From 74670bcd04cbc063a0835e3cae61eac1c2d0d148 Mon Sep 17 00:00:00 2001
From: koervcor <1015296415@qq.com>
Date: Fri, 17 Feb 2023 14:57:59 +0800
Subject: [PATCH 1/2] Update ConvNeXt

---
 .../ConvNeXt/1.5_requirements.txt             |   8 +
 .../ConvNeXt/1.8_requirements.txt             |   8 +
 .../cv/classification/ConvNeXt/README.md      | 249 +++++++++++++-----
 .../cv/classification/ConvNeXt/engine.py      |  17 +-
 .../cv/classification/ConvNeXt/main.py        |   3 +-
 5 files changed, 201 insertions(+), 84 deletions(-)
 create mode 100644 PyTorch/contrib/cv/classification/ConvNeXt/1.5_requirements.txt
 create mode 100644 PyTorch/contrib/cv/classification/ConvNeXt/1.8_requirements.txt

diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/1.5_requirements.txt b/PyTorch/contrib/cv/classification/ConvNeXt/1.5_requirements.txt
new file mode 100644
index 0000000000..da9be54422
--- /dev/null
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/1.5_requirements.txt
@@ -0,0 +1,8 @@
+six==1.15.0
+timm==0.3.2
+torchvision==0.2.2.post3
+pillow==8.4.0
+tensorboardX==2.5
+decorator
+protobuf==3.20.1
+sympy
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/1.8_requirements.txt b/PyTorch/contrib/cv/classification/ConvNeXt/1.8_requirements.txt
new file mode 100644
index 0000000000..f30b2205ef
--- /dev/null
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/1.8_requirements.txt
@@ -0,0 +1,8 @@
+six==1.15.0
+timm==0.3.2
+torchvision==0.9.1
+pillow==9.1.0
+tensorboardX==2.5
+decorator
+protobuf==3.20.1
+sympy
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/README.md b/PyTorch/contrib/cv/classification/ConvNeXt/README.md
index 9764a62ba4..adf519c6a7 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/README.md
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/README.md
@@ -1,69 +1,180 @@
-# ConvNext_for_PyTorch
-
-This implements training ConvNext of  on the ImageNet dataset, mainly modified from https://github.com/facebookresearch/ConvNeXt.git 
-
-## ConvNext_for_PyTorch Detail 
-
-As of the current date, Ascend-Pytorch is still inefficient for contiguous operations. 
-
-
-
-## Requirements 
-- pip install -r requirements.txt
-- pip install torch==1.8.1+ascend.rc2.20220505;torchvision==0.9.1;torch-npu 1.8.1rc2.post20220505;
-- Download the ImageNet dataset from http://www.image-net.org/
-    - Then, and move validation images to labeled subfolders, using [the following shell script](https://raw.githubusercontent.com/soumith/imagenetloader.torch/master/valprep.sh)
-## timm 
-将timm_need目录下的文件替换到timm的安装目录下
-```bash
-
-cd ../ConvNeXt
-/bin/cp -f timm_need/mixup.py ../timm/data/mixup.py
-/bin/cp -f timm_need/model_ema.py ../timm/utils/model_ema.py
-
-```
-## 软件包
-- 910版本
-- CANN toolkit_5.1.RC1
-- torch 1.8.1+ascend.rc2.20220505
-- 固件驱动 22.0.0
-
-## Training 
-
-To train a model, run `main.py` with the desired model architecture and the path to the ImageNet dataset:
-
-```bash
-# training 1p accuracy
-bash ./test/train_full_1p.sh --data_path=real_data_path
-
-# training 1p performance
-bash ./test/train_performance_1p.sh --data_path=real_data_path
-
-# training 8p accuracy
-bash ./test/train_full_8p.sh --data_path=real_data_path
-
-# training 8p performance
-bash ./test/train_performance_8p.sh --data_path=real_data_path
-
-#  eval 
-bash test/train_eval_8p.sh --data_path=real_data_path
-
-# finetuning
-bash test/train_finetune_1p.sh --data_path=real_data_path
-
-
-```
-
-## ConvNext_for_PyTorch training result 
-
-|  Acc@1  |   FPS    | Npu_nums | Epochs | AMP_Type |
-| :-----: | :------: | :------: | :----: | :------: |
-|    -    |    115.10      |    1     |  300   |    O1    |
-| 82.049 | 259.85 |    8     |  300   |    O1    |
-
-
-
-
-
-
-
+# ConvNeXt for PyTorch
+
+-   [概述](概述.md)
+-   [准备训练环境](准备训练环境.md)
+-   [开始训练](开始训练.md)
+-   [训练结果展示](训练结果展示.md)
+-   [版本说明](版本说明.md)
+
+
+
+# 概述
+
+## 简述
+
+ConvNeXt是基于ResNet-50网络结构，并参考Transformer网络的设计思想和技术引入到CNN网络模块中以结合这两种网络的优势，从而提高CNN网络的性能表现。其进行的优化设计主要有以下几点： 1.Macro design 2.ResNeXt 3.Inverted bottleneck 4. Large kernel size 5.Various layer-wise Micro designs
+
+- 参考实现：
+
+  ```
+  url=https://github.com/facebookresearch/ConvNeXt.git
+  commit_id=b9dd6b5a0885c381e7e7cd99ba64d1b61041bdfc
+  ```
+
+- 适配昇腾 AI 处理器的实现：
+
+  ```
+  url=https://gitee.com/ascend/ModelZoo-PyTorch.git
