diff --git a/.gitignore b/.gitignore
index e3fa5ac69f5b49bef3f079a9fdb4c1a4df4165d8..9376aa940a6060e88d9b2415909292a95a15ca7a 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,33 +1,5 @@
-paddle/operators/check_t.save
-paddle/operators/check_tensor.ls
-paddle/operators/tensor.save
-python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
-python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
-python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
*.DS_Store
*.vs
-build/
-build_doc/
*.user
-
-.vscode
-.idea
-.project
-.cproject
-.pydevproject
-.settings/
-
*.pyc
-CMakeSettings.json
-Makefile
-.test_env/
-third_party/
-
*~
-bazel-*
-third_party/
-
-build_*
-# clion workspace.
-cmake-build-*
-model_test
\ No newline at end of file
diff --git a/README.md b/README.md
index a8ab968086f8986fc792baa134c7b15079615316..98cef358dae468f9b16209299de2776210970a99 100644
--- a/README.md
+++ b/README.md
@@ -16,54 +16,56 @@ PaddlePaddle 提供了丰富的计算单元，使得用户可以采用模块化
## PaddleCV
模型|简介|模型优势|参考论文
--|:--:|:--:|:--:
-[AlexNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类经典模型|首次在CNN中成功的应用了ReLU、Dropout和LRN，并使用GPU进行运算加速|[ImageNet Classification with Deep Convolutional Neural Networks](https://www.researchgate.net/publication/267960550_ImageNet_Classification_with_Deep_Convolutional_Neural_Networks)
+[AlexNet](./fluid/PaddleCV/image_classification/models)|图像分类经典模型|首次在CNN中成功的应用了ReLU、Dropout和LRN，并使用GPU进行运算加速|[ImageNet Classification with Deep Convolutional Neural Networks](https://www.researchgate.net/publication/267960550_ImageNet_Classification_with_Deep_Convolutional_Neural_Networks)
[VGG](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类经典模型|在AlexNet的基础上使用3*3小卷积核，增加网络深度，具有很好的泛化能力|[Very Deep ConvNets for Large-Scale Inage Recognition](https://arxiv.org/pdf/1409.1556.pdf)
-[GoogleNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类经典模型|在不增加计算负载的前提下增加了网络的深度和宽度，性能更加优越|[Going deeper with convolutions](https://ieeexplore.ieee.org/document/7298594)
-[ResNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|残差网络|引入了新的残差结构，解决了随着网络加深，准确率下降的问题|[Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385)
-[Inception-v4](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类经典模型|更加deeper和wider的inception结构|[Inception-ResNet and the Impact of Residual Connections on Learning](http://arxiv.org/abs/1602.07261)
-[MobileNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|轻量级网络模型|为移动和嵌入式设备提出的高效模型|[MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861)
-[DPN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类模型|结合了DenseNet和ResNeXt的网络结构，对图像分类效果有所提升|[Dual Path Networks](https://arxiv.org/abs/1707.01629)
-[SE-ResNeXt](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/models)|图像分类模型|ResNeXt中加入了SE block，提高了模型准确率|[Squeeze-and-excitation networks](https://arxiv.org/abs/1709.01507)
-[SSD](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/object_detection/README_cn.md)|单阶段目标检测器|在不同尺度的特征图上检测对应尺度的目标,可以方便地插入到任何一种标准卷积网络中|[SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325)
-[Face Detector: PyramidBox](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/face_detection/README_cn.md)|基于SSD的单阶段人脸检测器|利用上下文信息解决困难人脸的检测问题，网络表达能力高，鲁棒性强|[PyramidBox: A Context-assisted Single Shot Face Detector](https://arxiv.org/pdf/1803.07737.pdf)
-[Faster RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/faster_rcnn/README_cn.md)|典型的两阶段目标检测器|创造性地采用卷积网络自行产生建议框，并且和目标检测网络共享卷积网络，建议框数目减少，质量提高|[Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks](https://arxiv.org/abs/1506.01497)
-[ICNet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/icnet)|图像实时语义分割模型|即考虑了速度，也考虑了准确性，在高分辨率图像的准确性和低复杂度网络的效率之间获得平衡|[ICNet for Real-Time Semantic Segmentation on High-Resolution Images](https://arxiv.org/abs/1704.08545)
-[DCGAN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/gan/c_gan)|图像生成模型|深度卷积生成对抗网络，将GAN和卷积网络结合起来，以解决GAN训练不稳定的问题|[Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks](https://arxiv.org/pdf/1511.06434.pdf)
-[ConditionalGAN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/gan/c_gan)|图像生成模型|条件生成对抗网络，一种带条件约束的GAN，使用额外信息对模型增加条件，可以指导数据生成过程|[Conditional Generative Adversarial Nets](https://arxiv.org/abs/1411.1784)
-[CycleGAN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/gan/cycle_gan)|图片转化模型|自动将某一类图片转换成另外一类图片，可用于风格迁移|[Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks](https://arxiv.org/abs/1703.10593)
-[CRNN-CTC模型](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/ocr_recognition)|场景文字识别模型|使用CTC model识别图片中单行英文字符|[Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks](https://www.researchgate.net/publication/221346365_Connectionist_temporal_classification_Labelling_unsegmented_sequence_data_with_recurrent_neural_'networks)
-[Attention模型](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/ocr_recognition)|场景文字识别模型|使用attention 识别图片中单行英文字符|[Recurrent Models of Visual Attention](https://arxiv.org/abs/1406.6247)
+[GoogleNet](./fluid/PaddleCV/image_classification/models)|图像分类经典模型|在不增加计算负载的前提下增加了网络的深度和宽度，性能更加优越|[Going deeper with convolutions](https://ieeexplore.ieee.org/document/7298594)
+[ResNet](./fluid/PaddleCV/image_classification/models)|残差网络|引入了新的残差结构，解决了随着网络加深，准确率下降的问题|[Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385)
+[Inception-v4](./fluid/PaddleCV/image_classification/models)|图像分类经典模型|更加deeper和wider的inception结构|[Inception-ResNet and the Impact of Residual Connections on Learning](http://arxiv.org/abs/1602.07261)
+[MobileNet](./fluid/PaddleCV/image_classification/models)|轻量级网络模型|为移动和嵌入式设备提出的高效模型|[MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications](https://arxiv.org/abs/1704.04861)
+[DPN](./fluid/PaddleCV/image_classification/models)|图像分类模型|结合了DenseNet和ResNeXt的网络结构，对图像分类效果有所提升|[Dual Path Networks](https://arxiv.org/abs/1707.01629)
+[SE-ResNeXt](./fluid/PaddleCV/image_classification/models)|图像分类模型|ResNeXt中加入了SE block，提高了模型准确率|[Squeeze-and-excitation networks](https://arxiv.org/abs/1709.01507)
+[SSD](./fluid/PaddleCV/object_detection/README_cn.md)|单阶段目标检测器|在不同尺度的特征图上检测对应尺度的目标,可以方便地插入到任何一种标准卷积网络中|[SSD: Single Shot MultiBox Detector](https://arxiv.org/abs/1512.02325)
