diff --git a/ACL_TensorFlow/contrib/cv/OSNet_ID1379_for_ACL/.keep b/ACL_TensorFlow/contrib/cv/OSNet_ID1379_for_ACL/.keep new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/.gitignore b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/.gitignore new file mode 100644 index 0000000000000000000000000000000000000000..a150ee4c9d6ccfa82fa878cf086d203fe2264f85 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/.gitignore @@ -0,0 +1,133 @@ +# Byte-compiled / optimized / DLL files +__pycache__/ +*.py[cod] +*$py.class + +# C extensions +*.so + +# Distribution / packaging +.Python +build/ +develop-eggs/ +dist/ +downloads/ +eggs/ +.eggs/ +lib/ +lib64/ +parts/ +sdist/ +var/ +wheels/ +pip-wheel-metadata/ +share/python-wheels/ +*.egg-info/ +.installed.cfg +*.egg +MANIFEST + +# PyInstaller +# Usually these files are written by a python script from a template +# before PyInstaller builds the exe, so as to inject date/other infos into it. +*.manifest +*.spec + +# Installer logs +pip-log.txt +pip-delete-this-directory.txt + +# Unit test / coverage reports +htmlcov/ +.tox/ +.nox/ +.coverage +.coverage.* +.cache +nosetests.xml +coverage.xml +*.cover +*.py,cover +.hypothesis/ +.pytest_cache/ + +# Translations +*.mo +*.pot + +# Django stuff: +*.log +local_settings.py +db.sqlite3 +db.sqlite3-journal + +# Flask stuff: +instance/ +.webassets-cache + +# Scrapy stuff: +.scrapy + +# Sphinx documentation +docs/_build/ + +# PyBuilder +target/ + +# Jupyter Notebook +.ipynb_checkpoints + +# IPython +profile_default/ +ipython_config.py + +# pyenv +.python-version + +# pipenv +# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control. +# However, in case of collaboration, if having platform-specific dependencies or dependencies +# having no cross-platform support, pipenv may install dependencies that don't work, or not +# install all needed dependencies. +#Pipfile.lock + +# PEP 582; used by e.g. github.com/David-OConnor/pyflow +__pypackages__/ + +# Celery stuff +celerybeat-schedule +celerybeat.pid + +# SageMath parsed files +*.sage.py + +# Environments +.env +.venv +env/ +venv/ +ENV/ +env.bak/ +venv.bak/ + +# Spyder project settings +.spyderproject +.spyproject + +# Rope project settings +.ropeproject + +# mkdocs documentation +/site + +# mypy +.mypy_cache/ +.dmypy.json +dmypy.json + +# Pyre type checker +.pyre/ + +.vscode/ + +model_checkpoint \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/.keep b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/.keep new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/README.md b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/README.md new file mode 100644 index 0000000000000000000000000000000000000000..711005ecdf9c67107c0525ba9311a80420143fad --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/README.md @@ -0,0 +1,221 @@ +- [基本信息](#基本信息.md) +- [概述](#概述.md) +- [训练环境准备](#训练环境准备.md) +- [快速上手](#快速上手.md) +- [训练结果](#训练结果.md) +- [高级参考](#高级参考.md) +

基本信息

+ +**发布者（Publisher）：Huawei** + +**应用领域（Application Domain）：Natural Language Processing** + +**版本（Version）：** + +**修改时间（Modified） ：2022.3.16** + +**大小（Size）：** + +**框架（Framework）：TensorFlow 1.15.0** + +**模型格式（Model Format）：h5** + +**精度（Precision）：Mixed** + +**处理器（Processor）：昇腾910** + +**应用级别（Categories）：Research** + +**描述（Description）：基于TensorFlow框架的OSNet网络训练代码** + +

概述

+ + CVPR 2019的文章Omni-Scale Feature Learning for Person Re-Identification，作者自己根据Reid任务重新设计了一种新的网络结构，比较轻便，对比shuffleNet，mobileNet，SqueezeNet等具有更好的性能。 + +- 参考论文： + + https://link.zhihu.com/?target=https%3A//arxiv.org/pdf/1905.00953.pdf + +- 参考实现： + + https://github.com/Purnay04/a2m_osnet_keras + +- 适配昇腾 AI 处理器的实现： + + + https://gitee.com/dw8023/modelzoo/tree/master/contrib/TensorFlow/Research/cv/OSNet_ID1379_for_TensorFlow + + + +- 通过Git获取对应commit\_id的代码方法如下： + + ``` + git clone {repository_url} # 克隆仓库的代码 + cd {repository_name} # 切换到模型的代码仓目录 + git checkout {branch} # 切换到对应分支 + git reset --hard ｛commit_id｝ # 代码设置到对应的commit_id + cd ｛code_path｝ # 切换到模型代码所在路径，若仓库下只有该模型，则无需切换 + ``` + +## 默认配置 + +- 训练数据集预处理： + + 数据增强： + (1)将图像随机裁剪出256128的尺寸然后按1.25的因子重新缩放； + (2)随机水平翻转； + (3)随机擦除 + - 图像的输入尺寸为128×64×3 + - 图像输入格式：.jpg + +- 测试数据集预处理： + + - 图像的输入尺寸为128×64×3 + - 图像输入格式：.jpg + +- 训练超参 + + - Batch size： 128 + - Train epoch: 100 + - initial learning rate：0.065 + + +## 支持特性 + +| 特性列表 | 是否支持 | +|-------|------| +| 分布式训练 | 否 | +| 混合精度 | 是 | +| 并行数据 | 否 | + +## 混合精度训练 + +昇腾910 AI处理器提供自动混合精度功能，可以针对全网中float32数据类型的算子，按照内置的优化策略，自动将部分float32的算子降低精度到float16，从而在精度损失很小的情况下提升系统性能并减少内存使用。 + +## 开启混合精度 + +脚本已默认开启混合精度，设置precision_mode参数的脚本参考如下。 + + ``` + custom_op = session_config.graph_options.rewrite_options.custom_optimizers.add() + custom_op.name = 'NpuOptimizer' + custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes(str(args.precision_mode)) + ``` + +

训练环境准备

+ +1. 硬件环境准备请参见各硬件产品文档"[驱动和固件安装升级指南]( https://support.huawei.com/enterprise/zh/category/ai-computing-platform-pid-1557196528909)"。需要在硬件设备上安装与CANN版本配套的固件与驱动。 +2. 宿主机上需要安装Docker并登录[Ascend Hub中心](https://ascendhub.huawei.com/#/detail?name=ascend-tensorflow-arm)获取镜像。 + + 当前模型支持的镜像列表如[表1](#zh-cn_topic_0000001074498056_table1519011227314)所示。 + + **表 1** 镜像列表 + + + + + + + + + + + + +

镜像名称

+

镜像版本

+

配套CANN版本

+
+

20.2.0

+

20.2

+
+ + +

快速上手

+ +- 数据集准备 +1. 模型训练使用Market-1501数据集，数据集obs链接如下： + + obs://osnet-id1379/dataset/ +## 模型训练 + +- 源码obs链接如下： + + obs://osnet-id1379/npu/ +- 启动训练之前，首先要配置程序运行相关环境变量。 + + 环境变量配置信息参见： + + [Ascend 910训练平台环境变量设置](https://gitee.com/ascend/modelzoo/wikis/Ascend%20910%E8%AE%AD%E7%BB%83%E5%B9%B3%E5%8F%B0%E7%8E%AF%E5%A2%83%E5%8F%98%E9%87%8F%E8%AE%BE%E7%BD%AE?sort_id=3148819) + +- 单卡训练 + + 1. 配置训练参数。 + + 首先在脚本train_full_1p.sh中，配置batch_size、num_epochs、train_image_dir等参数，请用户根据实际路径配置train_image_dir，或者在启动训练的命令行中以参数形式下发。 + + ``` + batch_size=128 + num_epochs=100 + initial_lr=0.065 + train_image_dir="../Market-1501-v15.09.15/bounding_box_train" + ``` + + 2. 启动训练。 + + 启动单卡训练 （脚本为LeNet_for_TensorFlow/test/train_full_1p.sh） + + ``` + bash train_full_1p.sh + ``` + +

