuda.is_available(): + raise ValueError( + "torch.cuda.is_available() should be True but is False. xformers' memory efficient attention is" + " only available for GPU " + ) + else: + try: + # Make sure we can run the memory efficient attention + _ = xformers.ops.memory_efficient_attention( + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + torch.randn((1, 2, 40), device="cuda"), + ) + except Exception as e: + raise e + + if is_lora: + # TODO (sayakpaul): should we throw a warning if someone wants to use the xformers + # variant when using PT 2.0 now that we have LoRAAttnProcessor2_0? + processor = LoRAXFormersAttnProcessor( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + processor = CustomDiffusionXFormersAttnProcessor( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + attention_op=attention_op, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + elif is_added_kv_processor: + # TODO(Patrick, Suraj, William) - currently xformers doesn't work for UnCLIP + # which uses this type of cross attention ONLY because the attention mask of format + # [0, ..., -10.000, ..., 0, ...,] is not supported + # throw warning + logger.info( + "Memory efficient attention with `xformers` might currently not work correctly if an attention mask is required for the attention operation." + ) + processor = XFormersAttnAddedKVProcessor(attention_op=attention_op) + else: + processor = XFormersAttnProcessor(attention_op=attention_op) + else: + if is_lora: + attn_processor_class = ( + LoRAAttnProcessor2_0 if hasattr(F, "scaled_dot_product_attention") else LoRAAttnProcessor + ) + processor = attn_processor_class( + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + rank=self.processor.rank, + ) + processor.load_state_dict(self.processor.state_dict()) + processor.to(self.processor.to_q_lora.up.weight.device) + elif is_custom_diffusion: + attn_processor_class = ( + CustomDiffusionAttnProcessor2_0 + if hasattr(F, "scaled_dot_product_attention") + else CustomDiffusionAttnProcessor + ) + processor = attn_processor_class( + train_kv=self.processor.train_kv, + train_q_out=self.processor.train_q_out, + hidden_size=self.processor.hidden_size, + cross_attention_dim=self.processor.cross_attention_dim, + ) + processor.load_state_dict(self.processor.state_dict()) + if hasattr(self.processor, "to_k_custom_diffusion"): + processor.to(self.processor.to_k_custom_diffusion.weight.device) + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() + if hasattr(F, "scaled_dot_product_attention") and self.scale_qk + else AttnProcessor() + ) + + self.set_processor(processor) + + def set_attention_slice(self, slice_size: int) -> None: + r""" + Set the slice size for attention computation. + + Args: + slice_size (`int`): + The slice size for attention computation. + """ + if slice_size is not None and slice_size > self.sliceable_head_dim: + raise ValueError(f"slice_size {slice_size} has to be smaller or equal to {self.sliceable_head_dim}.") + + if slice_size is not None and self.added_kv_proj_dim is not None: + processor = SlicedAttnAddedKVProcessor(slice_size) + elif slice_size is not None: + processor = SlicedAttnProcessor(slice_size) + elif self.added_kv_proj_dim is not None: + processor = AttnAddedKVProcessor() + else: + # set attention processor + # We use the AttnProcessor2_0 by default when torch 2.x is used which uses + # torch.nn.functional.scaled_dot_product_attention for native Flash/memory_efficient_attention + # but only if it has the default `scale` argument. TODO remove scale_qk check when we move to torch 2.1 + processor = ( + AttnProcessor2_0() if hasattr(F, "scaled_dot_product_attention") and self.scale_qk else AttnProcessor() + ) + + self.set_processor(processor) + + def set_processor(self, processor: "AttnProcessor") -> None: + r""" + Set the attention processor to use. + + Args: + processor (`AttnProcessor`): + The attention processor to use. + """ + # if current processor is in `self._modules` and if passed `processor` is not, we need to + # pop `processor` from `self._modules` + if ( + hasattr(self, "processor") + and isinstance(self.processor, torch.nn.Module) + and not isinstance(processor, torch.nn.Module) + ): + logger.info(f"You are removing possibly trained weights of {self.processor} with {processor}") + self._modules.pop("processor") + + self.processor = processor + + def get_processor(self, return_deprecated_lora: bool = False) -> "AttentionProcessor": + r""" + Get the attention processor in use. + + Args: + return_deprecated_lora (`bool`, *optional*, defaults to `False`): + Set to `True` to return the deprecated LoRA attention processor. + + Returns: + "AttentionProcessor": The attention processor in use. + """ + if not return_deprecated_lora: + return self.processor + + # TODO(Sayak, Patrick). The rest of the function is needed to ensure backwards compatible + # serialization format for LoRA Attention Processors. It should be deleted once the integration + # with PEFT is completed. + is_lora_activated = { + name: module.lora_layer is not None + for name, module in self.named_modules() + if hasattr(module, "lora_layer") + } + + # 1. if no layer has a LoRA activated we can return the processor as usual + if not any(is_lora_activated.values()): + return self.processor + + # If doesn't apply LoRA do `add_k_proj` or `add_v_proj` + is_lora_activated.pop("add_k_proj", None) + is_lora_activated.pop("add_v_proj", None) + # 2. else it is not posssible that only some layers have LoRA activated + if not all(is_lora_activated.values()): + raise ValueError( + f"Make sure that either all layers or no layers have LoRA activated, but have {is_lora_activated}" + ) + + # 3. And we need to merge the current LoRA layers into the corresponding LoRA attention processor + non_lora_processor_cls_name = self.processor.__class__.__name__ + lora_processor_cls = getattr(import_module(__name__), "LoRA" + non_lora_processor_cls_name) + + hidden_size = self.inner_dim + + # now create a LoRA attention processor from the LoRA layers + if lora_processor_cls in [LoRAAttnProcessor, LoRAAttnProcessor2_0, LoRAXFormersAttnProcessor]: + kwargs = { + "cross_attention_dim": self.cross_attention_dim, + "rank": self.to_q.lora_layer.rank, + "network_alpha": self.to_q.lora_layer.network_alpha, + "q_rank": self.to_q.lora_layer.rank, + "q_hidden_size": self.to_q.lora_layer.out_features, + "k_rank": self.to_k.lora_layer.rank, + "k_hidden_size": self.to_k.lora_layer.out_features, + "v_rank": self.to_v.lora_layer.rank, + "v_hidden_size": self.to_v.lora_layer.out_features, + "out_rank": self.to_out[0].lora_layer.rank, + "out_hidden_size": self.to_out[0].lora_layer.out_features, + } + + if hasattr(self.processor, "attention_op"): + kwargs["attention_op"] = self.processor.attention_op + + lora_processor = lora_processor_cls(hidden_size, **kwargs) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + elif lora_processor_cls == LoRAAttnAddedKVProcessor: + lora_processor = lora_processor_cls( + hidden_size, + cross_attention_dim=self.add_k_proj.weight.shape[0], + rank=self.to_q.lora_layer.rank, + network_alpha=self.to_q.lora_layer.network_alpha, + ) + lora_processor.to_q_lora.load_state_dict(self.to_q.lora_layer.state_dict()) + lora_processor.to_k_lora.load_state_dict(self.to_k.lora_layer.state_dict()) + lora_processor.to_v_lora.load_state_dict(self.to_v.lora_layer.state_dict()) + lora_processor.to_out_lora.load_state_dict(self.to_out[0].lora_layer.state_dict()) + + # only save if used + if self.add_k_proj.lora_layer is not None: + lora_processor.add_k_proj_lora.load_state_dict(self.add_k_proj.lora_layer.state_dict()) + lora_processor.add_v_proj_lora.load_state_dict(self.add_v_proj.lora_layer.state_dict()) + else: + lora_processor.add_k_proj_lora = None + lora_processor.add_v_proj_lora = None + else: + raise ValueError(f"{lora_processor_cls} does not exist.") + + return lora_processor + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + **cross_attention_kwargs, + ) -> torch.Tensor: + r""" + The forward method of the `Attention` class. + + Args: + hidden_states (`torch.Tensor`): + The hidden states of the query. + encoder_hidden_states (`torch.Tensor`, *optional*): + The hidden states of the encoder. + attention_mask (`torch.Tensor`, *optional*): + The attention mask to use. If `None`, no mask is applied. + **cross_attention_kwargs: + Additional keyword arguments to pass along to the cross attention. + + Returns: + `torch.Tensor`: The output of the attention layer. + """ + # The `Attention` class can call different attention processors / attention functions + # here we simply pass along all tensors to the selected processor class + # For standard processors that are defined here, `**cross_attention_kwargs` is empty + return self.processor( + self, + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + + def batch_to_head_dim(self, tensor: torch.Tensor) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size // heads, seq_len, dim * heads]`. `heads` + is the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size // head_size, head_size, seq_len, dim) + tensor = tensor.permute(0, 2, 1, 3).reshape(batch_size // head_size, seq_len, dim * head_size) + return tensor + + def head_to_batch_dim(self, tensor: torch.Tensor, out_dim: int = 3) -> torch.Tensor: + r""" + Reshape the tensor from `[batch_size, seq_len, dim]` to `[batch_size, seq_len, heads, dim // heads]` `heads` is + the number of heads initialized while constructing the `Attention` class. + + Args: + tensor (`torch.Tensor`): The tensor to reshape. + out_dim (`int`, *optional*, defaults to `3`): The output dimension of the tensor. If `3`, the tensor is + reshaped to `[batch_size * heads, seq_len, dim // heads]`. + + Returns: + `torch.Tensor`: The reshaped tensor. + """ + head_size = self.heads + batch_size, seq_len, dim = tensor.shape + tensor = tensor.reshape(batch_size, seq_len, head_size, dim // head_size) + tensor = tensor.permute(0, 2, 1, 3) + + if out_dim == 3: + tensor = tensor.reshape(batch_size * head_size, seq_len, dim // head_size) + + return tensor + + def get_attention_scores( + self, query: torch.Tensor, key: torch.Tensor, attention_mask: torch.Tensor = None + ) -> torch.Tensor: + r""" + Compute the attention scores. + + Args: + query (`torch.Tensor`): The query tensor. + key (`torch.Tensor`): The key tensor. + attention_mask (`torch.Tensor`, *optional*): The attention mask to use. If `None`, no mask is applied. + + Returns: + `torch.Tensor`: The attention probabilities/scores. + """ + dtype = query.dtype + if self.upcast_attention: + query = query.float() + key = key.float() + + if attention_mask is None: + baddbmm_input = torch.empty( + query.shape[0], query.shape[1], key.shape[1], dtype=query.dtype, device=query.device + ) + beta = 0 + else: + baddbmm_input = attention_mask + beta = 1 + + attention_scores = torch.baddbmm( + baddbmm_input, + query, + key.transpose(-1, -2), + beta=beta, + alpha=self.scale, + ) + del baddbmm_input + + if self.upcast_softmax: + attention_scores = attention_scores.float() + + attention_probs = attention_scores.softmax(dim=-1) + del attention_scores + + attention_probs = attention_probs.to(dtype) + + return attention_probs + + def prepare_attention_mask( + self, attention_mask: torch.Tensor, target_length: int, batch_size: int, out_dim: int = 3 + ) -> torch.Tensor: + r""" + Prepare the attention mask for the attention computation. + + Args: + attention_mask (`torch.Tensor`): + The attention mask to prepare. + target_length (`int`): + The target length of the attention mask. This is the length of the attention mask after padding. + batch_size (`int`): + The batch size, which is used to repeat the attention mask. + out_dim (`int`, *optional*, defaults to `3`): + The output dimension of the attention mask. Can be either `3` or `4`. + + Returns: + `torch.Tensor`: The prepared attention mask. + """ + head_size = self.heads + if attention_mask is None: + return attention_mask + + current_length: int = attention_mask.shape[-1] + if current_length != target_length: + if attention_mask.device.type == "mps": + # HACK: MPS: Does not support padding by greater than dimension of input tensor. + # Instead, we can manually construct the padding tensor. + padding_shape = (attention_mask.shape[0], attention_mask.shape[1], target_length) + padding = torch.zeros(padding_shape, dtype=attention_mask.dtype, device=attention_mask.device) + attention_mask = torch.cat([attention_mask, padding], dim=2) + else: + # TODO: for pipelines such as stable-diffusion, padding cross-attn mask: + # we want to instead pad by (0, remaining_length), where remaining_length is: + # remaining_length: int = target_length - current_length + # TODO: re-enable tests/models/test_models_unet_2d_condition.py#test_model_xattn_padding + attention_mask = F.pad(attention_mask, (0, target_length), value=0.0) + + if out_dim == 3: + if attention_mask.shape[0] < batch_size * head_size: + attention_mask = attention_mask.repeat_interleave(head_size, dim=0) + elif out_dim == 4: + attention_mask = attention_mask.unsqueeze(1) + attention_mask = attention_mask.repeat_interleave(head_size, dim=1) + + return attention_mask + + def norm_encoder_hidden_states(self, encoder_hidden_states: torch.Tensor) -> torch.Tensor: + r""" + Normalize the encoder hidden states. Requires `self.norm_cross` to be specified when constructing the + `Attention` class. + + Args: + encoder_hidden_states (`torch.Tensor`): Hidden states of the encoder. + + Returns: + `torch.Tensor`: The normalized encoder hidden states. + """ + assert self.norm_cross is not None, "self.norm_cross must be defined to call self.norm_encoder_hidden_states" + + if isinstance(self.norm_cross, nn.LayerNorm): + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + elif isinstance(self.norm_cross, nn.GroupNorm): + # Group norm norms along the channels dimension and expects + # input to be in the shape of (N, C, *). In this case, we want + # to norm along the hidden dimension, so we need to move + # (batch_size, sequence_length, hidden_size) -> + # (batch_size, hidden_size, sequence_length) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + encoder_hidden_states = self.norm_cross(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.transpose(1, 2) + else: + assert False + + return encoder_hidden_states + + @torch.no_grad() + def fuse_projections(self, fuse=True): + is_cross_attention = self.cross_attention_dim != self.query_dim + device = self.to_q.weight.data.device + dtype = self.to_q.weight.data.dtype + + if not is_cross_attention: + # fetch weight matrices. + concatenated_weights = torch.cat([self.to_q.weight.data, self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + # create a new single projection layer and copy over the weights. + self.to_qkv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_qkv.weight.copy_(concatenated_weights) + + else: + concatenated_weights = torch.cat([self.to_k.weight.data, self.to_v.weight.data]) + in_features = concatenated_weights.shape[1] + out_features = concatenated_weights.shape[0] + + self.to_kv = self.linear_cls(in_features, out_features, bias=False, device=device, dtype=dtype) + self.to_kv.weight.copy_(concatenated_weights) + + self.fused_projections = fuse + + +class AttnProcessor: + r""" + Default processor for performing attention-related computations. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + query = attn.head_to_batch_dim(query) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + +class AttnAddedKVProcessor: + r""" + Processor for performing attention-related computations with extra learnable key and value matrices for the text + encoder. + """ + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states, *args) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states, *args) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key) + value = attn.head_to_batch_dim(value) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=1) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=1) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + attention_probs = attn.get_attention_scores(query, key, attention_mask) + hidden_states = torch.bmm(attention_probs, value) + hidden_states = attn.batch_to_head_dim(hidden_states) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnAddedKVProcessor2_0: + r""" + Processor for performing scaled dot-product attention (enabled by default if you're using PyTorch 2.0), with extra + learnable key and value matrices for the text encoder. + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError( + "AttnAddedKVProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0." + ) + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.Tensor: + residual = hidden_states + + args = () if USE_PEFT_BACKEND else (scale,) + + hidden_states = hidden_states.view(hidden_states.shape[0], hidden_states.shape[1], -1).transpose(1, 2) + batch_size, sequence_length, _ = hidden_states.shape + + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size, out_dim=4) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + query = attn.to_q(hidden_states, *args) + query = attn.head_to_batch_dim(query, out_dim=4) + + encoder_hidden_states_key_proj = attn.add_k_proj(encoder_hidden_states) + encoder_hidden_states_value_proj = attn.add_v_proj(encoder_hidden_states) + encoder_hidden_states_key_proj = attn.head_to_batch_dim(encoder_hidden_states_key_proj, out_dim=4) + encoder_hidden_states_value_proj = attn.head_to_batch_dim(encoder_hidden_states_value_proj, out_dim=4) + + if not attn.only_cross_attention: + key = attn.to_k(hidden_states, *args) + value = attn.to_v(hidden_states, *args) + key = attn.head_to_batch_dim(key, out_dim=4) + value = attn.head_to_batch_dim(value, out_dim=4) + key = torch.cat([encoder_hidden_states_key_proj, key], dim=2) + value = torch.cat([encoder_hidden_states_value_proj, value], dim=2) + else: + key = encoder_hidden_states_key_proj + value = encoder_hidden_states_value_proj + + # the output of sdp = (batch, num_heads, seq_len, head_dim) + # TODO: add support for attn.scale when we move to Torch 2.1 + hidden_states = F.scaled_dot_product_attention( + query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + ) + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, residual.shape[1]) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + hidden_states = hidden_states.transpose(-1, -2).reshape(residual.shape) + hidden_states = hidden_states + residual + + return hidden_states + + +class AttnProcessor2_0: + r""" + Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0). + """ + + def __init__(self): + if not hasattr(F, "scaled_dot_product_attention"): + raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.") + + def __call__( + self, + attn: Attention, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + temb: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + residual = hidden_states + if attn.spatial_norm is not None: + hidden_states = attn.spatial_norm(hidden_states, temb) + + input_ndim = hidden_states.ndim + + if input_ndim == 4: + batch_size, channel, height, width = hidden_states.shape + hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2) + + batch_size, sequence_length, _ = ( + hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape + ) + + if attention_mask is not None: + attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size) + # scaled_dot_product_attention expects attention_mask shape to be + # (batch, heads, source_length, target_length) + attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1]) + + if attn.group_norm is not None: + hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2) + + args = () if USE_PEFT_BACKEND else (scale,) + query = attn.to_q(hidden_states, *args) + + if encoder_hidden_states is None: + encoder_hidden_states = hidden_states + elif attn.norm_cross: + encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states) + + key = attn.to_k(encoder_hidden_states, *args) + value = attn.to_v(encoder_hidden_states, *args) + + inner_dim = key.shape[-1] + head_dim = inner_dim // attn.heads + + query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2) + + hidden_states = torch_npu.npu_fusion_attention( + query, key, value, + atten_mask=attention_mask, + input_layout='BNSD', + scale=head_dim**-0.5, + head_num=attn.heads, + )[0] + # hidden_states = F.scaled_dot_product_attention( + # query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False + # ) + + hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim) + hidden_states = hidden_states.to(query.dtype) + + # linear proj + hidden_states = attn.to_out[0](hidden_states, *args) + # dropout + hidden_states = attn.to_out[1](hidden_states) + + if input_ndim == 4: + hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width) + + if attn.residual_connection: + hidden_states = hidden_states + residual + + hidden_states = hidden_states / attn.rescale_output_factor + + return hidden_states + + + +ADDED_KV_ATTENTION_PROCESSORS = ( + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +) + +CROSS_ATTENTION_PROCESSORS = ( + AttnProcessor, + AttnProcessor2_0, +) + +AttentionProcessor = Union[ + AttnProcessor, + AttnProcessor2_0, + AttnAddedKVProcessor, + AttnAddedKVProcessor2_0, +] diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/embeddings.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/embeddings.py new file mode 100644 index 0000000000000000000000000000000000000000..9f46aa5f3b03048bfb3affeafb20894982775c5d --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/embeddings.py @@ -0,0 +1,747 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch +from torch import nn + +from diffusers.utils import USE_PEFT_BACKEND, deprecate +from diffusers.models.lora import LoRACompatibleLinear + +from .activations import get_activation +from diffusers.models.embeddings import PatchEmbed, ImagePositionalEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + +def get_timestep_embedding( + timesteps: torch.Tensor, + embedding_dim: int, + flip_sin_to_cos: bool = False, + downscale_freq_shift: float = 1, + scale: float = 1, + max_period: int = 10000, +): + """ + This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings. + + :param timesteps: a 1-D Tensor of N indices, one per batch element. + These may be fractional. + :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the + embeddings. :return: an [N x dim] Tensor of positional embeddings. + """ + assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array" + + half_dim = embedding_dim // 2 + exponent = -math.log(max_period) * torch.arange( + start=0, end=half_dim, dtype=torch.float32, device=timesteps.device + ) + exponent = exponent / (half_dim - downscale_freq_shift) + + emb = torch.exp(exponent) + emb = timesteps[:, None].float() * emb[None, :] + + # scale embeddings + emb = scale * emb + + # concat sine and cosine embeddings + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1) + + # flip sine and cosine embeddings + if flip_sin_to_cos: + emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1) + + # zero pad + if embedding_dim % 2 == 1: + emb = torch.nn.functional.pad(emb, (0, 1, 0, 0)) + return emb + + +def get_2d_sincos_pos_embed( + embed_dim, grid_size, cls_token=False, extra_tokens=0, interpolation_scale=1.0, base_size=16 +): + """ + grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or + [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token) + """ + if isinstance(grid_size, int): + grid_size = (grid_size, grid_size) + + grid_h = np.arange(grid_size[0], dtype=np.float32) / (grid_size[0] / base_size) / interpolation_scale + grid_w = np.arange(grid_size[1], dtype=np.float32) / (grid_size[1] / base_size) / interpolation_scale + grid = np.meshgrid(grid_w, grid_h) # here w goes first + grid = np.stack(grid, axis=0) + + grid = grid.reshape([2, 1, grid_size[1], grid_size[0]]) + pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid) + if cls_token and extra_tokens > 0: + pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0) + return pos_embed + + +def get_2d_sincos_pos_embed_from_grid(embed_dim, grid): + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + # use half of dimensions to encode grid_h + emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2) + emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2) + + emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D) + return emb + + +def get_1d_sincos_pos_embed_from_grid(embed_dim, pos): + """ + embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D) + """ + if embed_dim % 2 != 0: + raise ValueError("embed_dim must be divisible by 2") + + omega = np.arange(embed_dim // 2, dtype=np.float64) + omega /= embed_dim / 2.0 + omega = 1.0 / 10000**omega # (D/2,) + + pos = pos.reshape(-1) # (M,) + out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product + + emb_sin = np.sin(out) # (M, D/2) + emb_cos = np.cos(out) # (M, D/2) + + emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D) + return emb + + +class TimestepEmbedding(nn.Module): + def __init__( + self, + in_channels: int, + time_embed_dim: int, + act_fn: str = "silu", + out_dim: int = None, + post_act_fn: Optional[str] = None, + cond_proj_dim=None, + sample_proj_bias=True, + ): + super().__init__() + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + self.linear_1 = linear_cls(in_channels, time_embed_dim, sample_proj_bias) + + if cond_proj_dim is not None: + self.cond_proj = nn.Linear(cond_proj_dim, in_channels, bias=False) + else: + self.cond_proj = None + + self.act = get_activation(act_fn) + + if out_dim is not None: + time_embed_dim_out = out_dim + else: + time_embed_dim_out = time_embed_dim + self.linear_2 = linear_cls(time_embed_dim, time_embed_dim_out, sample_proj_bias) + + if post_act_fn is None: + self.post_act = None + else: + self.post_act = get_activation(post_act_fn) + + def forward(self, sample, condition=None): + if condition is not None: + sample = sample + self.cond_proj(condition) + sample = self.linear_1(sample) + + if self.act is not None: + sample = self.act(sample) + + sample = self.linear_2(sample) + + if self.post_act is not None: + sample = self.post_act(sample) + return sample + + +class Timesteps(nn.Module): + def __init__(self, num_channels: int, flip_sin_to_cos: bool, downscale_freq_shift: float): + super().__init__() + self.num_channels = num_channels + self.flip_sin_to_cos = flip_sin_to_cos + self.downscale_freq_shift = downscale_freq_shift + + def forward(self, timesteps): + t_emb = get_timestep_embedding( + timesteps, + self.num_channels, + flip_sin_to_cos=self.flip_sin_to_cos, + downscale_freq_shift=self.downscale_freq_shift, + ) + return t_emb + + +class GaussianFourierProjection(nn.Module): + """Gaussian Fourier embeddings for noise levels.""" + + def __init__( + self, embedding_size: int = 256, scale: float = 1.0, set_W_to_weight=True, log=True, flip_sin_to_cos=False + ): + super().__init__() + self.weight = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + self.log = log + self.flip_sin_to_cos = flip_sin_to_cos + + if set_W_to_weight: + # to delete later + self.W = nn.Parameter(torch.randn(embedding_size) * scale, requires_grad=False) + + self.weight = self.W + + def forward(self, x): + if self.log: + x = torch.log(x) + + x_proj = x[:, None] * self.weight[None, :] * 2 * np.pi + + if self.flip_sin_to_cos: + out = torch.cat([torch.cos(x_proj), torch.sin(x_proj)], dim=-1) + else: + out = torch.cat([torch.sin(x_proj), torch.cos(x_proj)], dim=-1) + return out + + +class SinusoidalPositionalEmbedding(nn.Module): + """Apply positional information to a sequence of embeddings. + + Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to + them + + Args: + embed_dim: (int): Dimension of the positional embedding. + max_seq_length: Maximum sequence length to apply positional embeddings + + """ + + def __init__(self, embed_dim: int, max_seq_length: int = 32): + super().__init__() + position = torch.arange(max_seq_length).unsqueeze(1) + div_term = torch.exp(torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim)) + pe = torch.zeros(1, max_seq_length, embed_dim) + pe[0, :, 0::2] = torch.sin(position * div_term) + pe[0, :, 1::2] = torch.cos(position * div_term) + self.register_buffer("pe", pe) + + def forward(self, x): + _, seq_length, _ = x.shape + x = x + self.pe[:, :seq_length] + return x + + +class LabelEmbedding(nn.Module): + """ + Embeds class labels into vector representations. Also handles label dropout for classifier-free guidance. + + Args: + num_classes (`int`): The number of classes. + hidden_size (`int`): The size of the vector embeddings. + dropout_prob (`float`): The probability of dropping a label. + """ + + def __init__(self, num_classes, hidden_size, dropout_prob): + super().__init__() + use_cfg_embedding = dropout_prob > 0 + self.embedding_table = nn.Embedding(num_classes + use_cfg_embedding, hidden_size) + self.num_classes = num_classes + self.dropout_prob = dropout_prob + + def token_drop(self, labels, force_drop_ids=None): + """ + Drops labels to enable classifier-free guidance. + """ + if force_drop_ids is None: + drop_ids = torch.rand(labels.shape[0], device=labels.device) < self.dropout_prob + else: + drop_ids = torch.tensor(force_drop_ids == 1) + labels = torch.where(drop_ids, self.num_classes, labels) + return labels + + def forward(self, labels: torch.LongTensor, force_drop_ids=None): + use_dropout = self.dropout_prob > 0 + if (self.training and use_dropout) or (force_drop_ids is not None): + labels = self.token_drop(labels, force_drop_ids) + embeddings = self.embedding_table(labels) + return embeddings + + +class TextImageProjection(nn.Module): + def __init__( + self, + text_embed_dim: int = 1024, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 10, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.text_proj = nn.Linear(text_embed_dim, cross_attention_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + batch_size = text_embeds.shape[0] + + # image + image_text_embeds = self.image_embeds(image_embeds) + image_text_embeds = image_text_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + + # text + text_embeds = self.text_proj(text_embeds) + + return torch.cat([image_text_embeds, text_embeds], dim=1) + + +class ImageProjection(nn.Module): + def __init__( + self, + image_embed_dim: int = 768, + cross_attention_dim: int = 768, + num_image_text_embeds: int = 32, + ): + super().__init__() + + self.num_image_text_embeds = num_image_text_embeds + self.image_embeds = nn.Linear(image_embed_dim, self.num_image_text_embeds * cross_attention_dim) + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + batch_size = image_embeds.shape[0] + + # image + image_embeds = self.image_embeds(image_embeds) + image_embeds = image_embeds.reshape(batch_size, self.num_image_text_embeds, -1) + image_embeds = self.norm(image_embeds) + return image_embeds + + +class IPAdapterFullImageProjection(nn.Module): + def __init__(self, image_embed_dim=1024, cross_attention_dim=1024): + super().__init__() + from .attention import FeedForward + + self.ff = FeedForward(image_embed_dim, cross_attention_dim, mult=1, activation_fn="gelu") + self.norm = nn.LayerNorm(cross_attention_dim) + + def forward(self, image_embeds: torch.FloatTensor): + return self.norm(self.ff(image_embeds)) + + +class CombinedTimestepLabelEmbeddings(nn.Module): + def __init__(self, num_classes, embedding_dim, class_dropout_prob=0.1): + super().