+  code_path=PyTorch/contrib/cv/classification
+  ```
+
+
+# 准备训练环境
+
+## 准备环境
+
+- 当前模型支持的 PyTorch 版本和已知三方库依赖如下表所示。
+
+  **表 1**  版本支持表
+
+  | Torch_Version      | 三方库依赖版本                                 |
+  | :--------: | :----------------------------------------------------------: |
+  | PyTorch 1.5 | torchvision==0.2.2.post3；pillow==8.4.0 |
+  | PyTorch 1.8 | torchvision==0.9.1；pillow==9.1.0 |
+
+- 环境准备指导。
+
+  请参考《[Pytorch框架训练环境准备](https://www.hiascend.com/document/detail/zh/ModelZoo/pytorchframework/ptes)》。
+  
+- 安装依赖。
+
+  在模型源码包根目录下执行命令，安装模型对应PyTorch版本需要的依赖。
+  ```
+  pip install -r 1.5_requirements.txt  # PyTorch1.5版本
+  
+  pip install -r 1.8_requirements.txt  # PyTorch1.8版本
+  ```
+  > **说明：** 
+  >只需执行一条对应的PyTorch版本依赖安装命令。
+
+- timm 配置。
+
+  将 `timm_need` 目录下的文件替换到 `timm` 的安装目录下。
+  ```bash
+  cd ../ConvNeXt
+  /bin/cp -f timm_need/mixup.py ../timm/data/mixup.py
+  ```
+
+## 准备数据集
+
+1. 获取数据集。
+
+   用户自行获取 `ImageNet` 数据集，将数据集上传到服务器任意路径下并解压。
+
+   数据集目录结构参考如下所示。
+
+   ```
+   ├── ImageNet
+         ├──train
+              ├──类别1
+                    │──图片1
+                    │──图片2
+                    │   ...       
+              ├──类别2
+                    │──图片1
+                    │──图片2
+                    │   ...   
+              ├──...                     
+         ├──val  
+              ├──类别1
+                    │──图片1
+                    │──图片2
+                    │   ...       
+              ├──类别2
+                    │──图片1
+                    │──图片2
+                    │   ...              
+   ```
+   > **说明：** 
+   >该数据集的训练过程脚本只作为一种参考示例。
+
+# 开始训练
+
+## 训练模型
+
+1. 进入解压后的源码包根目录。
+
+   ```
+   cd /${模型文件夹名称} 
+   ```
+
+2. 运行训练脚本。
+
+   该模型支持单机单卡训练和单机8卡训练。
+
+   - 单机单卡训练
+
+     启动单卡训练。
+
+     ```
+     bash ./test/train_full_1p.sh --data_path=/data/xxx/  # 单卡精度
+     bash ./test/train_performance_1p.sh --data_path=/data/xxx/  # 单卡性能
+     ```
+
+   - 单机8卡训练
+
+     启动8卡训练。
+
+     ```
+     bash ./test/train_full_8p.sh --data_path=/data/xxx/  # 8卡精度
+     bash ./test/train_performance_8p.sh --data_path=/data/xxx/  # 8卡性能
+     ```
+
+   - 单机8卡评测
+
+     启动8卡评测。
+
+     ```
+     bash ./test/train_eval_8p.sh --data_path=/data/xxx/  # 8卡评测
+     ```
+
+   --data_path参数填写数据集路径，需写到数据集的一级目录。
+
+   模型训练脚本参数说明如下。
+
+   ```
+   公共参数：
+   --data                              //数据集路径
+   --model                             //模型训练名称，默认：convnext_tiny
+   --drop_path                         //防过拟合的比例，默认：0.1
+   --batch_size                        //训练批次大小，默认：128      
+   --lr                                //学习率，默认：0.004
+   --update_freq                       //梯度累积更新，默认：4
+   --use_amp                           //是否使用自动混合精度，默认：True
+   --model_ema                         //是否开启权重滑动更新，默认：True
+   --model_ema_eval                    //是否在训期间使用ema模型进行评估，默认：True
+   ```
+   
+   训练完成后，权重文件保存在当前路径下，并输出模型训练精度和性能信息。
+
+# 训练结果展示
+
+**表 2**  训练结果展示表
+
+| NAME    | Acc@1 |  FPS | Epochs | AMP_Type | Torch_Version |
+| :-----: | :---: | :--: | :----: | :------: | :-----------: |
+| 1p-竞品V| - | - |  1  | - | 1.5 |
+| 8p-竞品V| - | - | 300 | - | 1.5 |
+| 1p-NPU  | -     |  119 | 1      |       O2 |      1.8      |
+| 8p-NPU  | 82.05 |  261 | 300    |       O2 |      1.8      |
+
+
+# 版本说明
+
+## 变更
+
+2023.02.07：首次发布。
+
+## FAQ
+
+无。
\ No newline at end of file
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/engine.py b/PyTorch/contrib/cv/classification/ConvNeXt/engine.py
index eb815e594a..3377fb7fb0 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/engine.py
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/engine.py
@@ -80,14 +80,9 @@ def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
             if (data_iter_step + 1) % update_freq == 0:
                 optimizer.step()
                 optimizer.zero_grad()
-                # if model_ema is not None:
-                #     model_ema.update(model)
                 if model_ema is not None:
-                    if device.type == 'npu':
-                        params_fp32_fused = optimizer.get_model_combined_params()
-                        model_ema.update(model, params_fp32_fused[0],device)
-                    else:
-                        model_ema.update(model)
+                    model_ema.update(model)
+
                     