训练结果

+ +- 精度结果比对 + +|精度指标项|论文发布|GPU实测|NPU实测| +|---|---|---|---| +|mAP|xxx|56.0%|56.5%| + +- 性能结果比对 (暂无) + +|性能指标项|论文发布|GPU实测|NPU实测| +|---|---|---|---| +|FPS|XXX|YYY|ZZZ| + + +

高级参考

+ +## 脚本和示例代码 + +``` +├── train_cnn.py //网络训练与测试代码 +├── README.md //代码说明文档 +├── eval.py //精度测试代码 +├── data.py //数据集处理代码 +├── osnet.py //模型网络结构代码 +├── requirements.txt //训练python依赖列表 +├── train_full_1p.sh //训练启动脚本 +├── utils //训练与测试流程工具代码 +│ ├──general.py +│ ├──ranking.py +│ 、、、 +``` + +## 脚本参数 + +``` +--train_image_dir 数据集路径，默认：'/home/dingwei/osnet/dataset/Market-1501-v15.09.15/bounding_box_train/' +--batch_size 每个NPU的batch size，默认：128 +--initial_lr 初始学习率，默认：0.065 +--num_epoch 训练epcoh数量，默认：100 + +``` + +## 训练过程 + +1. 通过“模型训练”中的训练指令启动单卡卡训练。 + +2. 参考脚本的模型存储路径为./osnet.h5。 + + diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/__init__.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..8ca10a8fe0a73bb167cafcfcb0b1761eb0d92ee6 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/__init__.py @@ -0,0 +1,27 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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. diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/data.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/data.py new file mode 100644 index 0000000000000000000000000000000000000000..bb991cb8b6b15ae579d11e4fc1619d629c6cd687 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/data.py @@ -0,0 +1,71 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 numpy as np +import math +from tensorflow.keras.utils import Sequence, to_categorical +from tensorflow.keras.preprocessing.image import load_img +from sklearn.utils import shuffle as shuffle_tuple +from utils.general import seq + + +# Here, `x_set` is list of path to the images +# and `y_set` are the associated classes. + +class DataGenerator(Sequence): + + def __init__(self, x_set, y_set, batch_size, num_classes, shuffle=False, augment=False): + self.x, self.y = x_set, y_set + self.total_num_image = len(x_set) + self.batch_size = batch_size + self.num_classes = num_classes + self.shuffle = shuffle + self.augment = augment + + def __len__(self): + return math.ceil(len(self.x) / self.batch_size) + + def __getitem__(self, idx): + + batch_x = self.x[idx * self.batch_size: (idx + 1) * self.batch_size] + batch_y = self.y[idx * self.batch_size: (idx + 1) * self.batch_size] + + if self.shuffle: + batch_x, batch_y = shuffle_tuple(batch_x, batch_y) + + if self.augment: + batch_x = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in batch_x]).astype(np.uint8) + batch_x = seq.augment_images(batch_x) + batch_x = batch_x / 255. + else: + batch_x = np.array([np.asarray(load_img(file_name)) / 255. for file_name in batch_x]) + + batch_x = (batch_x - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + batch_y = to_categorical(np.array(batch_y), num_classes=self.num_classes) + + return batch_x, batch_y \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/eval.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/eval.py new file mode 100644 index 0000000000000000000000000000000000000000..3b578e80f96e136ec6f0ab20722786644238c471 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/eval.py @@ -0,0 +1,184 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 tensorflow.keras.models import load_model, Model +from tensorflow.keras.optimizers import SGD +from utils.triplet import generator_batch_test +from sklearn.preprocessing import normalize +import os +import argparse +import numpy as np +from osnet import OSNet +from utils.ranking import cmc, mean_ap +from utils.rerank import re_ranking + + +def pairwise_distance(mat1, mat2): + m = mat1.shape[0] # query number + n = mat2.shape[0] # gallery number + x = np.repeat(np.sum(np.square(mat1), axis=1, keepdims=True), n, axis=1) # mxn + y = np.repeat(np.sum(np.square(mat2), axis=1, keepdims=True), m, axis=1) # nxm + y = np.transpose(y) # mxn + return x + y - 2 * np.dot(mat1, mat2.T) + + +def evaluate(query_features, query_labels, query_cams, gallery_features, gallery_labels, + gallery_cams): + #query_feature: array, NxD + #query_cam: array, 1xN + #query_label: array, 1xN + #gallery_feature: array, MxD + #gallery_cam：array, 1xM + #gallery_label array, 1xM + distmat = pairwise_distance(query_features, gallery_features) + + print('Applying person re-ranking ...') + distmat_qq = pairwise_distance(query_features, query_features) + distmat_gg = pairwise_distance(gallery_features, gallery_features) + distmat = re_ranking(distmat, distmat_qq, distmat_gg) + # Compute mean AP + mAP = mean_ap(distmat, query_labels, gallery_labels, query_cams, gallery_cams) + print('Mean AP: {:4.1%}'.format(mAP)) + + cmc_configs = { + 'market1501': dict(separate_camera_set=False, + single_gallery_shot=False, + first_match_break=True),} + cmc_scores = {name: cmc(distmat, query_labels, gallery_labels, query_cams, + gallery_cams, **params) + for name, params in cmc_configs.items()} + cmc_topk = (1, 5, 10) + print('CMC Scores:') + for k in cmc_topk: + print(' top-{:<4}{:12.1%}'.format(k, cmc_scores['market1501'][k-1])) + + +def pairwise_distance(mat1, mat2): + m = mat1.shape[0] # query number + n = mat2.shape[0] # gallery number + x = np.repeat(np.sum(np.square(mat1), axis=1, keepdims=True), n, axis=1) # mxn + y = np.repeat(np.sum(np.square(mat2), axis=1, keepdims=True), m, axis=1) # nxm + y = np.transpose(y) # mxn + return x + y - 2 * np.dot(mat1, mat2.T) + + +def evaluate(query_features, query_labels, query_cams, gallery_features, gallery_labels, + gallery_cams): + #query_feature: array, NxD + #query_cam: array, 1xN + #query_label: array, 1xN + #gallery_feature: array, MxD + #gallery_cam：array, 1xM + #gallery_label array, 1xM + distmat = pairwise_distance(query_features, gallery_features) + + print('Applying person re-ranking ...') + distmat_qq = pairwise_distance(query_features, query_features) + distmat_gg = pairwise_distance(gallery_features, gallery_features) + distmat = re_ranking(distmat, distmat_qq, distmat_gg) + # Compute mean AP + mAP = mean_ap(distmat, query_labels, gallery_labels, query_cams, gallery_cams) + print('Final Accuracy accuracy: {:4.1%}'.format(mAP)) + + cmc_configs = { + 'market1501': dict(separate_camera_set=False, + single_gallery_shot=False, + first_match_break=True),} + cmc_scores = {name: cmc(distmat, query_labels, gallery_labels, query_cams, + gallery_cams, **params) + for name, params in cmc_configs.items()} + cmc_topk = (1, 5, 10) + print('CMC Scores:') + for k in cmc_topk: + print('top-{:<4}{:12.1%}'.format(k, cmc_scores['market1501'][k-1])) + + +def get_data_information(data_root): + img_path_list = [] + img_name_list = [] + img_cams_list = [] + image_names = os.listdir(data_root) #the best way is to use sorted list,i.e., sorted() + image_names = sorted(image_names)[:-1] + for item in image_names: + if item[-4:] == '.jpg': + img_path_list.append(os.path.join(data_root, item)) + img_name_list.append(item.split('_')[0]) + img_cams_list.append(item.split('c')[1][0]) + return img_path_list, np.array(img_name_list), np.array(img_cams_list) + + +def main(): + args = parser.parse_args() + # build model to extract features + model = OSNet(751).model + print("done") + print(model.summary()) + model.load_weights("osnet.h5") + model.summary() + dense_feature = model.get_layer('features').output + model_extract_features = Model(inputs=model.input, outputs=dense_feature) + model_extract_features.compile(loss=['categorical_crossentropy'], optimizer=SGD(lr=0.1), metrics=['accuracy']) + + # image_path, image_names, image_cams + query_dir = os.path.join(args.data_path, 'query') + gallery_dir = os.path.join(args.data_path, 'bounding_box_test') + query_img_list, query_name_list, query_cams_list = \ + get_data_information(query_dir) + gallery_img_list, gallery_name_list, gallery_cams_list = \ + get_data_information(gallery_dir) + + # obtain features + query_generator = generator_batch_test(query_img_list, args.img_width, + args.img_height, args.batch_size, shuffle=False) + query_features = model_extract_features.predict(query_generator, verbose=1, + steps=len(query_img_list)//args.batch_size if len(query_img_list)%args.batch_size==0 \ + else len(query_img_list)//args.batch_size+1) + query_features = normalize(query_features, norm='l2') + assert len(query_img_list) == query_features.shape[0], "something wrong with query samples" + + gallery_generator = generator_batch_test(gallery_img_list, args.img_width, args.img_height, + args.batch_size, shuffle=False) + gallery_features = model_extract_features.predict(gallery_generator,verbose=1, + steps=len(gallery_img_list)//args.batch_size if len(gallery_img_list)%args.batch_size==0 \ + else len(gallery_img_list)//args.batch_size+1) + gallery_features = normalize(gallery_features, norm='l2') + assert len(gallery_img_list) == gallery_features.shape[0], "something wrong with gallery samples" + #evaluate + evaluate(query_features, query_name_list, query_cams_list, + gallery_features, gallery_name_list, gallery_cams_list) + + +if __name__ == '__main__': + parser = argparse.ArgumentParser(description="Test") + parser.add_argument('--data_path', help='path to train_image', type=str, default='/home/dingwei/osnet/dataset/Market-1501-v15.09.15') + parser.add_argument('--img_width', type=int, default='64') + parser.add_argument('--img_height', type=int, default='128') + parser.add_argument('--learning_rate', type=float, default='0.01') + parser.add_argument('--batch_size', type=int, default='128') + parser.add_argument('--USE_Label_Smoothing', type=bool, default=True) + main() \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/osnet.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/osnet.py new file mode 100644 index 0000000000000000000000000000000000000000..8dcb9121ddb3ee403ffd4f97a2459d434b3ab79f --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/osnet.py @@ -0,0 +1,323 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 tensorflow as tf +import tensorflow.keras.backend as K +from tensorflow.keras import backend +from tensorflow.keras import layers +from tensorflow.keras import models, Sequential +from tensorflow.keras import utils as keras_utils +from tensorflow.keras.layers import Input +import numpy as np +from tensorflow_core.python.keras.layers import Activation +from tensorflow.keras.models import Model + + +def get_aggregation_gate(in_filters, reduction=16): + gate = Sequential([ + layers.GlobalAveragePooling2D(), + layers.Dense(in_filters // reduction, use_bias=False), + layers.BatchNormalization(), + layers.Activation('relu'), + layers.Dense(in_filters), + layers.Activation('sigmoid'), + layers.Reshape((1, 1, -1)) # reshape as (H, W, C) + ]) + return gate + + +def construct_fc_layer(feature_dims, input_dim, dropout_p = None, name="features"): + if feature_dims is None or feature_dims < 0: + feature_dims = input_dim + return None + + if isinstance(feature_dims, int): + feature_dims = [feature_dims] + + new_layers = [] + for dim in feature_dims: + new_layers.append(layers.Dense(dim)) + new_layers.append(layers.BatchNormalization()) + new_layers.append(layers.Activation('relu')) + if dropout_p is not None: + new_layers.append(layers.Dropout(dropout_p)) + input_dim = dim + feature_dims = feature_dims[-1] + + return Sequential(new_layers, name=name) + +class conv2d_bn(tf.keras.Model): + def __init__(self, + filters, + kernel_size = (3, 3), + padding = 'same', + strides = (1, 1), + activation = 'relu', + **kwargs): + super(conv2d_bn, self).__init__(**kwargs) + self.filters = filters + self.kernel_size = kernel_size + self.padding = padding + self.strides = strides + self.activation = activation + + self.bn_axis = -1 + self.conv = layers.Conv2D(self.filters, self.kernel_size, strides=self.strides, padding = self.padding, use_bias=False) + self.bn_norm = layers.BatchNormalization(axis=self.bn_axis) + + def call(self, x): + x = self.conv(x) + x = self.bn_norm(x) + if self.activation is not None: + x = layers.Activation(self.activation)(x) + return x + +class Light_conv3x3_bn(tf.keras.Model): + def __init__(self, filters): + super(Light_conv3x3_bn, self).__init__() + self.bn_axis = 3 + self.filters = filters + + self.conv = layers.Conv2D(self.filters, kernel_size=1, strides=1, padding='same', use_bias=False) + self.depth_conv = layers.DepthwiseConv2D(kernel_size=3, strides=1, padding='same', use_bias=False) + self.bn_norm = layers.BatchNormalization(axis=self.bn_axis) + self.activation = layers.Activation('relu') + + def call(self, x): + x = self.conv(x) + x = self.depth_conv(x) + x = self.bn_norm(x) + x = self.activation(x) + return x + +class os_bottleneck(tf.keras.Model): + def __init__(self,out_filters, bottleneck_reduction=4, **kwargs): + super(os_bottleneck, self).__init__(**kwargs) + self.out_filters = out_filters + self.mid_filters = self.out_filters // bottleneck_reduction + + self.conv1 = conv2d_bn(self.mid_filters, kernel_size=(1, 1)) + + self.conv2a = Light_conv3x3_bn(self.mid_filters) + self.conv2b = Sequential([ + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters) + ]) + self.conv2c = Sequential([ + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters) + ]) + self.conv2d = Sequential([ + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters), + Light_conv3x3_bn(self.mid_filters) + ]) + + self.conv3 = conv2d_bn(self.out_filters, kernel_size=(1, 1), activation=None) + self.ident_conv = conv2d_bn(self.out_filters, kernel_size=(1, 1), activation=None) + self.gate = get_aggregation_gate(self.mid_filters) + + self.Adder = layers.Add() + self.Multiply = layers.Multiply() + self.activation = layers.Activation('relu') + + def call(self, x): + in_filters = x.shape[-1] + identity = x + + x1 = self.conv1(x) + + #CONV2a + branch1 = self.conv2a(x1) + + #CONV2b + branch2 = self.conv2b(x1) + + #CONV2c + branch3 = self.conv2c(x1) + + #CONV2d + branch4 = self.conv2d(x1) + + x2 = self.Adder([ + self.Multiply([branch1, self.gate(branch1)]), + self.Multiply([branch2, self.gate(branch2)]), + self.Multiply([branch3, self.gate(branch3)]), + self.Multiply([branch4, self.gate(branch4)]) + ]) + + x3 = self.conv3(x2) + + if in_filters != self.out_filters: + #print("before",identity.shape, in_filters ,self.out_filters) + identity = self.ident_conv(identity) + #print("after",identity.shape) + + out = self.Adder([identity, x3]) # residual connection, out = x3 + identity in Pytorch + out = self.activation(out) + return out + + +class OSNet(tf.keras.Model): + def __init__(self, + classes, + include_top = False, + input_tensor = None, + pooling = None, + feature_dim = 512, + loss_type = {'xent'}, + **kwargs): + + super(OSNet, self).