__init__() + + self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=1) + self.timestep_embedder = TimestepEmbedding(in_channels=256, time_embed_dim=embedding_dim) + self.class_embedder = LabelEmbedding(num_classes, embedding_dim, class_dropout_prob) + + def forward(self, timestep, class_labels, hidden_dtype=None): + timesteps_proj = self.time_proj(timestep) + timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_dtype)) # (N, D) + + class_labels = self.class_embedder(class_labels) # (N, D) + + conditioning = timesteps_emb + class_labels # (N, D) + + return conditioning + + +class TextTimeEmbedding(nn.Module): + def __init__(self, encoder_dim: int, time_embed_dim: int, num_heads: int = 64): + super().__init__() + self.norm1 = nn.LayerNorm(encoder_dim) + self.pool = AttentionPooling(num_heads, encoder_dim) + self.proj = nn.Linear(encoder_dim, time_embed_dim) + self.norm2 = nn.LayerNorm(time_embed_dim) + + def forward(self, hidden_states): + hidden_states = self.norm1(hidden_states) + hidden_states = self.pool(hidden_states) + hidden_states = self.proj(hidden_states) + hidden_states = self.norm2(hidden_states) + return hidden_states + + +class TextImageTimeEmbedding(nn.Module): + def __init__(self, text_embed_dim: int = 768, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.text_proj = nn.Linear(text_embed_dim, time_embed_dim) + self.text_norm = nn.LayerNorm(time_embed_dim) + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + + def forward(self, text_embeds: torch.FloatTensor, image_embeds: torch.FloatTensor): + # text + time_text_embeds = self.text_proj(text_embeds) + time_text_embeds = self.text_norm(time_text_embeds) + + # image + time_image_embeds = self.image_proj(image_embeds) + + return time_image_embeds + time_text_embeds + + +class ImageTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + + def forward(self, image_embeds: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + return time_image_embeds + + +class ImageHintTimeEmbedding(nn.Module): + def __init__(self, image_embed_dim: int = 768, time_embed_dim: int = 1536): + super().__init__() + self.image_proj = nn.Linear(image_embed_dim, time_embed_dim) + self.image_norm = nn.LayerNorm(time_embed_dim) + self.input_hint_block = nn.Sequential( + nn.Conv2d(3, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 16, 3, padding=1), + nn.SiLU(), + nn.Conv2d(16, 32, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(32, 32, 3, padding=1), + nn.SiLU(), + nn.Conv2d(32, 96, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(96, 96, 3, padding=1), + nn.SiLU(), + nn.Conv2d(96, 256, 3, padding=1, stride=2), + nn.SiLU(), + nn.Conv2d(256, 4, 3, padding=1), + ) + + def forward(self, image_embeds: torch.FloatTensor, hint: torch.FloatTensor): + # image + time_image_embeds = self.image_proj(image_embeds) + time_image_embeds = self.image_norm(time_image_embeds) + hint = self.input_hint_block(hint) + return time_image_embeds, hint + + +class AttentionPooling(nn.Module): + # Copied from https://github.com/deep-floyd/IF/blob/2f91391f27dd3c468bf174be5805b4cc92980c0b/deepfloyd_if/model/nn.py#L54 + + def __init__(self, num_heads, embed_dim, dtype=None): + super().__init__() + self.dtype = dtype + self.positional_embedding = nn.Parameter(torch.randn(1, embed_dim) / embed_dim**0.5) + self.k_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.q_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.v_proj = nn.Linear(embed_dim, embed_dim, dtype=self.dtype) + self.num_heads = num_heads + self.dim_per_head = embed_dim // self.num_heads + + def forward(self, x): + bs, length, width = x.size() + + def shape(x): + # (bs, length, width) --> (bs, length, n_heads, dim_per_head) + x = x.view(bs, -1, self.num_heads, self.dim_per_head) + # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head) + x = x.transpose(1, 2) + # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head) + x = x.reshape(bs * self.num_heads, -1, self.dim_per_head) + # (bs*n_heads, length, dim_per_head) --> (bs*n_heads, dim_per_head, length) + x = x.transpose(1, 2) + return x + + class_token = x.mean(dim=1, keepdim=True) + self.positional_embedding.to(x.dtype) + x = torch.cat([class_token, x], dim=1) # (bs, length+1, width) + + # (bs*n_heads, class_token_length, dim_per_head) + q = shape(self.q_proj(class_token)) + # (bs*n_heads, length+class_token_length, dim_per_head) + k = shape(self.k_proj(x)) + v = shape(self.v_proj(x)) + + # (bs*n_heads, class_token_length, length+class_token_length): + scale = 1 / math.sqrt(math.sqrt(self.dim_per_head)) + weight = torch.einsum("bct,bcs->bts", q * scale, k * scale) # More stable with f16 than dividing afterwards + weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype) + + # (bs*n_heads, dim_per_head, class_token_length) + a = torch.einsum("bts,bcs->bct", weight, v) + + # (bs, length+1, width) + a = a.reshape(bs, -1, 1).transpose(1, 2) + + return a[:, 0, :] # cls_token + + +def get_fourier_embeds_from_boundingbox(embed_dim, box): + """ + Args: + embed_dim: int + box: a 3-D tensor [B x N x 4] representing the bounding boxes for GLIGEN pipeline + Returns: + [B x N x embed_dim] tensor of positional embeddings + """ + + batch_size, num_boxes = box.shape[:2] + + emb = 100 ** (torch.arange(embed_dim) / embed_dim) + emb = emb[None, None, None].to(device=box.device, dtype=box.dtype) + emb = emb * box.unsqueeze(-1) + + emb = torch.stack((emb.sin(), emb.cos()), dim=-1) + emb = emb.permute(0, 1, 3, 4, 2).reshape(batch_size, num_boxes, embed_dim * 2 * 4) + + return emb + + +class GLIGENTextBoundingboxProjection(nn.Module): + def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8): + super().__init__() + self.positive_len = positive_len + self.out_dim = out_dim + + self.fourier_embedder_dim = fourier_freqs + self.position_dim = fourier_freqs * 2 * 4 # 2: sin/cos, 4: xyxy + + if isinstance(out_dim, tuple): + out_dim = out_dim[0] + + if feature_type == "text-only": + self.linears = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_positive_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + elif feature_type == "text-image": + self.linears_text = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.linears_image = nn.Sequential( + nn.Linear(self.positive_len + self.position_dim, 512), + nn.SiLU(), + nn.Linear(512, 512), + nn.SiLU(), + nn.Linear(512, out_dim), + ) + self.null_text_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + self.null_image_feature = torch.nn.Parameter(torch.zeros([self.positive_len])) + + self.null_position_feature = torch.nn.Parameter(torch.zeros([self.position_dim])) + + def forward( + self, + boxes, + masks, + positive_embeddings=None, + phrases_masks=None, + image_masks=None, + phrases_embeddings=None, + image_embeddings=None, + ): + masks = masks.unsqueeze(-1) + + # embedding position (it may includes padding as placeholder) + xyxy_embedding = get_fourier_embeds_from_boundingbox(self.fourier_embedder_dim, boxes) # B*N*4 -> B*N*C + + # learnable null embedding + xyxy_null = self.null_position_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + xyxy_embedding = xyxy_embedding * masks + (1 - masks) * xyxy_null + + # positionet with text only information + if positive_embeddings is not None: + # learnable null embedding + positive_null = self.null_positive_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + positive_embeddings = positive_embeddings * masks + (1 - masks) * positive_null + + objs = self.linears(torch.cat([positive_embeddings, xyxy_embedding], dim=-1)) + + # positionet with text and image infomation + else: + phrases_masks = phrases_masks.unsqueeze(-1) + image_masks = image_masks.unsqueeze(-1) + + # learnable null embedding + text_null = self.null_text_feature.view(1, 1, -1) + image_null = self.null_image_feature.view(1, 1, -1) + + # replace padding with learnable null embedding + phrases_embeddings = phrases_embeddings * phrases_masks + (1 - phrases_masks) * text_null + image_embeddings = image_embeddings * image_masks + (1 - image_masks) * image_null + + objs_text = self.linears_text(torch.cat([phrases_embeddings, xyxy_embedding], dim=-1)) + objs_image = self.linears_image(torch.cat([image_embeddings, xyxy_embedding], dim=-1)) + objs = torch.cat([objs_text, objs_image], dim=1) + + return objs + + +class PixArtAlphaTextProjection(nn.Module): + """ + Projects caption embeddings. Also handles dropout for classifier-free guidance. + + Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py + """ + + def __init__(self, in_features, hidden_size, num_tokens=120): + super().__init__() + self.linear_1 = nn.Linear(in_features=in_features, out_features=hidden_size, bias=True) + self.act_1 = nn.GELU(approximate="tanh") + self.linear_2 = nn.Linear(in_features=hidden_size, out_features=hidden_size, bias=True) + + def forward(self, caption): + hidden_states = self.linear_1(caption) + hidden_states = self.act_1(hidden_states) + hidden_states = self.linear_2(hidden_states) + return hidden_states + + +class IPAdapterPlusImageProjection(nn.Module): + """Resampler of IP-Adapter Plus. + + Args: + ---- + embed_dims (int): The feature dimension. Defaults to 768. + output_dims (int): The number of output channels, that is the same + number of the channels in the + `unet.config.cross_attention_dim`. Defaults to 1024. + hidden_dims (int): The number of hidden channels. Defaults to 1280. + depth (int): The number of blocks. Defaults to 8. + dim_head (int): The number of head channels. Defaults to 64. + heads (int): Parallel attention heads. Defaults to 16. + num_queries (int): The number of queries. Defaults to 8. + ffn_ratio (float): The expansion ratio of feedforward network hidden + layer channels. Defaults to 4. + """ + + def __init__( + self, + embed_dims: int = 768, + output_dims: int = 1024, + hidden_dims: int = 1280, + depth: int = 4, + dim_head: int = 64, + heads: int = 16, + num_queries: int = 8, + ffn_ratio: float = 4, + ) -> None: + super().__init__() + from .attention import FeedForward # Lazy import to avoid circular import + + self.latents = nn.Parameter(torch.randn(1, num_queries, hidden_dims) / hidden_dims**0.5) + + self.proj_in = nn.Linear(embed_dims, hidden_dims) + + self.proj_out = nn.Linear(hidden_dims, output_dims) + self.norm_out = nn.LayerNorm(output_dims) + + self.layers = nn.ModuleList([]) + for _ in range(depth): + self.layers.append( + nn.ModuleList( + [ + nn.LayerNorm(hidden_dims), + nn.LayerNorm(hidden_dims), + Attention( + query_dim=hidden_dims, + dim_head=dim_head, + heads=heads, + out_bias=False, + ), + nn.Sequential( + nn.LayerNorm(hidden_dims), + FeedForward(hidden_dims, hidden_dims, activation_fn="gelu", mult=ffn_ratio, bias=False), + ), + ] + ) + ) + + def forward(self, x: torch.Tensor) -> torch.Tensor: + """Forward pass. + + Args: + ---- + x (torch.Tensor): Input Tensor. + + Returns: + ------- + torch.Tensor: Output Tensor. + """ + latents = self.latents.repeat(x.size(0), 1, 1) + + x = self.proj_in(x) + + for ln0, ln1, attn, ff in self.layers: + residual = latents + + encoder_hidden_states = ln0(x) + latents = ln1(latents) + encoder_hidden_states = torch.cat([encoder_hidden_states, latents], dim=-2) + latents = attn(latents, encoder_hidden_states) + residual + latents = ff(latents) + latents + + latents = self.proj_out(latents) + return self.norm_out(latents) + + +class MultiIPAdapterImageProjection(nn.Module): + def __init__(self, IPAdapterImageProjectionLayers: Union[List[nn.Module], Tuple[nn.Module]]): + super().__init__() + self.image_projection_layers = nn.ModuleList(IPAdapterImageProjectionLayers) + + def forward(self, image_embeds: List[torch.FloatTensor]): + projected_image_embeds = [] + + # currently, we accept `image_embeds` as + # 1. a tensor (deprecated) with shape [batch_size, embed_dim] or [batch_size, sequence_length, embed_dim] + # 2. list of `n` tensors where `n` is number of ip-adapters, each tensor can hae shape [batch_size, num_images, embed_dim] or [batch_size, num_images, sequence_length, embed_dim] + if not isinstance(image_embeds, list): + deprecation_message = ( + "You have passed a tensor as `image_embeds`.This is deprecated and will be removed in a future release." + " Please make sure to update your script to pass `image_embeds` as a list of tensors to supress this warning." + ) + deprecate("image_embeds not a list", "1.0.0", deprecation_message, standard_warn=False) + image_embeds = [image_embeds.unsqueeze(1)] + + if len(image_embeds) != len(self.image_projection_layers): + raise ValueError( + f"image_embeds must have the same length as image_projection_layers, got {len(image_embeds)} and {len(self.image_projection_layers)}" + ) + + for image_embed, image_projection_layer in zip(image_embeds, self.image_projection_layers): + batch_size, num_images = image_embed.shape[0], image_embed.shape[1] + image_embed = image_embed.reshape((batch_size * num_images,) + image_embed.shape[2:]) + image_embed = image_projection_layer(image_embed) + image_embed = image_embed.reshape((batch_size, num_images) + image_embed.shape[1:]) + + projected_image_embeds.append(image_embed) + + return projected_image_embeds diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/normalization.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/normalization.py new file mode 100644 index 0000000000000000000000000000000000000000..5acced9ea6a6c10c800e2560c6390e3134bc82a5 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/layers/normalization.py @@ -0,0 +1,254 @@ +# coding=utf-8 +# Copyright 2023 HuggingFace Inc. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import numbers +from typing import Dict, Optional, Tuple + +import torch +import torch.nn as nn +import torch.nn.functional as F + +from diffusers.utils import is_torch_version +from .activations import get_activation +from .embeddings import CombinedTimestepLabelEmbeddings, PixArtAlphaCombinedTimestepSizeEmbeddings + + +class AdaLayerNorm(nn.Module): + r""" + Norm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + self.emb = nn.Embedding(num_embeddings, embedding_dim) + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, embedding_dim * 2) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False) + + def forward(self, x: torch.Tensor, timestep: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(self.emb(timestep))) + scale, shift = torch.chunk(emb, 2) + x = self.norm(x) * (1 + scale) + shift + return x + + +class AdaLayerNormZero(nn.Module): + r""" + Norm layer adaptive layer norm zero (adaLN-Zero). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + """ + + def __init__(self, embedding_dim: int, num_embeddings: int): + super().__init__() + + self.emb = CombinedTimestepLabelEmbeddings(num_embeddings, embedding_dim) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + self.norm = nn.LayerNorm(embedding_dim, elementwise_affine=False, eps=1e-6) + + def forward( + self, + x: torch.Tensor, + timestep: torch.Tensor, + class_labels: torch.LongTensor, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + emb = self.linear(self.silu(self.emb(timestep, class_labels, hidden_dtype=hidden_dtype))) + shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = emb.chunk(6, dim=1) + x = self.norm(x) * (1 + scale_msa[:, None]) + shift_msa[:, None] + return x, gate_msa, shift_mlp, scale_mlp, gate_mlp + + +class AdaLayerNormSingle(nn.Module): + r""" + Norm layer adaptive layer norm single (adaLN-single). + + As proposed in PixArt-Alpha (see: https://arxiv.org/abs/2310.00426; Section 2.3). + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + use_additional_conditions (`bool`): To use additional conditions for normalization or not. + """ + + def __init__(self, embedding_dim: int, use_additional_conditions: bool = False): + super().__init__() + + self.emb = PixArtAlphaCombinedTimestepSizeEmbeddings( + embedding_dim, size_emb_dim=embedding_dim // 3, use_additional_conditions=use_additional_conditions + ) + + self.silu = nn.SiLU() + self.linear = nn.Linear(embedding_dim, 6 * embedding_dim, bias=True) + + def forward( + self, + timestep: torch.Tensor, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + batch_size: Optional[int] = None, + hidden_dtype: Optional[torch.dtype] = None, + ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]: + # No modulation happening here. + embedded_timestep = self.emb(timestep, **added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_dtype) + return self.linear(self.silu(embedded_timestep)), embedded_timestep + + +class AdaGroupNorm(nn.Module): + r""" + GroupNorm layer modified to incorporate timestep embeddings. + + Parameters: + embedding_dim (`int`): The size of each embedding vector. + num_embeddings (`int`): The size of the embeddings dictionary. + num_groups (`int`): The number of groups to separate the channels into. + act_fn (`str`, *optional*, defaults to `None`): The activation function to use. + eps (`float`, *optional*, defaults to `1e-5`): The epsilon value to use for numerical stability. + """ + + def __init__( + self, embedding_dim: int, out_dim: int, num_groups: int, act_fn: Optional[str] = None, eps: float = 1e-5 + ): + super().__init__() + self.num_groups = num_groups + self.eps = eps + + if act_fn is None: + self.act = None + else: + self.act = get_activation(act_fn) + + self.linear = nn.Linear(embedding_dim, out_dim * 2) + + def forward(self, x: torch.Tensor, emb: torch.Tensor) -> torch.Tensor: + if self.act: + emb = self.act(emb) + emb = self.linear(emb) + emb = emb[:, :, None, None] + scale, shift = emb.chunk(2, dim=1) + + x = F.group_norm(x, self.num_groups, eps=self.eps) + x = x * (1 + scale) + shift + return x + + +class AdaLayerNormContinuous(nn.Module): + def __init__( + self, + embedding_dim: int, + conditioning_embedding_dim: int, + # NOTE: It is a bit weird that the norm layer can be configured to have scale and shift parameters + # because the output is immediately scaled and shifted by the projected conditioning embeddings. + # Note that AdaLayerNorm does not let the norm layer have scale and shift parameters. + # However, this is how it was implemented in the original code, and it's rather likely you should + # set `elementwise_affine` to False. + elementwise_affine=True, + eps=1e-5, + bias=True, + norm_type="layer_norm", + ): + super().__init__() + self.silu = nn.SiLU() + self.linear = nn.Linear(conditioning_embedding_dim, embedding_dim * 2, bias=bias) + if norm_type == "layer_norm": + self.norm = LayerNorm(embedding_dim, eps, elementwise_affine, bias) + elif norm_type == "rms_norm": + self.norm = RMSNorm(embedding_dim, eps, elementwise_affine) + else: + raise ValueError(f"unknown norm_type {norm_type}") + + def forward(self, x: torch.Tensor, conditioning_embedding: torch.Tensor) -> torch.Tensor: + emb = self.linear(self.silu(conditioning_embedding)) + scale, shift = torch.chunk(emb, 2, dim=1) + x = self.norm(x) * (1 + scale)[:, None, :] + shift[:, None, :] + return x + + +if is_torch_version(">=", "2.1.0"): + LayerNorm = nn.LayerNorm +else: + # Has optional bias parameter compared to torch layer norm + # TODO: replace with torch layernorm once min required torch version >= 2.1 + class LayerNorm(nn.Module): + def __init__(self, dim, eps: float = 1e-5, elementwise_affine: bool = True, bias: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + self.bias = nn.Parameter(torch.zeros(dim)) if bias else None + else: + self.weight = None + self.bias = None + + def forward(self, input): + return F.layer_norm(input, self.dim, self.weight, self.bias, self.eps) + + +class RMSNorm(nn.Module): + def __init__(self, dim, eps: float, elementwise_affine: bool = True): + super().__init__() + + self.eps = eps + + if isinstance(dim, numbers.Integral): + dim = (dim,) + + self.dim = torch.Size(dim) + + if elementwise_affine: + self.weight = nn.Parameter(torch.ones(dim)) + else: + self.weight = None + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.eps) + + if self.weight is not None: + # convert into half-precision if necessary + if self.weight.dtype in [torch.float16, torch.bfloat16]: + hidden_states = hidden_states.to(self.weight.dtype) + hidden_states = hidden_states * self.weight + else: + hidden_states = hidden_states.to(input_dtype) + + return hidden_states + + +class GlobalResponseNorm(nn.Module): + # Taken from https://github.com/facebookresearch/ConvNeXt-V2/blob/3608f67cc1dae164790c5d0aead7bf2d73d9719b/models/utils.py#L105 + def __init__(self, dim): + super().__init__() + self.gamma = nn.Parameter(torch.zeros(1, 1, 1, dim)) + self.beta = nn.Parameter(torch.zeros(1, 1, 1, dim)) + + def forward(self, x): + gx = torch.norm(x, p=2, dim=(1, 2), keepdim=True) + nx = gx / (gx.mean(dim=-1, keepdim=True) + 1e-6) + return self.gamma * (x * nx) + self.beta + x diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..5af450275b73656a52deb912765cab1461ffd090 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/__init__.py @@ -0,0 +1 @@ +from .unets.unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/model_utils.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/model_utils.py new file mode 100644 index 0000000000000000000000000000000000000000..0c9d557c2e7ca0446054d8ec36a71ee1f9de7d80 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/model_utils.py @@ -0,0 +1 @@ +from diffusers.models.modeling_utils import ModelMixin \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..f5429a94ed6f940af278e7a7b7050fadf912f627 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/__init__.py @@ -0,0 +1,2 @@ +from .dual_transformer_2d import DualTransformer2DModel +from .transformer_2d import Transformer2DModel \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/dual_transformer_2d.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/dual_transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..21b135c2eb86a6cc1ec2799c61c094e814893267 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/dual_transformer_2d.py @@ -0,0 +1,155 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Optional + +from torch import nn + +from .transformer_2d import Transformer2DModel, Transformer2DModelOutput + + +class DualTransformer2DModel(nn.Module): + """ + Dual transformer wrapper that combines two `Transformer2DModel`s for mixed inference. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + Pass if the input is continuous. The number of channels in the input and output. + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.1): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of encoder_hidden_states dimensions to use. + sample_size (`int`, *optional*): Pass if the input is discrete. The width of the latent images. + Note that this is fixed at training time as it is used for learning a number of position embeddings. See + `ImagePositionalEmbeddings`. + num_vector_embeds (`int`, *optional*): + Pass if the input is discrete. The number of classes of the vector embeddings of the latent pixels. + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): Pass if at least one of the norm_layers is `AdaLayerNorm`. + The number of diffusion steps used during training. Note that this is fixed at training time as it is used + to learn a number of embeddings that are added to the hidden states. During inference, you can denoise for + up to but not more than steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the TransformerBlocks' attention should contain a bias parameter. + """ + + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + ): + super().__init__() + self.transformers = nn.ModuleList( + [ + Transformer2DModel( + num_attention_heads=num_attention_heads, + attention_head_dim=attention_head_dim, + in_channels=in_channels, + num_layers=num_layers, + dropout=dropout, + norm_num_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim, + attention_bias=attention_bias, + sample_size=sample_size, + num_vector_embeds=num_vector_embeds, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + ) + for _ in range(2) + ] + ) + + # Variables that can be set by a pipeline: + + # The ratio of transformer1 to transformer2's output states to be combined during inference + self.mix_ratio = 0.5 + + # The shape of `encoder_hidden_states` is expected to be + # `(batch_size, condition_lengths[0]+condition_lengths[1], num_features)` + self.condition_lengths = [77, 257] + + # Which transformer to use to encode which condition. + # E.g. `(1, 0)` means that we'll use `transformers[1](conditions[0])` and `transformers[0](conditions[1])` + self.transformer_index_for_condition = [1, 0] + + def forward( + self, + hidden_states, + encoder_hidden_states, + timestep=None, + attention_mask=None, + cross_attention_kwargs=None, + return_dict: bool = True, + ): + """ + Args: + hidden_states ( When discrete, `torch.LongTensor` of shape `(batch size, num latent pixels)`. + When continuous, `torch.FloatTensor` of shape `(batch size, channel, height, width)`): Input + hidden_states. + encoder_hidden_states ( `torch.LongTensor` of shape `(batch size, encoder_hidden_states dim)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.long`, *optional*): + Optional timestep to be applied as an embedding in AdaLayerNorm's. Used to indicate denoising step. + attention_mask (`torch.FloatTensor`, *optional*): + Optional attention mask to be applied in Attention. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain tuple. + + Returns: + [`~models.transformer_2d.Transformer2DModelOutput`] or `tuple`: + [`~models.transformer_2d.Transformer2DModelOutput`] if `return_dict` is True, otherwise a `tuple`. When + returning a tuple, the first element is the sample tensor. + """ + input_states = hidden_states + + encoded_states = [] + tokens_start = 0 + # attention_mask is not used yet + for i in range(2): + # for each of the two transformers, pass the corresponding condition tokens + condition_state = encoder_hidden_states[:, tokens_start : tokens_start + self.condition_lengths[i]] + transformer_index = self.transformer_index_for_condition[i] + encoded_state = self.transformers[transformer_index]( + input_states, + encoder_hidden_states=condition_state, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + return_dict=False, + )[0] + encoded_states.append(encoded_state - input_states) + tokens_start += self.condition_lengths[i] + + output_states = encoded_states[0] * self.mix_ratio + encoded_states[1] * (1 - self.mix_ratio) + output_states = output_states + input_states + + if not return_dict: + return (output_states,) + + return Transformer2DModelOutput(sample=output_states) diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/transformer_2d.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/transformer_2d.py new file mode 100644 index 0000000000000000000000000000000000000000..e33621f2497a737e92cd1e6e5d7386ea1f83b76c --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/transformers/transformer_2d.py @@ -0,0 +1,459 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, Optional + +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version +from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear + +from stablediffusion.layers.attention import BasicTransformerBlock +from stablediffusion.layers.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection +from stablediffusion.layers.normalization import AdaLayerNormSingle +from stablediffusion.models.model_utils import ModelMixin + + +@dataclass +class Transformer2DModelOutput(BaseOutput): + """ + The output of [`Transformer2DModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete): + The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability + distributions for the unnoised latent pixels. + """ + + sample: torch.FloatTensor + + +class Transformer2DModel(ModelMixin, ConfigMixin): + """ + A 2D Transformer model for image-like data. + + Parameters: + num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention. + attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head. + in_channels (`int`, *optional*): + The number of channels in the input and output (specify if the input is **continuous**). + num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use. + sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**). + This is fixed during training since it is used to learn a number of position embeddings. + num_vector_embeds (`int`, *optional*): + The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**). + Includes the class for the masked latent pixel. + activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward. + num_embeds_ada_norm ( `int`, *optional*): + The number of diffusion steps used during training. Pass if at least one of the norm_layers is + `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are + added to the hidden states. + + During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`. + attention_bias (`bool`, *optional*): + Configure if the `TransformerBlocks` attention should contain a bias parameter. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + num_attention_heads: int = 16, + attention_head_dim: int = 88, + in_channels: Optional[int] = None, + out_channels: Optional[int] = None, + num_layers: int = 1, + dropout: float = 0.0, + norm_num_groups: int = 32, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + sample_size: Optional[int] = None, + num_vector_embeds: Optional[int] = None, + patch_size: Optional[int] = None, + activation_fn: str = "geglu", + num_embeds_ada_norm: Optional[int] = None, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + double_self_attention: bool = False, + upcast_attention: bool = False, + norm_type: str = "layer_norm", + norm_elementwise_affine: bool = True, + norm_eps: float = 1e-5, + attention_type: str = "default", + caption_channels: int = None, + ): + super().