+[Face Detector: PyramidBox](./fluid/PaddleCV/face_detection/README_cn.md)|基于SSD的单阶段人脸检测器|利用上下文信息解决困难人脸的检测问题，网络表达能力高，鲁棒性强|[PyramidBox: A Context-assisted Single Shot Face Detector](https://arxiv.org/pdf/1803.07737.pdf)
+[Faster RCNN](./fluid/PaddleCV/rcnn/README_cn.md)|典型的两阶段目标检测器|创造性地采用卷积网络自行产生建议框，并且和目标检测网络共享卷积网络，建议框数目减少，质量提高|[Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks](https://arxiv.org/abs/1506.01497)
+[Mask RCNN](./fluid/PaddleCV/rcnn/README_cn.md)|基于Faster RCNN模型的经典实例分割模型|在原有Faster RCNN模型基础上添加分割分支，得到掩码结果，实现了掩码和类别预测关系的解藕。|[Mask R-CNN](https://arxiv.org/abs/1703.06870)
+[ICNet](./fluid/PaddleCV/icnet)|图像实时语义分割模型|即考虑了速度，也考虑了准确性，在高分辨率图像的准确性和低复杂度网络的效率之间获得平衡|[ICNet for Real-Time Semantic Segmentation on High-Resolution Images](https://arxiv.org/abs/1704.08545)
+[DCGAN](./fluid/PaddleCV/gan/c_gan)|图像生成模型|深度卷积生成对抗网络，将GAN和卷积网络结合起来，以解决GAN训练不稳定的问题|[Unsupervised Representation Learning with Deep Convolutional Generative Adversarial Networks](https://arxiv.org/pdf/1511.06434.pdf)
+[ConditionalGAN](./fluid/PaddleCV/gan/c_gan)|图像生成模型|条件生成对抗网络，一种带条件约束的GAN，使用额外信息对模型增加条件，可以指导数据生成过程|[Conditional Generative Adversarial Nets](https://arxiv.org/abs/1411.1784)
+[CycleGAN](./fluid/PaddleCV/gan/cycle_gan)|图片转化模型|自动将某一类图片转换成另外一类图片，可用于风格迁移|[Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks](https://arxiv.org/abs/1703.10593)
+[CRNN-CTC模型](./fluid/PaddleCV/ocr_recognition)|场景文字识别模型|使用CTC model识别图片中单行英文字符|[Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks](https://www.researchgate.net/publication/221346365_Connectionist_temporal_classification_Labelling_unsegmented_sequence_data_with_recurrent_neural_'networks)
+[Attention模型](./fluid/PaddleCV/ocr_recognition)|场景文字识别模型|使用attention 识别图片中单行英文字符|[Recurrent Models of Visual Attention](https://arxiv.org/abs/1406.6247)
[Metric Learning](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/metric_learning)|度量学习模型|能够用于分析对象时间的关联、比较关系，可应用于辅助分类、聚类问题，也广泛用于图像检索、人脸识别等领域|-
-[TSN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/video_classification)|视频分类模型|基于长范围时间结构建模，结合了稀疏时间采样策略和视频级监督来保证使用整段视频时学习得有效和高效|[Temporal Segment Networks: Towards Good Practices for Deep Action Recognition](https://arxiv.org/abs/1608.00859)
-[caffe2fluid](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/caffe2fluid)|将Caffe模型转换为Paddle Fluid配置和模型文件工具|-|-
+[TSN](./fluid/PaddleCV/video_classification)|视频分类模型|基于长范围时间结构建模，结合了稀疏时间采样策略和视频级监督来保证使用整段视频时学习得有效和高效|[Temporal Segment Networks: Towards Good Practices for Deep Action Recognition](https://arxiv.org/abs/1608.00859)
+[视频模型库](./fluid/PaddleCV/video)|视频模型库|给开发者提供基于PaddlePaddle的便捷、高效的使用深度学习算法解决视频理解、视频编辑、视频生成等一系列模型||
+[caffe2fluid](./fluid/PaddleCV/caffe2fluid)|将Caffe模型转换为Paddle Fluid配置和模型文件工具|-|-
## PaddleNLP
模型|简介|模型优势|参考论文
--|:--:|:--:|:--:
-[Transformer](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md)|机器翻译模型|基于self-attention，计算复杂度小，并行度高，容易学习长程依赖，翻译效果更好|[Attention Is All You Need](https://arxiv.org/abs/1706.03762)
+[Transformer](./fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md)|机器翻译模型|基于self-attention，计算复杂度小，并行度高，容易学习长程依赖，翻译效果更好|[Attention Is All You Need](https://arxiv.org/abs/1706.03762)
[LAC](https://github.com/baidu/lac/blob/master/README.md)|联合的词法分析模型|能够整体性地完成中文分词、词性标注、专名识别任务|[Chinese Lexical Analysis with Deep Bi-GRU-CRF Network](https://arxiv.org/abs/1807.01882)
[Senta](https://github.com/baidu/Senta/blob/master/README.md)|情感倾向分析模型集|百度AI开放平台中情感倾向分析模型|-
-[DAM](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleNLP/deep_attention_matching_net)|语义匹配模型|百度自然语言处理部发表于ACL-2018的工作,用于检索式聊天机器人多轮对话中应答的选择|[Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network](http://aclweb.org/anthology/P18-1103)
+[DAM](./fluid/PaddleNLP/deep_attention_matching_net)|语义匹配模型|百度自然语言处理部发表于ACL-2018的工作,用于检索式聊天机器人多轮对话中应答的选择|[Multi-Turn Response Selection for Chatbots with Deep Attention Matching Network](http://aclweb.org/anthology/P18-1103)
[SimNet](https://github.com/baidu/AnyQ/blob/master/tools/simnet/train/paddle/README.md)|语义匹配框架|使用SimNet构建出的模型可以便捷的加入AnyQ系统中，增强AnyQ系统的语义匹配能力|-
-[DuReader](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleNLP/machine_reading_comprehension/README.md)|阅读理解模型|百度MRC数据集上的机器阅读理解模型|-
-[Bi-GRU-CRF](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleNLP/sequence_tagging_for_ner/README.md)|命名实体识别|结合了CRF和双向GRU的命名实体识别模型|-
+[DuReader](./fluid/PaddleNLP/machine_reading_comprehension/README.md)|阅读理解模型|百度MRC数据集上的机器阅读理解模型|-
+[Bi-GRU-CRF](./fluid/PaddleNLP/sequence_tagging_for_ner/README.md)|命名实体识别|结合了CRF和双向GRU的命名实体识别模型|-
## PaddleRec
模型|简介|模型优势|参考论文
--|:--:|:--:|:--:
-[TagSpace](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/tagspace)|文本及标签的embedding表示学习模型|应用于工业级的标签推荐，具体应用场景有feed新闻标签推荐等|[#TagSpace: Semantic embeddings from hashtags](https://www.bibsonomy.org/bibtex/0ed4314916f8e7c90d066db45c293462)
-[GRU4Rec](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/gru4rec)|个性化推荐模型|首次将RNN（GRU）运用于session-based推荐，相比传统的KNN和矩阵分解，效果有明显的提升|[Session-based Recommendations with Recurrent Neural Networks](https://arxiv.org/abs/1511.06939)
-[SSR](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/ssr)|序列语义检索推荐模型|使用参考论文中的思想，使用多种时间粒度进行用户行为预测|[Multi-Rate Deep Learning for Temporal Recommendation](https://dl.acm.org/citation.cfm?id=2914726)
-[DeepCTR](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleRec/ctr/README.cn.md)|点击率预估模型|只实现了DeepFM论文中介绍的模型的DNN部分，DeepFM会在其他例子中给出|[DeepFM: A Factorization-Machine based Neural Network for CTR Prediction](https://arxiv.org/abs/1703.04247)
-[Multiview-Simnet](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleRec/multiview_simnet)|个性化推荐模型|基于多元视图，将用户和项目的多个功能视图合并为一个统一模型|[A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems](http://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/frp1159-songA.pdf)
+[TagSpace](./fluid/PaddleRec/tagspace)|文本及标签的embedding表示学习模型|应用于工业级的标签推荐，具体应用场景有feed新闻标签推荐等|[#TagSpace: Semantic embeddings from hashtags](https://www.bibsonomy.org/bibtex/0ed4314916f8e7c90d066db45c293462)
+[GRU4Rec](./fluid/PaddleRec/gru4rec)|个性化推荐模型|首次将RNN（GRU）运用于session-based推荐，相比传统的KNN和矩阵分解，效果有明显的提升|[Session-based Recommendations with Recurrent Neural Networks](https://arxiv.org/abs/1511.06939)
+[SSR](./fluid/PaddleRec/ssr)|序列语义检索推荐模型|使用参考论文中的思想，使用多种时间粒度进行用户行为预测|[Multi-Rate Deep Learning for Temporal Recommendation](https://dl.acm.org/citation.cfm?id=2914726)