__init__() + + self.include_top = include_top + self.input_tensor = input_tensor + self.pooling = pooling + self.feature_dim = feature_dim + self.loss_type = loss_type + + #Network component + + self.conv1 = conv2d_bn(64, (7, 7), strides=(2, 2), name="conv1") # conv1: 128x64x64 + self.maxpool = layers.MaxPooling2D(pool_size=(3, 3), strides=(2, 2), padding='same', name="1p") # 1p: 64x32x64 + + self.a2 = os_bottleneck(256, name="2a") # 2a: 64x32x256 + self.b2 = os_bottleneck(256, name="2b") # 2b: 64x32x256 + self.t2 = conv2d_bn(256, (1, 1), name="2t") # 2t: 64x32x256 + self.p2 = layers.AveragePooling2D(pool_size=(2, 2), strides=(2, 2), name="2p") # 2p: 32x16x256 + + self.a3 = os_bottleneck(384, name="3a") # 3a: 16x8x384 + self.b3 = os_bottleneck(384, name="3b") # 3b: 16x8x384 + self.t3 = conv2d_bn(384, (1, 1), name="3t") # 3t: 16x8x384 + self.p3 = layers.AveragePooling2D(pool_size = (2, 2), strides = (2, 2), name="3p") # 3p: 8x4x384 + + self.a4 = os_bottleneck(512, name="4a") # 4a: 8x4x512 + self.b4 = os_bottleneck(512, name="4b") # 4b: 8x4x512 + self.t4 = conv2d_bn(512, (1, 1), name="4t") # 4t: 8x4x512 + + self.global_avgpool = layers.GlobalAveragePooling2D(name="Avg_pooling") + self.aft_dropout = layers.Dropout(0.2) + self.fc = construct_fc_layer(self.feature_dim, 512) + + # self.classifier = layers.Dense(self.classes, name="classifier") + self.activation = layers.Activation("softmax", name="softmax") + self.model = self.build(Input(shape=(128, 64, 3)), classes) + + def build(self, input_tensor, classes, training=True, return_featuremaps = False): + """ + self.img_input = layers.Input(tensor=self.input_tensor, shape=self.input_shape) + if self.input_tensor is None: + if self.input_shape is None: + raise ValueError('neither input_tensor nor input_shape is given') + else: + if not backend.is_keras_tensor(self.input_tensor): + self.img_input = layers.Input(tensor=self.input_tensor, shape=self.input_shape) + else: + self.img_input = self.input_tensor + """ + + + x = self.conv1(input_tensor) + x = self.maxpool(x) + + x = self.a2(x) + x = self.b2(x) + x = self.t2(x) + x = self.p2(x) + + x = self.a3(x) + x = self.b3(x) + x = self.t3(x) + x = self.p3(x) + + x = self.a4(x) + x = self.b4(x) + x = self.t4(x) + + if return_featuremaps: + return x + + v = self.global_avgpool(x) + v = self.aft_dropout(v) + + if self.fc is not None: + v = self.fc(v) + # return v + + if not training: + return v + y = layers.Dense(classes, name="classifier")(v) + if self.loss_type == {'xent'}: + w = self.activation(y) + # return layers.Model(Input(shape=(128, 64, 3)), output=w, name='osnet') + return Model(input_tensor, w) + elif self.loss_type == 'triplet': + return y, v + + else: + raise KeyError("Unsupported Loss: {}".format(self.loss)) + + +if __name__ == "__main__": + # from keras.utils.vis_utils import plot_model + # model = OSNet(751) + # plot_model(model, to_file="model.png", show_shapes=True); + + + raw_input = (1, 256, 128, 3) + model = OSNet(751).model + # output = model(Input(shape=(256, 128, 3)), training=True) + print("done") + + loss = tf.keras.losses.CategoricalCrossentropy(label_smoothing=0.1) + optimizer = tf.keras.optimizers.SGD(lr=0.065, momentum=0.9) + model.compile( + optimizer = optimizer, + loss = loss, + metrics = ['accuracy']) + # print(output) + # model.build(input_shape=raw_input) + print(model.summary()) + # https://stackoverflow.com/questions/62242330/error-when-subclassing-the-model-class-you-should-implement-a-call-method + + names = [layer.name for layer in model.layers] + print(names, len(names)) + + # tf.keras.utils.plot_model( + # model, # here is the trick (for now) + # to_file='model.png', dpi=96, # saving + # show_shapes=True, show_layer_names=True, # show shapes and layer name + # expand_nested=True # will show nested block + # ) + + # trainable_count = np.sum([K.count_params(w) for w in model.trainable_weights]) + # print(trainable_count) + + diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/requirements.txt b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/requirements.txt new file mode 100644 index 0000000000000000000000000000000000000000..00379840ef885bfb36af0942893de75f18de1757 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/requirements.txt @@ -0,0 +1,4 @@ +imgaug +tensorflow-gpu +scikit-learn +Pillow \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train.py new file mode 100644 index 0000000000000000000000000000000000000000..faf2bc5f178cf70d6f8b9e06d30dafd405f7deed --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train.py @@ -0,0 +1,129 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 os +from sklearn.preprocessing import LabelEncoder, normalize +from sklearn.model_selection import train_test_split +import tensorflow as tf +import tensorflow.keras.backend as K +from tensorflow.keras.callbacks import LearningRateScheduler +from utils.general import categorical_crossentropy_label_smoothing +from tensorflow.keras.optimizers import SGD +from data import DataGenerator +from osnet import OSNet +from tensorflow.core.protobuf.rewriter_config_pb2 import RewriterConfig +import argparse +from npu_bridge.npu_init import * +# Set training parameters +image_shape = (128, 64, 3) # h x w x c +use_label_smoothing = True + + +def parse_args(args): + """ + Parse the arguments. + """ + parser = argparse.ArgumentParser(description='Simple training script for training OSNet.') + parser.add_argument('--initial_lr', help='initial learning rate', type=float, default=0.065) + parser.add_argument('--batch_size', help='batch size', type=int, default=128) + parser.add_argument('--num_epoch', help='total epoch to train', type=int, default=100) + parser.add_argument('--train_image_dir', help='path to train_image', type=str, default='/home/dingwei/osnet/dataset/Market-1501-v15.09.15/bounding_box_train') + # parser.add_argument('--output_path', help='path to output', type=str, default='/home/dingwei/osnet/osnet_tf/output') + print(vars(parser.parse_args(args))) + return parser.parse_args(args) + + +def main(args=None): + # Npu setting + sess_config = tf.ConfigProto() + custom_op = sess_config.graph_options.rewrite_options.custom_optimizers.add() + custom_op.name = "NpuOptimizer" + custom_op.parameter_map["precision_mode"].s = tf.compat.as_bytes("allow_mix_precision") + custom_op.parameter_map["dynamic_input"].b = True + custom_op.parameter_map["dynamic_graph_execute_mode"].s = tf.compat.as_bytes("lazy_recompile") + custom_op.parameter_map["use_off_line"].b = True # 必须显式开启，在昇腾AI处理器执行训练 + sess_config.graph_options.rewrite_options.remapping = RewriterConfig.OFF # 必须显式关闭remap + sess_config.graph_options.rewrite_options.memory_optimization = RewriterConfig.OFF + sess = tf.Session(config=sess_config) + K.set_session(sess) + + args = parse_args(args) + # Preprocess data + train_image_filenames = sorted([filename for filename in os.listdir(args.train_image_dir) if filename.endswith(".jpg")]) + train_image_paths = [os.path.join(args.train_image_dir, name) for name in train_image_filenames] + train_person_ids = [name[:4] for name in train_image_filenames] + label_encoder = LabelEncoder() + label_encoder.fit(train_person_ids) + train_person_ids_encoded = label_encoder.transform(train_person_ids) + num_person_ids = len(set(train_person_ids_encoded)) + + train_img_paths, val_img_paths, train_person_ids, val_person_ids = train_test_split( + train_image_paths, train_person_ids_encoded, test_size=0.1, random_state=2021, + stratify=train_person_ids_encoded) + print( + f"# train images: {len(train_img_paths)}, # val images: {len(val_img_paths)}, # image labels: {num_person_ids}") + + # Contruct and show model + baseline_model = OSNet(751).model + print(baseline_model.summary()) + + loss = categorical_crossentropy_label_smoothing if use_label_smoothing else "categorical_crossentropy" + # loss = tf.keras.losses.SparseCategoricalCrossentropy() + optimizer = SGD(lr=args.initial_lr, momentum=0.9) + baseline_model.compile(optimizer=optimizer, loss=loss, metrics=["accuracy"]) + + # Set lr decay + def scheduler(epoch): + # 每隔100个epoch，学习率减小为原来的9/10 + if epoch % 30 == 0 and epoch != 0: + lr = K.get_value(baseline_model.optimizer.lr) + K.set_value(baseline_model.optimizer.lr, lr * 0.9) + print("lr changed to {}".format(lr * 0.9)) + return K.get_value(baseline_model.optimizer.lr) + reduce_lr = LearningRateScheduler(scheduler) + + train_generator = DataGenerator(train_img_paths, train_person_ids, batch_size=args.batch_size, + num_classes=num_person_ids, shuffle=True, augment=True) + val_generator = DataGenerator(val_img_paths, val_person_ids, batch_size=args.batch_size, num_classes=num_person_ids) + + # Train model + baseline_model.fit( + train_generator, + epochs=args.num_epoch, + # validation_data=val_generator, + callbacks=[reduce_lr], + shuffle=True, + ) + + # Save model + baseline_model.save_weights("osnet.h5") + print("Training completed and model saved.") + sess.close() + +if __name__ == "__main__": + main() diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train_full_1p.sh b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train_full_1p.sh new file mode 100644 index 0000000000000000000000000000000000000000..8297653997da9bd0d1231b6c4393114d778aa998 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/train_full_1p.sh @@ -0,0 +1,187 @@ +#!