         else: # full precision
             loss /= update_freq
@@ -95,14 +90,8 @@ def train_one_epoch(model: torch.nn.Module, criterion: torch.nn.Module,
             if (data_iter_step + 1) % update_freq == 0:
                 optimizer.step()
                 optimizer.zero_grad()
-                # if model_ema is not None:
-                #     model_ema.update(model)
                 if model_ema is not None:
-                    if device.type == 'npu':
-                        params_fp32_fused = optimizer.get_model_combined_params()
-                        model_ema.update(model, params_fp32_fused[0])
-                    else:
-                        model_ema.update(model)
+                    model_ema.update(model)
 
         torch.npu.synchronize()
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/main.py b/PyTorch/contrib/cv/classification/ConvNeXt/main.py
index 919d2d666e..1f3cd9d32e 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/main.py
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/main.py
@@ -11,7 +11,8 @@ import datetime
 import numpy as np
 import time
 import torch
-import torch_npu
+if torch.__version__ >= '1.8':
+    import torch_npu
 import torch.nn as nn
 import torch.backends.cudnn as cudnn
 import json
-- 
Gitee


From 166704b2fedfce2dab9422bb4a981ed544a78556 Mon Sep 17 00:00:00 2001
From: koervcor <1015296415@qq.com>
Date: Sun, 23 Apr 2023 15:47:15 +0800
Subject: [PATCH 2/2] =?UTF-8?q?=E4=BF=AE=E6=94=B9batchsize=E8=AE=A1?=
 =?UTF-8?q?=E7=AE=97=E6=96=B9=E6=B3=95?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 PyTorch/contrib/cv/classification/ConvNeXt/README.md     | 9 +++++----
 PyTorch/contrib/cv/classification/ConvNeXt/main.py       | 4 ++--
 .../cv/classification/ConvNeXt/test/train_finetune_1p.sh | 2 +-
 .../cv/classification/ConvNeXt/test/train_full_1p.sh     | 2 +-
 .../cv/classification/ConvNeXt/test/train_full_8p.sh     | 2 +-
 .../classification/ConvNeXt/test/train_performance_1p.sh | 2 +-
 .../classification/ConvNeXt/test/train_performance_8p.sh | 2 +-
 7 files changed, 12 insertions(+), 11 deletions(-)

diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/README.md b/PyTorch/contrib/cv/classification/ConvNeXt/README.md
index adf519c6a7..5bf0bb3689 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/README.md
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/README.md
@@ -165,16 +165,17 @@ ConvNeXt是基于ResNet-50网络结构，并参考Transformer网络的设计思
 | :-----: | :---: | :--: | :----: | :------: | :-----------: |
 | 1p-竞品V| - | - |  1  | - | 1.5 |
 | 8p-竞品V| - | - | 300 | - | 1.5 |
-| 1p-NPU  | -     |  119 | 1      |       O2 |      1.8      |
-| 8p-NPU  | 82.05 |  261 | 300    |       O2 |      1.8      |
+| 1p-NPU  | -     |  370.60 | 1      |       O2 |      1.8      |
+| 8p-NPU  | 82.05 |  2968.12 | 300    |       O2 |      1.8      |
 
 
 # 版本说明
 
 ## 变更
 
-2023.02.07：首次发布。
+2023.04.23: 更新内容重新发布。
+2022.08.09: 首次发布。
 
 ## FAQ
 
-无。
\ No newline at end of file
+无。
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/main.py b/PyTorch/contrib/cv/classification/ConvNeXt/main.py
index 1f3cd9d32e..fdcf153c9b 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/main.py
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/main.py
@@ -8,6 +8,8 @@
 
 import argparse
 import datetime
+import json
+import os
 import numpy as np
 import time
 import torch
@@ -15,8 +17,6 @@ if torch.__version__ >= '1.8':
     import torch_npu
 import torch.nn as nn
 import torch.backends.cudnn as cudnn
-import json
-import os
 
 from pathlib import Path
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_finetune_1p.sh b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_finetune_1p.sh
index 61dc55f6cc..28396e76f7 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_finetune_1p.sh
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_finetune_1p.sh
@@ -14,7 +14,7 @@ data_path=""
 Network="ConvNext"
 
 #训练batch_size,,需要模型审视修改
-batch_size=16
+batch_size=128
 
 
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_1p.sh b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_1p.sh
index d638c75e5b..a64769abf6 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_1p.sh
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_1p.sh
@@ -14,7 +14,7 @@ data_path=""
 Network="ConvNext"
 
 #训练batch_size,,需要模型审视修改
-batch_size=16
+batch_size=128
 
 
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_8p.sh b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_8p.sh
index 76a658166c..0d1e79974e 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_8p.sh
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_full_8p.sh
@@ -94,7 +94,7 @@ e2e_time=$(( $end_time - $start_time ))
 #最后一个迭代FPS值
 step_time=`grep -a 'time:'  ${cur_path}/test/output/$ASCEND_DEVICE_ID/train_full_8p.log|awk -F "time: " '{print $NF}'|awk 'END {print}'| awk -F "  data:" '{print $1}'`
 Acc=`grep -a 'Acc@1' ${cur_path}/test/output/$ASCEND_DEVICE_ID/train_full_8p.log | awk 'END {print}'| awk -F " " '{print $3}'`
-FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${step_time}'}'`
+FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}' * 8/'${step_time}'}'`
 #最后一个迭代loss值
 loss=`grep -a 'loss:'  ${cur_path}/test/output/$ASCEND_DEVICE_ID/train_full_8p.log | awk -F "loss:" '{print $NF}'| awk 'END {print}' | awk -F "(" '{print $1}'`
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_1p.sh b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_1p.sh
index 27bdf5e293..3dcd1507da 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_1p.sh
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_1p.sh
@@ -14,7 +14,7 @@ data_path=""
 Network="ConvNext"
 
 #训练batch_size,,需要模型审视修改
-batch_size=16
+batch_size=128
 
 
 
diff --git a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_8p.sh b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_8p.sh
index 2a38738b3a..6c8c9c13f8 100644
--- a/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_8p.sh
+++ b/PyTorch/contrib/cv/classification/ConvNeXt/test/train_performance_8p.sh
@@ -98,7 +98,7 @@ e2e_time=$(( $end_time - $start_time ))
 #最后一个迭代FPS值
 step_time=`grep -a 'time:'  ${cur_path}/test/output/$ASCEND_DEVICE_ID/train_perf_8p.log|awk -F "time: " '{print $NF}'|awk 'END {print}'| awk -F "  data:" '{print $1}'`
 
-FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${step_time}'}'`
+FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}' * 8/'${step_time}'}'`
 #最后一个迭代loss值
 loss=`grep -a 'loss:'  ${cur_path}/test/output/$ASCEND_DEVICE_ID/train_perf_8p.log | awk -F "loss:" '{print $NF}'| awk 'END {print}' | awk -F "(" '{print $1}'`
 
-- 
Gitee