__init__() + self.use_linear_projection = use_linear_projection + self.num_attention_heads = num_attention_heads + self.attention_head_dim = attention_head_dim + inner_dim = num_attention_heads * attention_head_dim + + conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv + linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear + + # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)` + # Define whether input is continuous or discrete depending on configuration + self.is_input_continuous = (in_channels is not None) and (patch_size is None) + self.is_input_vectorized = num_vector_embeds is not None + self.is_input_patches = in_channels is not None and patch_size is not None + + if norm_type == "layer_norm" and num_embeds_ada_norm is not None: + deprecation_message = ( + f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or" + " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config." + " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect" + " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it" + " would be very nice if you could open a Pull request for the `transformer/config.json` file" + ) + deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False) + norm_type = "ada_norm" + + if self.is_input_continuous and self.is_input_vectorized: + raise ValueError( + f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make" + " sure that either `in_channels` or `num_vector_embeds` is None." + ) + elif self.is_input_vectorized and self.is_input_patches: + raise ValueError( + f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make" + " sure that either `num_vector_embeds` or `num_patches` is None." + ) + elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches: + raise ValueError( + f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:" + f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None." + ) + + # 2. Define input layers + if self.is_input_continuous: + self.in_channels = in_channels + + self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True) + if use_linear_projection: + self.proj_in = linear_cls(in_channels, inner_dim) + else: + self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size" + assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed" + + self.height = sample_size + self.width = sample_size + self.num_vector_embeds = num_vector_embeds + self.num_latent_pixels = self.height * self.width + + self.latent_image_embedding = ImagePositionalEmbeddings( + num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width + ) + elif self.is_input_patches: + assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size" + + self.height = sample_size + self.width = sample_size + + self.patch_size = patch_size + interpolation_scale = self.config.sample_size // 64 # => 64 (= 512 pixart) has interpolation scale 1 + interpolation_scale = max(interpolation_scale, 1) + self.pos_embed = PatchEmbed( + height=sample_size, + width=sample_size, + patch_size=patch_size, + in_channels=in_channels, + embed_dim=inner_dim, + interpolation_scale=interpolation_scale, + ) + + # 3. Define transformers blocks + self.transformer_blocks = nn.ModuleList( + [ + BasicTransformerBlock( + inner_dim, + num_attention_heads, + attention_head_dim, + dropout=dropout, + cross_attention_dim=cross_attention_dim, + activation_fn=activation_fn, + num_embeds_ada_norm=num_embeds_ada_norm, + attention_bias=attention_bias, + only_cross_attention=only_cross_attention, + double_self_attention=double_self_attention, + upcast_attention=upcast_attention, + norm_type=norm_type, + norm_elementwise_affine=norm_elementwise_affine, + norm_eps=norm_eps, + attention_type=attention_type, + ) + for d in range(num_layers) + ] + ) + + # 4. Define output layers + self.out_channels = in_channels if out_channels is None else out_channels + if self.is_input_continuous: + # TODO: should use out_channels for continuous projections + if use_linear_projection: + self.proj_out = linear_cls(inner_dim, in_channels) + else: + self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0) + elif self.is_input_vectorized: + self.norm_out = nn.LayerNorm(inner_dim) + self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1) + elif self.is_input_patches and norm_type != "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim) + self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + elif self.is_input_patches and norm_type == "ada_norm_single": + self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6) + self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5) + self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels) + + # 5. PixArt-Alpha blocks. + self.adaln_single = None + self.use_additional_conditions = False + if norm_type == "ada_norm_single": + self.use_additional_conditions = self.config.sample_size == 128 + # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use + # additional conditions until we find better name + self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions) + + self.caption_projection = None + if caption_channels is not None: + self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim) + + self.gradient_checkpointing = False + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def forward( + self, + hidden_states: torch.Tensor, + encoder_hidden_states: Optional[torch.Tensor] = None, + timestep: Optional[torch.LongTensor] = None, + added_cond_kwargs: Dict[str, torch.Tensor] = None, + class_labels: Optional[torch.LongTensor] = None, + cross_attention_kwargs: Dict[str, Any] = None, + attention_mask: Optional[torch.Tensor] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ): + """ + The [`Transformer2DModel`] forward method. + + Args: + hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous): + Input `hidden_states`. + encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*): + Conditional embeddings for cross attention layer. If not given, cross-attention defaults to + self-attention. + timestep ( `torch.LongTensor`, *optional*): + Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`. + class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*): + Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in + `AdaLayerZeroNorm`. + cross_attention_kwargs ( `Dict[str, Any]`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + attention_mask ( `torch.Tensor`, *optional*): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + encoder_attention_mask ( `torch.Tensor`, *optional*): + Cross-attention mask applied to `encoder_hidden_states`. Two formats supported: + + * Mask `(batch, sequence_length)` True = keep, False = discard. + * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard. + + If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format + above. This bias will be added to the cross-attention scores. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + + Returns: + If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a + `tuple` where the first element is the sample tensor. + """ + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension. + # we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward. + # we can tell by counting dims; if ndim == 2: it's a mask rather than a bias. + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None and attention_mask.ndim == 2: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2: + encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # Retrieve lora scale. + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + # 1. Input + if self.is_input_continuous: + batch, _, height, width = hidden_states.shape + residual = hidden_states + + hidden_states = self.norm(hidden_states) + if not self.use_linear_projection: + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + else: + inner_dim = hidden_states.shape[1] + hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim) + hidden_states = ( + self.proj_in(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_in(hidden_states) + ) + + elif self.is_input_vectorized: + hidden_states = self.latent_image_embedding(hidden_states) + elif self.is_input_patches: + height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size + hidden_states = self.pos_embed(hidden_states) + + if self.adaln_single is not None: + if self.use_additional_conditions and added_cond_kwargs is None: + raise ValueError( + "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`." + ) + batch_size = hidden_states.shape[0] + timestep, embedded_timestep = self.adaln_single( + timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype + ) + + # 2. Blocks + if self.caption_projection is not None: + batch_size = hidden_states.shape[0] + encoder_hidden_states = self.caption_projection(encoder_hidden_states) + encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1]) + + for block in self.transformer_blocks: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + attention_mask, + encoder_hidden_states, + encoder_attention_mask, + timestep, + cross_attention_kwargs, + class_labels, + **ckpt_kwargs, + ) + else: + hidden_states = block( + hidden_states, + attention_mask=attention_mask, + encoder_hidden_states=encoder_hidden_states, + encoder_attention_mask=encoder_attention_mask, + timestep=timestep, + cross_attention_kwargs=cross_attention_kwargs, + class_labels=class_labels, + ) + + # 3. Output + if self.is_input_continuous: + if not self.use_linear_projection: + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + else: + hidden_states = ( + self.proj_out(hidden_states, scale=lora_scale) + if not USE_PEFT_BACKEND + else self.proj_out(hidden_states) + ) + hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous() + + output = hidden_states + residual + elif self.is_input_vectorized: + hidden_states = self.norm_out(hidden_states) + logits = self.out(hidden_states) + # (batch, self.num_vector_embeds - 1, self.num_latent_pixels) + logits = logits.permute(0, 2, 1) + + # log(p(x_0)) + output = F.log_softmax(logits.double(), dim=1).float() + + if self.is_input_patches: + if self.config.norm_type != "ada_norm_single": + conditioning = self.transformer_blocks[0].norm1.emb( + timestep, class_labels, hidden_dtype=hidden_states.dtype + ) + shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None] + hidden_states = self.proj_out_2(hidden_states) + elif self.config.norm_type == "ada_norm_single": + shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1) + hidden_states = self.norm_out(hidden_states) + # Modulation + hidden_states = hidden_states * (1 + scale) + shift + hidden_states = self.proj_out(hidden_states) + hidden_states = hidden_states.squeeze(1) + + # unpatchify + if self.adaln_single is None: + height = width = int(hidden_states.shape[1] ** 0.5) + hidden_states = hidden_states.reshape( + shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels) + ) + hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states) + output = hidden_states.reshape( + shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size) + ) + + if not return_dict: + return (output,) + + return Transformer2DModelOutput(sample=output) diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..1d3437072dac140d915d4e7922e1edbf31ed34de --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/__init__.py @@ -0,0 +1 @@ +from .unet_2d_condition import UNet2DConditionModel \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_blocks.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_blocks.py new file mode 100644 index 0000000000000000000000000000000000000000..b7d60b8bf0ce78a88f3f6cf29093e54812cc6cb4 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_blocks.py @@ -0,0 +1,3592 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Any, Dict, Optional, Tuple, Union + +import numpy as np +import torch +import torch.nn.functional as F +from torch import nn + +from diffusers.utils import is_torch_version, logging +from diffusers.utils.torch_utils import apply_freeu +from diffusers.models.resnet import ( + Downsample2D, + FirDownsample2D, + FirUpsample2D, + KDownsample2D, + KUpsample2D, + ResnetBlock2D, + ResnetBlockCondNorm2D, + Upsample2D, +) + +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0 +from stablediffusion.layers.normalization import AdaGroupNorm +from ..transformers.dual_transformer_2d import DualTransformer2DModel +from ..transformers.transformer_2d import Transformer2DModel + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +def get_down_block( + down_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + temb_channels: int, + add_downsample: bool, + resnet_eps: float, + resnet_act_fn: str, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + downsample_padding: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + downsample_type: Optional[str] = None, + dropout: float = 0.0, +): + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type + if down_block_type == "DownBlock2D": + return DownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "ResnetDownsampleBlock2D": + return ResnetDownsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif down_block_type == "AttnDownBlock2D": + if add_downsample is False: + downsample_type = None + else: + downsample_type = downsample_type or "conv" # default to 'conv' + return AttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + downsample_type=downsample_type, + ) + elif down_block_type == "CrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D") + return CrossAttnDownBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif down_block_type == "SimpleCrossAttnDownBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D") + return SimpleCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif down_block_type == "SkipDownBlock2D": + return SkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnSkipDownBlock2D": + return AttnSkipDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "DownEncoderBlock2D": + return DownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "AttnDownEncoderBlock2D": + return AttnDownEncoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + downsample_padding=downsample_padding, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif down_block_type == "KDownBlock2D": + return KDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif down_block_type == "KCrossAttnDownBlock2D": + return KCrossAttnDownBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + dropout=dropout, + add_downsample=add_downsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + add_self_attention=True if not add_downsample else False, + ) + raise ValueError(f"{down_block_type} does not exist.") + + +def get_up_block( + up_block_type: str, + num_layers: int, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + add_upsample: bool, + resnet_eps: float, + resnet_act_fn: str, + resolution_idx: Optional[int] = None, + transformer_layers_per_block: int = 1, + num_attention_heads: Optional[int] = None, + resnet_groups: Optional[int] = None, + cross_attention_dim: Optional[int] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + attention_type: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: float = 1.0, + cross_attention_norm: Optional[str] = None, + attention_head_dim: Optional[int] = None, + upsample_type: Optional[str] = None, + dropout: float = 0.0, +) -> nn.Module: + # If attn head dim is not defined, we default it to the number of heads + if attention_head_dim is None: + logger.warn( + f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}." + ) + attention_head_dim = num_attention_heads + + up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type + if up_block_type == "UpBlock2D": + return UpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "ResnetUpsampleBlock2D": + return ResnetUpsampleBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + ) + elif up_block_type == "CrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D") + return CrossAttnUpBlock2D( + num_layers=num_layers, + transformer_layers_per_block=transformer_layers_per_block, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + num_attention_heads=num_attention_heads, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + ) + elif up_block_type == "SimpleCrossAttnUpBlock2D": + if cross_attention_dim is None: + raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D") + return SimpleCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + output_scale_factor=resnet_out_scale_factor, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif up_block_type == "AttnUpBlock2D": + if add_upsample is False: + upsample_type = None + else: + upsample_type = upsample_type or "conv" # default to 'conv' + + return AttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + upsample_type=upsample_type, + ) + elif up_block_type == "SkipUpBlock2D": + return SkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "AttnSkipUpBlock2D": + return AttnSkipUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + prev_output_channel=prev_output_channel, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + ) + elif up_block_type == "UpDecoderBlock2D": + return UpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "AttnUpDecoderBlock2D": + return AttnUpDecoderBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + resnet_groups=resnet_groups, + attention_head_dim=attention_head_dim, + resnet_time_scale_shift=resnet_time_scale_shift, + temb_channels=temb_channels, + ) + elif up_block_type == "KUpBlock2D": + return KUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + ) + elif up_block_type == "KCrossAttnUpBlock2D": + return KCrossAttnUpBlock2D( + num_layers=num_layers, + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + resolution_idx=resolution_idx, + dropout=dropout, + add_upsample=add_upsample, + resnet_eps=resnet_eps, + resnet_act_fn=resnet_act_fn, + cross_attention_dim=cross_attention_dim, + attention_head_dim=attention_head_dim, + ) + + raise ValueError(f"{up_block_type} does not exist.") + + +class AutoencoderTinyBlock(nn.Module): + """ + Tiny Autoencoder block used in [`AutoencoderTiny`]. It is a mini residual module consisting of plain conv + ReLU + blocks. + + Args: + in_channels (`int`): The number of input channels. + out_channels (`int`): The number of output channels. + act_fn (`str`): + ` The activation function to use. Supported values are `"swish"`, `"mish"`, `"gelu"`, and `"relu"`. + + Returns: + `torch.FloatTensor`: A tensor with the same shape as the input tensor, but with the number of channels equal to + `out_channels`. + """ + + def __init__(self, in_channels: int, out_channels: int, act_fn: str): + super().__init__() + act_fn = get_activation(act_fn) + self.conv = nn.Sequential( + nn.Conv2d(in_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + act_fn, + nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1), + ) + self.skip = ( + nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False) + if in_channels != out_channels + else nn.Identity() + ) + self.fuse = nn.ReLU() + + def forward(self, x: torch.FloatTensor) -> torch.FloatTensor: + return self.fuse(self.conv(x) + self.skip(x)) + + +class UNetMidBlock2D(nn.Module): + """ + A 2D UNet mid-block [`UNetMidBlock2D`] with multiple residual blocks and optional attention blocks. + + Args: + in_channels (`int`): The number of input channels. + temb_channels (`int`): The number of temporal embedding channels. + dropout (`float`, *optional*, defaults to 0.0): The dropout rate. + num_layers (`int`, *optional*, defaults to 1): The number of residual blocks. + resnet_eps (`float`, *optional*, 1e-6 ): The epsilon value for the resnet blocks. + resnet_time_scale_shift (`str`, *optional*, defaults to `default`): + The type of normalization to apply to the time embeddings. This can help to improve the performance of the + model on tasks with long-range temporal dependencies. + resnet_act_fn (`str`, *optional*, defaults to `swish`): The activation function for the resnet blocks. + resnet_groups (`int`, *optional*, defaults to 32): + The number of groups to use in the group normalization layers of the resnet blocks. + attn_groups (`Optional[int]`, *optional*, defaults to None): The number of groups for the attention blocks. + resnet_pre_norm (`bool`, *optional*, defaults to `True`): + Whether to use pre-normalization for the resnet blocks. + add_attention (`bool`, *optional*, defaults to `True`): Whether to add attention blocks. + attention_head_dim (`int`, *optional*, defaults to 1): + Dimension of a single attention head. The number of attention heads is determined based on this value and + the number of input channels. + output_scale_factor (`float`, *optional*, defaults to 1.0): The output scale factor. + + Returns: + `torch.FloatTensor`: The output of the last residual block, which is a tensor of shape `(batch_size, + in_channels, height, width)`. + + """ + + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + attn_groups: Optional[int] = None, + resnet_pre_norm: bool = True, + add_attention: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + ): + super().__init__() + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + self.add_attention = add_attention + + if attn_groups is None: + attn_groups = resnet_groups if resnet_time_scale_shift == "default" else None + + # there is always at least one resnet + if resnet_time_scale_shift == "spatial": + resnets = [ + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ] + else: + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {in_channels}." + ) + attention_head_dim = in_channels + + for _ in range(num_layers): + if self.add_attention: + attentions.append( + Attention( + in_channels, + heads=in_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=attn_groups, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + else: + attentions.append(None) + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward(self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None) -> torch.FloatTensor: + hidden_states = self.resnets[0](hidden_states, temb) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if attn is not None: + hidden_states = attn(hidden_states, temb=temb) + hidden_states = resnet(hidden_states, temb) + + return hidden_states + + +class UNetMidBlock2DCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + # support for variable transformer layers per block + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ] + attentions = [] + + for i in range(num_layers): + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + in_channels // num_attention_heads, + in_channels=in_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + else: + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class UNetMidBlock2DSimpleCrossAttn(nn.Module): + def __init__( + self, + in_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + cross_attention_dim: int = 1280, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + self.attention_head_dim = attention_head_dim + resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32) + + self.num_heads = in_channels // self.attention_head_dim + + # there is always at least one resnet + resnets = [ + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ] + attentions = [] + + for _ in range(num_layers): + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=in_channels, + cross_attention_dim=in_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=in_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + hidden_states = self.resnets[0](hidden_states, temb, scale=lora_scale) + for attn, resnet in zip(self.attentions, self.resnets[1:]): + # attn + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + # resnet + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class AttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + downsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + self.downsample_type = downsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if downsample_type == "conv": + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + elif downsample_type == "resnet": + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + cross_attention_kwargs.update({"scale": lora_scale}) + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + if self.downsample_type == "resnet": + hidden_states = downsampler(hidden_states, temb=temb, scale=lora_scale) + else: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states += (hidden_states,) + + return hidden_states, output_states + + +class CrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + downsample_padding: int = 1, + add_downsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + additional_residuals: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + blocks = list(zip(self.resnets, self.attentions)) + + for i, (resnet, attn) in enumerate(blocks): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + # apply additional residuals to the output of the last pair of resnet and attention blocks + if i == len(blocks) - 1 and additional_residuals is not None: + hidden_states = hidden_states + additional_residuals + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale=scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class DownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=None, scale=scale) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnDownEncoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=None, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + Downsample2D( + out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op" + ) + ] + ) + else: + self.downsamplers = None + + def forward(self, hidden_states: torch.FloatTensor, scale: float = 1.0) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=None, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, scale) + + return hidden_states + + +class AttnSkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale=scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class SkipDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_downsample: bool = True, + downsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + self.resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(in_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if add_downsample: + self.resnet_down = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + down=True, + kernel="fir", + ) + self.downsamplers = nn.ModuleList([FirDownsample2D(out_channels, out_channels=out_channels)]) + self.skip_conv = nn.Conv2d(3, out_channels, kernel_size=(1, 1), stride=(1, 1)) + else: + self.resnet_down = None + self.downsamplers = None + self.skip_conv = None + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + skip_sample: Optional[torch.FloatTensor] = None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...], torch.FloatTensor]: + output_states = () + + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb, scale) + output_states += (hidden_states,) + + if self.downsamplers is not None: + hidden_states = self.resnet_down(hidden_states, temb, scale) + for downsampler in self.downsamplers: + skip_sample = downsampler(skip_sample) + + hidden_states = self.skip_conv(skip_sample) + hidden_states + + output_states += (hidden_states,) + + return hidden_states, output_states, skip_sample + + +class ResnetDownsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class SimpleCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_downsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + + self.has_cross_attention = True + + resnets = [] + attentions = [] + + self.attention_head_dim = attention_head_dim + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + resnets.append( + ResnetBlock2D( + in_channels=in_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + self.downsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + down=True, + ) + ] + ) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + output_states = output_states + (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states, temb, scale=lora_scale) + + output_states = output_states + (hidden_states,) + + return hidden_states, output_states + + +class KDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + add_downsample: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_downsample: + # YiYi's comments- might be able to use FirDownsample2D, look into details later + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale) + + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class KCrossAttnDownBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + cross_attention_dim: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_group_size: int = 32, + add_downsample: bool = True, + attention_head_dim: int = 64, + add_self_attention: bool = False, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + + for i in range(num_layers): + in_channels = in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + dropout=dropout, + temb_channels=temb_channels, + groups=groups, + groups_out=groups_out, + eps=resnet_eps, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + out_channels, + out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + group_size=resnet_group_size, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_downsample: + self.downsamplers = nn.ModuleList([KDownsample2D()]) + else: + self.downsamplers = None + + self.gradient_checkpointing = False + + def forward( + self, + hidden_states: torch.FloatTensor, + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> Tuple[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]: + output_states = () + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.downsamplers is None: + output_states += (None,) + else: + output_states += (hidden_states,) + + if self.downsamplers is not None: + for downsampler in self.downsamplers: + hidden_states = downsampler(hidden_states) + + return hidden_states, output_states + + +class AttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: int = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + upsample_type: str = "conv", + ): + super().__init__() + resnets = [] + attentions = [] + + self.upsample_type = upsample_type + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `in_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if upsample_type == "conv": + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + elif upsample_type == "resnet": + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + if self.upsample_type == "resnet": + hidden_states = upsampler(hidden_states, temb=temb, scale=scale) + else: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class CrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + transformer_layers_per_block: Union[int, Tuple[int]] = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + num_attention_heads: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + only_cross_attention: bool = False, + upcast_attention: bool = False, + attention_type: str = "default", + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.num_attention_heads = num_attention_heads + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * num_layers + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + if not dual_cross_attention: + attentions.append( + Transformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=transformer_layers_per_block[i], + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + ) + else: + attentions.append( + DualTransformer2DModel( + num_attention_heads, + out_channels // num_attention_heads, + in_channels=out_channels, + num_layers=1, + cross_attention_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet, attn in zip(self.resnets, self.attentions): + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + return_dict=False, + )[0] + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=lora_scale) + + return hidden_states + + +class UpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + is_freeu_enabled = ( + getattr(self, "s1", None) + and getattr(self, "s2", None) + and getattr(self, "b1", None) + and getattr(self, "b2", None) + ) + + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + + # FreeU: Only operate on the first two stages + if is_freeu_enabled: + hidden_states, res_hidden_states = apply_freeu( + self.resolution_idx, + hidden_states, + res_hidden_states, + s1=self.s1, + s2=self.s2, + b1=self.b1, + b2=self.b2, + ) + + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, upsample_size, scale=scale) + + return hidden_states + + +class UpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", # default, spatial + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet in self.resnets: + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class AttnUpDecoderBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + temb_channels: Optional[int] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + for i in range(num_layers): + input_channels = in_channels if i == 0 else out_channels + + if resnet_time_scale_shift == "spatial": + resnets.append( + ResnetBlockCondNorm2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm="spatial", + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + ) + ) + else: + resnets.append( + ResnetBlock2D( + in_channels=input_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=resnet_groups if resnet_time_scale_shift != "spatial" else None, + spatial_norm_dim=temb_channels if resnet_time_scale_shift == "spatial" else None, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)]) + else: + self.upsamplers = None + + self.resolution_idx = resolution_idx + + def forward( + self, hidden_states: torch.FloatTensor, temb: Optional[torch.FloatTensor] = None, scale: float = 1.0 + ) -> torch.FloatTensor: + for resnet, attn in zip(self.resnets, self.attentions): + hidden_states = resnet(hidden_states, temb=temb, scale=scale) + cross_attention_kwargs = {"scale": scale} + hidden_states = attn(hidden_states, temb=temb, **cross_attention_kwargs) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, scale=scale) + + return hidden_states + + +class AttnSkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + ): + super().