/bin/bash + +########################################################## +#########第3行 至 100行，请一定不要、不要、不要修改########## +#########第3行 至 100行，请一定不要、不要、不要修改########## +#########第3行 至 100行，请一定不要、不要、不要修改########## +########################################################## +# shell脚本所在路径 +cur_path=`echo $(cd $(dirname $0);pwd)` + +# 判断当前shell是否是performance +perf_flag=`echo $0 | grep performance | wc -l` + +# 当前执行网络的名称 +Network=`echo $(cd $(dirname $0);pwd) | awk -F"/" '{print $(NF-1)}'` + +export RANK_SIZE=1 +export RANK_ID=0 +export JOB_ID=10087 + +# 路径参数初始化 +data_path='/home/dingwei/osnet/dataset/Market-1501-v15.09.15' +output_path='/home/dingwei/osnet/osnet_tf/output' + +# 帮助信息，不需要修改 +if [[ $1 == --help || $1 == -h ]];then + echo"usage:./train_performance_1P.sh " + echo " " + echo "parameter explain: + --data_path # dataset of training + --output_path # output of training + --train_steps # max_step for training + --train_epochs # max_epoch for training + --batch_size # batch size + -h/--help show help message + " + exit 1 +fi + +# 参数校验，不需要修改 +for para in $* +do + if [[ $para == --data_path* ]];then + data_path=`echo ${para#*=}` + elif [[ $para == --output_path* ]];then + output_path=`echo ${para#*=}` + elif [[ $para == --train_steps* ]];then + train_steps=`echo ${para#*=}` + elif [[ $para == --train_epochs* ]];then + train_epochs=`echo ${para#*=}` + elif [[ $para == --batch_size* ]];then + batch_size=`echo ${para#*=}` + fi +done + +# 校验是否传入data_path,不需要修改 +if [[ $data_path == "" ]];then + echo "[Error] para \"data_path\" must be config" + exit 1 +fi + +# 校验是否传入output_path,不需要修改 +if [[ $output_path == "" ]];then + output_path="./test/output/${ASCEND_DEVICE_ID}" +fi + +# 设置打屏日志文件名，请保留，文件名为${print_log} +print_log="./test/output/${ASCEND_DEVICE_ID}/train_${ASCEND_DEVICE_ID}.log" +modelarts_flag=${MODELARTS_MODEL_PATH} +if [ x"${modelarts_flag}" != x ]; +then + echo "running without etp..." + print_log_name=`ls /home/ma-user/modelarts/log/ | grep proc-rank` + print_log="/home/ma-user/modelarts/log/${print_log_name}" +fi +echo "### get your log here : ${print_log}" + +CaseName="" +function get_casename() +{ + if [ x"${perf_flag}" = x1 ]; + then + CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'perf' + else + CaseName=${Network}_bs${batch_size}_${RANK_SIZE}'p'_'acc' + fi +} + +# 跳转到code目录 +cd ${cur_path}/../ +rm -rf ./test/output/${ASCEND_DEVICE_ID} +mkdir -p ./test/output/${ASCEND_DEVICE_ID} + +# 训练开始时间记录，不需要修改 +start_time=$(date +%s) +########################################################## +#########第3行 至 100行，请一定不要、不要、不要修改########## +#########第3行 至 100行，请一定不要、不要、不要修改########## +#########第3行 至 100行，请一定不要、不要、不要修改########## +########################################################## + +#========================================================= +#========================================================= +#========训练执行命令，需要根据您的网络进行修改============== +#========================================================= +#========================================================= +# 基础参数，需要模型审视修改 +# 您的训练数据集在${data_path}路径下，请直接使用这个变量获取 +# 您的训练输出目录在${output_path}路径下，请直接使用这个变量获取 +# 您的其他基础参数，可以自定义增加，但是batch_size请保留，并且设置正确的值 +batch_size=128 + + +if [ x"${modelarts_flag}" != x ]; +then + python3.7 ./train.py --train_image_dir=${data_path}/bounding_box_train --num_epoch=3 + python3.7 ./eval.py --data_path=${data_path} +else + python3.7 ./train.py --train_image_dir=${data_path}/bounding_box_train --num_epoch=1 1>>${print_log} 2>&1 + python3.7 ./eval.py --data_path=${data_path} 1>>${print_log} 2>&1 +fi + +# 性能相关数据计算 +StepTime=`grep "ms/step :" ${print_log} | tail -n 10 | awk '{print $NF}' | awk '{sum+=$1} END {print sum/NR}'` +FPS=`awk 'BEGIN{printf "%.2f\n", '${batch_size}'/'${StepTime}*1000'}'` + +# 精度相关数据计算 +train_accuracy=`grep "Final Accuracy accuracy" ${print_log} | awk '{print $NF}'` +# 提取所有loss打印信息 +grep "loss :" ${print_log} | awk -F ":" '{print $4}' | awk -F "-" '{print $1}' > ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt + + +########################################################### +#########后面的所有内容请不要修改########################### +#########后面的所有内容请不要修改########################### +#########后面的所有内容请不要修改########################### +########################################################### + +# 判断本次执行是否正确使用Ascend NPU +use_npu_flag=`grep "The model has been compiled on the Ascend AI processor" ${print_log} | wc -l` +if [ x"${use_npu_flag}" == x0 ]; +then + echo "------------------ ERROR NOTICE START ------------------" + echo "ERROR, your task haven't used Ascend NPU, please check your npu Migration." + echo "------------------ ERROR NOTICE END------------------" +else + echo "------------------ INFO NOTICE START------------------" + echo "INFO, your task have used Ascend NPU, please check your result." + echo "------------------ INFO NOTICE END------------------" +fi + +# 获取最终的casename，请保留，case文件名为${CaseName} +get_casename + +# 重命名loss文件 +if [ -f ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ]; +then + mv ./test/output/${ASCEND_DEVICE_ID}/my_output_loss.txt ./test/output/${ASCEND_DEVICE_ID}/${CaseName}_loss.txt +fi + +# 训练端到端耗时 +end_time=$(date +%s) +e2e_time=$(( $end_time - $start_time )) + +echo "------------------ Final result ------------------" +# 输出性能FPS/单step耗时/端到端耗时 +echo "Final Performance images/sec : $FPS" +echo "Final Performance ms/step : $StepTime" +echo "E2E Training Duration sec : $e2e_time" + +# 输出训练精度 +echo "Final Train Accuracy : ${train_accuracy}" + +# 最后一个迭代loss值，不需要修改 +ActualLoss=(`awk 'END {print $NF}' $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}_loss.txt`) + +#关键信息打印到${CaseName}.log中，不需要修改 +echo "Network = ${Network}" > $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "RankSize = ${RANK_SIZE}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "BatchSize = ${batch_size}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "DeviceType = `uname -m`" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "CaseName = ${CaseName}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "ActualFPS = ${FPS}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "TrainingTime = ${StepTime}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "ActualLoss = ${ActualLoss}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "E2ETrainingTime = ${e2e_time}" >> $cur_path/output/$ASCEND_DEVICE_ID/${CaseName}.log +echo "TrainAccuracy = ${train_accuracy}" >> $cur_path/test/output/$ASCEND_DEVICE_ID/${CaseName}.log diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/.keep b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/.keep new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/__init__.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/__init__.py new file mode 100644 index 0000000000000000000000000000000000000000..cdb22dca64c3c8c63e6a5c2972abbc36d50db629 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/__init__.py @@ -0,0 +1,27 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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. diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/general.