+[DeepCTR](./fluid/PaddleRec/ctr/README.cn.md)|点击率预估模型|只实现了DeepFM论文中介绍的模型的DNN部分，DeepFM会在其他例子中给出|[DeepFM: A Factorization-Machine based Neural Network for CTR Prediction](https://arxiv.org/abs/1703.04247)
+[Multiview-Simnet](./fluid/PaddleRec/multiview_simnet)|个性化推荐模型|基于多元视图，将用户和项目的多个功能视图合并为一个统一模型|[A Multi-View Deep Learning Approach for Cross Domain User Modeling in Recommendation Systems](http://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/frp1159-songA.pdf)
## Other Models
模型|简介|模型优势|参考论文
--|:--:|:--:|:--:
-[DeepASR](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepASR/README_cn.md)|语音识别系统|利用Fluid框架完成语音识别中声学模型的配置和训练，并集成 Kaldi 的解码器|-
-[DQN](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepQNetwork/README_cn.md)|深度Q网络|value based强化学习算法，第一个成功地将深度学习和强化学习结合起来的模型|[Human-level control through deep reinforcement learning](https://www.nature.com/articles/nature14236)
-[DoubleDQN](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepQNetwork/README_cn.md)|DQN的变体|将Double Q的想法应用在DQN上，解决过优化问题|[Font Size: Deep Reinforcement Learning with Double Q-Learning](https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewPaper/12389)
-[DuelingDQN](https://github.com/PaddlePaddle/models/blob/develop/fluid/DeepQNetwork/README_cn.md)|DQN的变体|改进了DQN模型，提高了模型的性能|[Dueling Network Architectures for Deep Reinforcement Learning](http://proceedings.mlr.press/v48/wangf16.html)
+[DeepASR](./fluid/DeepASR/README_cn.md)|语音识别系统|利用Fluid框架完成语音识别中声学模型的配置和训练，并集成 Kaldi 的解码器|-
+[DQN](./fluid/DeepQNetwork/README_cn.md)|深度Q网络|value based强化学习算法，第一个成功地将深度学习和强化学习结合起来的模型|[Human-level control through deep reinforcement learning](https://www.nature.com/articles/nature14236)
+[DoubleDQN](./fluid/DeepQNetwork/README_cn.md)|DQN的变体|将Double Q的想法应用在DQN上，解决过优化问题|[Font Size: Deep Reinforcement Learning with Double Q-Learning](https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewPaper/12389)
+[DuelingDQN](./fluid/DeepQNetwork/README_cn.md)|DQN的变体|改进了DQN模型，提高了模型的性能|[Dueling Network Architectures for Deep Reinforcement Learning](http://proceedings.mlr.press/v48/wangf16.html)
## License
This tutorial is contributed by [PaddlePaddle](https://github.com/PaddlePaddle/Paddle) and licensed under the [Apache-2.0 license](LICENSE).
diff --git a/fluid/AutoDL/LRC/README.md b/fluid/AutoDL/LRC/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..df9af47d4a3876371673cbbfef0ad2553768b9a5
--- /dev/null
+++ b/fluid/AutoDL/LRC/README.md
@@ -0,0 +1,74 @@
+# LRC Local Rademachar Complexity Regularization
+Regularization of Deep Neural Networks(DNNs) for the sake of improving their generalization capability is important and chllenging. This directory contains image classification model based on a novel regularizer rooted in Local Rademacher Complexity (LRC). We appreciate the contribution by [DARTS](https://arxiv.org/abs/1806.09055) for our research. The regularization by LRC and DARTS are combined in this model on CIFAR-10 dataset. Code accompanying the paper
+> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
+> Yingzhen Yang, Xingjian Li, Jun Huan.\
+> _arXiv:1902.00873_.
+
+---
+# Table of Contents
+
+- [Installation](#installation)
+- [Data preparation](#data-preparation)
+- [Training](#training)
+
+## Installation
+
+Running sample code in this directory requires PaddelPaddle Fluid v.1.2.0 and later. If the PaddlePaddle on your device is lower than this version, please follow the instructions in [installation document](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle) and make an update.
+
+## Data preparation
+
+When you want to use the cifar-10 dataset for the first time, you can download the dataset as:
+
+ sh ./dataset/download.sh
+
+Please make sure your environment has an internet connection.
+
+The dataset will be downloaded to `dataset/cifar/cifar-10-batches-py` in the same directory as the `train.py`. If automatic download fails, you can download cifar-10-python.tar.gz from https://www.cs.toronto.edu/~kriz/cifar.html and decompress it to the location mentioned above.
+
+
+## Training
+
+After data preparation, one can start the training step by:
+
+ python -u train_mixup.py \
+ --batch_size=80 \
+ --auxiliary \
+ --weight_decay=0.0003 \
+ --learning_rate=0.025 \
+ --lrc_loss_lambda=0.7 \
+ --cutout
+- Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to train.
+- For more help on arguments:
+
+ python train_mixup.py --help
+
+**data reader introduction:**
+
+* Data reader is defined in `reader.py`.
+* Reshape the images to 32 * 32.
+* In training stage, images are padding to 40 * 40 and cropped randomly to the original size.
+* In training stage, images are horizontally random flipped.
+* Images are standardized to (0, 1).
+* In training stage, cutout images randomly.
+* Shuffle the order of the input images during training.
+
+**model configuration:**
+
+* Use auxiliary loss and auxiliary\_weight=0.4.
+* Use dropout and drop\_path\_prob=0.2.
+* Set lrc\_loss\_lambda=0.7.
+
+**training strategy:**
+
+* Use momentum optimizer with momentum=0.9.
+* Weight decay is 0.0003.
+* Use cosine decay with init\_lr=0.025.
+* Total epoch is 600.
+* Use Xaiver initalizer to weight in conv2d, Constant initalizer to weight in batch norm and Normal initalizer to weight in fc.
+* Initalize bias in batch norm and fc to zero constant and do not add bias to conv2d.
+
+
+## Reference
+
+ - DARTS: Differentiable Architecture Search [`paper`](https://arxiv.org/abs/1806.09055)
+ - Differentiable architecture search in PyTorch [`code`](https://github.com/quark0/darts)
diff --git a/fluid/AutoDL/LRC/README_cn.md b/fluid/AutoDL/LRC/README_cn.md
new file mode 100644
index 0000000000000000000000000000000000000000..06dc937074de199af31db97ee200e7690443b1b0
--- /dev/null
+++ b/fluid/AutoDL/LRC/README_cn.md
@@ -0,0 +1,71 @@
+# LRC 局部Rademachar复杂度正则化
+为了在深度神经网络中提升泛化能力，正则化的选择十分重要也具有挑战性。本目录包括了一种基于局部rademacher复杂度的新型正则（LRC）的图像分类模型。十分感谢[DARTS](https://arxiv.org/abs/1806.09055)模型对本研究提供的帮助。该模型将LRC正则和DARTS网络相结合，在CIFAR-10数据集中得到了很出色的效果。代码和文章一同发布
+> [An Empirical Study on Regularization of Deep Neural Networks by Local Rademacher Complexity](https://arxiv.org/abs/1902.00873)\
+> Yingzhen Yang, Xingjian Li, Jun Huan.\
+> _arXiv:1902.00873_.