__init__() + self.attentions = nn.ModuleList([]) + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(resnet_in_channels + res_skip_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + if attention_head_dim is None: + logger.warn( + f"It is not recommend to pass `attention_head_dim=None`. Defaulting `attention_head_dim` to `out_channels`: {out_channels}." + ) + attention_head_dim = out_channels + + self.attentions.append( + Attention( + out_channels, + heads=out_channels // attention_head_dim, + dim_head=attention_head_dim, + rescale_output_factor=output_scale_factor, + eps=resnet_eps, + norm_num_groups=32, + residual_connection=True, + bias=True, + upcast_softmax=True, + _from_deprecated_attn_block=True, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + cross_attention_kwargs = {"scale": scale} + hidden_states = self.attentions[0](hidden_states, **cross_attention_kwargs) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class SkipUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_pre_norm: bool = True, + output_scale_factor: float = np.sqrt(2.0), + add_upsample: bool = True, + upsample_padding: int = 1, + ): + super().__init__() + self.resnets = nn.ModuleList([]) + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + self.resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min((resnet_in_channels + res_skip_channels) // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + ) + ) + + self.upsampler = FirUpsample2D(in_channels, out_channels=out_channels) + if add_upsample: + self.resnet_up = ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=min(out_channels // 4, 32), + groups_out=min(out_channels // 4, 32), + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + use_in_shortcut=True, + up=True, + kernel="fir", + ) + self.skip_conv = nn.Conv2d(out_channels, 3, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) + self.skip_norm = torch.nn.GroupNorm( + num_groups=min(out_channels // 4, 32), num_channels=out_channels, eps=resnet_eps, affine=True + ) + self.act = nn.SiLU() + else: + self.resnet_up = None + self.skip_conv = None + self.skip_norm = None + self.act = None + + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + skip_sample=None, + scale: float = 1.0, + ) -> Tuple[torch.FloatTensor, torch.FloatTensor]: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + hidden_states = resnet(hidden_states, temb, scale=scale) + + if skip_sample is not None: + skip_sample = self.upsampler(skip_sample) + else: + skip_sample = 0 + + if self.resnet_up is not None: + skip_sample_states = self.skip_norm(hidden_states) + skip_sample_states = self.act(skip_sample_states) + skip_sample_states = self.skip_conv(skip_sample_states) + + skip_sample = skip_sample + skip_sample_states + + hidden_states = self.resnet_up(hidden_states, temb, scale=scale) + + return hidden_states, skip_sample + + +class ResnetUpsampleBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + prev_output_channel: int, + out_channels: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + ): + super().__init__() + resnets = [] + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + for resnet in self.resnets: + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=scale) + + return hidden_states + + +class SimpleCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + prev_output_channel: int, + temb_channels: int, + resolution_idx: Optional[int] = None, + dropout: float = 0.0, + num_layers: int = 1, + resnet_eps: float = 1e-6, + resnet_time_scale_shift: str = "default", + resnet_act_fn: str = "swish", + resnet_groups: int = 32, + resnet_pre_norm: bool = True, + attention_head_dim: int = 1, + cross_attention_dim: int = 1280, + output_scale_factor: float = 1.0, + add_upsample: bool = True, + skip_time_act: bool = False, + only_cross_attention: bool = False, + cross_attention_norm: Optional[str] = None, + ): + super().__init__() + resnets = [] + attentions = [] + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + self.num_heads = out_channels // self.attention_head_dim + + for i in range(num_layers): + res_skip_channels = in_channels if (i == num_layers - 1) else out_channels + resnet_in_channels = prev_output_channel if i == 0 else out_channels + + resnets.append( + ResnetBlock2D( + in_channels=resnet_in_channels + res_skip_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + ) + ) + + processor = ( + AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() + ) + + attentions.append( + Attention( + query_dim=out_channels, + cross_attention_dim=out_channels, + heads=self.num_heads, + dim_head=self.attention_head_dim, + added_kv_proj_dim=cross_attention_dim, + norm_num_groups=resnet_groups, + bias=True, + upcast_softmax=True, + only_cross_attention=only_cross_attention, + cross_attention_norm=cross_attention_norm, + processor=processor, + ) + ) + self.attentions = nn.ModuleList(attentions) + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList( + [ + ResnetBlock2D( + in_channels=out_channels, + out_channels=out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=resnet_groups, + dropout=dropout, + time_embedding_norm=resnet_time_scale_shift, + non_linearity=resnet_act_fn, + output_scale_factor=output_scale_factor, + pre_norm=resnet_pre_norm, + skip_time_act=skip_time_act, + up=True, + ) + ] + ) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + lora_scale = cross_attention_kwargs.get("scale", 1.0) + if attention_mask is None: + # if encoder_hidden_states is defined: we are doing cross-attn, so we should use cross-attn mask. + mask = None if encoder_hidden_states is None else encoder_attention_mask + else: + # when attention_mask is defined: we don't even check for encoder_attention_mask. + # this is to maintain compatibility with UnCLIP, which uses 'attention_mask' param for cross-attn masks. + # TODO: UnCLIP should express cross-attn mask via encoder_attention_mask param instead of via attention_mask. + # then we can simplify this whole if/else block to: + # mask = attention_mask if encoder_hidden_states is None else encoder_attention_mask + mask = attention_mask + + for resnet, attn in zip(self.resnets, self.attentions): + # resnet + # pop res hidden states + res_hidden_states = res_hidden_states_tuple[-1] + res_hidden_states_tuple = res_hidden_states_tuple[:-1] + hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1) + + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + hidden_states = torch.utils.checkpoint.checkpoint(create_custom_forward(resnet), hidden_states, temb) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=mask, + **cross_attention_kwargs, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states, temb, scale=lora_scale) + + return hidden_states + + +class KUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 5, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: Optional[int] = 32, + add_upsample: bool = True, + ): + super().__init__() + resnets = [] + k_in_channels = 2 * out_channels + k_out_channels = in_channels + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=k_out_channels if (i == num_layers - 1) else out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + + self.resnets = nn.ModuleList(resnets) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + upsample_size: Optional[int] = None, + scale: float = 1.0, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + for resnet in self.resnets: + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module): + def custom_forward(*inputs): + return module(*inputs) + + return custom_forward + + if is_torch_version(">=", "1.11.0"): + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb, use_reentrant=False + ) + else: + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), hidden_states, temb + ) + else: + hidden_states = resnet(hidden_states, temb, scale=scale) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +class KCrossAttnUpBlock2D(nn.Module): + def __init__( + self, + in_channels: int, + out_channels: int, + temb_channels: int, + resolution_idx: int, + dropout: float = 0.0, + num_layers: int = 4, + resnet_eps: float = 1e-5, + resnet_act_fn: str = "gelu", + resnet_group_size: int = 32, + attention_head_dim: int = 1, # attention dim_head + cross_attention_dim: int = 768, + add_upsample: bool = True, + upcast_attention: bool = False, + ): + super().__init__() + resnets = [] + attentions = [] + + is_first_block = in_channels == out_channels == temb_channels + is_middle_block = in_channels != out_channels + add_self_attention = True if is_first_block else False + + self.has_cross_attention = True + self.attention_head_dim = attention_head_dim + + # in_channels, and out_channels for the block (k-unet) + k_in_channels = out_channels if is_first_block else 2 * out_channels + k_out_channels = in_channels + + num_layers = num_layers - 1 + + for i in range(num_layers): + in_channels = k_in_channels if i == 0 else out_channels + groups = in_channels // resnet_group_size + groups_out = out_channels // resnet_group_size + + if is_middle_block and (i == num_layers - 1): + conv_2d_out_channels = k_out_channels + else: + conv_2d_out_channels = None + + resnets.append( + ResnetBlockCondNorm2D( + in_channels=in_channels, + out_channels=out_channels, + conv_2d_out_channels=conv_2d_out_channels, + temb_channels=temb_channels, + eps=resnet_eps, + groups=groups, + groups_out=groups_out, + dropout=dropout, + non_linearity=resnet_act_fn, + time_embedding_norm="ada_group", + conv_shortcut_bias=False, + ) + ) + attentions.append( + KAttentionBlock( + k_out_channels if (i == num_layers - 1) else out_channels, + k_out_channels // attention_head_dim + if (i == num_layers - 1) + else out_channels // attention_head_dim, + attention_head_dim, + cross_attention_dim=cross_attention_dim, + temb_channels=temb_channels, + attention_bias=True, + add_self_attention=add_self_attention, + cross_attention_norm="layer_norm", + upcast_attention=upcast_attention, + ) + ) + + self.resnets = nn.ModuleList(resnets) + self.attentions = nn.ModuleList(attentions) + + if add_upsample: + self.upsamplers = nn.ModuleList([KUpsample2D()]) + else: + self.upsamplers = None + + self.gradient_checkpointing = False + self.resolution_idx = resolution_idx + + def forward( + self, + hidden_states: torch.FloatTensor, + res_hidden_states_tuple: Tuple[torch.FloatTensor, ...], + temb: Optional[torch.FloatTensor] = None, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + upsample_size: Optional[int] = None, + attention_mask: Optional[torch.FloatTensor] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + res_hidden_states_tuple = res_hidden_states_tuple[-1] + if res_hidden_states_tuple is not None: + hidden_states = torch.cat([hidden_states, res_hidden_states_tuple], dim=1) + + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + for resnet, attn in zip(self.resnets, self.attentions): + if self.training and self.gradient_checkpointing: + + def create_custom_forward(module, return_dict=None): + def custom_forward(*inputs): + if return_dict is not None: + return module(*inputs, return_dict=return_dict) + else: + return module(*inputs) + + return custom_forward + + ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {} + hidden_states = torch.utils.checkpoint.checkpoint( + create_custom_forward(resnet), + hidden_states, + temb, + **ckpt_kwargs, + ) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + hidden_states = resnet(hidden_states, temb, scale=lora_scale) + hidden_states = attn( + hidden_states, + encoder_hidden_states=encoder_hidden_states, + emb=temb, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + + if self.upsamplers is not None: + for upsampler in self.upsamplers: + hidden_states = upsampler(hidden_states) + + return hidden_states + + +# can potentially later be renamed to `No-feed-forward` attention +class KAttentionBlock(nn.Module): + r""" + A basic Transformer block. + + Parameters: + dim (`int`): The number of channels in the input and output. + num_attention_heads (`int`): The number of heads to use for multi-head attention. + attention_head_dim (`int`): The number of channels in each head. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention. + attention_bias (`bool`, *optional*, defaults to `False`): + Configure if the attention layers should contain a bias parameter. + upcast_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to upcast the attention computation to `float32`. + temb_channels (`int`, *optional*, defaults to 768): + The number of channels in the token embedding. + add_self_attention (`bool`, *optional*, defaults to `False`): + Set to `True` to add self-attention to the block. + cross_attention_norm (`str`, *optional*, defaults to `None`): + The type of normalization to use for the cross attention. Can be `None`, `layer_norm`, or `group_norm`. + group_size (`int`, *optional*, defaults to 32): + The number of groups to separate the channels into for group normalization. + """ + + def __init__( + self, + dim: int, + num_attention_heads: int, + attention_head_dim: int, + dropout: float = 0.0, + cross_attention_dim: Optional[int] = None, + attention_bias: bool = False, + upcast_attention: bool = False, + temb_channels: int = 768, # for ada_group_norm + add_self_attention: bool = False, + cross_attention_norm: Optional[str] = None, + group_size: int = 32, + ): + super().__init__() + self.add_self_attention = add_self_attention + + # 1. Self-Attn + if add_self_attention: + self.norm1 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn1 = Attention( + query_dim=dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + cross_attention_dim=None, + cross_attention_norm=None, + ) + + # 2. Cross-Attn + self.norm2 = AdaGroupNorm(temb_channels, dim, max(1, dim // group_size)) + self.attn2 = Attention( + query_dim=dim, + cross_attention_dim=cross_attention_dim, + heads=num_attention_heads, + dim_head=attention_head_dim, + dropout=dropout, + bias=attention_bias, + upcast_attention=upcast_attention, + cross_attention_norm=cross_attention_norm, + ) + + def _to_3d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 3, 1).reshape(hidden_states.shape[0], height * weight, -1) + + def _to_4d(self, hidden_states: torch.FloatTensor, height: int, weight: int) -> torch.FloatTensor: + return hidden_states.permute(0, 2, 1).reshape(hidden_states.shape[0], -1, height, weight) + + def forward( + self, + hidden_states: torch.FloatTensor, + encoder_hidden_states: Optional[torch.FloatTensor] = None, + # TODO: mark emb as non-optional (self.norm2 requires it). + # requires assessing impact of change to positional param interface. + emb: Optional[torch.FloatTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + encoder_attention_mask: Optional[torch.FloatTensor] = None, + ) -> torch.FloatTensor: + cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {} + + # 1. Self-Attention + if self.add_self_attention: + norm_hidden_states = self.norm1(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + + attn_output = self.attn1( + norm_hidden_states, + encoder_hidden_states=None, + attention_mask=attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + # 2. Cross-Attention/None + norm_hidden_states = self.norm2(hidden_states, emb) + + height, weight = norm_hidden_states.shape[2:] + norm_hidden_states = self._to_3d(norm_hidden_states, height, weight) + attn_output = self.attn2( + norm_hidden_states, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask if encoder_hidden_states is None else encoder_attention_mask, + **cross_attention_kwargs, + ) + attn_output = self._to_4d(attn_output, height, weight) + + hidden_states = attn_output + hidden_states + + return hidden_states diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_condition.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_condition.py new file mode 100644 index 0000000000000000000000000000000000000000..bbe078d297cf740e02198ff0a8b02133b5d7b00b --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/models/unets/unet_2d_condition.py @@ -0,0 +1,1218 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from dataclasses import dataclass +from typing import Any, Dict, List, Optional, Tuple, Union + +import torch +import torch.nn as nn +import torch.utils.checkpoint + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin +from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers +from stablediffusion.layers.activations import get_activation +from stablediffusion.layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from stablediffusion.layers.embeddings import ( + GaussianFourierProjection, + GLIGENTextBoundingboxProjection, + ImageHintTimeEmbedding, + ImageProjection, + ImageTimeEmbedding, + TextImageProjection, + TextImageTimeEmbedding, + TextTimeEmbedding, + TimestepEmbedding, + Timesteps, +) +from .unet_2d_blocks import ( + UNetMidBlock2D, + UNetMidBlock2DCrossAttn, + UNetMidBlock2DSimpleCrossAttn, + get_down_block, + get_up_block, +) +from stablediffusion.models.model_utils import ModelMixin + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + + +@dataclass +class UNet2DConditionOutput(BaseOutput): + """ + The output of [`UNet2DConditionModel`]. + + Args: + sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`): + The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model. + """ + + sample: torch.FloatTensor = None + + +class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin): + r""" + A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample + shaped output. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`): + Height and width of input/output sample. + in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample. + out_channels (`int`, *optional*, defaults to 4): Number of channels in the output. + center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample. + flip_sin_to_cos (`bool`, *optional*, defaults to `False`): + Whether to flip the sin to cos in the time embedding. + freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding. + down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`): + The tuple of downsample blocks to use. + mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`): + Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or + `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`): + The tuple of upsample blocks to use. + only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`): + Whether to include self-attention in the basic transformer blocks, see + [`~models.attention.BasicTransformerBlock`]. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`): + The tuple of output channels for each block. + layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. + downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution. + mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block. + dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization. + If `None`, normalization and activation layers is skipped in post-processing. + norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization. + cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280): + The dimension of the cross attention features. + transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None): + The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling + blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for + [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`], + [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`]. + encoder_hid_dim (`int`, *optional*, defaults to None): + If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim` + dimension to `cross_attention_dim`. + encoder_hid_dim_type (`str`, *optional*, defaults to `None`): + If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text + embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`. + attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads. + num_attention_heads (`int`, *optional*): + The number of attention heads. If not defined, defaults to `attention_head_dim` + resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config + for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`. + class_embed_type (`str`, *optional*, defaults to `None`): + The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`, + `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`. + addition_embed_type (`str`, *optional*, defaults to `None`): + Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or + "text". "text" will use the `TextTimeEmbedding` layer. + addition_time_embed_dim: (`int`, *optional*, defaults to `None`): + Dimension for the timestep embeddings. + num_class_embeds (`int`, *optional*, defaults to `None`): + Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing + class conditioning with `class_embed_type` equal to `None`. + time_embedding_type (`str`, *optional*, defaults to `positional`): + The type of position embedding to use for timesteps. Choose from `positional` or `fourier`. + time_embedding_dim (`int`, *optional*, defaults to `None`): + An optional override for the dimension of the projected time embedding. + time_embedding_act_fn (`str`, *optional*, defaults to `None`): + Optional activation function to use only once on the time embeddings before they are passed to the rest of + the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`. + timestep_post_act (`str`, *optional*, defaults to `None`): + The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`. + time_cond_proj_dim (`int`, *optional*, defaults to `None`): + The dimension of `cond_proj` layer in the timestep embedding. + conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer. conv_out_kernel (`int`, + *optional*, default to `3`): The kernel size of `conv_out` layer. projection_class_embeddings_input_dim (`int`, + *optional*): The dimension of the `class_labels` input when + `class_embed_type="projection"`. Required when `class_embed_type="projection"`. + class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time + embeddings with the class embeddings. + mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`): + Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If + `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the + `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False` + otherwise. + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + sample_size: Optional[int] = None, + in_channels: int = 4, + out_channels: int = 4, + center_input_sample: bool = False, + flip_sin_to_cos: bool = True, + freq_shift: int = 0, + down_block_types: Tuple[str] = ( + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "CrossAttnDownBlock2D", + "DownBlock2D", + ), + mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn", + up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"), + only_cross_attention: Union[bool, Tuple[bool]] = False, + block_out_channels: Tuple[int] = (320, 640, 1280, 1280), + layers_per_block: Union[int, Tuple[int]] = 2, + downsample_padding: int = 1, + mid_block_scale_factor: float = 1, + dropout: float = 0.0, + act_fn: str = "silu", + norm_num_groups: Optional[int] = 32, + norm_eps: float = 1e-5, + cross_attention_dim: Union[int, Tuple[int]] = 1280, + transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1, + reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None, + encoder_hid_dim: Optional[int] = None, + encoder_hid_dim_type: Optional[str] = None, + attention_head_dim: Union[int, Tuple[int]] = 8, + num_attention_heads: Optional[Union[int, Tuple[int]]] = None, + dual_cross_attention: bool = False, + use_linear_projection: bool = False, + class_embed_type: Optional[str] = None, + addition_embed_type: Optional[str] = None, + addition_time_embed_dim: Optional[int] = None, + num_class_embeds: Optional[int] = None, + upcast_attention: bool = False, + resnet_time_scale_shift: str = "default", + resnet_skip_time_act: bool = False, + resnet_out_scale_factor: int = 1.0, + time_embedding_type: str = "positional", + time_embedding_dim: Optional[int] = None, + time_embedding_act_fn: Optional[str] = None, + timestep_post_act: Optional[str] = None, + time_cond_proj_dim: Optional[int] = None, + conv_in_kernel: int = 3, + conv_out_kernel: int = 3, + projection_class_embeddings_input_dim: Optional[int] = None, + attention_type: str = "default", + class_embeddings_concat: bool = False, + mid_block_only_cross_attention: Optional[bool] = None, + cross_attention_norm: Optional[str] = None, + addition_embed_type_num_heads=64, + ): + super().__init__() + + self.sample_size = sample_size + + if num_attention_heads is not None: + raise ValueError( + "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19." + ) + + # If `num_attention_heads` is not defined (which is the case for most models) + # it will default to `attention_head_dim`. This looks weird upon first reading it and it is. + # The reason for this behavior is to correct for incorrectly named variables that were introduced + # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131 + # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking + # which is why we correct for the naming here. + num_attention_heads = num_attention_heads or attention_head_dim + + # Check inputs + if len(down_block_types) != len(up_block_types): + raise ValueError( + f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}." + ) + + if len(block_out_channels) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}." + ) + + if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}." + ) + + if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types): + raise ValueError( + f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}." + ) + if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None: + for layer_number_per_block in transformer_layers_per_block: + if isinstance(layer_number_per_block, list): + raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.") + + # input + conv_in_padding = (conv_in_kernel - 1) // 2 + self.conv_in = nn.Conv2d( + in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding + ) + + # time + if time_embedding_type == "fourier": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 2 + if time_embed_dim % 2 != 0: + raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.") + self.time_proj = GaussianFourierProjection( + time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos + ) + timestep_input_dim = time_embed_dim + elif time_embedding_type == "positional": + time_embed_dim = time_embedding_dim or block_out_channels[0] * 4 + + self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift) + timestep_input_dim = block_out_channels[0] + else: + raise ValueError( + f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`." + ) + + self.time_embedding = TimestepEmbedding( + timestep_input_dim, + time_embed_dim, + act_fn=act_fn, + post_act_fn=timestep_post_act, + cond_proj_dim=time_cond_proj_dim, + ) + + if encoder_hid_dim_type is None and encoder_hid_dim is not None: + encoder_hid_dim_type = "text_proj" + self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type) + logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.") + + if encoder_hid_dim is None and encoder_hid_dim_type is not None: + raise ValueError( + f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}." + ) + + if encoder_hid_dim_type == "text_proj": + self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim) + elif encoder_hid_dim_type == "text_image_proj": + # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image_proj"` (Kadinsky 2.1)` + self.encoder_hid_proj = TextImageProjection( + text_embed_dim=encoder_hid_dim, + image_embed_dim=cross_attention_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 + self.encoder_hid_proj = ImageProjection( + image_embed_dim=encoder_hid_dim, + cross_attention_dim=cross_attention_dim, + ) + elif encoder_hid_dim_type is not None: + raise ValueError( + f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'." + ) + else: + self.encoder_hid_proj = None + + # class embedding + if class_embed_type is None and num_class_embeds is not None: + self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim) + elif class_embed_type == "timestep": + self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn) + elif class_embed_type == "identity": + self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim) + elif class_embed_type == "projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set" + ) + # The projection `class_embed_type` is the same as the timestep `class_embed_type` except + # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings + # 2. it projects from an arbitrary input dimension. + # + # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations. + # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings. + # As a result, `TimestepEmbedding` can be passed arbitrary vectors. + self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif class_embed_type == "simple_projection": + if projection_class_embeddings_input_dim is None: + raise ValueError( + "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set" + ) + self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim) + else: + self.class_embedding = None + + if addition_embed_type == "text": + if encoder_hid_dim is not None: + text_time_embedding_from_dim = encoder_hid_dim + else: + text_time_embedding_from_dim = cross_attention_dim + + self.add_embedding = TextTimeEmbedding( + text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads + ) + elif addition_embed_type == "text_image": + # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much + # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use + # case when `addition_embed_type == "text_image"` (Kadinsky 2.1)` + self.add_embedding = TextImageTimeEmbedding( + text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim + ) + elif addition_embed_type == "text_time": + self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift) + self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim) + elif addition_embed_type == "image": + # Kandinsky 2.2 + self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type == "image_hint": + # Kandinsky 2.2 ControlNet + self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim) + elif addition_embed_type is not None: + raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.") + + if time_embedding_act_fn is None: + self.time_embed_act = None + else: + self.time_embed_act = get_activation(time_embedding_act_fn) + + self.down_blocks = nn.ModuleList([]) + self.up_blocks = nn.ModuleList([]) + + if isinstance(only_cross_attention, bool): + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = only_cross_attention + + only_cross_attention = [only_cross_attention] * len(down_block_types) + + if mid_block_only_cross_attention is None: + mid_block_only_cross_attention = False + + if isinstance(num_attention_heads, int): + num_attention_heads = (num_attention_heads,) * len(down_block_types) + + if isinstance(attention_head_dim, int): + attention_head_dim = (attention_head_dim,) * len(down_block_types) + + if isinstance(cross_attention_dim, int): + cross_attention_dim = (cross_attention_dim,) * len(down_block_types) + + if isinstance(layers_per_block, int): + layers_per_block = [layers_per_block] * len(down_block_types) + + if isinstance(transformer_layers_per_block, int): + transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types) + + if class_embeddings_concat: + # The time embeddings are concatenated with the class embeddings. The dimension of the + # time embeddings passed to the down, middle, and up blocks is twice the dimension of the + # regular time embeddings + blocks_time_embed_dim = time_embed_dim * 2 + else: + blocks_time_embed_dim = time_embed_dim + + # down + output_channel = block_out_channels[0] + for i, down_block_type in enumerate(down_block_types): + input_channel = output_channel + output_channel = block_out_channels[i] + is_final_block = i == len(block_out_channels) - 1 + + down_block = get_down_block( + down_block_type, + num_layers=layers_per_block[i], + transformer_layers_per_block=transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + temb_channels=blocks_time_embed_dim, + add_downsample=not is_final_block, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resnet_groups=norm_num_groups, + cross_attention_dim=cross_attention_dim[i], + num_attention_heads=num_attention_heads[i], + downsample_padding=downsample_padding, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.down_blocks.append(down_block) + + # mid + if mid_block_type == "UNetMidBlock2DCrossAttn": + self.mid_block = UNetMidBlock2DCrossAttn( + transformer_layers_per_block=transformer_layers_per_block[-1], + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_time_scale_shift=resnet_time_scale_shift, + cross_attention_dim=cross_attention_dim[-1], + num_attention_heads=num_attention_heads[-1], + resnet_groups=norm_num_groups, + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + upcast_attention=upcast_attention, + attention_type=attention_type, + ) + elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn": + self.mid_block = UNetMidBlock2DSimpleCrossAttn( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + cross_attention_dim=cross_attention_dim[-1], + attention_head_dim=attention_head_dim[-1], + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + skip_time_act=resnet_skip_time_act, + only_cross_attention=mid_block_only_cross_attention, + cross_attention_norm=cross_attention_norm, + ) + elif mid_block_type == "UNetMidBlock2D": + self.mid_block = UNetMidBlock2D( + in_channels=block_out_channels[-1], + temb_channels=blocks_time_embed_dim, + dropout=dropout, + num_layers=0, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + output_scale_factor=mid_block_scale_factor, + resnet_groups=norm_num_groups, + resnet_time_scale_shift=resnet_time_scale_shift, + add_attention=False, + ) + elif mid_block_type is None: + self.mid_block = None + else: + raise ValueError(f"unknown mid_block_type : {mid_block_type}") + + # count how many layers upsample the images + self.num_upsamplers = 0 + + # up + reversed_block_out_channels = list(reversed(block_out_channels)) + reversed_num_attention_heads = list(reversed(num_attention_heads)) + reversed_layers_per_block = list(reversed(layers_per_block)) + reversed_cross_attention_dim = list(reversed(cross_attention_dim)) + reversed_transformer_layers_per_block = ( + list(reversed(transformer_layers_per_block)) + if reverse_transformer_layers_per_block is None + else reverse_transformer_layers_per_block + ) + only_cross_attention = list(reversed(only_cross_attention)) + + output_channel = reversed_block_out_channels[0] + for i, up_block_type in enumerate(up_block_types): + is_final_block = i == len(block_out_channels) - 1 + + prev_output_channel = output_channel + output_channel = reversed_block_out_channels[i] + input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)] + + # add upsample block for all BUT final layer + if not is_final_block: + add_upsample = True + self.num_upsamplers += 1 + else: + add_upsample = False + + up_block = get_up_block( + up_block_type, + num_layers=reversed_layers_per_block[i] + 1, + transformer_layers_per_block=reversed_transformer_layers_per_block[i], + in_channels=input_channel, + out_channels=output_channel, + prev_output_channel=prev_output_channel, + temb_channels=blocks_time_embed_dim, + add_upsample=add_upsample, + resnet_eps=norm_eps, + resnet_act_fn=act_fn, + resolution_idx=i, + resnet_groups=norm_num_groups, + cross_attention_dim=reversed_cross_attention_dim[i], + num_attention_heads=reversed_num_attention_heads[i], + dual_cross_attention=dual_cross_attention, + use_linear_projection=use_linear_projection, + only_cross_attention=only_cross_attention[i], + upcast_attention=upcast_attention, + resnet_time_scale_shift=resnet_time_scale_shift, + attention_type=attention_type, + resnet_skip_time_act=resnet_skip_time_act, + resnet_out_scale_factor=resnet_out_scale_factor, + cross_attention_norm=cross_attention_norm, + attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel, + dropout=dropout, + ) + self.up_blocks.append(up_block) + prev_output_channel = output_channel + + # out + if norm_num_groups is not None: + self.conv_norm_out = nn.GroupNorm( + num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps + ) + + self.conv_act = get_activation(act_fn) + + else: + self.conv_norm_out = None + self.conv_act = None + + conv_out_padding = (conv_out_kernel - 1) // 2 + self.conv_out = nn.Conv2d( + block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding + ) + + if attention_type in ["gated", "gated-text-image"]: + positive_len = 768 + if isinstance(cross_attention_dim, int): + positive_len = cross_attention_dim + elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list): + positive_len = cross_attention_dim[0] + + feature_type = "text-only" if attention_type == "gated" else "text-image" + self.position_net = GLIGENTextBoundingboxProjection( + positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type + ) + + @property + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + def set_attention_slice(self, slice_size): + r""" + Enable sliced attention computation. + + When this option is enabled, the attention module splits the input tensor in slices to compute attention in + several steps. This is useful for saving some memory in exchange for a small decrease in speed. + + Args: + slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`): + When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If + `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is + provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim` + must be a multiple of `slice_size`. + """ + sliceable_head_dims = [] + + def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module): + if hasattr(module, "set_attention_slice"): + sliceable_head_dims.append(module.sliceable_head_dim) + + for child in module.children(): + fn_recursive_retrieve_sliceable_dims(child) + + # retrieve number of attention layers + for module in self.children(): + fn_recursive_retrieve_sliceable_dims(module) + + num_sliceable_layers = len(sliceable_head_dims) + + if slice_size == "auto": + # half the attention head size is usually a good trade-off between + # speed and memory + slice_size = [dim // 2 for dim in sliceable_head_dims] + elif slice_size == "max": + # make smallest slice possible + slice_size = num_sliceable_layers * [1] + + slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size + + if len(slice_size) != len(sliceable_head_dims): + raise ValueError( + f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different" + f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}." + ) + + for i in range(len(slice_size)): + size = slice_size[i] + dim = sliceable_head_dims[i] + if size is not None and size > dim: + raise ValueError(f"size {size} has to be smaller or equal to {dim}.") + + # Recursively walk through all the children. + # Any children which exposes the set_attention_slice method + # gets the message + def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]): + if hasattr(module, "set_attention_slice"): + module.set_attention_slice(slice_size.pop()) + + for child in module.children(): + fn_recursive_set_attention_slice(child, slice_size) + + reversed_slice_size = list(reversed(slice_size)) + for module in self.children(): + fn_recursive_set_attention_slice(module, reversed_slice_size) + + def _set_gradient_checkpointing(self, module, value=False): + if hasattr(module, "gradient_checkpointing"): + module.gradient_checkpointing = value + + def enable_freeu(self, s1, s2, b1, b2): + r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stage blocks where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that + are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate the "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + for i, upsample_block in enumerate(self.up_blocks): + setattr(upsample_block, "s1", s1) + setattr(upsample_block, "s2", s2) + setattr(upsample_block, "b1", b1) + setattr(upsample_block, "b2", b2) + + def disable_freeu(self): + """Disables the FreeU mechanism.""" + freeu_keys = {"s1", "s2", "b1", "b2"} + for i, upsample_block in enumerate(self.up_blocks): + for k in freeu_keys: + if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None: + setattr(upsample_block, k, None) + + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) + + def unload_lora(self): + """Unloads LoRA weights.""" + deprecate( + "unload_lora", + "0.28.0", + "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().", + ) + for module in self.modules(): + if hasattr(module, "set_lora_layer"): + module.set_lora_layer(None) + + def forward( + self, + sample: torch.FloatTensor, + timestep: Union[torch.Tensor, float, int], + encoder_hidden_states: torch.Tensor, + class_labels: Optional[torch.Tensor] = None, + timestep_cond: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None, + down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + mid_block_additional_residual: Optional[torch.Tensor] = None, + down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None, + encoder_attention_mask: Optional[torch.Tensor] = None, + return_dict: bool = True, + ) -> Union[UNet2DConditionOutput, Tuple]: + r""" + The [`UNet2DConditionModel`] forward method. + + Args: + sample (`torch.FloatTensor`): + The noisy input tensor with the following shape `(batch, channel, height, width)`. + timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input. + encoder_hidden_states (`torch.FloatTensor`): + The encoder hidden states with shape `(batch, sequence_length, feature_dim)`. + class_labels (`torch.Tensor`, *optional*, defaults to `None`): + Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. + timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`): + Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed + through the `self.time_embedding` layer to obtain the timestep embeddings. + attention_mask (`torch.Tensor`, *optional*, defaults to `None`): + An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask + is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large + negative values to the attention scores corresponding to "discard" tokens. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under + `self.processor` in + [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*): + A tuple of tensors that if specified are added to the residuals of down unet blocks. + mid_block_additional_residual: (`torch.Tensor`, *optional*): + A tensor that if specified is added to the residual of the middle unet block. + encoder_attention_mask (`torch.Tensor`): + A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If + `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias, + which adds large negative values to the attention scores corresponding to "discard" tokens. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain + tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttnProcessor`]. + added_cond_kwargs: (`dict`, *optional*): + A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that + are passed along to the UNet blocks. + down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added to UNet long skip connections from down blocks to up blocks for + example from ControlNet side model(s) + mid_block_additional_residual (`torch.Tensor`, *optional*): + additional residual to be added to UNet mid block output, for example from ControlNet side model + down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*): + additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) + + Returns: + [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`: + If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise + a `tuple` is returned where the first element is the sample tensor. + """ + # By default samples have to be AT least a multiple of the overall upsampling factor. + # The overall upsampling factor is equal to 2 ** (# num of upsampling layers). + # However, the upsampling interpolation output size can be forced to fit any upsampling size + # on the fly if necessary. + default_overall_up_factor = 2**self.num_upsamplers + + # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor` + forward_upsample_size = False + upsample_size = None + + for dim in sample.shape[-2:]: + if dim % default_overall_up_factor != 0: + # Forward upsample size to force interpolation output size. + forward_upsample_size = True + break + + # ensure attention_mask is a bias, and give it a singleton query_tokens dimension + # expects mask of shape: + # [batch, key_tokens] + # adds singleton query_tokens dimension: + # [batch, 1, key_tokens] + # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes: + # [batch, heads, query_tokens, key_tokens] (e.g. torch sdp attn) + # [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn) + if attention_mask is not None: + # assume that mask is expressed as: + # (1 = keep, 0 = discard) + # convert mask into a bias that can be added to attention scores: + # (keep = +0, discard = -10000.0) + attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0 + attention_mask = attention_mask.unsqueeze(1) + + # convert encoder_attention_mask to a bias the same way we do for attention_mask + if encoder_attention_mask is not None: + encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0 + encoder_attention_mask = encoder_attention_mask.unsqueeze(1) + + # 0. center input if necessary + if self.config.center_input_sample: + sample = 2 * sample - 1.0 + + # 1. time + timesteps = timestep + if not torch.is_tensor(timesteps): + # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can + # This would be a good case for the `match` statement (Python 3.10+) + is_mps = sample.device.type == "mps" + if isinstance(timestep, float): + dtype = torch.float32 if is_mps else torch.float64 + else: + dtype = torch.int32 if is_mps else torch.int64 + timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device) + elif len(timesteps.shape) == 0: + timesteps = timesteps[None].to(sample.device) + + # broadcast to batch dimension in a way that's compatible with ONNX/Core ML + timesteps = timesteps.expand(sample.shape[0]) + + t_emb = self.time_proj(timesteps) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # but time_embedding might actually be running in fp16. so we need to cast here. + # there might be better ways to encapsulate this. + t_emb = t_emb.to(dtype=sample.dtype) + + emb = self.time_embedding(t_emb, timestep_cond) + aug_emb = None + + if self.class_embedding is not None: + if class_labels is None: + raise ValueError("class_labels should be provided when num_class_embeds > 0") + + if self.config.class_embed_type == "timestep": + class_labels = self.time_proj(class_labels) + + # `Timesteps` does not contain any weights and will always return f32 tensors + # there might be better ways to encapsulate this. + class_labels = class_labels.to(dtype=sample.dtype) + + class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype) + + if self.config.class_embeddings_concat: + emb = torch.cat([emb, class_emb], dim=-1) + else: + emb = emb + class_emb + + if self.config.addition_embed_type == "text": + aug_emb = self.add_embedding(encoder_hidden_states) + elif self.config.addition_embed_type == "text_image": + # Kandinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + + image_embs = added_cond_kwargs.get("image_embeds") + text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states) + aug_emb = self.add_embedding(text_embs, image_embs) + elif self.config.addition_embed_type == "text_time": + # SDXL - style + if "text_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`" + ) + text_embeds = added_cond_kwargs.get("text_embeds") + if "time_ids" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`" + ) + time_ids = added_cond_kwargs.get("time_ids") + time_embeds = self.add_time_proj(time_ids.flatten()) + time_embeds = time_embeds.reshape((text_embeds.shape[0], -1)) + add_embeds = torch.concat([text_embeds, time_embeds], dim=-1) + add_embeds = add_embeds.to(emb.dtype) + aug_emb = self.add_embedding(add_embeds) + elif self.config.addition_embed_type == "image": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + aug_emb = self.add_embedding(image_embs) + elif self.config.addition_embed_type == "image_hint": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`" + ) + image_embs = added_cond_kwargs.get("image_embeds") + hint = added_cond_kwargs.get("hint") + aug_emb, hint = self.add_embedding(image_embs, hint) + sample = torch.cat([sample, hint], dim=1) + + emb = emb + aug_emb if aug_emb is not None else emb + + if self.time_embed_act is not None: + emb = self.time_embed_act(emb) + + if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj": + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj": + # Kadinsky 2.1 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj": + # Kandinsky 2.2 - style + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + encoder_hidden_states = self.encoder_hid_proj(image_embeds) + elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj": + if "image_embeds" not in added_cond_kwargs: + raise ValueError( + f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_conditions`" + ) + image_embeds = added_cond_kwargs.get("image_embeds") + image_embeds = self.encoder_hid_proj(image_embeds) + encoder_hidden_states = (encoder_hidden_states, image_embeds) + + # 2. pre-process + sample = self.conv_in(sample) + + # 2.5 GLIGEN position net + if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None: + cross_attention_kwargs = cross_attention_kwargs.copy() + gligen_args = cross_attention_kwargs.pop("gligen") + cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)} + + # 3. down + lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0 + if USE_PEFT_BACKEND: + # weight the lora layers by setting `lora_scale` for each PEFT layer + scale_lora_layers(self, lora_scale) + + is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None + # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets + is_adapter = down_intrablock_additional_residuals is not None + # maintain backward compatibility for legacy usage, where + # T2I-Adapter and ControlNet both use down_block_additional_residuals arg + # but can only use one or the other + if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None: + deprecate( + "T2I should not use down_block_additional_residuals", + "1.3.0", + "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \ + and will be removed in diffusers 1.3.0. `down_block_additional_residuals` should only be used \ + for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ", + standard_warn=False, + ) + down_intrablock_additional_residuals = down_block_additional_residuals + is_adapter = True + + down_block_res_samples = (sample,) + for downsample_block in self.down_blocks: + if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention: + # For t2i-adapter CrossAttnDownBlock2D + additional_residuals = {} + if is_adapter and len(down_intrablock_additional_residuals) > 0: + additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0) + + sample, res_samples = downsample_block( + hidden_states=sample, + temb=emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + **additional_residuals, + ) + else: + sample, res_samples = downsample_block(hidden_states=sample, temb=emb, scale=lora_scale) + if is_adapter and len(down_intrablock_additional_residuals) > 0: + sample += down_intrablock_additional_residuals.pop(0) + + down_block_res_samples += res_samples + + if is_controlnet: + new_down_block_res_samples = () + + for down_block_res_sample, down_block_additional_residual in zip( + down_block_res_samples, down_block_additional_residuals + ): + down_block_res_sample = down_block_res_sample + down_block_additional_residual + new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,) + + down_block_res_samples = new_down_block_res_samples + + # 4. mid + if self.mid_block is not None: + if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention: + sample = self.mid_block( + sample, + emb, + encoder_hidden_states=encoder_hidden_states, + attention_mask=attention_mask, + cross_attention_kwargs=cross_attention_kwargs, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = self.mid_block(sample, emb) + + # To support T2I-Adapter-XL + if ( + is_adapter + and len(down_intrablock_additional_residuals) > 0 + and sample.shape == down_intrablock_additional_residuals[0].shape + ): + sample += down_intrablock_additional_residuals.pop(0) + + if is_controlnet: + sample = sample + mid_block_additional_residual + + # 5. up + for i, upsample_block in enumerate(self.up_blocks): + is_final_block = i == len(self.up_blocks) - 1 + + res_samples = down_block_res_samples[-len(upsample_block.resnets) :] + down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)] + + # if we have not reached the final block and need to forward the + # upsample size, we do it here + if not is_final_block and forward_upsample_size: + upsample_size = down_block_res_samples[-1].shape[2:] + + if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + encoder_hidden_states=encoder_hidden_states, + cross_attention_kwargs=cross_attention_kwargs, + upsample_size=upsample_size, + attention_mask=attention_mask, + encoder_attention_mask=encoder_attention_mask, + ) + else: + sample = upsample_block( + hidden_states=sample, + temb=emb, + res_hidden_states_tuple=res_samples, + upsample_size=upsample_size, + scale=lora_scale, + ) + + # 6. post-process + if self.conv_norm_out: + sample = self.conv_norm_out(sample) + sample = self.conv_act(sample) + sample = self.conv_out(sample) + + if USE_PEFT_BACKEND: + # remove `lora_scale` from each PEFT layer + unscale_lora_layers(self, lora_scale) + + if not return_dict: + return (sample,) + + return UNet2DConditionOutput(sample=sample) diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..efd186f76b5e7023389c7d4c5f6358e44492d874 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/__init__.py @@ -0,0 +1 @@ +from .pipeline_stable_diffusion import StableDiffusionPipeline, DiffusionPipeline \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/pipeline_stable_diffusion.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/pipeline_stable_diffusion.py new file mode 100644 index 0000000000000000000000000000000000000000..5d00932bcf2274801db295625c428aade682b856 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/pipeline/pipeline_stable_diffusion.py @@ -0,0 +1,1088 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +import inspect +from typing import Any, Callable, Dict, List, Optional, Union + +import torch +from packaging import version +from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection + +from diffusers.configuration_utils import FrozenDict +from diffusers.image_processor import PipelineImageInput, VaeImageProcessor +from diffusers.loaders import FromSingleFileMixin, IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin +from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel +from diffusers.models.attention_processor import FusedAttnProcessor2_0 +from diffusers.models.lora import adjust_lora_scale_text_encoder +from diffusers.schedulers import KarrasDiffusionSchedulers +from diffusers.utils import ( + USE_PEFT_BACKEND, + deprecate, + logging, + replace_example_docstring, + scale_lora_layers, + unscale_lora_layers, +) +from diffusers.utils.torch_utils import randn_tensor +from diffusers.pipelines.pipeline_utils import DiffusionPipeline +from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput +from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker + + +logger = logging.get_logger(__name__) # pylint: disable=invalid-name + +EXAMPLE_DOC_STRING = """ + Examples: + ```py + >>> import torch + >>> from diffusers import StableDiffusionPipeline + + >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16) + >>> pipe = pipe.to("cuda") + + >>> prompt = "a photo of an astronaut riding a horse on mars" + >>> image = pipe(prompt).images[0] + ``` +""" + + +def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0): + """ + Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4 + """ + std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True) + std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True) + # rescale the results from guidance (fixes overexposure) + noise_pred_rescaled = noise_cfg * (std_text / std_cfg) + # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images + noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg + return noise_cfg + + +def retrieve_timesteps( + scheduler, + num_inference_steps: Optional[int] = None, + device: Optional[Union[str, torch.device]] = None, + timesteps: Optional[List[int]] = None, + **kwargs, +): + """ + Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles + custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`. + + Args: + scheduler (`SchedulerMixin`): + The scheduler to get timesteps from. + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. If used, + `timesteps` must be `None`. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + timesteps (`List[int]`, *optional*): + Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default + timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps` + must be `None`. + + Returns: + `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the + second element is the number of inference steps. + """ + if timesteps is not None: + accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys()) + if not accepts_timesteps: + raise ValueError( + f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom" + f" timestep schedules. Please check whether you are using the correct scheduler." + ) + scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs) + timesteps = scheduler.timesteps + num_inference_steps = len(timesteps) + else: + scheduler.set_timesteps(num_inference_steps, device=device, **kwargs) + timesteps = scheduler.timesteps + return timesteps, num_inference_steps + + +class StableDiffusionPipeline( + DiffusionPipeline, TextualInversionLoaderMixin, LoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin +): + r""" + Pipeline for text-to-image generation using Stable Diffusion. + + This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods + implemented for all pipelines (downloading, saving, running on a particular device, etc.). + + The pipeline also inherits the following loading methods: + - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings + - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights + - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights + - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files + - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters + + Args: + vae ([`AutoencoderKL`]): + Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations. + text_encoder ([`~transformers.CLIPTextModel`]): + Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)). + tokenizer ([`~transformers.CLIPTokenizer`]): + A `CLIPTokenizer` to tokenize text. + unet ([`UNet2DConditionModel`]): + A `UNet2DConditionModel` to denoise the encoded image latents. + scheduler ([`SchedulerMixin`]): + A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of + [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`]. + safety_checker ([`StableDiffusionSafetyChecker`]): + Classification module that estimates whether generated images could be considered offensive or harmful. + Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details + about a model's potential harms. + feature_extractor ([`~transformers.CLIPImageProcessor`]): + A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`. + """ + + model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae" + _optional_components = ["safety_checker", "feature_extractor", "image_encoder"] + _exclude_from_cpu_offload = ["safety_checker"] + _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"] + + def __init__( + self, + vae: AutoencoderKL, + text_encoder: CLIPTextModel, + tokenizer: CLIPTokenizer, + unet: UNet2DConditionModel, + scheduler: KarrasDiffusionSchedulers, + safety_checker: StableDiffusionSafetyChecker, + feature_extractor: CLIPImageProcessor, + image_encoder: CLIPVisionModelWithProjection = None, + requires_safety_checker: bool = True, + ): + super().