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/general.py new file mode 100644 index 0000000000000000000000000000000000000000..070a2182db41fe6be8a0bc78f2249c78e093791a --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/general.py @@ -0,0 +1,59 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 tensorflow.keras.losses import categorical_crossentropy + +def categorical_crossentropy_label_smoothing(y_true, y_pred): + label_smoothing = 0.1 + return categorical_crossentropy(y_true, y_pred, label_smoothing=label_smoothing) + +from imgaug import augmenters as iaa +seq = iaa.Sequential([ + iaa.Fliplr(0.5), # horizontal flips + iaa.Crop(percent=(0, 0.1)), # random crops + # Small gaussian blur with random sigma between 0 and 0.5. + # But we only blur about 50% of all images. + iaa.Sometimes( + 0.5, + iaa.GaussianBlur(sigma=(0, 0.5)) + ), + # Random Erase + iaa.Sometimes( + 0.5, + iaa.Cutout(nb_iterations=1, size=[0.3, 0.4], squared=False) + ), + # Apply affine transformations to each image. + # Scale/zoom them, translate/move them, rotate them and shear them. + iaa.Sometimes( + 0.3, + iaa.Affine( + rotate=(-10, 10), + shear=(-8, 8) + ) + ), +], random_order=True, random_state=2021) \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/ranking.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/ranking.py new file mode 100644 index 0000000000000000000000000000000000000000..d3f9eff4dfe96b64e1bb147447fb60ecbc0b2f89 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/ranking.py @@ -0,0 +1,140 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 __future__ import absolute_import +from collections import defaultdict + +import numpy as np +from sklearn.metrics import average_precision_score + + + +def _unique_sample(ids_dict, num): + mask = np.zeros(num, dtype=np.bool) + for _, indices in ids_dict.items(): + i = np.random.choice(indices) + mask[i] = True + return mask + + +def cmc(distmat, query_ids=None, gallery_ids=None, + query_cams=None, gallery_cams=None, topk=100, + separate_camera_set=False, + single_gallery_shot=False, + first_match_break=False): + m, n = distmat.shape + # Fill up default values + if query_ids is None: + query_ids = np.arange(m) + if gallery_ids is None: + gallery_ids = np.arange(n) + if query_cams is None: + query_cams = np.zeros(m).astype(np.int32) + if gallery_cams is None: + gallery_cams = np.ones(n).astype(np.int32) + # Ensure numpy array + query_ids = np.asarray(query_ids) + gallery_ids = np.asarray(gallery_ids) + query_cams = np.asarray(query_cams) + gallery_cams = np.asarray(gallery_cams) + # Sort and find correct matches + indices = np.argsort(distmat, axis=1) + matches = (gallery_ids[indices] == query_ids[:, np.newaxis]) + # Compute CMC for each query + ret = np.zeros(topk) + num_valid_queries = 0 + for i in range(m): + # Filter out the same id and same camera + valid = ((gallery_ids[indices[i]] != query_ids[i]) | + (gallery_cams[indices[i]] != query_cams[i])) + if separate_camera_set: + # Filter out samples from same camera + valid &= (gallery_cams[indices[i]] != query_cams[i]) + if not np.any(matches[i, valid]): continue + if single_gallery_shot: + repeat = 10 + gids = gallery_ids[indices[i][valid]] + inds = np.where(valid)[0] + ids_dict = defaultdict(list) + for j, x in zip(inds, gids): + ids_dict[x].append(j) + else: + repeat = 1 + for _ in range(repeat): + if single_gallery_shot: + # Randomly choose one instance for each id + sampled = (valid & _unique_sample(ids_dict, len(valid))) + index = np.nonzero(matches[i, sampled])[0] + else: + index = np.nonzero(matches[i, valid])[0] + delta = 1. / (len(index) * repeat) + for j, k in enumerate(index): + if k - j >= topk: break + if first_match_break: + ret[k - j] += 1 + break + ret[k - j] += delta + num_valid_queries += 1 + if num_valid_queries == 0: + raise RuntimeError("No valid query") + return ret.cumsum() / num_valid_queries + + +def mean_ap(distmat, query_ids=None, gallery_ids=None, + query_cams=None, gallery_cams=None): + m, n = distmat.shape + # Fill up default values + if query_ids is None: + query_ids = np.arange(m) + if gallery_ids is None: + gallery_ids = np.arange(n) + if query_cams is None: + query_cams = np.zeros(m).astype(np.int32) + if gallery_cams is None: + gallery_cams = np.ones(n).astype(np.int32) + # Ensure numpy array + query_ids = np.asarray(query_ids) + gallery_ids = np.asarray(gallery_ids) + query_cams = np.asarray(query_cams) + gallery_cams = np.asarray(gallery_cams) + # Sort and find correct matches + indices = np.argsort(distmat, axis=1) + matches = (gallery_ids[indices] == query_ids[:, np.newaxis]) + # Compute AP for each query + aps = [] + for i in range(m): + # Filter out the same id and same camera + valid = ((gallery_ids[indices[i]] != query_ids[i]) | + (gallery_cams[indices[i]] != query_cams[i])) + y_true = matches[i, valid] + y_score = -distmat[i][indices[i]][valid] + if not np.any(y_true): continue + aps.append(average_precision_score(y_true, y_score)) + if len(aps) == 0: + raise RuntimeError("No valid query") + return np.mean(aps) \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/rerank.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/rerank.py new file mode 100644 index 0000000000000000000000000000000000000000..519ae9c796cca7685315d3764b39e13bc69470a4 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/rerank.py @@ -0,0 +1,125 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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. + +#!/usr/bin/env python2/python3 +# -*- coding: utf-8 -*- +""" +Source: https://github.com/zhunzhong07/person-re-ranking +Created on Mon Jun 26 14:46:56 2017 +@author: luohao +Modified by Houjing Huang, 2017-12-22. +- This version accepts distance matrix instead of raw features. +- The difference of `/` division between python 2 and 3 is handled. +- numpy.float16 is replaced by numpy.float32 for numerical precision. +CVPR2017 paper:Zhong Z, Zheng L, Cao D, et al. Re-ranking Person Re-identification with k-reciprocal Encoding[J]. 2017. +url:http://openaccess.thecvf.com/content_cvpr_2017/papers/Zhong_Re-Ranking_Person_Re-Identification_CVPR_2017_paper.pdf +Matlab version: https://github.com/zhunzhong07/person-re-ranking +API +q_g_dist: query-gallery distance matrix, numpy array, shape [num_query, num_gallery] +q_q_dist: query-query distance matrix, numpy array, shape [num_query, num_query] +g_g_dist: gallery-gallery distance matrix, numpy array, shape [num_gallery, num_gallery] +k1, k2, lambda_value: parameters, the original paper is (k1=20, k2=6, lambda_value=0.3) +Returns: + final_dist: re-ranked distance, numpy array, shape [num_query, num_gallery] +""" +from __future__ import absolute_import +from __future__ import print_function +from __future__ import division + +__all__ = ['re_ranking'] + +import numpy as np + + +def re_ranking(q_g_dist, q_q_dist, g_g_dist, k1=20, k2=6, lambda_value=0.3): + + # The following naming, e.g. gallery_num, is different from outer scope. + # Don't care about it. + + original_dist = np.concatenate( + [np.concatenate([q_q_dist, q_g_dist], axis=1), + np.concatenate([q_g_dist.T, g_g_dist], axis=1)], + axis=0) + original_dist = np.power(original_dist, 2).astype(np.float32) + original_dist = np.transpose(1. * original_dist/np.max(original_dist,axis = 0)) + V = np.zeros_like(original_dist).astype(np.float32) + initial_rank = np.argsort(original_dist).astype(np.int32) + + query_num = q_g_dist.shape[0] + gallery_num = q_g_dist.shape[0] + q_g_dist.shape[1] + all_num = gallery_num + + for i in range(all_num): + # k-reciprocal neighbors + forward_k_neigh_index = initial_rank[i,:k1+1] + backward_k_neigh_index = initial_rank[forward_k_neigh_index,:k1+1] + fi = np.where(backward_k_neigh_index==i)[0] + k_reciprocal_index = forward_k_neigh_index[fi] + k_reciprocal_expansion_index = k_reciprocal_index + for j in range(len(k_reciprocal_index)): + candidate = k_reciprocal_index[j] + candidate_forward_k_neigh_index = initial_rank[candidate,:int(np.around(k1/2.))