+
+---
+# 内容
+
+- [安装](#安装)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+
+## 安装
+
+在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.2.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html#paddlepaddle)中的说明来更新PaddlePaddle。
+
+## 数据准备
+
+第一次使用CIFAR-10数据集时，您可以通过如果命令下载：
+
+ sh ./dataset/download.sh
+
+请确保您的环境有互联网连接。数据会下载到`train.py`同目录下的`dataset/cifar/cifar-10-batches-py`。如果下载失败，您可以自行从https://www.cs.toronto.edu/~kriz/cifar.html上下载cifar-10-python.tar.gz并解压到上述位置。
+
+## 模型训练
+
+数据准备好后，可以通过如下命令开始训练：
+
+ python -u train_mixup.py \
+ --batch_size=80 \
+ --auxiliary \
+ --weight_decay=0.0003 \
+ --learning_rate=0.025 \
+ --lrc_loss_lambda=0.7 \
+ --cutout
+- 通过设置 ```export CUDA_VISIBLE_DEVICES=0```指定单张GPU训练。
+- 可选参数见：
+
+ python train_mixup.py --help
+
+**数据读取器说明：**
+
+* 数据读取器定义在`reader.py`中
+* 输入图像尺寸统一变换为32 * 32
+* 训练时将图像填充为40 * 40然后随机剪裁为原输入图像大小
+* 训练时图像随机水平翻转
+* 对图像每个像素做归一化处理
+* 训练时对图像做随机遮挡
+* 训练时对输入图像做随机洗牌
+
+**模型配置：**
+
+* 使用辅助损失，辅助损失权重为0.4
+* 使用dropout，随机丢弃率为0.2
+* 设置lrc\_loss\_lambda为0.7
+
+**训练策略：**
+
+* 采用momentum优化算法训练，momentum=0.9
+* 权重衰减系数为0.0001
+* 采用正弦学习率衰减，初始学习率为0.025
+* 总共训练600轮
+* 对卷积权重采用Xaiver初始化，对batch norm权重采用固定初始化，对全连接层权重采用高斯初始化
+* 对batch norm和全连接层偏差采用固定初始化，不对卷积设置偏差
+
+
+## 引用
+
+ - DARTS: Differentiable Architecture Search [`论文`](https://arxiv.org/abs/1806.09055)
+ - Differentiable Architecture Search in PyTorch [`代码`](https://github.com/quark0/darts)
diff --git a/fluid/AutoDL/LRC/dataset/download.sh b/fluid/AutoDL/LRC/dataset/download.sh
new file mode 100644
index 0000000000000000000000000000000000000000..0981c3b6878421f80d392f314fd0ae836644a63c
--- /dev/null
+++ b/fluid/AutoDL/LRC/dataset/download.sh
@@ -0,0 +1,10 @@
+DIR="$( cd "$(dirname "$0")" ; pwd -P )"
+cd "$DIR"
+mkdir cifar
+cd cifar
+# Download the data.
+echo "Downloading..."
+wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
+# Extract the data.
+echo "Extracting..."
+tar zvxf cifar-10-python.tar.gz
diff --git a/fluid/AutoDL/LRC/genotypes.py b/fluid/AutoDL/LRC/genotypes.py
new file mode 100644
index 0000000000000000000000000000000000000000..349fbd2478a7c2d1bb4cc3dd901b470de3c8b906
--- /dev/null
+++ b/fluid/AutoDL/LRC/genotypes.py
@@ -0,0 +1,116 @@
+# Copyright (c) 2019 PaddlePaddle 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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+
+from collections import namedtuple
+
+Genotype = namedtuple('Genotype', 'normal normal_concat reduce reduce_concat')
+
+PRIMITIVES = [
+ 'none', 'max_pool_3x3', 'avg_pool_3x3', 'skip_connect', 'sep_conv_3x3',
+ 'sep_conv_5x5', 'dil_conv_3x3', 'dil_conv_5x5'
+]
+
+NASNet = Genotype(
+ normal=[
+ ('sep_conv_5x5', 1),
+ ('sep_conv_3x3', 0),
+ ('sep_conv_5x5', 0),
+ ('sep_conv_3x3', 0),
+ ('avg_pool_3x3', 1),
+ ('skip_connect', 0),
+ ('avg_pool_3x3', 0),
+ ('avg_pool_3x3', 0),
+ ('sep_conv_3x3', 1),
+ ('skip_connect', 1),
+ ],
+ normal_concat=[2, 3, 4, 5, 6],
+ reduce=[
+ ('sep_conv_5x5', 1),
+ ('sep_conv_7x7', 0),
+ ('max_pool_3x3', 1),
+ ('sep_conv_7x7', 0),
+ ('avg_pool_3x3', 1),
+ ('sep_conv_5x5', 0),
+ ('skip_connect', 3),
+ ('avg_pool_3x3', 2),
+ ('sep_conv_3x3', 2),
+ ('max_pool_3x3', 1),
+ ],
+ reduce_concat=[4, 5, 6], )
+
+AmoebaNet = Genotype(
+ normal=[
+ ('avg_pool_3x3', 0),
+ ('max_pool_3x3', 1),
+ ('sep_conv_3x3', 0),
+ ('sep_conv_5x5', 2),
+ ('sep_conv_3x3', 0),
+ ('avg_pool_3x3', 3),
+ ('sep_conv_3x3', 1),
+ ('skip_connect', 1),
+ ('skip_connect', 0),
+ ('avg_pool_3x3', 1),
+ ],
+ normal_concat=[4, 5, 6],
+ reduce=[
+ ('avg_pool_3x3', 0),
+ ('sep_conv_3x3', 1),
+ ('max_pool_3x3', 0),
+ ('sep_conv_7x7', 2),
+ ('sep_conv_7x7', 0),
+ ('avg_pool_3x3', 1),
+ ('max_pool_3x3', 0),
+ ('max_pool_3x3', 1),
+ ('conv_7x1_1x7', 0),
+ ('sep_conv_3x3', 5),
+ ],
+ reduce_concat=[3, 4, 6])
+
+DARTS_V1 = Genotype(
+ normal=[('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('skip_connect', 0),
+ ('sep_conv_3x3', 1), ('skip_connect', 0), ('sep_conv_3x3', 1),
+ ('sep_conv_3x3', 0), ('skip_connect', 2)],
+ normal_concat=[2, 3, 4, 5],
+ reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2),
+ ('max_pool_3x3', 0), ('max_pool_3x3', 0), ('skip_connect', 2),
+ ('skip_connect', 2), ('avg_pool_3x3', 0)],
+ reduce_concat=[2, 3, 4, 5])
+DARTS_V2 = Genotype(
+ normal=[('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0),
+ ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0),
+ ('skip_connect', 0), ('dil_conv_3x3', 2)],
+ normal_concat=[2, 3, 4, 5],
+ reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2),
+ ('max_pool_3x3', 1), ('max_pool_3x3', 0), ('skip_connect', 2),
+ ('skip_connect', 2), ('max_pool_3x3', 1)],
+ reduce_concat=[2, 3, 4, 5])
+
+MY_DARTS = Genotype(
+ normal=[('sep_conv_3x3', 0), ('skip_connect', 1), ('skip_connect', 0),
+ ('dil_conv_5x5', 1), ('skip_connect', 0), ('sep_conv_3x3', 1),
+ ('skip_connect', 0), ('sep_conv_3x3', 1)],
+ normal_concat=range(2, 6),
+ reduce=[('max_pool_3x3', 0), ('max_pool_3x3', 1), ('max_pool_3x3', 0),
+ ('skip_connect', 2), ('max_pool_3x3', 0), ('skip_connect', 2),
+ ('skip_connect', 2), ('skip_connect', 3)],
+ reduce_concat=range(2, 6))
+
+DARTS = MY_DARTS
diff --git a/fluid/AutoDL/LRC/learning_rate.py b/fluid/AutoDL/LRC/learning_rate.py
new file mode 100644
index 0000000000000000000000000000000000000000..3965171b487884d36e4a7447f10f312204803bf8
--- /dev/null
+++ b/fluid/AutoDL/LRC/learning_rate.py
@@ -0,0 +1,43 @@
+# Copyright (c) 2019 PaddlePaddle 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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.layers.ops as ops
+from paddle.fluid.layers.learning_rate_scheduler import _decay_step_counter
+import math
+from paddle.fluid.initializer import init_on_cpu
+
+
+def cosine_decay(learning_rate, num_epoch, steps_one_epoch):
+ """Applies cosine decay to the learning rate.