__init__() + + if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`" + f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure " + "to update the config accordingly as leaving `steps_offset` might led to incorrect results" + " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub," + " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`" + " file" + ) + deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["steps_offset"] = 1 + scheduler._internal_dict = FrozenDict(new_config) + + if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True: + deprecation_message = ( + f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`." + " `clip_sample` should be set to False in the configuration file. Please make sure to update the" + " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in" + " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very" + " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file" + ) + deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(scheduler.config) + new_config["clip_sample"] = False + scheduler._internal_dict = FrozenDict(new_config) + + if safety_checker is None and requires_safety_checker: + logger.warning( + f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure" + " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered" + " results in services or applications open to the public. Both the diffusers team and Hugging Face" + " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling" + " it only for use-cases that involve analyzing network behavior or auditing its results. For more" + " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ." + ) + + if safety_checker is not None and feature_extractor is None: + raise ValueError( + "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety" + " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead." + ) + + is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse( + version.parse(unet.config._diffusers_version).base_version + ) < version.parse("0.9.0.dev0") + is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64 + if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64: + deprecation_message = ( + "The configuration file of the unet has set the default `sample_size` to smaller than" + " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the" + " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-" + " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5" + " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the" + " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`" + " in the config might lead to incorrect results in future versions. If you have downloaded this" + " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for" + " the `unet/config.json` file" + ) + deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False) + new_config = dict(unet.config) + new_config["sample_size"] = 64 + unet._internal_dict = FrozenDict(new_config) + + self.register_modules( + vae=vae, + text_encoder=text_encoder, + tokenizer=tokenizer, + unet=unet, + scheduler=scheduler, + safety_checker=safety_checker, + feature_extractor=feature_extractor, + image_encoder=image_encoder, + ) + self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) + self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor) + self.register_to_config(requires_safety_checker=requires_safety_checker) + + def enable_vae_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.vae.enable_slicing() + + def disable_vae_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_slicing() + + def enable_vae_tiling(self): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.vae.enable_tiling() + + def disable_vae_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to + computing decoding in one step. + """ + self.vae.disable_tiling() + + def _encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + **kwargs, + ): + deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple." + deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False) + + prompt_embeds_tuple = self.encode_prompt( + prompt=prompt, + device=device, + num_images_per_prompt=num_images_per_prompt, + do_classifier_free_guidance=do_classifier_free_guidance, + negative_prompt=negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + **kwargs, + ) + + # concatenate for backwards comp + prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]]) + + return prompt_embeds + + def encode_prompt( + self, + prompt, + device, + num_images_per_prompt, + do_classifier_free_guidance, + negative_prompt=None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + lora_scale: Optional[float] = None, + clip_skip: Optional[int] = None, + ): + r""" + Encodes the prompt into text encoder hidden states. + + Args: + prompt (`str` or `List[str]`, *optional*): + prompt to be encoded + device: (`torch.device`): + torch device + num_images_per_prompt (`int`): + number of images that should be generated per prompt + do_classifier_free_guidance (`bool`): + whether to use classifier free guidance or not + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts not to guide the image generation. If not defined, one has to pass + `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is + less than `1`). + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not + provided, text embeddings will be generated from `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt + weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input + argument. + lora_scale (`float`, *optional*): + A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + """ + # set lora scale so that monkey patched LoRA + # function of text encoder can correctly access it + if lora_scale is not None and isinstance(self, LoraLoaderMixin): + self._lora_scale = lora_scale + + # dynamically adjust the LoRA scale + if not USE_PEFT_BACKEND: + adjust_lora_scale_text_encoder(self.text_encoder, lora_scale) + else: + scale_lora_layers(self.text_encoder, lora_scale) + + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + if prompt_embeds is None: + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + prompt = self.maybe_convert_prompt(prompt, self.tokenizer) + + text_inputs = self.tokenizer( + prompt, + padding="max_length", + max_length=self.tokenizer.model_max_length, + truncation=True, + return_tensors="pt", + ) + text_input_ids = text_inputs.input_ids + untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids + + if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal( + text_input_ids, untruncated_ids + ): + removed_text = self.tokenizer.batch_decode( + untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1] + ) + logger.warning( + "The following part of your input was truncated because CLIP can only handle sequences up to" + f" {self.tokenizer.model_max_length} tokens: {removed_text}" + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = text_inputs.attention_mask.to(device) + else: + attention_mask = None + + if clip_skip is None: + prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask) + prompt_embeds = prompt_embeds[0] + else: + prompt_embeds = self.text_encoder( + text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True + ) + # Access the `hidden_states` first, that contains a tuple of + # all the hidden states from the encoder layers. Then index into + # the tuple to access the hidden states from the desired layer. + prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)] + # We also need to apply the final LayerNorm here to not mess with the + # representations. The `last_hidden_states` that we typically use for + # obtaining the final prompt representations passes through the LayerNorm + # layer. + prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds) + + if self.text_encoder is not None: + prompt_embeds_dtype = self.text_encoder.dtype + elif self.unet is not None: + prompt_embeds_dtype = self.unet.dtype + else: + prompt_embeds_dtype = prompt_embeds.dtype + + prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + bs_embed, seq_len, _ = prompt_embeds.shape + # duplicate text embeddings for each generation per prompt, using mps friendly method + prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) + prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1) + + # get unconditional embeddings for classifier free guidance + if do_classifier_free_guidance and negative_prompt_embeds is None: + uncond_tokens: List[str] + if negative_prompt is None: + uncond_tokens = [""] * batch_size + elif prompt is not None and type(prompt) is not type(negative_prompt): + raise TypeError( + f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !=" + f" {type(prompt)}." + ) + elif isinstance(negative_prompt, str): + uncond_tokens = [negative_prompt] + elif batch_size != len(negative_prompt): + raise ValueError( + f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:" + f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches" + " the batch size of `prompt`." + ) + else: + uncond_tokens = negative_prompt + + # textual inversion: procecss multi-vector tokens if necessary + if isinstance(self, TextualInversionLoaderMixin): + uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer) + + max_length = prompt_embeds.shape[1] + uncond_input = self.tokenizer( + uncond_tokens, + padding="max_length", + max_length=max_length, + truncation=True, + return_tensors="pt", + ) + + if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask: + attention_mask = uncond_input.attention_mask.to(device) + else: + attention_mask = None + + negative_prompt_embeds = self.text_encoder( + uncond_input.input_ids.to(device), + attention_mask=attention_mask, + ) + negative_prompt_embeds = negative_prompt_embeds[0] + + if do_classifier_free_guidance: + # duplicate unconditional embeddings for each generation per prompt, using mps friendly method + seq_len = negative_prompt_embeds.shape[1] + + negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device) + + negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1) + negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) + + if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND: + # Retrieve the original scale by scaling back the LoRA layers + unscale_lora_layers(self.text_encoder, lora_scale) + + return prompt_embeds, negative_prompt_embeds + + def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None): + dtype = next(self.image_encoder.parameters()).dtype + + if not isinstance(image, torch.Tensor): + image = self.feature_extractor(image, return_tensors="pt").pixel_values + + image = image.to(device=device, dtype=dtype) + if output_hidden_states: + image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2] + image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_enc_hidden_states = self.image_encoder( + torch.zeros_like(image), output_hidden_states=True + ).hidden_states[-2] + uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave( + num_images_per_prompt, dim=0 + ) + return image_enc_hidden_states, uncond_image_enc_hidden_states + else: + image_embeds = self.image_encoder(image).image_embeds + image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0) + uncond_image_embeds = torch.zeros_like(image_embeds) + + return image_embeds, uncond_image_embeds + + def prepare_ip_adapter_image_embeds(self, ip_adapter_image, device, num_images_per_prompt): + if not isinstance(ip_adapter_image, list): + ip_adapter_image = [ip_adapter_image] + + if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers): + raise ValueError( + f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters." + ) + + image_embeds = [] + for single_ip_adapter_image, image_proj_layer in zip( + ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers + ): + output_hidden_state = not isinstance(image_proj_layer, ImageProjection) + single_image_embeds, single_negative_image_embeds = self.encode_image( + single_ip_adapter_image, device, 1, output_hidden_state + ) + single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0) + single_negative_image_embeds = torch.stack([single_negative_image_embeds] * num_images_per_prompt, dim=0) + + if self.do_classifier_free_guidance: + single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds]) + single_image_embeds = single_image_embeds.to(device) + + image_embeds.append(single_image_embeds) + + return image_embeds + + def run_safety_checker(self, image, device, dtype): + if self.safety_checker is None: + has_nsfw_concept = None + else: + if torch.is_tensor(image): + feature_extractor_input = self.image_processor.postprocess(image, output_type="pil") + else: + feature_extractor_input = self.image_processor.numpy_to_pil(image) + safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device) + image, has_nsfw_concept = self.safety_checker( + images=image, clip_input=safety_checker_input.pixel_values.to(dtype) + ) + return image, has_nsfw_concept + + def decode_latents(self, latents): + deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead" + deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False) + + latents = 1 / self.vae.config.scaling_factor * latents + image = self.vae.decode(latents, return_dict=False)[0] + image = (image / 2 + 0.5).clamp(0, 1) + # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16 + image = image.cpu().permute(0, 2, 3, 1).float().numpy() + return image + + def prepare_extra_step_kwargs(self, generator, eta): + # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature + # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers. + # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502 + # and should be between [0, 1] + + accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys()) + extra_step_kwargs = {} + if accepts_eta: + extra_step_kwargs["eta"] = eta + + # check if the scheduler accepts generator + accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys()) + if accepts_generator: + extra_step_kwargs["generator"] = generator + return extra_step_kwargs + + def check_inputs( + self, + prompt, + height, + width, + callback_steps, + negative_prompt=None, + prompt_embeds=None, + negative_prompt_embeds=None, + callback_on_step_end_tensor_inputs=None, + ): + if height % 8 != 0 or width % 8 != 0: + raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.") + + if callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0): + raise ValueError( + f"`callback_steps` has to be a positive integer but is {callback_steps} of type" + f" {type(callback_steps)}." + ) + if callback_on_step_end_tensor_inputs is not None and not all( + k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs + ): + raise ValueError( + f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" + ) + + if prompt is not None and prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" + " only forward one of the two." + ) + elif prompt is None and prompt_embeds is None: + raise ValueError( + "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." + ) + elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): + raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") + + if negative_prompt is not None and negative_prompt_embeds is not None: + raise ValueError( + f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:" + f" {negative_prompt_embeds}. Please make sure to only forward one of the two." + ) + + if prompt_embeds is not None and negative_prompt_embeds is not None: + if prompt_embeds.shape != negative_prompt_embeds.shape: + raise ValueError( + "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but" + f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`" + f" {negative_prompt_embeds.shape}." + ) + + def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None): + shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor) + if isinstance(generator, list) and len(generator) != batch_size: + raise ValueError( + f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" + f" size of {batch_size}. Make sure the batch size matches the length of the generators." + ) + + if latents is None: + latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) + else: + latents = latents.to(device) + + # scale the initial noise by the standard deviation required by the scheduler + latents = latents * self.scheduler.init_noise_sigma + return latents + + def enable_freeu(self, s1: float, s2: float, b1: float, b2: float): + r"""Enables the FreeU mechanism as in https://arxiv.org/abs/2309.11497. + + The suffixes after the scaling factors represent the stages where they are being applied. + + Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of the values + that are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL. + + Args: + s1 (`float`): + Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + s2 (`float`): + Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to + mitigate "oversmoothing effect" in the enhanced denoising process. + b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features. + b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features. + """ + if not hasattr(self, "unet"): + raise ValueError("The pipeline must have `unet` for using FreeU.") + self.unet.enable_freeu(s1=s1, s2=s2, b1=b1, b2=b2) + + def disable_freeu(self): + """Disables the FreeU mechanism if enabled.""" + self.unet.disable_freeu() + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.fuse_qkv_projections + def fuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + """ + self.fusing_unet = False + self.fusing_vae = False + + if unet: + self.fusing_unet = True + self.unet.fuse_qkv_projections() + self.unet.set_attn_processor(FusedAttnProcessor2_0()) + + if vae: + if not isinstance(self.vae, AutoencoderKL): + raise ValueError("`fuse_qkv_projections()` is only supported for the VAE of type `AutoencoderKL`.") + + self.fusing_vae = True + self.vae.fuse_qkv_projections() + self.vae.set_attn_processor(FusedAttnProcessor2_0()) + + # Copied from diffusers.pipelines.stable_diffusion_xl.pipeline_stable_diffusion_xl.StableDiffusionXLPipeline.unfuse_qkv_projections + def unfuse_qkv_projections(self, unet: bool = True, vae: bool = True): + """Disable QKV projection fusion if enabled. + + + + This API is 🧪 experimental. + + + + Args: + unet (`bool`, defaults to `True`): To apply fusion on the UNet. + vae (`bool`, defaults to `True`): To apply fusion on the VAE. + + """ + if unet: + if not self.fusing_unet: + logger.warning("The UNet was not initially fused for QKV projections. Doing nothing.") + else: + self.unet.unfuse_qkv_projections() + self.fusing_unet = False + + if vae: + if not self.fusing_vae: + logger.warning("The VAE was not initially fused for QKV projections. Doing nothing.") + else: + self.vae.unfuse_qkv_projections() + self.fusing_vae = False + + # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding + def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32): + """ + See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298 + + Args: + timesteps (`torch.Tensor`): + generate embedding vectors at these timesteps + embedding_dim (`int`, *optional*, defaults to 512): + dimension of the embeddings to generate + dtype: + data type of the generated embeddings + + Returns: + `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)` + """ + assert len(w.shape) == 1 + w = w * 1000.0 + + half_dim = embedding_dim // 2 + emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1) + emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb) + emb = w.to(dtype)[:, None] * emb[None, :] + emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1) + if embedding_dim % 2 == 1: # zero pad + emb = torch.nn.functional.pad(emb, (0, 1)) + assert emb.shape == (w.shape[0], embedding_dim) + return emb + + @property + def guidance_scale(self): + return self._guidance_scale + + @property + def guidance_rescale(self): + return self._guidance_rescale + + @property + def clip_skip(self): + return self._clip_skip + + # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2) + # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1` + # corresponds to doing no classifier free guidance. + @property + def do_classifier_free_guidance(self): + return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None + + @property + def cross_attention_kwargs(self): + return self._cross_attention_kwargs + + @property + def num_timesteps(self): + return self._num_timesteps + + @property + def interrupt(self): + return self._interrupt + + @torch.no_grad() + @replace_example_docstring(EXAMPLE_DOC_STRING) + def __call__( + self, + prompt: Union[str, List[str]] = None, + height: Optional[int] = None, + width: Optional[int] = None, + num_inference_steps: int = 50, + timesteps: List[int] = None, + guidance_scale: float = 7.5, + negative_prompt: Optional[Union[str, List[str]]] = None, + num_images_per_prompt: Optional[int] = 1, + eta: float = 0.0, + generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None, + latents: Optional[torch.FloatTensor] = None, + prompt_embeds: Optional[torch.FloatTensor] = None, + negative_prompt_embeds: Optional[torch.FloatTensor] = None, + ip_adapter_image: Optional[PipelineImageInput] = None, + output_type: Optional[str] = "pil", + return_dict: bool = True, + cross_attention_kwargs: Optional[Dict[str, Any]] = None, + guidance_rescale: float = 0.0, + clip_skip: Optional[int] = None, + callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None, + callback_on_step_end_tensor_inputs: List[str] = ["latents"], + **kwargs, + ): + r""" + The call function to the pipeline for generation. + + Args: + prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`. + height (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The height in pixels of the generated image. + width (`int`, *optional*, defaults to `self.unet.config.sample_size * self.vae_scale_factor`): + The width in pixels of the generated image. + num_inference_steps (`int`, *optional*, defaults to 50): + The number of denoising steps. More denoising steps usually lead to a higher quality image at the + expense of slower inference. + timesteps (`List[int]`, *optional*): + Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument + in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is + passed will be used. Must be in descending order. + guidance_scale (`float`, *optional*, defaults to 7.5): + A higher guidance scale value encourages the model to generate images closely linked to the text + `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`. + negative_prompt (`str` or `List[str]`, *optional*): + The prompt or prompts to guide what to not include in image generation. If not defined, you need to + pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 1`). + num_images_per_prompt (`int`, *optional*, defaults to 1): + The number of images to generate per prompt. + eta (`float`, *optional*, defaults to 0.0): + Corresponds to parameter eta (η) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies + to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers. + generator (`torch.Generator` or `List[torch.Generator]`, *optional*): + A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make + generation deterministic. + latents (`torch.FloatTensor`, *optional*): + Pre-generated noisy latents sampled from a Gaussian distribution, to be used as inputs for image + generation. Can be used to tweak the same generation with different prompts. If not provided, a latents + tensor is generated by sampling using the supplied random `generator`. + prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not + provided, text embeddings are generated from the `prompt` input argument. + negative_prompt_embeds (`torch.FloatTensor`, *optional*): + Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If + not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument. + ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters. + output_type (`str`, *optional*, defaults to `"pil"`): + The output format of the generated image. Choose between `PIL.Image` or `np.array`. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a + plain tuple. + cross_attention_kwargs (`dict`, *optional*): + A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in + [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py). + guidance_rescale (`float`, *optional*, defaults to 0.0): + Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are + Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when + using zero terminal SNR. + clip_skip (`int`, *optional*): + Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that + the output of the pre-final layer will be used for computing the prompt embeddings. + callback_on_step_end (`Callable`, *optional*): + A function that calls at the end of each denoising steps during the inference. The function is called + with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, + callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by + `callback_on_step_end_tensor_inputs`. + callback_on_step_end_tensor_inputs (`List`, *optional*): + The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list + will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the + `._callback_tensor_inputs` attribute of your pipeline class. + + Examples: + + Returns: + [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`: + If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned, + otherwise a `tuple` is returned where the first element is a list with the generated images and the + second element is a list of `bool`s indicating whether the corresponding generated image contains + "not-safe-for-work" (nsfw) content. + """ + + callback = kwargs.pop("callback", None) + callback_steps = kwargs.pop("callback_steps", None) + + if callback is not None: + deprecate( + "callback", + "1.0.0", + "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + if callback_steps is not None: + deprecate( + "callback_steps", + "1.0.0", + "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`", + ) + + # 0. Default height and width to unet + height = height or self.unet.config.sample_size * self.vae_scale_factor + width = width or self.unet.config.sample_size * self.vae_scale_factor + # to deal with lora scaling and other possible forward hooks + + # 1. Check inputs. Raise error if not correct + self.check_inputs( + prompt, + height, + width, + callback_steps, + negative_prompt, + prompt_embeds, + negative_prompt_embeds, + callback_on_step_end_tensor_inputs, + ) + + self._guidance_scale = guidance_scale + self._guidance_rescale = guidance_rescale + self._clip_skip = clip_skip + self._cross_attention_kwargs = cross_attention_kwargs + self._interrupt = False + + # 2. Define call parameters + if prompt is not None and isinstance(prompt, str): + batch_size = 1 + elif prompt is not None and isinstance(prompt, list): + batch_size = len(prompt) + else: + batch_size = prompt_embeds.shape[0] + + device = self._execution_device + + # 3. Encode input prompt + lora_scale = ( + self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None + ) + + prompt_embeds, negative_prompt_embeds = self.encode_prompt( + prompt, + device, + num_images_per_prompt, + self.do_classifier_free_guidance, + negative_prompt, + prompt_embeds=prompt_embeds, + negative_prompt_embeds=negative_prompt_embeds, + lora_scale=lora_scale, + clip_skip=self.clip_skip, + ) + + # For classifier free guidance, we need to do two forward passes. + # Here we concatenate the unconditional and text embeddings into a single batch + # to avoid doing two forward passes + if self.do_classifier_free_guidance: + prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds]) + + if ip_adapter_image is not None: + image_embeds = self.prepare_ip_adapter_image_embeds( + ip_adapter_image, device, batch_size * num_images_per_prompt + ) + + # 4. Prepare timesteps + timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps) + + # 5. Prepare latent variables + num_channels_latents = self.unet.config.in_channels + latents = self.prepare_latents( + batch_size * num_images_per_prompt, + num_channels_latents, + height, + width, + prompt_embeds.dtype, + device, + generator, + latents, + ) + + # 6. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline + extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta) + + # 6.1 Add image embeds for IP-Adapter + added_cond_kwargs = {"image_embeds": image_embeds} if ip_adapter_image is not None else None + + # 6.2 Optionally get Guidance Scale Embedding + timestep_cond = None + if self.unet.config.time_cond_proj_dim is not None: + guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt) + timestep_cond = self.get_guidance_scale_embedding( + guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim + ).to(device=device, dtype=latents.dtype) + + # 7. Denoising loop + num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order + self._num_timesteps = len(timesteps) + with self.progress_bar(total=num_inference_steps) as progress_bar: + for i, t in enumerate(timesteps): + if self.interrupt: + continue + + # expand the latents if we are doing classifier free guidance + latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents + latent_model_input = self.scheduler.scale_model_input(latent_model_input, t) + + # predict the noise residual + noise_pred = self.unet( + latent_model_input, + t, + encoder_hidden_states=prompt_embeds, + timestep_cond=timestep_cond, + cross_attention_kwargs=self.cross_attention_kwargs, + added_cond_kwargs=added_cond_kwargs, + return_dict=False, + )[0] + + # perform guidance + if self.do_classifier_free_guidance: + noise_pred_uncond, noise_pred_text = noise_pred.chunk(2) + noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond) + + if self.do_classifier_free_guidance and self.guidance_rescale > 0.0: + # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf + noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale) + + # compute the previous noisy sample x_t -> x_t-1 + latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0] + + if callback_on_step_end is not None: + callback_kwargs = {} + for k in callback_on_step_end_tensor_inputs: + callback_kwargs[k] = locals()[k] + callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) + + latents = callback_outputs.pop("latents", latents) + prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) + negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds) + + # call the callback, if provided + if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): + progress_bar.update() + if callback is not None and i % callback_steps == 0: + step_idx = i // getattr(self.scheduler, "order", 1) + callback(step_idx, t, latents) + + if not output_type == "latent": + image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[ + 0 + ] + image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype) + else: + image = latents + has_nsfw_concept = None + + if has_nsfw_concept is None: + do_denormalize = [True] * image.shape[0] + else: + do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept] + + image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize) + + # Offload all models + self.maybe_free_model_hooks() + + if not return_dict: + return (image, has_nsfw_concept) + + return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept) diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..6f853e7d075eb2528f2ef9e4e230a082af79fb80 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/__init__.py @@ -0,0 +1 @@ +from .scheduling_pndm import PNDMScheduler, SchedulerMixin \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/scheduling_pndm.