+1] + candidate_backward_k_neigh_index = initial_rank[candidate_forward_k_neigh_index,:int(np.around(k1/2.))+1] + fi_candidate = np.where(candidate_backward_k_neigh_index == candidate)[0] + candidate_k_reciprocal_index = candidate_forward_k_neigh_index[fi_candidate] + if len(np.intersect1d(candidate_k_reciprocal_index,k_reciprocal_index))> 2./3*len(candidate_k_reciprocal_index): + k_reciprocal_expansion_index = np.append(k_reciprocal_expansion_index,candidate_k_reciprocal_index) + + k_reciprocal_expansion_index = np.unique(k_reciprocal_expansion_index) + weight = np.exp(-original_dist[i,k_reciprocal_expansion_index]) + V[i,k_reciprocal_expansion_index] = 1.*weight/np.sum(weight) + original_dist = original_dist[:query_num,] + if k2 != 1: + V_qe = np.zeros_like(V,dtype=np.float32) + for i in range(all_num): + V_qe[i,:] = np.mean(V[initial_rank[i,:k2],:],axis=0) + V = V_qe + del V_qe + del initial_rank + invIndex = [] + for i in range(gallery_num): + invIndex.append(np.where(V[:,i] != 0)[0]) + + jaccard_dist = np.zeros_like(original_dist,dtype = np.float32) + + + for i in range(query_num): + temp_min = np.zeros(shape=[1,gallery_num],dtype=np.float32) + indNonZero = np.where(V[i,:] != 0)[0] + indImages = [] + indImages = [invIndex[ind] for ind in indNonZero] + for j in range(len(indNonZero)): + temp_min[0,indImages[j]] = temp_min[0,indImages[j]]+ np.minimum(V[i,indNonZero[j]],V[indImages[j],indNonZero[j]]) + jaccard_dist[i] = 1-temp_min/(2.-temp_min) + + final_dist = jaccard_dist*(1-lambda_value) + original_dist*lambda_value + del original_dist + del V + del jaccard_dist + final_dist = final_dist[:query_num,query_num:] + return final_dist \ No newline at end of file diff --git a/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/triplet.py b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/triplet.py new file mode 100644 index 0000000000000000000000000000000000000000..5073703489fa4870d6eded9e4f02d3bcc3ae8c22 --- /dev/null +++ b/TensorFlow/contrib/cv/OSNet_ID1379_for_TensorFlow/utils/triplet.py @@ -0,0 +1,358 @@ +# Copyright 2017 The TensorFlow Authors. 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. +# ============================================================================ +# Copyright 2021 Huawei Technologies Co., Ltd +# +# 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 numpy as np +import math +import random +from sklearn.utils import shuffle as shuffle_tuple +from tensorflow.keras.utils import Sequence, to_categorical +from tensorflow.keras.preprocessing.image import load_img +from tensorflow.python.keras.applications.mobilenet_v2 import MobileNetV2 +from tensorflow.python.keras.applications.resnet_v2 import ResNet50V2 +from tensorflow.keras.layers import Dense, Activation, Lambda, BatchNormalization, Input, concatenate, Embedding +from tensorflow.keras.models import Model, model_from_json +from tensorflow.keras import backend as K +import cv2 +from .general import seq + +random.seed(2021) + + +def create_model(image_shape, num_person_ids, show_model_summary=False): + anchor_input = Input(image_shape, name="anchor_input") + positive_input = Input(image_shape, name="positive_input") + negative_input = Input(image_shape, name="negative_input") + + cnn_model = MobileNetV2(input_shape=image_shape, alpha=0.5, include_top=False, pooling="max") + cnn_model.trainable = False + + anchor_embedding = cnn_model(anchor_input) + positive_embedding = cnn_model(positive_input) + negative_embedding = cnn_model(negative_input) + + merged_vector = concatenate([anchor_embedding, positive_embedding, negative_embedding], axis=-1, name="triplet") + + dense_anchor = Dense(num_person_ids)(anchor_embedding) + softmax_anchor_output = Activation("softmax", name="softmax")(dense_anchor) + + triplet_model = Model([anchor_input, positive_input, negative_input], [merged_vector, softmax_anchor_output]) + + if show_model_summary: + triplet_model.summary() + + return triplet_model + + +def create_semi_hard_triplet_model(image_shape, num_person_ids, show_model_summary=False, resnet=True, last_stride_reduce=True, bn=True, center_loss=True, average_pooling=True): + if resnet: + print("Using model ResNet50V2\n") + cnn_model = ResNet50V2(input_shape=image_shape, include_top=False, pooling=("avg" if average_pooling else "max")) + if last_stride_reduce: + cnn_model.get_layer("conv4_block6_2_conv").strides = (1,1) + cnn_model.get_layer("max_pooling2d_2").strides = (1,1) + cnn_model = model_from_json(cnn_model.to_json()) + else: + print("Using model MobileNetV2\n") + cnn_model = MobileNetV2(input_shape=image_shape, alpha=0.5, include_top=False, pooling=("avg" if average_pooling else "max")) + if last_stride_reduce: + cnn_model.get_layer("block_13_pad").padding = (1,1) + cnn_model.get_layer("block_13_depthwise").strides = (1,1) + cnn_model = model_from_json(cnn_model.to_json()) + + global_pool = cnn_model.layers[-1].output + cnn_model.layers[-1]._name = "triplet" + + if bn: + features_bn = BatchNormalization(name="features_bn")(global_pool) + dense = Dense(num_person_ids)(features_bn) + else: + dense = Dense(num_person_ids)(global_pool) + softmax_output = Activation("softmax", name="softmax")(dense) + + if center_loss: + input_target = Input(shape=(1,)) + centers = Embedding(num_person_ids, global_pool.shape[-1], name="embedding_center")(input_target) + center_loss = Lambda(lambda x: 0.5 * K.sum(K.square((x[0] - x[1])), axis=1, keepdims=True), name="center")((global_pool, centers)) + triplet_model = Model([cnn_model.input, input_target], [global_pool, softmax_output, center_loss]) + else: + triplet_model = Model(cnn_model.input, [global_pool, softmax_output]) + + if show_model_summary: + triplet_model.summary() + + return triplet_model + +#### +# Self defined triplet loss +#### +def triplet_loss(y_true, y_pred, alpha=0.3): + y_pred = K.l2_normalize(y_pred, axis=1) + batch_num = y_pred.shape.as_list()[-1] // 3 + + anchor = y_pred[:, 0:batch_num] + positive = y_pred[:, batch_num:2*batch_num] + negative = y_pred[:, 2*batch_num:3*batch_num] + + pos_dist = K.sum(K.square(anchor - positive), axis=1) + neg_dist = K.sum(K.square(anchor - negative), axis=1) + + loss = K.maximum(pos_dist - neg_dist + alpha, 0.0) + + return loss + +#### +# Learning rate from paper +#### +def lr_decay_warmup(epoch, initial_rate): + epoch += 1 + if epoch < 11: + return 3.5e-4 * epoch / 10 + elif epoch < 41: + return 3.5e-4 + elif epoch < 71: + return 3.5e-5 + else: + return 3.5e-6 + + + +# Train data generator for cnn +class DataGenerator(Sequence): + + # Here, `x_set` is list of path to the images + # and `y_set` are the associated classes. + def __init__(self, x_set, y_set, batch_size, num_classes, shuffle=False, augment=False): + self.x, self.y = x_set, y_set + self.batch_size = batch_size + self.num_classes = num_classes + self.shuffle = shuffle + self.augment = augment + + def __len__(self): + return math.ceil(len(self.x) / self.batch_size) + + def __getitem__(self, idx): + + batch_x = self.x[idx * self.batch_size: (idx + 1) * self.batch_size] + batch_y = self.y[idx * self.batch_size: (idx + 1) * self.batch_size] + + if self.shuffle: + batch_x, batch_y = shuffle_tuple(batch_x, batch_y) + + if self.augment: + batch_x = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in batch_x]).astype(np.uint8) + batch_x = seq.augment_images(batch_x) + batch_x = batch_x / 255. + else: + batch_x = np.array([np.asarray(load_img(file_path)) / 255. for file_path in batch_x]) + + batch_x = (batch_x - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + batch_y = to_categorical(np.array(batch_y), num_classes=self.num_classes) + + return batch_x, batch_y + +# Train data generator for self defined triplet loss model +class DataGeneratorTriplet(Sequence): + def __init__(self, x_set, y_set, batch_size, num_classes, shuffle=False, augment=False): + self.x, self.y = x_set, y_set + + # Make dict with key -> person_id, value -> list of associated images + self.image_to_label = {} + for image_path, image_label in zip(self.x, self.y): + self.image_to_label.setdefault(image_label, []).append(image_path) + + # Get only anchor_id with more than 1 image + self.anchor_filtered = [k for k, v in self.image_to_label.items() if len(v) > 1] + + self.batch_size = batch_size + self.num_classes = num_classes + self.shuffle = shuffle + self.augment = augment + + def __len__(self): + return math.ceil(len(self.x) / self.batch_size) + + def __getitem__(self, idx): + + if self.shuffle: + random.shuffle(self.anchor_filtered) + + # Get random sample of anchor_ids; amount: batch_size + anchor_ids_sampled = random.sample(self.anchor_filtered, k=self.batch_size) + # Get candidates of nagetive sample ids + negative_id_cands = list(set(self.image_to_label.keys()) - set(anchor_ids_sampled)) + + # Get anchor and positive image paths + anchor_positive_list = [tuple(random.sample(self.image_to_label[id], k=2)) for id in anchor_ids_sampled] + anchor_img_paths, positive_img_paths = zip(*anchor_positive_list) + + # Get negative image_paths + negative_id_sampled = random.sample(negative_id_cands, k=self.batch_size) + negative_img_paths = [random.choice(self.image_to_label[id]) for id in negative_id_sampled] + + if self.augment: + anchor_X_batch = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in anchor_img_paths]).astype(np.uint8) + anchor_X_batch = seq.augment_images(anchor_X_batch) + + positive_X_batch = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in positive_img_paths]).astype(np.uint8) + positive_X_batch = seq.augment_images(positive_X_batch) + + negative_X_batch = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in negative_img_paths]).astype(np.uint8) + negative_X_batch = seq.augment_images(negative_X_batch) + + else: + anchor_X_batch = np.array([np.asarray(load_img(file_path)) for file_path in anchor_img_paths]) + positive_X_batch = np.array([np.asarray(load_img(file_path)) for file_path in positive_img_paths]) + negative_X_batch = np.array([np.asarray(load_img(file_path)) for file_path in negative_img_paths]) + + anchor_X_batch = anchor_X_batch / 255. + positive_X_batch = positive_X_batch / 255. + negative_X_batch = negative_X_batch / 255. + + # Minus mean, devide by standard_deviation + anchor_X_batch = (anchor_X_batch - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + positive_X_batch = (positive_X_batch - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + negative_X_batch = (negative_X_batch - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + + anchor_Y_batch = to_categorical(np.array(anchor_ids_sampled), num_classes=self.num_classes) + + return ([anchor_X_batch, positive_X_batch, negative_X_batch], [anchor_Y_batch, anchor_Y_batch]) + + +# Train data generator for tensorflow-addons semihardtriplet loss model +class DataGeneratorHardTriplet(Sequence): + def __init__(self, x_set, y_set, person_id_num, image_per_person_id, num_classes, shuffle=False, augment=False, center_loss=True): + self.x, self.y = x_set, y_set + + # Make dict with key -> person_id, value -> list of associated images + self.image_to_label = {} + for image_path, image_label in zip(self.x, self.y): + self.image_to_label.setdefault(image_label, []).append(image_path) + + # Get only anchor_id with at least `image_per_person_id` + self.y_filtered = [k for k, v in self.image_to_label.items() if len(v) >= image_per_person_id] + + self.person_id_num = person_id_num + self.image_per_person_id = image_per_person_id + self.num_classes = num_classes + self.shuffle = shuffle + self.augment = augment + self.center_loss = center_loss + + def __len__(self): + return math.ceil(len(self.x) / (self.person_id_num * self.image_per_person_id)) + + def __getitem__(self, idx): + + if self.shuffle: + random.shuffle(self.y_filtered) + + # Get random sample of ids; amount: `person_id_num` + person_ids_chosen = random.sample(self.y_filtered, k=self.person_id_num) + # For each id, get random sample of associate images; amount: `image_per_person_id` + img_paths_sampled = [random.sample(self.image_to_label[id], k=self.image_per_person_id) for id in person_ids_chosen] + img_paths_sampled = [path for paths in img_paths_sampled for path in paths] # Flattening `img_paths_sampled` + + # Expand person_ids_chosen by `image_per_person_id` times to map with `img_paths_sampled` + label_sampled = [[id] * self.image_per_person_id for id in person_ids_chosen] + label_sampled = np.array([label for labels in label_sampled for label in labels]) # Flattening `label_sampled` + + if self.augment: + X_batch = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in img_paths_sampled]).astype(np.uint8) + X_batch = seq.augment_images(X_batch) + else: + X_batch = np.array([np.asarray(load_img(file_path)) for file_path in img_paths_sampled]) + + X_batch = X_batch / 255. + X_batch = (X_batch - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + + Y_batch = to_categorical(np.array(label_sampled), num_classes=self.num_classes) + + if self.center_loss: + return ([X_batch, label_sampled], [label_sampled, Y_batch, label_sampled]) + else: + return (X_batch, [label_sampled, Y_batch]) + + +# Test data generator +class DataGeneratorPredict(Sequence): + + def __init__(self, x_set, batch_size, shuffle=False, augment=False): + self.x = x_set + self.batch_size = batch_size + self.shuffle = shuffle + self.augment = augment + + def __len__(self): + return math.ceil(len(self.x) / self.batch_size) + + def __getitem__(self, idx): + + batch_x = self.x[idx * self.batch_size: (idx + 1) * self.batch_size] + + if self.augment: + batch_x = np.array([np.asarray(load_img(file_path)).astype(np.uint8) for file_path in batch_x]).astype(np.uint8) + batch_x = seq.augment_images(batch_x) + batch_x = batch_x / 255. + else: + batch_x = np.array([np.asarray(load_img(file_path)) / 255. for file_path in batch_x]) + + batch_x = (batch_x - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + + return batch_x + + +def generator_batch_test(img_path_list, img_width, img_height, batch_size=32, shuffle=False): + N = len(img_path_list) + + if shuffle: + img_path_list = shuffle_tuple(img_path_list) + + batch_index = 0 # indicates batch_size + + while True: + current_index = (batch_index*batch_size) % N #the first index for each batch per epoch + if N >= (current_index+batch_size): # judge whether the current end index is over the train num + current_batch_size = batch_size + batch_index += 1 # indicates the next batch_size + else: + current_batch_size = N - current_index + batch_index = 0 + img_batch_list = img_path_list[current_index:current_index + current_batch_size] + + X_batch = np.zeros((current_batch_size, img_height, img_width, 3)) + for i, img_path in enumerate(img_batch_list): + img = cv2.imread(img_path) + if img.shape[:2] != (img_height, img_width): + img = cv2.resize(img, (img_width, img_height)) + img[:, :, [0, 1, 2]] = img[:, :, [2, 1, 0]] + X_batch[i, :, :, :] = img + # normalization + X_batch = X_batch / 255. + X_batch = (X_batch - np.array([0.485, 0.456, 0.406])) / np.array([0.229, 0.224, 0.225]) + yield X_batch \ No newline at end of file