+ lr = 0.5 * (math.cos(epoch * (math.pi / 120)) + 1)
+ """
+ global_step = _decay_step_counter()
+
+ with init_on_cpu():
+ decayed_lr = learning_rate * \
+ (ops.cos((global_step / steps_one_epoch) \
+ * math.pi / num_epoch) + 1)/2
+ return decayed_lr
diff --git a/fluid/AutoDL/LRC/model.py b/fluid/AutoDL/LRC/model.py
new file mode 100644
index 0000000000000000000000000000000000000000..45a403495ecc0b7cc0ac3b541d75702adbef31b2
--- /dev/null
+++ b/fluid/AutoDL/LRC/model.py
@@ -0,0 +1,313 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import sys
+import numpy as np
+import time
+import functools
+import paddle
+import paddle.fluid as fluid
+from operations import *
+
+
+class Cell():
+ def __init__(self, genotype, C_prev_prev, C_prev, C, reduction,
+ reduction_prev):
+ print(C_prev_prev, C_prev, C)
+
+ if reduction_prev:
+ self.preprocess0 = functools.partial(FactorizedReduce, C_out=C)
+ else:
+ self.preprocess0 = functools.partial(
+ ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
+ self.preprocess1 = functools.partial(
+ ReLUConvBN, C_out=C, kernel_size=1, stride=1, padding=0)
+ if reduction:
+ op_names, indices = zip(*genotype.reduce)
+ concat = genotype.reduce_concat
+ else:
+ op_names, indices = zip(*genotype.normal)
+ concat = genotype.normal_concat
+ print(op_names, indices, concat, reduction)
+ self._compile(C, op_names, indices, concat, reduction)
+
+ def _compile(self, C, op_names, indices, concat, reduction):
+ assert len(op_names) == len(indices)
+ self._steps = len(op_names) // 2
+ self._concat = concat
+ self.multiplier = len(concat)
+
+ self._ops = []
+ for name, index in zip(op_names, indices):
+ stride = 2 if reduction and index < 2 else 1
+ op = functools.partial(OPS[name], C=C, stride=stride, affine=True)
+ self._ops += [op]
+ self._indices = indices
+
+ def forward(self, s0, s1, drop_prob, is_train, name):
+ self.training = is_train
+ preprocess0_name = name + 'preprocess0.'
+ preprocess1_name = name + 'preprocess1.'
+ s0 = self.preprocess0(s0, name=preprocess0_name)
+ s1 = self.preprocess1(s1, name=preprocess1_name)
+ out = [s0, s1]
+ for i in range(self._steps):
+ h1 = out[self._indices[2 * i]]
+ h2 = out[self._indices[2 * i + 1]]
+ op1 = self._ops[2 * i]
+ op2 = self._ops[2 * i + 1]
+ h3 = op1(h1, name=name + '_ops.' + str(2 * i) + '.')
+ h4 = op2(h2, name=name + '_ops.' + str(2 * i + 1) + '.')
+ if self.training and drop_prob > 0.:
+ if h3 != h1:
+ h3 = fluid.layers.dropout(
+ h3,
+ drop_prob,
+ dropout_implementation='upscale_in_train')
+ if h4 != h2:
+ h4 = fluid.layers.dropout(
+ h4,
+ drop_prob,
+ dropout_implementation='upscale_in_train')
+ s = h3 + h4
+ out += [s]
+ return fluid.layers.concat([out[i] for i in self._concat], axis=1)
+
+
+def AuxiliaryHeadCIFAR(input, num_classes, aux_name='auxiliary_head'):
+ relu_a = fluid.layers.relu(input)
+ pool_a = fluid.layers.pool2d(relu_a, 5, 'avg', 3)
+ conv2d_a = fluid.layers.conv2d(
+ pool_a,
+ 128,
+ 1,
+ name=aux_name + '.features.2',
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=aux_name + '.features.2.weight'),
+ bias_attr=False)
+ bn_a_name = aux_name + '.features.3'
+ bn_a = fluid.layers.batch_norm(
+ conv2d_a,
+ act='relu',
+ name=bn_a_name,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=bn_a_name + '.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=bn_a_name + '.bias'),
+ moving_mean_name=bn_a_name + '.running_mean',
+ moving_variance_name=bn_a_name + '.running_var')
+ conv2d_b = fluid.layers.conv2d(
+ bn_a,
+ 768,
+ 2,
+ name=aux_name + '.features.5',
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=aux_name + '.features.5.weight'),
+ bias_attr=False)
+ bn_b_name = aux_name + '.features.6'
+ bn_b = fluid.layers.batch_norm(
+ conv2d_b,
+ act='relu',
+ name=bn_b_name,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=bn_b_name + '.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=bn_b_name + '.bias'),
+ moving_mean_name=bn_b_name + '.running_mean',
+ moving_variance_name=bn_b_name + '.running_var')
+ fc_name = aux_name + '.classifier'
+ fc = fluid.layers.fc(bn_b,
+ num_classes,
+ name=fc_name,
+ param_attr=ParamAttr(
+ initializer=Normal(scale=1e-3),
+ name=fc_name + '.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=fc_name + '.bias'))
+ return fc
+
+
+def StemConv(input, C_out, kernel_size, padding):
+ conv_a = fluid.layers.conv2d(
+ input,
+ C_out,
+ kernel_size,
+ padding=padding,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0), name='stem.0.weight'),
+ bias_attr=False)
+ bn_a = fluid.layers.batch_norm(
+ conv_a,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name='stem.1.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name='stem.1.bias'),
+ moving_mean_name='stem.1.running_mean',
+ moving_variance_name='stem.1.running_var')
+ return bn_a
+
+
+class NetworkCIFAR(object):
+ def __init__(self, C, class_num, layers, auxiliary, genotype):
+ self.class_num = class_num
+ self._layers = layers
+ self._auxiliary = auxiliary
+
+ stem_multiplier = 3
+ self.drop_path_prob = 0
+ C_curr = stem_multiplier * C
+
+ C_prev_prev, C_prev, C_curr = C_curr, C_curr, C
+ self.cells = []
+ reduction_prev = False
+ for i in range(layers):
+ if i in [layers // 3, 2 * layers // 3]:
+ C_curr *= 2
+ reduction = True
+ else:
+ reduction = False
+ cell = Cell(genotype, C_prev_prev, C_prev, C_curr, reduction,
+ reduction_prev)
+ reduction_prev = reduction
+ self.cells += [cell]
+ C_prev_prev, C_prev = C_prev, cell.multiplier * C_curr
+ if i == 2 * layers // 3:
+ C_to_auxiliary = C_prev
+
+ def forward(self, init_channel, is_train):
+ self.training = is_train
+ self.logits_aux = None
+ num_channel = init_channel * 3
+ s0 = StemConv(self.image, num_channel, kernel_size=3, padding=1)
+ s1 = s0
+ for i, cell in enumerate(self.cells):
+ name = 'cells.' + str(i) + '.'