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/scheduling_pndm.py new file mode 100644 index 0000000000000000000000000000000000000000..7d50d01e9602ee2bfa4b612baadcfdc8128deff1 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/schedulers/scheduling_pndm.py @@ -0,0 +1,478 @@ +# Copyright 2023 Zhejiang University Team and The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. + +# DISCLAIMER: This file is strongly influenced by https://github.com/ermongroup/ddim + +import math +from typing import List, Optional, Tuple, Union + +import numpy as np +import torch + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.schedulers.scheduling_utils import KarrasDiffusionSchedulers, SchedulerMixin, SchedulerOutput + + +# Copied from diffusers.schedulers.scheduling_ddpm.betas_for_alpha_bar +def betas_for_alpha_bar( + num_diffusion_timesteps, + max_beta=0.999, + alpha_transform_type="cosine", +): + """ + Create a beta schedule that discretizes the given alpha_t_bar function, which defines the cumulative product of + (1-beta) over time from t = [0,1]. + + Contains a function alpha_bar that takes an argument t and transforms it to the cumulative product of (1-beta) up + to that part of the diffusion process. + + + Args: + num_diffusion_timesteps (`int`): the number of betas to produce. + max_beta (`float`): the maximum beta to use; use values lower than 1 to + prevent singularities. + alpha_transform_type (`str`, *optional*, default to `cosine`): the type of noise schedule for alpha_bar. + Choose from `cosine` or `exp` + + Returns: + betas (`np.ndarray`): the betas used by the scheduler to step the model outputs + """ + if alpha_transform_type == "cosine": + + def alpha_bar_fn(t): + return math.cos((t + 0.008) / 1.008 * math.pi / 2) ** 2 + + elif alpha_transform_type == "exp": + + def alpha_bar_fn(t): + return math.exp(t * -12.0) + + else: + raise ValueError(f"Unsupported alpha_tranform_type: {alpha_transform_type}") + + betas = [] + for i in range(num_diffusion_timesteps): + t1 = i / num_diffusion_timesteps + t2 = (i + 1) / num_diffusion_timesteps + betas.append(min(1 - alpha_bar_fn(t2) / alpha_bar_fn(t1), max_beta)) + return torch.tensor(betas, dtype=torch.float32) + + +class PNDMScheduler(SchedulerMixin, ConfigMixin): + """ + `PNDMScheduler` uses pseudo numerical methods for diffusion models such as the Runge-Kutta and linear multi-step + method. + + This model inherits from [`SchedulerMixin`] and [`ConfigMixin`]. Check the superclass documentation for the generic + methods the library implements for all schedulers such as loading and saving. + + Args: + num_train_timesteps (`int`, defaults to 1000): + The number of diffusion steps to train the model. + beta_start (`float`, defaults to 0.0001): + The starting `beta` value of inference. + beta_end (`float`, defaults to 0.02): + The final `beta` value. + beta_schedule (`str`, defaults to `"linear"`): + The beta schedule, a mapping from a beta range to a sequence of betas for stepping the model. Choose from + `linear`, `scaled_linear`, or `squaredcos_cap_v2`. + trained_betas (`np.ndarray`, *optional*): + Pass an array of betas directly to the constructor to bypass `beta_start` and `beta_end`. + skip_prk_steps (`bool`, defaults to `False`): + Allows the scheduler to skip the Runge-Kutta steps defined in the original paper as being required before + PLMS steps. + set_alpha_to_one (`bool`, defaults to `False`): + Each diffusion step uses the alphas product value at that step and at the previous one. For the final step + there is no previous alpha. When this option is `True` the previous alpha product is fixed to `1`, + otherwise it uses the alpha value at step 0. + prediction_type (`str`, defaults to `epsilon`, *optional*): + Prediction type of the scheduler function; can be `epsilon` (predicts the noise of the diffusion process) + or `v_prediction` (see section 2.4 of [Imagen Video](https://imagen.research.google/video/paper.pdf) + paper). + timestep_spacing (`str`, defaults to `"leading"`): + The way the timesteps should be scaled. Refer to Table 2 of the [Common Diffusion Noise Schedules and + Sample Steps are Flawed](https://huggingface.co/papers/2305.08891) for more information. + steps_offset (`int`, defaults to 0): + An offset added to the inference steps. You can use a combination of `offset=1` and + `set_alpha_to_one=False` to make the last step use step 0 for the previous alpha product like in Stable + Diffusion. + """ + + _compatibles = [e.name for e in KarrasDiffusionSchedulers] + order = 1 + + @register_to_config + def __init__( + self, + num_train_timesteps: int = 1000, + beta_start: float = 0.0001, + beta_end: float = 0.02, + beta_schedule: str = "linear", + trained_betas: Optional[Union[np.ndarray, List[float]]] = None, + skip_prk_steps: bool = False, + set_alpha_to_one: bool = False, + prediction_type: str = "epsilon", + timestep_spacing: str = "leading", + steps_offset: int = 0, + ): + if trained_betas is not None: + self.betas = torch.tensor(trained_betas, dtype=torch.float32) + elif beta_schedule == "linear": + self.betas = torch.linspace(beta_start, beta_end, num_train_timesteps, dtype=torch.float32) + elif beta_schedule == "scaled_linear": + # this schedule is very specific to the latent diffusion model. + self.betas = torch.linspace(beta_start**0.5, beta_end**0.5, num_train_timesteps, dtype=torch.float32) ** 2 + elif beta_schedule == "squaredcos_cap_v2": + # Glide cosine schedule + self.betas = betas_for_alpha_bar(num_train_timesteps) + else: + raise NotImplementedError(f"{beta_schedule} does is not implemented for {self.__class__}") + + self.alphas = 1.0 - self.betas + self.alphas_cumprod = torch.cumprod(self.alphas, dim=0) + + self.final_alpha_cumprod = torch.tensor(1.0) if set_alpha_to_one else self.alphas_cumprod[0] + + # standard deviation of the initial noise distribution + self.init_noise_sigma = 1.0 + + # For now we only support F-PNDM, i.e. the runge-kutta method + # For more information on the algorithm please take a look at the paper: https://arxiv.org/pdf/2202.09778.pdf + # mainly at formula (9), (12), (13) and the Algorithm 2. + self.pndm_order = 4 + + # running values + self.cur_model_output = 0 + self.counter = 0 + self.cur_sample = None + self.ets = [] + + # setable values + self.num_inference_steps = None + self._timesteps = np.arange(0, num_train_timesteps)[::-1].copy() + self.prk_timesteps = None + self.plms_timesteps = None + self.timesteps = None + + def set_timesteps(self, num_inference_steps: int, device: Union[str, torch.device] = None): + """ + Sets the discrete timesteps used for the diffusion chain (to be run before inference). + + Args: + num_inference_steps (`int`): + The number of diffusion steps used when generating samples with a pre-trained model. + device (`str` or `torch.device`, *optional*): + The device to which the timesteps should be moved to. If `None`, the timesteps are not moved. + """ + + self.num_inference_steps = num_inference_steps + # "linspace", "leading", "trailing" corresponds to annotation of Table 2. of https://arxiv.org/abs/2305.08891 + if self.config.timestep_spacing == "linspace": + self._timesteps = ( + np.linspace(0, self.config.num_train_timesteps - 1, num_inference_steps).round().astype(np.int64) + ) + elif self.config.timestep_spacing == "leading": + step_ratio = self.config.num_train_timesteps // self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = (np.arange(0, num_inference_steps) * step_ratio).round() + self._timesteps += self.config.steps_offset + elif self.config.timestep_spacing == "trailing": + step_ratio = self.config.num_train_timesteps / self.num_inference_steps + # creates integer timesteps by multiplying by ratio + # casting to int to avoid issues when num_inference_step is power of 3 + self._timesteps = np.round(np.arange(self.config.num_train_timesteps, 0, -step_ratio))[::-1].astype( + np.int64 + ) + self._timesteps -= 1 + else: + raise ValueError( + f"{self.config.timestep_spacing} is not supported. Please make sure to choose one of 'linspace', 'leading' or 'trailing'." + ) + + if self.config.skip_prk_steps: + # for some models like stable diffusion the prk steps can/should be skipped to + # produce better results. When using PNDM with `self.config.skip_prk_steps` the implementation + # is based on crowsonkb's PLMS sampler implementation: https://github.com/CompVis/latent-diffusion/pull/51 + self.prk_timesteps = np.array([]) + self.plms_timesteps = np.concatenate([self._timesteps[:-1], self._timesteps[-2:-1], self._timesteps[-1:]])[ + ::-1 + ].copy() + else: + prk_timesteps = np.array(self._timesteps[-self.pndm_order :]).repeat(2) + np.tile( + np.array([0, self.config.num_train_timesteps // num_inference_steps // 2]), self.pndm_order + ) + self.prk_timesteps = (prk_timesteps[:-1].repeat(2)[1:-1])[::-1].copy() + self.plms_timesteps = self._timesteps[:-3][ + ::-1 + ].copy() # we copy to avoid having negative strides which are not supported by torch.from_numpy + + timesteps = np.concatenate([self.prk_timesteps, self.plms_timesteps]).astype(np.int64) + self.timesteps = torch.from_numpy(timesteps).to(device) + + self.ets = [] + self.counter = 0 + self.cur_model_output = 0 + + def step( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the diffusion + process from the learned model outputs (most often the predicted noise), and calls [`~PNDMScheduler.step_prk`] + or [`~PNDMScheduler.step_plms`] depending on the internal variable `counter`. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.counter < len(self.prk_timesteps) and not self.config.skip_prk_steps: + return self.step_prk(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + else: + return self.step_plms(model_output=model_output, timestep=timestep, sample=sample, return_dict=return_dict) + + def step_prk( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the Runge-Kutta method. It performs four forward passes to approximate the solution to the differential + equation. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + diff_to_prev = 0 if self.counter % 2 else self.config.num_train_timesteps // self.num_inference_steps // 2 + prev_timestep = timestep - diff_to_prev + timestep = self.prk_timesteps[self.counter // 4 * 4] + + if self.counter % 4 == 0: + self.cur_model_output += 1 / 6 * model_output + self.ets.append(model_output) + self.cur_sample = sample + elif (self.counter - 1) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 2) % 4 == 0: + self.cur_model_output += 1 / 3 * model_output + elif (self.counter - 3) % 4 == 0: + model_output = self.cur_model_output + 1 / 6 * model_output + self.cur_model_output = 0 + + # cur_sample should not be `None` + cur_sample = self.cur_sample if self.cur_sample is not None else sample + + prev_sample = self._get_prev_sample(cur_sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def step_plms( + self, + model_output: torch.FloatTensor, + timestep: int, + sample: torch.FloatTensor, + return_dict: bool = True, + ) -> Union[SchedulerOutput, Tuple]: + """ + Predict the sample from the previous timestep by reversing the SDE. This function propagates the sample with + the linear multistep method. It performs one forward pass multiple times to approximate the solution. + + Args: + model_output (`torch.FloatTensor`): + The direct output from learned diffusion model. + timestep (`int`): + The current discrete timestep in the diffusion chain. + sample (`torch.FloatTensor`): + A current instance of a sample created by the diffusion process. + return_dict (`bool`): + Whether or not to return a [`~schedulers.scheduling_utils.SchedulerOutput`] or tuple. + + Returns: + [`~schedulers.scheduling_utils.SchedulerOutput`] or `tuple`: + If return_dict is `True`, [`~schedulers.scheduling_utils.SchedulerOutput`] is returned, otherwise a + tuple is returned where the first element is the sample tensor. + + """ + if self.num_inference_steps is None: + raise ValueError( + "Number of inference steps is 'None', you need to run 'set_timesteps' after creating the scheduler" + ) + + if not self.config.skip_prk_steps and len(self.ets) < 3: + raise ValueError( + f"{self.__class__} can only be run AFTER scheduler has been run " + "in 'prk' mode for at least 12 iterations " + "See: https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/pipeline_pndm.py " + "for more information." + ) + + prev_timestep = timestep - self.config.num_train_timesteps // self.num_inference_steps + + if self.counter != 1: + self.ets = self.ets[-3:] + self.ets.append(model_output) + else: + prev_timestep = timestep + timestep = timestep + self.config.num_train_timesteps // self.num_inference_steps + + if len(self.ets) == 1 and self.counter == 0: + model_output = model_output + self.cur_sample = sample + elif len(self.ets) == 1 and self.counter == 1: + model_output = (model_output + self.ets[-1]) / 2 + sample = self.cur_sample + self.cur_sample = None + elif len(self.ets) == 2: + model_output = (3 * self.ets[-1] - self.ets[-2]) / 2 + elif len(self.ets) == 3: + model_output = (23 * self.ets[-1] - 16 * self.ets[-2] + 5 * self.ets[-3]) / 12 + else: + model_output = (1 / 24) * (55 * self.ets[-1] - 59 * self.ets[-2] + 37 * self.ets[-3] - 9 * self.ets[-4]) + + prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) + self.counter += 1 + + if not return_dict: + return (prev_sample,) + + return SchedulerOutput(prev_sample=prev_sample) + + def scale_model_input(self, sample: torch.FloatTensor, *args, **kwargs) -> torch.FloatTensor: + """ + Ensures interchangeability with schedulers that need to scale the denoising model input depending on the + current timestep. + + Args: + sample (`torch.FloatTensor`): + The input sample. + + Returns: + `torch.FloatTensor`: + A scaled input sample. + """ + return sample + + def _get_prev_sample(self, sample, timestep, prev_timestep, model_output): + # See formula (9) of PNDM paper https://arxiv.org/pdf/2202.09778.pdf + # this function computes x_(t−δ) using the formula of (9) + # Note that x_t needs to be added to both sides of the equation + + # Notation ( -> + # alpha_prod_t -> α_t + # alpha_prod_t_prev -> α_(t−δ) + # beta_prod_t -> (1 - α_t) + # beta_prod_t_prev -> (1 - α_(t−δ)) + # sample -> x_t + # model_output -> e_θ(x_t, t) + # prev_sample -> x_(t−δ) + alpha_prod_t = self.alphas_cumprod[timestep] + alpha_prod_t_prev = self.alphas_cumprod[prev_timestep] if prev_timestep >= 0 else self.final_alpha_cumprod + beta_prod_t = 1 - alpha_prod_t + beta_prod_t_prev = 1 - alpha_prod_t_prev + + if self.config.prediction_type == "v_prediction": + model_output = (alpha_prod_t**0.5) * model_output + (beta_prod_t**0.5) * sample + elif self.config.prediction_type != "epsilon": + raise ValueError( + f"prediction_type given as {self.config.prediction_type} must be one of `epsilon` or `v_prediction`" + ) + + # corresponds to (α_(t−δ) - α_t) divided by + # denominator of x_t in formula (9) and plus 1 + # Note: (α_(t−δ) - α_t) / (sqrt(α_t) * (sqrt(α_(t−δ)) + sqr(α_t))) = + # sqrt(α_(t−δ)) / sqrt(α_t)) + sample_coeff = (alpha_prod_t_prev / alpha_prod_t) ** (0.5) + + # corresponds to denominator of e_θ(x_t, t) in formula (9) + model_output_denom_coeff = alpha_prod_t * beta_prod_t_prev ** (0.5) + ( + alpha_prod_t * beta_prod_t * alpha_prod_t_prev + ) ** (0.5) + + # full formula (9) + prev_sample = ( + sample_coeff * sample - (alpha_prod_t_prev - alpha_prod_t) * model_output / model_output_denom_coeff + ) + + return prev_sample + + # Copied from diffusers.schedulers.scheduling_ddpm.DDPMScheduler.add_noise + def add_noise( + self, + original_samples: torch.FloatTensor, + noise: torch.FloatTensor, + timesteps: torch.IntTensor, + ) -> torch.FloatTensor: + # Make sure alphas_cumprod and timestep have same device and dtype as original_samples + # Move the self.alphas_cumprod to device to avoid redundant CPU to GPU data movement + # for the subsequent add_noise calls + self.alphas_cumprod = self.alphas_cumprod.to(device=original_samples.device) + alphas_cumprod = self.alphas_cumprod.to(dtype=original_samples.dtype) + timesteps = timesteps.to(original_samples.device) + + sqrt_alpha_prod = alphas_cumprod[timesteps] ** 0.5 + sqrt_alpha_prod = sqrt_alpha_prod.flatten() + while len(sqrt_alpha_prod.shape) < len(original_samples.shape): + sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1) + + sqrt_one_minus_alpha_prod = (1 - alphas_cumprod[timesteps]) ** 0.5 + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten() + while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape): + sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1) + + noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise + return noisy_samples + + def __len__(self): + return self.config.num_train_timesteps diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/__init__.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..58bbb8f14ee8e394705159b0ebf877de963c143a --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/__init__.py @@ -0,0 +1 @@ +from .autoencoder_kl import AutoencoderKL \ No newline at end of file diff --git a/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/autoencoder_kl.py b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/autoencoder_kl.py new file mode 100644 index 0000000000000000000000000000000000000000..0a9d02770b66dd1874224d14fad5f6d46f7f0370 --- /dev/null +++ b/MindIE/MindIE-Torch/built-in/foundation/stable_diffusion/stable_diffusion/stablediffusion/vae/autoencoder_kl.py @@ -0,0 +1,488 @@ +# Copyright 2023 The HuggingFace Team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import Dict, Optional, Tuple, Union + +import torch +import torch.nn as nn + +from diffusers.configuration_utils import ConfigMixin, register_to_config +from diffusers.loaders import FromOriginalVAEMixin +from diffusers.utils.accelerate_utils import apply_forward_hook +from ..layers.attention_processor import ( + ADDED_KV_ATTENTION_PROCESSORS, + CROSS_ATTENTION_PROCESSORS, + Attention, + AttentionProcessor, + AttnAddedKVProcessor, + AttnProcessor, +) +from diffusers.models.modeling_outputs import AutoencoderKLOutput +from diffusers.models.autoencoders.vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder + +from stablediffusion.models.model_utils import ModelMixin + + +class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin): + r""" + A VAE model with KL loss for encoding images into latents and decoding latent representations into images. + + This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented + for all models (such as downloading or saving). + + Parameters: + in_channels (int, *optional*, defaults to 3): Number of channels in the input image. + out_channels (int, *optional*, defaults to 3): Number of channels in the output. + down_block_types (`Tuple[str]`, *optional*, defaults to `("DownEncoderBlock2D",)`): + Tuple of downsample block types. + up_block_types (`Tuple[str]`, *optional*, defaults to `("UpDecoderBlock2D",)`): + Tuple of upsample block types. + block_out_channels (`Tuple[int]`, *optional*, defaults to `(64,)`): + Tuple of block output channels. + act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. + latent_channels (`int`, *optional*, defaults to 4): Number of channels in the latent space. + sample_size (`int`, *optional*, defaults to `32`): Sample input size. + scaling_factor (`float`, *optional*, defaults to 0.18215): + The component-wise standard deviation of the trained latent space computed using the first batch of the + training set. This is used to scale the latent space to have unit variance when training the diffusion + model. The latents are scaled with the formula `z = z * scaling_factor` before being passed to the + diffusion model. When decoding, the latents are scaled back to the original scale with the formula: `z = 1 + / scaling_factor * z`. For more details, refer to sections 4.3.2 and D.1 of the [High-Resolution Image + Synthesis with Latent Diffusion Models](https://arxiv.org/abs/2112.10752) paper. + force_upcast (`bool`, *optional*, default to `True`): + If enabled it will force the VAE to run in float32 for high image resolution pipelines, such as SD-XL. VAE + can be fine-tuned / trained to a lower range without loosing too much precision in which case + `force_upcast` can be set to `False` - see: https://huggingface.co/madebyollin/sdxl-vae-fp16-fix + """ + + _supports_gradient_checkpointing = True + + @register_to_config + def __init__( + self, + in_channels: int = 3, + out_channels: int = 3, + down_block_types: Tuple[str] = ("DownEncoderBlock2D",), + up_block_types: Tuple[str] = ("UpDecoderBlock2D",), + block_out_channels: Tuple[int] = (64,), + layers_per_block: int = 1, + act_fn: str = "silu", + latent_channels: int = 4, + norm_num_groups: int = 32, + sample_size: int = 32, + scaling_factor: float = 0.18215, + force_upcast: float = True, + ): + super().__init__() + + # pass init params to Encoder + self.encoder = Encoder( + in_channels=in_channels, + out_channels=latent_channels, + down_block_types=down_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + act_fn=act_fn, + norm_num_groups=norm_num_groups, + double_z=True, + ) + + # pass init params to Decoder + self.decoder = Decoder( + in_channels=latent_channels, + out_channels=out_channels, + up_block_types=up_block_types, + block_out_channels=block_out_channels, + layers_per_block=layers_per_block, + norm_num_groups=norm_num_groups, + act_fn=act_fn, + ) + + self.quant_conv = nn.Conv2d(2 * latent_channels, 2 * latent_channels, 1) + self.post_quant_conv = nn.Conv2d(latent_channels, latent_channels, 1) + + self.use_slicing = False + self.use_tiling = False + + # only relevant if vae tiling is enabled + self.tile_sample_min_size = self.config.sample_size + sample_size = ( + self.config.sample_size[0] + if isinstance(self.config.sample_size, (list, tuple)) + else self.config.sample_size + ) + self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1))) + self.tile_overlap_factor = 0.25 + + def _set_gradient_checkpointing(self, module, value=False): + if isinstance(module, (Encoder, Decoder)): + module.gradient_checkpointing = value + + def enable_tiling(self, use_tiling: bool = True): + r""" + Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to + compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow + processing larger images. + """ + self.use_tiling = use_tiling + + def disable_tiling(self): + r""" + Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.enable_tiling(False) + + def enable_slicing(self): + r""" + Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to + compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. + """ + self.use_slicing = True + + def disable_slicing(self): + r""" + Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing + decoding in one step. + """ + self.use_slicing = False + + @property + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.attn_processors + def attn_processors(self) -> Dict[str, AttentionProcessor]: + r""" + Returns: + `dict` of attention processors: A dictionary containing all attention processors used in the model with + indexed by its weight name. + """ + # set recursively + processors = {} + + def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]): + if hasattr(module, "get_processor"): + processors[f"{name}.processor"] = module.get_processor(return_deprecated_lora=True) + + for sub_name, child in module.named_children(): + fn_recursive_add_processors(f"{name}.{sub_name}", child, processors) + + return processors + + for name, module in self.named_children(): + fn_recursive_add_processors(name, module, processors) + + return processors + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_attn_processor + def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]): + r""" + Sets the attention processor to use to compute attention. + + Parameters: + processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`): + The instantiated processor class or a dictionary of processor classes that will be set as the processor + for **all** `Attention` layers. + + If `processor` is a dict, the key needs to define the path to the corresponding cross attention + processor. This is strongly recommended when setting trainable attention processors. + + """ + count = len(self.attn_processors.keys()) + + if isinstance(processor, dict) and len(processor) != count: + raise ValueError( + f"A dict of processors was passed, but the number of processors {len(processor)} does not match the" + f" number of attention layers: {count}. Please make sure to pass {count} processor classes." + ) + + def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor): + if hasattr(module, "set_processor"): + if not isinstance(processor, dict): + module.set_processor(processor) + else: + module.set_processor(processor.pop(f"{name}.processor")) + + for sub_name, child in module.named_children(): + fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor) + + for name, module in self.named_children(): + fn_recursive_attn_processor(name, module, processor) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor + def set_default_attn_processor(self): + """ + Disables custom attention processors and sets the default attention implementation. + """ + if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnAddedKVProcessor() + elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): + processor = AttnProcessor() + else: + raise ValueError( + f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}" + ) + + self.set_attn_processor(processor) + + @apply_forward_hook + def encode( + self, x: torch.FloatTensor, return_dict: bool = True + ) -> Union[AutoencoderKLOutput, Tuple[DiagonalGaussianDistribution]]: + """ + Encode a batch of images into latents. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + The latent representations of the encoded images. If `return_dict` is True, a + [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. + """ + if self.use_tiling and (x.shape[-1] > self.tile_sample_min_size or x.shape[-2] > self.tile_sample_min_size): + return self.tiled_encode(x, return_dict=return_dict) + + if self.use_slicing and x.shape[0] > 1: + encoded_slices = [self.encoder(x_slice) for x_slice in x.split(1)] + h = torch.cat(encoded_slices) + else: + h = self.encoder(x) + + moments = self.quant_conv(h) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def _decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + if self.use_tiling and (z.shape[-1] > self.tile_latent_min_size or z.shape[-2] > self.tile_latent_min_size): + return self.tiled_decode(z, return_dict=return_dict) + + z = self.post_quant_conv(z) + dec = self.decoder(z) + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + @apply_forward_hook + def decode( + self, z: torch.FloatTensor, return_dict: bool = True, generator=None + ) -> Union[DecoderOutput, torch.FloatTensor]: + """ + Decode a batch of images. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + + """ + if self.use_slicing and z.shape[0] > 1: + decoded_slices = [self._decode(z_slice).sample for z_slice in z.split(1)] + decoded = torch.cat(decoded_slices) + else: + decoded = self._decode(z).sample + + if not return_dict: + return (decoded,) + + return DecoderOutput(sample=decoded) + + def blend_v(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[2], b.shape[2], blend_extent) + for y in range(blend_extent): + b[:, :, y, :] = a[:, :, -blend_extent + y, :] * (1 - y / blend_extent) + b[:, :, y, :] * (y / blend_extent) + return b + + def blend_h(self, a: torch.Tensor, b: torch.Tensor, blend_extent: int) -> torch.Tensor: + blend_extent = min(a.shape[3], b.shape[3], blend_extent) + for x in range(blend_extent): + b[:, :, :, x] = a[:, :, :, -blend_extent + x] * (1 - x / blend_extent) + b[:, :, :, x] * (x / blend_extent) + return b + + def tiled_encode(self, x: torch.FloatTensor, return_dict: bool = True) -> AutoencoderKLOutput: + r"""Encode a batch of images using a tiled encoder. + + When this option is enabled, the VAE will split the input tensor into tiles to compute encoding in several + steps. This is useful to keep memory use constant regardless of image size. The end result of tiled encoding is + different from non-tiled encoding because each tile uses a different encoder. To avoid tiling artifacts, the + tiles overlap and are blended together to form a smooth output. You may still see tile-sized changes in the + output, but they should be much less noticeable. + + Args: + x (`torch.FloatTensor`): Input batch of images. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. + + Returns: + [`~models.autoencoder_kl.AutoencoderKLOutput`] or `tuple`: + If return_dict is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain + `tuple` is returned. + """ + overlap_size = int(self.tile_sample_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_latent_min_size * self.tile_overlap_factor) + row_limit = self.tile_latent_min_size - blend_extent + + # Split the image into 512x512 tiles and encode them separately. + rows = [] + for i in range(0, x.shape[2], overlap_size): + row = [] + for j in range(0, x.shape[3], overlap_size): + tile = x[:, :, i : i + self.tile_sample_min_size, j : j + self.tile_sample_min_size] + tile = self.encoder(tile) + tile = self.quant_conv(tile) + row.append(tile) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + moments = torch.cat(result_rows, dim=2) + posterior = DiagonalGaussianDistribution(moments) + + if not return_dict: + return (posterior,) + + return AutoencoderKLOutput(latent_dist=posterior) + + def tiled_decode(self, z: torch.FloatTensor, return_dict: bool = True) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Decode a batch of images using a tiled decoder. + + Args: + z (`torch.FloatTensor`): Input batch of latent vectors. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. + + Returns: + [`~models.vae.DecoderOutput`] or `tuple`: + If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is + returned. + """ + overlap_size = int(self.tile_latent_min_size * (1 - self.tile_overlap_factor)) + blend_extent = int(self.tile_sample_min_size * self.tile_overlap_factor) + row_limit = self.tile_sample_min_size - blend_extent + + # Split z into overlapping 64x64 tiles and decode them separately. + # The tiles have an overlap to avoid seams between tiles. + rows = [] + for i in range(0, z.shape[2], overlap_size): + row = [] + for j in range(0, z.shape[3], overlap_size): + tile = z[:, :, i : i + self.tile_latent_min_size, j : j + self.tile_latent_min_size] + tile = self.post_quant_conv(tile) + decoded = self.decoder(tile) + row.append(decoded) + rows.append(row) + result_rows = [] + for i, row in enumerate(rows): + result_row = [] + for j, tile in enumerate(row): + # blend the above tile and the left tile + # to the current tile and add the current tile to the result row + if i > 0: + tile = self.blend_v(rows[i - 1][j], tile, blend_extent) + if j > 0: + tile = self.blend_h(row[j - 1], tile, blend_extent) + result_row.append(tile[:, :, :row_limit, :row_limit]) + result_rows.append(torch.cat(result_row, dim=3)) + + dec = torch.cat(result_rows, dim=2) + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + def forward( + self, + sample: torch.FloatTensor, + sample_posterior: bool = False, + return_dict: bool = True, + generator: Optional[torch.Generator] = None, + ) -> Union[DecoderOutput, torch.FloatTensor]: + r""" + Args: + sample (`torch.FloatTensor`): Input sample. + sample_posterior (`bool`, *optional*, defaults to `False`): + Whether to sample from the posterior. + return_dict (`bool`, *optional*, defaults to `True`): + Whether or not to return a [`DecoderOutput`] instead of a plain tuple. + """ + x = sample + posterior = self.encode(x).latent_dist + if sample_posterior: + z = posterior.sample(generator=generator) + else: + z = posterior.mode() + dec = self.decode(z).sample + + if not return_dict: + return (dec,) + + return DecoderOutput(sample=dec) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.fuse_qkv_projections + def fuse_qkv_projections(self): + """ + Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, + key, value) are fused. For cross-attention modules, key and value projection matrices are fused. + + + + This API is 🧪 experimental. + + + """ + self.original_attn_processors = None + + for _, attn_processor in self.attn_processors.items(): + if "Added" in str(attn_processor.__class__.