+ s0, s1 = s1, cell.forward(s0, s1, self.drop_path_prob, is_train,
+ name)
+ if i == int(2 * self._layers // 3):
+ if self._auxiliary and self.training:
+ self.logits_aux = AuxiliaryHeadCIFAR(s1, self.class_num)
+ out = fluid.layers.adaptive_pool2d(s1, (1, 1), "avg")
+ self.logits = fluid.layers.fc(out,
+ size=self.class_num,
+ param_attr=ParamAttr(
+ initializer=Normal(scale=1e-3),
+ name='classifier.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ name='classifier.bias'))
+ return self.logits, self.logits_aux
+
+ def build_input(self, image_shape, batch_size, is_train):
+ if is_train:
+ py_reader = fluid.layers.py_reader(
+ capacity=64,
+ shapes=[[-1] + image_shape, [-1, 1], [-1, 1], [-1, 1], [-1, 1],
+ [-1, 1], [-1, batch_size, self.class_num - 1]],
+ lod_levels=[0, 0, 0, 0, 0, 0, 0],
+ dtypes=[
+ "float32", "int64", "int64", "float32", "int32", "int32",
+ "float32"
+ ],
+ use_double_buffer=True,
+ name='train_reader')
+ else:
+ py_reader = fluid.layers.py_reader(
+ capacity=64,
+ shapes=[[-1] + image_shape, [-1, 1]],
+ lod_levels=[0, 0],
+ dtypes=["float32", "int64"],
+ use_double_buffer=True,
+ name='test_reader')
+ return py_reader
+
+ def train_model(self, py_reader, init_channels, aux, aux_w, batch_size,
+ loss_lambda):
+ self.image, self.ya, self.yb, self.lam, self.label_reshape,\
+ self.non_label_reshape, self.rad_var = fluid.layers.read_file(py_reader)
+ self.logits, self.logits_aux = self.forward(init_channels, True)
+ self.mixup_loss = self.mixup_loss(aux, aux_w)
+ self.lrc_loss = self.lrc_loss(batch_size)
+ return self.mixup_loss + loss_lambda * self.lrc_loss
+
+ def test_model(self, py_reader, init_channels):
+ self.image, self.ya = fluid.layers.read_file(py_reader)
+ self.logits, _ = self.forward(init_channels, False)
+ prob = fluid.layers.softmax(self.logits, use_cudnn=False)
+ loss = fluid.layers.cross_entropy(prob, self.ya)
+ acc_1 = fluid.layers.accuracy(self.logits, self.ya, k=1)
+ acc_5 = fluid.layers.accuracy(self.logits, self.ya, k=5)
+ return loss, acc_1, acc_5
+
+ def mixup_loss(self, auxiliary, auxiliary_weight):
+ prob = fluid.layers.softmax(self.logits, use_cudnn=False)
+ loss_a = fluid.layers.cross_entropy(prob, self.ya)
+ loss_b = fluid.layers.cross_entropy(prob, self.yb)
+ loss_a_mean = fluid.layers.reduce_mean(loss_a)
+ loss_b_mean = fluid.layers.reduce_mean(loss_b)
+ loss = self.lam * loss_a_mean + (1 - self.lam) * loss_b_mean
+ if auxiliary:
+ prob_aux = fluid.layers.softmax(self.logits_aux, use_cudnn=False)
+ loss_a_aux = fluid.layers.cross_entropy(prob_aux, self.ya)
+ loss_b_aux = fluid.layers.cross_entropy(prob_aux, self.yb)
+ loss_a_aux_mean = fluid.layers.reduce_mean(loss_a_aux)
+ loss_b_aux_mean = fluid.layers.reduce_mean(loss_b_aux)
+ loss_aux = self.lam * loss_a_aux_mean + (1 - self.lam
+ ) * loss_b_aux_mean
+ return loss + auxiliary_weight * loss_aux
+
+ def lrc_loss(self, batch_size):
+ y_diff_reshape = fluid.layers.reshape(self.logits, shape=(-1, 1))
+ label_reshape = fluid.layers.squeeze(self.label_reshape, axes=[1])
+ non_label_reshape = fluid.layers.squeeze(
+ self.non_label_reshape, axes=[1])
+ label_reshape.stop_gradient = True
+ non_label_reshape.stop_graident = True
+
+ y_diff_label_reshape = fluid.layers.gather(y_diff_reshape,
+ label_reshape)
+ y_diff_non_label_reshape = fluid.layers.gather(y_diff_reshape,
+ non_label_reshape)
+ y_diff_label = fluid.layers.reshape(
+ y_diff_label_reshape, shape=(-1, batch_size, 1))
+ y_diff_non_label = fluid.layers.reshape(
+ y_diff_non_label_reshape,
+ shape=(-1, batch_size, self.class_num - 1))
+ y_diff_ = y_diff_non_label - y_diff_label
+
+ y_diff_ = fluid.layers.transpose(y_diff_, perm=[1, 2, 0])
+ rad_var_trans = fluid.layers.transpose(self.rad_var, perm=[1, 2, 0])
+ rad_y_diff_trans = rad_var_trans * y_diff_
+ lrc_loss_sum = fluid.layers.reduce_sum(rad_y_diff_trans, dim=[0, 1])
+ lrc_loss_ = fluid.layers.abs(lrc_loss_sum) / (batch_size *
+ (self.class_num - 1))
+ lrc_loss_mean = fluid.layers.reduce_mean(lrc_loss_)
+
+ return lrc_loss_mean
diff --git a/fluid/AutoDL/LRC/operations.py b/fluid/AutoDL/LRC/operations.py
new file mode 100644
index 0000000000000000000000000000000000000000..b015722a1bc5dbf682c90812a971f3dbb2cd8c9a
--- /dev/null
+++ b/fluid/AutoDL/LRC/operations.py
@@ -0,0 +1,349 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import os
+import sys
+import numpy as np
+import time
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid.param_attr import ParamAttr
+from paddle.fluid.initializer import Xavier
+from paddle.fluid.initializer import Normal
+from paddle.fluid.initializer import Constant
+
+OPS = {
+ 'none' : lambda input, C, stride, name, affine: Zero(input, stride, name),
+ 'avg_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'avg', pool_stride=stride, pool_padding=1, name=name),
+ 'max_pool_3x3' : lambda input, C, stride, name, affine: fluid.layers.pool2d(input, 3, 'max', pool_stride=stride, pool_padding=1, name=name),
+ 'skip_connect' : lambda input,C, stride, name, affine: Identity(input, name) if stride == 1 else FactorizedReduce(input, C, name=name, affine=affine),
+ 'sep_conv_3x3' : lambda input,C, stride, name, affine: SepConv(input, C, C, 3, stride, 1, name=name, affine=affine),
+ 'sep_conv_5x5' : lambda input,C, stride, name, affine: SepConv(input, C, C, 5, stride, 2, name=name, affine=affine),
+ 'sep_conv_7x7' : lambda input,C, stride, name, affine: SepConv(input, C, C, 7, stride, 3, name=name, affine=affine),
+ 'dil_conv_3x3' : lambda input,C, stride, name, affine: DilConv(input, C, C, 3, stride, 2, 2, name=name, affine=affine),
+ 'dil_conv_5x5' : lambda input,C, stride, name, affine: DilConv(input, C, C, 5, stride, 4, 2, name=name, affine=affine),
+ 'conv_7x1_1x7' : lambda input,C, stride, name, affine: SevenConv(input, C, name=name, affine=affine)
+}
+
+
+def ReLUConvBN(input, C_out, kernel_size, stride, padding, name='',
+ affine=True):
+ relu_a = fluid.layers.relu(input)
+ conv2d_a = fluid.layers.conv2d(
+ relu_a,
+ C_out,
+ kernel_size,
+ stride,
+ padding,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.1.weight'),
+ bias_attr=False)
+ if affine:
+ reluconvbn_out = fluid.layers.batch_norm(
+ conv2d_a,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'op.2.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'op.2.bias'),
+ moving_mean_name=name + 'op.2.running_mean',
+ moving_variance_name=name + 'op.2.running_var')
+ else:
+ reluconvbn_out = fluid.layers.batch_norm(
+ conv2d_a,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'op.2.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'op.2.bias'),
+ moving_mean_name=name + 'op.2.running_mean',
+ moving_variance_name=name + 'op.2.running_var')
+ return reluconvbn_out
+
+
+def DilConv(input,
+ C_in,
+ C_out,
+ kernel_size,
+ stride,
+ padding,
+ dilation,
+ name='',
+ affine=True):
+ relu_a = fluid.layers.relu(input)
+ conv2d_a = fluid.layers.conv2d(
+ relu_a,
+ C_in,
+ kernel_size,
+ stride,
+ padding,
+ dilation,
+ groups=C_in,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.1.weight'),
+ bias_attr=False,
+ use_cudnn=False)
+ conv2d_b = fluid.layers.conv2d(
+ conv2d_a,
+ C_out,
+ 1,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.2.weight'),
+ bias_attr=False)
+ if affine:
+ dilconv_out = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+ else:
+ dilconv_out = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+ return dilconv_out
+
+
+def SepConv(input,
+ C_in,
+ C_out,
+ kernel_size,
+ stride,
+ padding,
+ name='',
+ affine=True):
+ relu_a = fluid.layers.relu(input)
+ conv2d_a = fluid.layers.conv2d(
+ relu_a,
+ C_in,
+ kernel_size,
+ stride,
+ padding,
+ groups=C_in,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.1.weight'),
+ bias_attr=False,
+ use_cudnn=False)
+ conv2d_b = fluid.layers.conv2d(
+ conv2d_a,
+ C_in,
+ 1,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.2.weight'),
+ bias_attr=False)
+ if affine:
+ bn_a = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+ else:
+ bn_a = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+
+ relu_b = fluid.layers.relu(bn_a)
+ conv2d_d = fluid.layers.conv2d(
+ relu_b,
+ C_in,
+ kernel_size,
+ 1,
+ padding,
+ groups=C_in,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.5.weight'),
+ bias_attr=False,
+ use_cudnn=False)
+ conv2d_e = fluid.layers.conv2d(
+ conv2d_d,
+ C_out,
+ 1,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.6.weight'),
+ bias_attr=False)
+ if affine:
+ sepconv_out = fluid.layers.batch_norm(
+ conv2d_e,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'op.7.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'op.7.bias'),
+ moving_mean_name=name + 'op.7.running_mean',
+ moving_variance_name=name + 'op.7.running_var')
+ else:
+ sepconv_out = fluid.layers.batch_norm(
+ conv2d_e,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'op.7.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'op.7.bias'),
+ moving_mean_name=name + 'op.7.running_mean',
+ moving_variance_name=name + 'op.7.running_var')
+ return sepconv_out
+
+
+def SevenConv(input, C_out, stride, name='', affine=True):
+ relu_a = fluid.layers.relu(input)
+ conv2d_a = fluid.layers.conv2d(
+ relu_a,
+ C_out, (1, 7), (1, stride), (0, 3),
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.1.weight'),
+ bias_attr=False)
+ conv2d_b = fluid.layers.conv2d(
+ conv2d_a,
+ C_out, (7, 1), (stride, 1), (3, 0),
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'op.2.weight'),
+ bias_attr=False)
+ if affine:
+ out = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+ else:
+ out = fluid.layers.batch_norm(
+ conv2d_b,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'op.3.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'op.3.bias'),
+ moving_mean_name=name + 'op.3.running_mean',
+ moving_variance_name=name + 'op.3.running_var')
+
+
+def Identity(input, name=''):
+ return input
+
+
+def Zero(input, stride, name=''):
+ ones = np.ones(input.shape[-2:])
+ ones[::stride, ::stride] = 0
+ ones = fluid.layers.assign(ones)
+ return input * ones
+
+
+def FactorizedReduce(input, C_out, name='', affine=True):
+ relu_a = fluid.layers.relu(input)
+ conv2d_a = fluid.layers.conv2d(
+ relu_a,
+ C_out // 2,
+ 1,
+ 2,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'conv_1.weight'),
+ bias_attr=False)
+ h_end = relu_a.shape[2]
+ w_end = relu_a.shape[3]
+ slice_a = fluid.layers.slice(relu_a, [2, 3], [1, 1], [h_end, w_end])
+ conv2d_b = fluid.layers.conv2d(
+ slice_a,
+ C_out // 2,
+ 1,
+ 2,
+ param_attr=ParamAttr(
+ initializer=Xavier(
+ uniform=False, fan_in=0),
+ name=name + 'conv_2.weight'),
+ bias_attr=False)
+ out = fluid.layers.concat([conv2d_a, conv2d_b], axis=1)
+ if affine:
+ out = fluid.layers.batch_norm(
+ out,
+ param_attr=ParamAttr(
+ initializer=Constant(1.), name=name + 'bn.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.), name=name + 'bn.bias'),
+ moving_mean_name=name + 'bn.running_mean',
+ moving_variance_name=name + 'bn.running_var')
+ else:
+ out = fluid.layers.batch_norm(
+ out,
+ param_attr=ParamAttr(
+ initializer=Constant(1.),
+ learning_rate=0.,
+ name=name + 'bn.weight'),
+ bias_attr=ParamAttr(
+ initializer=Constant(0.),
+ learning_rate=0.,
+ name=name + 'bn.bias'),
+ moving_mean_name=name + 'bn.running_mean',
+ moving_variance_name=name + 'bn.running_var')
+ return out
diff --git a/fluid/AutoDL/LRC/reader.py b/fluid/AutoDL/LRC/reader.py
new file mode 100644
index 0000000000000000000000000000000000000000..20b32b504e9245c4ff3892f08736d800080daab4
--- /dev/null
+++ b/fluid/AutoDL/LRC/reader.py
@@ -0,0 +1,187 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rig hts 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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+"""
+CIFAR-10 dataset.
+This module will download dataset from
+https://www.cs.toronto.edu/~kriz/cifar.html and parse train/test set into
+paddle reader creators.
+The CIFAR-10 dataset consists of 60000 32x32 colour images in 10 classes,
+with 6000 images per class. There are 50000 training images and 10000 test images.