__name__): + raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.") + + self.original_attn_processors = self.attn_processors + + for module in self.modules(): + if isinstance(module, Attention): + module.fuse_projections(fuse=True) + + # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.unfuse_qkv_projections + def unfuse_qkv_projections(self): + """Disables the fused QKV projection if enabled. + + + + This API is 🧪 experimental. + + + + """ + if self.original_attn_processors is not None: + self.set_attn_processor(self.original_attn_processors) diff --git a/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/test_modeling_flax_{{cookiecutter.lowercase_modelname}}.py b/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/test_modeling_flax_{{cookiecutter.lowercase_modelname}}.py deleted file mode 100644 index 69b0a7fae2012908567036c3b727669f745cc543..0000000000000000000000000000000000000000 --- a/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/test_modeling_flax_{{cookiecutter.lowercase_modelname}}.py +++ /dev/null @@ -1,669 +0,0 @@ -# coding=utf-8 -# Copyright 2022 The HuggingFace Inc. team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -{% if cookiecutter.is_encoder_decoder_model == "False" %} - -import unittest - -from transformers import is_flax_available, {{cookiecutter.camelcase_modelname}}Config -from transformers.testing_utils import require_flax, slow - -from ..test_configuration_common import ConfigTester -from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor - -if is_flax_available(): - import numpy as np - from transformers import ( - Flax{{cookiecutter.camelcase_modelname}}ForCausalLM, - Flax{{cookiecutter.camelcase_modelname}}ForMaskedLM, - Flax{{cookiecutter.camelcase_modelname}}ForMultipleChoice, - Flax{{cookiecutter.camelcase_modelname}}ForQuestionAnswering, - Flax{{cookiecutter.camelcase_modelname}}ForSequenceClassification, - Flax{{cookiecutter.camelcase_modelname}}ForTokenClassification, - Flax{{cookiecutter.camelcase_modelname}}Model, - ) - - -class Flax{{cookiecutter.camelcase_modelname}}ModelTester: - def __init__( - self, - parent, - batch_size=13, - seq_length=7, - is_training=True, - use_input_mask=True, - use_token_type_ids=True, - use_labels=True, - vocab_size=99, - hidden_size=32, - num_hidden_layers=5, - num_attention_heads=4, - intermediate_size=37, - hidden_act="gelu", - hidden_dropout_prob=0.1, - attention_probs_dropout_prob=0.1, - max_position_embeddings=512, - type_vocab_size=16, - type_sequence_label_size=2, - initializer_range=0.02, - num_labels=3, - num_choices=4, - scope=None, - ): - self.parent = parent - self.batch_size = 13 - self.seq_length = 7 - self.is_training = True - self.use_input_mask = True - self.use_token_type_ids = True - self.use_labels = True - self.vocab_size = 99 - self.hidden_size = 32 - self.num_hidden_layers = 5 - self.num_attention_heads = 4 - self.intermediate_size = 37 - self.hidden_act = "gelu" - self.hidden_dropout_prob = 0.1 - self.attention_probs_dropout_prob = 0.1 - self.max_position_embeddings = 512 - self.type_vocab_size = 16 - self.type_sequence_label_size = 2 - self.initializer_range = 0.02 - self.num_labels = 3 - self.num_choices = 4 - self.scope = None - - def prepare_config_and_inputs(self): - input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) - - input_mask = None - if self.use_input_mask: - input_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2) - - token_type_ids = None - if self.use_token_type_ids: - token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size) - - sequence_labels = None - token_labels = None - choice_labels = None - if self.use_labels: - sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size) - token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels) - choice_labels = ids_tensor([self.batch_size], self.num_choices) - - config = {{cookiecutter.camelcase_modelname}}Config( - vocab_size=self.vocab_size, - hidden_size=self.hidden_size, - num_hidden_layers=self.num_hidden_layers, - num_attention_heads=self.num_attention_heads, - intermediate_size=self.intermediate_size, - hidden_act=self.hidden_act, - hidden_dropout_prob=self.hidden_dropout_prob, - attention_probs_dropout_prob=self.attention_probs_dropout_prob, - max_position_embeddings=self.max_position_embeddings, - type_vocab_size=self.type_vocab_size, - initializer_range=self.initializer_range, - return_dict=True, - ) - - return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - - def create_and_check_model( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - model = Flax{{cookiecutter.camelcase_modelname}}Model(config=config) - inputs = {"input_ids": input_ids, "attention_mask": input_mask, "token_type_ids": token_type_ids} - - inputs = [input_ids, input_mask] - - result = model(*inputs) - - self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size)) - - def create_and_check_lm_head( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - config.is_decoder = True - model = Flax{{cookiecutter.camelcase_modelname}}ForCausalLM(config=config) - inputs = { - "input_ids": input_ids, - "attention_mask": input_mask, - "token_type_ids": token_type_ids, - } - prediction_scores = model(**inputs)["logits"] - self.parent.assertListEqual( - list(prediction_scores.shape), [self.batch_size, self.seq_length, self.vocab_size] - ) - - def create_and_check_for_masked_lm( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - model = Flax{{cookiecutter.camelcase_modelname}}ForMaskedLM(config=config) - inputs = { - "input_ids": input_ids, - "attention_mask": input_mask, - "token_type_ids": token_type_ids, - } - result = model(**inputs) - self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size)) - - def create_and_check_for_sequence_classification( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - config.num_labels = self.num_labels - model = Flax{{cookiecutter.camelcase_modelname}}ForSequenceClassification(config=config) - inputs = { - "input_ids": input_ids, - "attention_mask": input_mask, - "token_type_ids": token_type_ids, - } - - result = model(**inputs) - self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels)) - - def create_and_check_for_multiple_choice( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - config.num_choices = self.num_choices - model = Flax{{cookiecutter.camelcase_modelname}}ForMultipleChoice(config=config) - multiple_choice_inputs_ids = np.tile(np.expand_dims(input_ids, 1), (1, self.num_choices, 1)) - multiple_choice_input_mask = np.tile(np.expand_dims(input_mask, 1), (1, self.num_choices, 1)) - multiple_choice_token_type_ids = np.tile(np.expand_dims(token_type_ids, 1), (1, self.num_choices, 1)) - inputs = { - "input_ids": multiple_choice_inputs_ids, - "attention_mask": multiple_choice_input_mask, - "token_type_ids": multiple_choice_token_type_ids, - } - result = model(**inputs) - self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices)) - - def create_and_check_for_token_classification( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - config.num_labels = self.num_labels - model = Flax{{cookiecutter.camelcase_modelname}}ForTokenClassification(config=config) - inputs = { - "input_ids": input_ids, - "attention_mask": input_mask, - "token_type_ids": token_type_ids, - } - result = model(**inputs) - self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels)) - - def create_and_check_for_question_answering( - self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels - ): - model = Flax{{cookiecutter.camelcase_modelname}}ForQuestionAnswering(config=config) - inputs = { - "input_ids": input_ids, - "attention_mask": input_mask, - "token_type_ids": token_type_ids, - } - - result = model(**inputs) - self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length)) - self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length)) - - def prepare_config_and_inputs_for_common(self): - config_and_inputs = self.prepare_config_and_inputs() - ( - config, - input_ids, - token_type_ids, - input_mask, - sequence_labels, - token_labels, - choice_labels, - ) = config_and_inputs - inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask} - return config, inputs_dict - - -@require_flax -class Flax{{cookiecutter.camelcase_modelname}}ModelTest(FlaxModelTesterMixin, unittest.TestCase): - - all_model_classes = ( - ( - Flax{{cookiecutter.camelcase_modelname}}Model, - Flax{{cookiecutter.camelcase_modelname}}ForCausalLM, - Flax{{cookiecutter.camelcase_modelname}}ForMaskedLM, - Flax{{cookiecutter.camelcase_modelname}}ForQuestionAnswering, - Flax{{cookiecutter.camelcase_modelname}}ForSequenceClassification, - Flax{{cookiecutter.camelcase_modelname}}ForTokenClassification, - Flax{{cookiecutter.camelcase_modelname}}ForMultipleChoice, - ) - if is_flax_available() - else () - ) - - test_head_masking = False - test_onnx = False - - def setUp(self): - self.model_tester = Flax{{cookiecutter.camelcase_modelname}}ModelTester(self) - self.config_tester = ConfigTester(self, config_class={{cookiecutter.camelcase_modelname}}Config, hidden_size=37) - - def test_config(self): - self.config_tester.run_common_tests() - - def test_model(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_model(*config_and_inputs) - - def test_for_masked_lm(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_for_masked_lm(*config_and_inputs) - - def test_for_causal_lm(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_lm_head(*config_and_inputs) - - def test_for_multiple_choice(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs) - - def test_for_question_answering(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_for_question_answering(*config_and_inputs) - - def test_for_sequence_classification(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs) - - def test_for_token_classification(self): - config_and_inputs = self.model_tester.prepare_config_and_inputs() - self.model_tester.create_and_check_for_token_classification(*config_and_inputs) - - @slow - def test_model_from_pretrained(self): - model = Flax{{cookiecutter.camelcase_modelname}}Model.from_pretrained("{{cookiecutter.checkpoint_identifier}}") - self.assertIsNotNone(model) - - -def _assert_tensors_equal(a, b, atol=1e-12, prefix=""): - """If tensors not close, or a and b arent both tensors, raise a nice Assertion error.""" - if a is None and b is None: - return True - try: - if _assert_tensors_equal(a, b, atol=atol): - return True - raise - except Exception: - if len(prefix) > 0: - prefix = f"{prefix}: " - raise AssertionError(f"{prefix}{a} != {b}") - - -@require_flax -class Flax{{cookiecutter.camelcase_modelname}}ModelIntegrationTest(unittest.TestCase): - @slow - def test_inference_masked_lm(self): - model = Flax{{cookiecutter.camelcase_modelname}}ForMaskedLM.from_pretrained("{{cookiecutter.checkpoint_identifier}}") - input_ids = np.array([[0, 1, 2, 3, 4, 5]]) - output = model(input_ids)[0] - - # TODO Replace vocab size - vocab_size = 32000 - - expected_shape = [1, 6, vocab_size] - self.assertEqual(output.shape, expected_shape) - - print(output[:, :3, :3]) - - # TODO Replace values below with what was printed above. - expected_slice = np.array( - [ - [ - [-0.05243197, -0.04498899, 0.05512108], - [-0.07444685, -0.01064632, 0.04352357], - [-0.05020351, 0.05530146, 0.00700043], - ] - ] - ) - _assert_tensors_equal(output[:, :3, :3], expected_slice, atol=1e-4) - -{% else %} -import unittest - -from transformers import ( - is_flax_available, - {{cookiecutter.camelcase_modelname}}Config, - {{cookiecutter.camelcase_modelname}}Tokenizer, -) -from transformers.testing_utils import require_sentencepiece, require_flax, require_tokenizers, slow - -from ..test_configuration_common import ConfigTester -from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor - - -if is_flax_available(): - import numpy as np - import jax.numpy as jnp - from transformers import ( - Flax{{cookiecutter.camelcase_modelname}}ForConditionalGeneration, - Flax{{cookiecutter.camelcase_modelname}}ForQuestionAnswering, - Flax{{cookiecutter.camelcase_modelname}}ForSequenceClassification, - Flax{{cookiecutter.camelcase_modelname}}Model, - ) - - -@require_flax -class Flax{{cookiecutter.camelcase_modelname}}ModelTester: - config_cls = {{cookiecutter.camelcase_modelname}}Config - config_updates = {} - hidden_act = "gelu" - - def __init__( - self, - parent, - batch_size=13, - seq_length=7, - is_training=True, - use_labels=False, - vocab_size=99, - hidden_size=32, - num_hidden_layers=5, - num_attention_heads=4, - intermediate_size=37, - hidden_dropout_prob=0.1, - attention_probs_dropout_prob=0.1, - max_position_embeddings=20, - eos_token_id=2, - pad_token_id=1, - bos_token_id=0, - ): - self.parent = parent - self.batch_size = batch_size - self.seq_length = seq_length - self.is_training = is_training - self.use_labels = use_labels - self.vocab_size = vocab_size - self.hidden_size = hidden_size - self.num_hidden_layers = num_hidden_layers - self.num_attention_heads = num_attention_heads - self.intermediate_size = intermediate_size - - self.hidden_dropout_prob = hidden_dropout_prob - self.attention_probs_dropout_prob = attention_probs_dropout_prob - self.max_position_embeddings = max_position_embeddings - self.eos_token_id = eos_token_id - self.pad_token_id = pad_token_id - self.bos_token_id = bos_token_id - - def prepare_config_and_inputs_for_common(self): - input_ids = ids_tensor([self.batch_size, self.seq_length - 1], self.vocab_size).clip(3, self.vocab_size) - eos_tensor = np.expand_dims(np.array([self.eos_token_id] * self.batch_size), 1) - input_ids = np.concatenate([input_ids, eos_tensor], axis=1) - - decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size) - - config = self.config_cls( - vocab_size=self.vocab_size, - d_model=self.hidden_size, - encoder_layers=self.num_hidden_layers, - decoder_layers=self.num_hidden_layers, - encoder_attention_heads=self.num_attention_heads, - decoder_attention_heads=self.num_attention_heads, - encoder_ffn_dim=self.intermediate_size, - decoder_ffn_dim=self.intermediate_size, - dropout=self.hidden_dropout_prob, - attention_dropout=self.attention_probs_dropout_prob, - max_position_embeddings=self.max_position_embeddings, - eos_token_ids=[2], - bos_token_id=self.bos_token_id, - pad_token_id=self.pad_token_id, - decoder_start_token_id=self.pad_token_id, - **self.config_updates, - ) - inputs_dict = prepare_{{cookiecutter.lowercase_modelname}}_inputs_dict(config, input_ids, decoder_input_ids) - return config, inputs_dict - - def check_use_cache_forward(self, model_class_name, config, inputs_dict): - max_decoder_length = 20 - model = model_class_name(config) - - encoder_outputs = model.encode(inputs_dict["input_ids"]) - - decoder_input_ids, decoder_attention_mask = ( - inputs_dict["decoder_input_ids"], - inputs_dict["decoder_attention_mask"], - ) - - past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs) - decoder_attention_mask = jnp.ones((decoder_input_ids.shape[0], max_decoder_length), dtype="i4") - - decoder_position_ids = jnp.broadcast_to( - jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :], - (decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1), - ) - outputs_cache = model.decode( - decoder_input_ids[:, :-1], - encoder_outputs, - decoder_attention_mask=decoder_attention_mask, - past_key_values=past_key_values, - decoder_position_ids=decoder_position_ids, - ) - - decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4") - outputs_cache_next = model.decode( - decoder_input_ids[:, -1:], - encoder_outputs, - decoder_attention_mask=decoder_attention_mask, - past_key_values=outputs_cache.past_key_values, - decoder_position_ids=decoder_position_ids, - ) - - outputs = model.decode(decoder_input_ids, encoder_outputs) - - diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))) - self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}") - - def check_use_cache_forward_with_attn_mask(self, model_class_name, config, inputs_dict): - max_decoder_length = 20 - model = model_class_name(config) - - encoder_outputs = model.encode(inputs_dict["input_ids"]) - - decoder_input_ids, decoder_attention_mask = ( - inputs_dict["decoder_input_ids"], - inputs_dict["decoder_attention_mask"], - ) - - decoder_attention_mask_cache = jnp.concatenate( - [ - decoder_attention_mask, - jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])), - ], - axis=-1, - ) - - past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs) - decoder_position_ids = jnp.broadcast_to( - jnp.arange(decoder_input_ids.shape[-1] - 1)[None, :], - (decoder_input_ids.shape[0], decoder_input_ids.shape[-1] - 1), - ) - - outputs_cache = model.decode( - decoder_input_ids[:, :-1], - encoder_outputs, - decoder_attention_mask=decoder_attention_mask_cache, - past_key_values=past_key_values, - decoder_position_ids=decoder_position_ids, - ) - decoder_position_ids = jnp.array(decoder_input_ids.shape[0] * [[decoder_input_ids.shape[-1] - 1]], dtype="i4") - outputs_cache_next = model.decode( - decoder_input_ids[:, -1:], - encoder_outputs, - past_key_values=outputs_cache.past_key_values, - decoder_attention_mask=decoder_attention_mask_cache, - decoder_position_ids=decoder_position_ids, - ) - - outputs = model.decode(decoder_input_ids, encoder_outputs, decoder_attention_mask=decoder_attention_mask) - - diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5]))) - self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}") - - -def prepare_{{cookiecutter.lowercase_modelname}}_inputs_dict( - config, - input_ids, - decoder_input_ids, - attention_mask=None, - decoder_attention_mask=None, -): - if attention_mask is None: - attention_mask = np.not_equal(input_ids, config.pad_token_id).astype(np.int8) - if decoder_attention_mask is None: - decoder_attention_mask = np.concatenate([np.ones(decoder_input_ids[:, :1].shape, dtype=np.int8), np.not_equal(decoder_input_ids[:, 1:], config.pad_token_id).astype(np.int8)], axis=-1) - return { - "input_ids": input_ids, - "decoder_input_ids": decoder_input_ids, - "attention_mask": attention_mask, - "decoder_attention_mask": decoder_attention_mask, - } - - -@require_flax -class Flax{{cookiecutter.camelcase_modelname}}ModelTest(FlaxModelTesterMixin, unittest.TestCase): - all_model_classes = ( - ( - Flax{{cookiecutter.camelcase_modelname}}ForConditionalGeneration, - Flax{{cookiecutter.camelcase_modelname}}ForQuestionAnswering, - Flax{{cookiecutter.camelcase_modelname}}ForSequenceClassification, - Flax{{cookiecutter.camelcase_modelname}}Model, - ) if is_flax_available() - else () - ) - all_generative_model_classes = (Flax{{cookiecutter.camelcase_modelname}}ForConditionalGeneration,) if is_flax_available() else () - is_encoder_decoder = True - test_pruning = False - test_head_masking = False - test_onnx = False - - def setUp(self): - self.model_tester = Flax{{cookiecutter.camelcase_modelname}}ModelTester(self) - self.config_tester = ConfigTester(self, config_class={{cookiecutter.camelcase_modelname}}Config) - - def test_config(self): - self.config_tester.run_common_tests() - - def test_use_cache_forward(self): - config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() - for model_class in self.all_model_classes: - self.model_tester.check_use_cache_forward(model_class, config, inputs_dict) - - def test_use_cache_forward_with_attn_mask(self): - config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common() - for model_class in self.all_model_classes: - self.model_tester.check_use_cache_forward_with_attn_mask(model_class, config, inputs_dict) - - -def _assert_tensors_equal(a, b, atol=1e-12, prefix=""): - """If tensors not close, or a and b arent both tensors, raise a nice Assertion error.""" - if a is None and b is None: - return True - try: - if _assert_tensors_equal(a, b, atol=atol): - return True - raise - except Exception: - if len(prefix) > 0: - prefix = f"{prefix}: " - raise AssertionError(f"{prefix}{a} != {b}") - - -def _long_tensor(tok_lst): - return np.array(tok_lst, dtype=np.int32) - - -TOLERANCE = 1e-4 - - -@slow -@require_sentencepiece -@require_tokenizers -@require_flax -class Flax{{cookiecutter.camelcase_modelname}}ModelIntegrationTest(unittest.TestCase): - def test_inference_no_head(self): - model = Flax{{cookiecutter.camelcase_modelname}}Model.from_pretrained('{{cookiecutter.checkpoint_identifier}}') - # change to intended input here - input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) - decoder_input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) - inputs_dict = prepare_{{cookiecutter.lowercase_modelname}}_inputs_dict(model.config, input_ids, decoder_input_ids) - output = model(**inputs_dict)[0] - expected_shape = (1, 11, 1024) - self.assertEqual(output.shape, expected_shape) - # change to expected output here - expected_slice = np.array( - [[0.7144, 0.8143, -1.2813], [0.7144, 0.8143, -1.2813], [-0.0467, 2.5911, -2.1845]], - ) - _assert_tensors_equal(output[:, :3, :3], expected_slice, atol=TOLERANCE) - - def test_inference_with_head(self): - model = Flax{{cookiecutter.camelcase_modelname}}ForConditionalGeneration.from_pretrained('{{cookiecutter.checkpoint_identifier}}') - # change to intended input here - input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) - decoder_input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]]) - inputs_dict = prepare_{{cookiecutter.lowercase_modelname}}_inputs_dict(model.config, input_ids, decoder_input_ids) - output = model(**inputs_dict)[0] - expected_shape = (1, 11, 1024) - self.assertEqual(output.shape, expected_shape) - # change to expected output here - expected_slice = np.array( - [[0.7144, 0.8143, -1.2813], [0.7144, 0.8143, -1.2813], [-0.0467, 2.5911, -2.1845]], - ) - _assert_tensors_equal(output[:, :3, :3], expected_slice, atol=TOLERANCE) - - def test_seq_to_seq_generation(self): - hf = Flax{{cookiecutter.camelcase_modelname}}ForConditionalGeneration.from_pretrained('{{cookiecutter.checkpoint_identifier}}') - tok = {{cookiecutter.camelcase_modelname}}Tokenizer.from_pretrained('{{cookiecutter.checkpoint_identifier}}') - - batch_input = [ - # string 1, - # string 2, - # string 3, - # string 4, - ] - - # The below article tests that we don't add any hypotheses outside of the top n_beams - dct = tok.batch_encode_plus( - batch_input, - max_length=512, - padding="max_length", - truncation_strategy="only_first", - truncation=True, - return_tensors="np", - ) - - hypotheses_batch = hf.generate( - input_ids=dct["input_ids"], - attention_mask=dct["attention_mask"], - num_beams=2, - ) - - EXPECTED = [ - # here expected 1, - # here expected 2, - # here expected 3, - # here expected 4, - ] - - generated = tok.batch_decode( - hypotheses_batch.tolist(), clean_up_tokenization_spaces=True, skip_special_tokens=True - ) - assert generated == EXPECTED -{%- endif %} diff --git a/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/tokenization_fast_{{cookiecutter.lowercase_modelname}}.py b/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/tokenization_fast_{{cookiecutter.lowercase_modelname}}.py deleted file mode 100644 index 6e6c93698367feabaf050af25355e8386089bebc..0000000000000000000000000000000000000000 --- a/PyTorch/built-in/nlp/Bert-text-classification_for_PyTorch/transformers/templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/tokenization_fast_{{cookiecutter.lowercase_modelname}}.py +++ /dev/null @@ -1,201 +0,0 @@ -# coding=utf-8 -# Copyright 2022 {{cookiecutter.authors}} and The HuggingFace Inc. team. All rights reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. -"""Tokenization classes for {{cookiecutter.modelname}}.""" - -{%- if cookiecutter.tokenizer_type == "Based on BERT" %} -from ...utils import logging -from ..bert.tokenization_bert_fast import BertTokenizerFast -from .tokenization_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Tokenizer - - -logger = logging.get_logger(__name__) - -VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt"} - -PRETRAINED_VOCAB_FILES_MAP = { - "vocab_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/vocab.txt", - } -} - -PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { - "{{cookiecutter.checkpoint_identifier}}": 512, -} - - -PRETRAINED_INIT_CONFIGURATION = { - "{{cookiecutter.checkpoint_identifier}}": {"do_lower_case": False}, -} - - -class {{cookiecutter.camelcase_modelname}}TokenizerFast(BertTokenizerFast): - r""" - Construct a "fast" {{cookiecutter.modelname}} tokenizer (backed by HuggingFace's *tokenizers* library). - - [`~{{cookiecutter.camelcase_modelname}}TokenizerFast`] is identical to [`BertTokenizerFast`] and runs - end-to-end tokenization: punctuation splitting and wordpiece. - - Refer to superclass [`BertTokenizerFast`] for usage examples and documentation concerning - parameters. - """ - - vocab_files_names = VOCAB_FILES_NAMES - pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP - max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES - pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION - slow_tokenizer_class = {{cookiecutter.camelcase_modelname}}Tokenizer - -{%- elif cookiecutter.tokenizer_type == "Based on BART" %} -from ...utils import logging -from ..bart.tokenization_bart_fast import BartTokenizerFast -from .tokenization_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Tokenizer - - -logger = logging.get_logger(__name__) - -VOCAB_FILES_NAMES = {"vocab_file": "vocab.json", "merges_file": "merges.txt", "tokenizer_file": "tokenizer.json"} - -PRETRAINED_VOCAB_FILES_MAP = { - "vocab_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/vocab.json", - }, - "merges_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/merges.txt", - }, - "tokenizer_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/tokenizer.json", - }, -} - -PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { - "{{cookiecutter.checkpoint_identifier}}": 1024, -} - - -class {{cookiecutter.camelcase_modelname}}TokenizerFast(BartTokenizerFast): - r""" - Construct a "fast" {{cookiecutter.modelname}} tokenizer (backed by HuggingFace's *tokenizers* library). - - [`~{{cookiecutter.camelcase_modelname}}TokenizerFast`] is identical to [`BartTokenizerFast`] and runs - end-to-end tokenization: punctuation splitting and wordpiece. - - Refer to superclass [`BartTokenizerFast`] for usage examples and documentation concerning - parameters. - """ - - vocab_files_names = VOCAB_FILES_NAMES - pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP - max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES - slow_tokenizer_class = {{cookiecutter.camelcase_modelname}}Tokenizer - -{%- elif cookiecutter.tokenizer_type == "Standalone" %} -from typing import List, Optional - -from tokenizers import ByteLevelBPETokenizer - -from ...tokenization_utils_fast import PreTrainedTokenizerFast -from ...utils import logging -from .tokenization_{{cookiecutter.lowercase_modelname}} import {{cookiecutter.camelcase_modelname}}Tokenizer - - -logger = logging.get_logger(__name__) - -VOCAB_FILES_NAMES = {"vocab_file": "vocab.txt", "tokenizer_file": "tokenizer.json"} - -PRETRAINED_VOCAB_FILES_MAP = { - "vocab_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/vocab.txt", - }, - "tokenizer_file": { - "{{cookiecutter.checkpoint_identifier}}": "https://huggingface.co/{{cookiecutter.checkpoint_identifier}}/resolve/main/tokenizer.json", - }, -} - -PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = { - "{{cookiecutter.checkpoint_identifier}}": 1024, -} - -class {{cookiecutter.camelcase_modelname}}TokenizerFast(PreTrainedTokenizerFast): - """ - Construct a "fast" {{cookiecutter.modelname}} tokenizer (backed by HuggingFace's *tokenizers* library). - - Args: - vocab_file (`str`): - Path to the vocabulary file. - """ - - vocab_files_names = VOCAB_FILES_NAMES - pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP - max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES - slow_tokenizer_class = {{cookiecutter.camelcase_modelname}}Tokenizer - - def __init__( - self, - vocab_file, - merges_file, - unk_token="<|endoftext|>", - bos_token="<|endoftext|>", - eos_token="<|endoftext|>", - add_prefix_space=False, - trim_offsets=True, - **kwargs - ): - super().__init__( - ByteLevelBPETokenizer( - vocab_file=vocab_file, - merges_file=merges_file, - add_prefix_space=add_prefix_space, - trim_offsets=trim_offsets, - ), - bos_token=bos_token, - eos_token=eos_token, - unk_token=unk_token, - **kwargs, - ) - self.add_prefix_space = add_prefix_space - - def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None): - output = [self.bos_token_id] + token_ids_0 + [self.eos_token_id] - if token_ids_1 is None: - return output - - return output + [self.eos_token_id] + token_ids_1 + [self.eos_token_id] - - - def create_token_type_ids_from_sequences( - self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None - ) -> List[int]: - """ - Create a mask from the two sequences passed to be used in a sequence-pair classification task. - {{cookiecutter.modelname}} does not make use of token type ids, therefore a list of zeros is returned. - - Args: - token_ids_0 (`List[int]`): - List of IDs. - token_ids_1 (`List[int]`, *optional*): - Optional second list of IDs for sequence pairs. - - Returns: - `List[int]`: List of zeros. - """ - sep = [self.sep_token_id] - cls = [self.cls_token_id] - - if token_ids_1 is None: - return len(cls + token_ids_0 + sep) * [0] - return len(cls + token_ids_0 + sep + sep + token_ids_1 + sep) * [0] - - -{% endif %} diff --git a/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/COCO-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_200k_bs64_crop_640_640_coco_dsconv.yaml b/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/COCO-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_200k_bs64_crop_640_640_coco_dsconv.yaml deleted file mode 100644 index 6944c6fdf3dcaafdc0a740188610fe604cb7d3be..0000000000000000000000000000000000000000 --- a/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/COCO-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_200k_bs64_crop_640_640_coco_dsconv.yaml +++ /dev/null @@ -1,42 +0,0 @@ -_BASE_: ../Cityscapes-PanopticSegmentation/Base-PanopticDeepLab-OS16.yaml -MODEL: - WEIGHTS: "detectron2://DeepLab/R-52.pkl" - PIXEL_MEAN: [123.675, 116.280, 103.530] - PIXEL_STD: [58.395, 57.120, 57.375] - BACKBONE: - NAME: "build_resnet_deeplab_backbone" - RESNETS: - DEPTH: 50 - NORM: "SyncBN" - RES5_MULTI_GRID: [1, 2, 4] - STEM_TYPE: "deeplab" - STEM_OUT_CHANNELS: 128 - STRIDE_IN_1X1: False - SEM_SEG_HEAD: - NUM_CLASSES: 133 - LOSS_TOP_K: 1.0 - USE_DEPTHWISE_SEPARABLE_CONV: True - PANOPTIC_DEEPLAB: - STUFF_AREA: 4096 - NMS_KERNEL: 41 - SIZE_DIVISIBILITY: 640 - USE_DEPTHWISE_SEPARABLE_CONV: True -DATASETS: - TRAIN: ("coco_2017_train_panoptic",) - TEST: ("coco_2017_val_panoptic",) -SOLVER: - BASE_LR: 0.0005 - MAX_ITER: 200000 - IMS_PER_BATCH: 64 -INPUT: - FORMAT: "RGB" - GAUSSIAN_SIGMA: 8 - MIN_SIZE_TRAIN: !!python/object/apply:eval ["[int(x * 0.1 * 640) for x in range(5, 16)]"] - MIN_SIZE_TRAIN_SAMPLING: "choice" - MIN_SIZE_TEST: 640 - MAX_SIZE_TRAIN: 960 - MAX_SIZE_TEST: 640 - CROP: - ENABLED: True - TYPE: "absolute" - SIZE: (640, 640) diff --git a/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024.yaml b/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024.yaml deleted file mode 100644 index fde902bb2a87ccaf2c6fea4e79be4144ca44e239..0000000000000000000000000000000000000000 --- a/PyTorch/dev/cv/image_classification/SlowFast_ID0646_for_PyTorch/detectron2/projects/Panoptic-DeepLab/configs/Cityscapes-PanopticSegmentation/panoptic_deeplab_R_52_os16_mg124_poly_90k_bs32_crop_512_1024.yaml +++ /dev/null @@ -1,20 +0,0 @@ -_BASE_: Base-PanopticDeepLab-OS16.yaml -MODEL: - WEIGHTS: "detectron2://DeepLab/R-52.pkl" - PIXEL_MEAN: [123.675, 116.280, 103.530] - PIXEL_STD: [58.395, 57.120, 57.375] - BACKBONE: - NAME: "build_resnet_deeplab_backbone" - RESNETS: - DEPTH: 50 - NORM: "SyncBN" - RES5_MULTI_GRID: [1, 2, 4] - STEM_TYPE: "deeplab" - STEM_OUT_CHANNELS: 128 - STRIDE_IN_1X1: False -SOLVER: - MAX_ITER: 90000 -INPUT: - FORMAT: "RGB" - CROP: - SIZE: (512, 1024)