+"""
+
+from PIL import Image
+from PIL import ImageOps
+import numpy as np
+
+import cPickle
+import random
+import utils
+import paddle.fluid as fluid
+import time
+import os
+import functools
+import paddle.reader
+
+__all__ = ['train10', 'test10']
+
+image_size = 32
+image_depth = 3
+half_length = 8
+
+CIFAR_MEAN = [0.4914, 0.4822, 0.4465]
+CIFAR_STD = [0.24703233, 0.24348505, 0.26158768]
+
+
+def generate_reshape_label(label, batch_size, CIFAR_CLASSES=10):
+ reshape_label = np.zeros((batch_size, 1), dtype='int32')
+ reshape_non_label = np.zeros(
+ (batch_size * (CIFAR_CLASSES - 1), 1), dtype='int32')
+ num = 0
+ for i in range(batch_size):
+ label_i = label[i]
+ reshape_label[i] = label_i + i * CIFAR_CLASSES
+ for j in range(CIFAR_CLASSES):
+ if label_i != j:
+ reshape_non_label[num] = \
+ j + i * CIFAR_CLASSES
+ num += 1
+ return reshape_label, reshape_non_label
+
+
+def generate_bernoulli_number(batch_size, CIFAR_CLASSES=10):
+ rcc_iters = 50
+ rad_var = np.zeros((rcc_iters, batch_size, CIFAR_CLASSES - 1))
+ for i in range(rcc_iters):
+ bernoulli_num = np.random.binomial(size=batch_size, n=1, p=0.5)
+ bernoulli_map = np.array([])
+ ones = np.ones((CIFAR_CLASSES - 1, 1))
+ for batch_id in range(batch_size):
+ num = bernoulli_num[batch_id]
+ var_id = 2 * ones * num - 1
+ bernoulli_map = np.append(bernoulli_map, var_id)
+ rad_var[i] = bernoulli_map.reshape((batch_size, CIFAR_CLASSES - 1))
+ return rad_var.astype('float32')
+
+
+def preprocess(sample, is_training, args):
+ image_array = sample.reshape(3, image_size, image_size)
+ rgb_array = np.transpose(image_array, (1, 2, 0))
+ img = Image.fromarray(rgb_array, 'RGB')
+
+ if is_training:
+ # pad and ramdom crop
+ img = ImageOps.expand(img, (4, 4, 4, 4), fill=0) # pad to 40 * 40 * 3
+ left_top = np.random.randint(9, size=2) # rand 0 - 8
+ img = img.crop((left_top[0], left_top[1], left_top[0] + image_size,
+ left_top[1] + image_size))
+ if np.random.randint(2):
+ img = img.transpose(Image.FLIP_LEFT_RIGHT)
+
+ img = np.array(img).astype(np.float32)
+
+ # per_image_standardization
+ img_float = img / 255.0
+ img = (img_float - CIFAR_MEAN) / CIFAR_STD
+
+ if is_training and args.cutout:
+ center = np.random.randint(image_size, size=2)
+ offset_width = max(0, center[0] - half_length)
+ offset_height = max(0, center[1] - half_length)
+ target_width = min(center[0] + half_length, image_size)
+ target_height = min(center[1] + half_length, image_size)
+
+ for i in range(offset_height, target_height):
+ for j in range(offset_width, target_width):
+ img[i][j][:] = 0.0
+
+ img = np.transpose(img, (2, 0, 1))
+ return img
+
+
+def reader_creator_filepath(filename, sub_name, is_training, args):
+ files = os.listdir(filename)
+ names = [each_item for each_item in files if sub_name in each_item]
+ names.sort()
+ datasets = []
+ for name in names:
+ print("Reading file " + name)
+ batch = cPickle.load(open(filename + name, 'rb'))
+ data = batch['data']
+ labels = batch.get('labels', batch.get('fine_labels', None))
+ assert labels is not None
+ dataset = zip(data, labels)
+ datasets.extend(dataset)
+ random.shuffle(datasets)
+
+ def read_batch(datasets, args):
+ for sample, label in datasets:
+ im = preprocess(sample, is_training, args)
+ yield im, [int(label)]
+
+ def reader():
+ batch_data = []
+ batch_label = []
+ for data, label in read_batch(datasets, args):
+ batch_data.append(data)
+ batch_label.append(label)
+ if len(batch_data) == args.batch_size:
+ batch_data = np.array(batch_data, dtype='float32')
+ batch_label = np.array(batch_label, dtype='int64')
+ if is_training:
+ flatten_label, flatten_non_label = \
+ generate_reshape_label(batch_label, args.batch_size)
+ rad_var = generate_bernoulli_number(args.batch_size)
+ mixed_x, y_a, y_b, lam = utils.mixup_data(
+ batch_data, batch_label, args.batch_size,
+ args.mix_alpha)
+ batch_out = [[mixed_x, y_a, y_b, lam, flatten_label, \
+ flatten_non_label, rad_var]]
+ yield batch_out
+ else:
+ batch_out = [[batch_data, batch_label]]
+ yield batch_out
+ batch_data = []
+ batch_label = []
+
+ return reader
+
+
+def train10(args):
+ """
+ CIFAR-10 training set creator.
+ It returns a reader creator, each sample in the reader is image pixels in
+ [0, 1] and label in [0, 9].
+ :return: Training reader creator
+ :rtype: callable
+ """
+
+ return reader_creator_filepath(args.data, 'data_batch', True, args)
+
+
+def test10(args):
+ """
+ CIFAR-10 test set creator.
+ It returns a reader creator, each sample in the reader is image pixels in
+ [0, 1] and label in [0, 9].
+ :return: Test reader creator.
+ :rtype: callable
+ """
+ return reader_creator_filepath(args.data, 'test_batch', False, args)
diff --git a/fluid/AutoDL/LRC/run.sh b/fluid/AutoDL/LRC/run.sh
new file mode 100644
index 0000000000000000000000000000000000000000..9f1a045d49789c3e9aebbc2a73b84b11da471b5a
--- /dev/null
+++ b/fluid/AutoDL/LRC/run.sh
@@ -0,0 +1,8 @@
+CUDA_VISIBLE_DEVICES=0 python -u train_mixup.py \
+--batch_size=80 \
+--auxiliary \
+--weight_decay=0.0003 \
+--learning_rate=0.025 \
+--lrc_loss_lambda=0.7 \
+--cutout
+
diff --git a/fluid/AutoDL/LRC/train_mixup.py b/fluid/AutoDL/LRC/train_mixup.py
new file mode 100644
index 0000000000000000000000000000000000000000..de752c84bcf9276aa83540d60370517e66c0704f
--- /dev/null
+++ b/fluid/AutoDL/LRC/train_mixup.py
@@ -0,0 +1,247 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from learning_rate import cosine_decay
+import numpy as np
+import argparse
+from model import NetworkCIFAR as Network
+import reader
+import sys
+import os
+import time
+import logging
+import genotypes
+import paddle.fluid as fluid
+import shutil
+import utils
+import cPickle as cp
+
+parser = argparse.ArgumentParser("cifar")
+parser.add_argument(
+ '--data',
+ type=str,
+ default='./dataset/cifar/cifar-10-batches-py/',
+ help='location of the data corpus')
+parser.add_argument('--batch_size', type=int, default=96, help='batch size')
+parser.add_argument(
+ '--learning_rate', type=float, default=0.025, help='init learning rate')
+parser.add_argument('--momentum', type=float, default=0.9, help='momentum')
+parser.add_argument(
+ '--weight_decay', type=float, default=3e-4, help='weight decay')
+parser.add_argument(
+ '--report_freq', type=float, default=50, help='report frequency')
+parser.add_argument(
+ '--epochs', type=int, default=600, help='num of training epochs')
+parser.add_argument(
+ '--init_channels', type=int, default=36, help='num of init channels')
+parser.add_argument(
+ '--layers', type=int, default=20, help='total number of layers')
+parser.add_argument(
+ '--model_path',
+ type=str,
+ default='saved_models',
+ help='path to save the model')
+parser.add_argument(
+ '--auxiliary',
+ action='store_true',
+ default=False,
+ help='use auxiliary tower')
+parser.add_argument(
+ '--auxiliary_weight',
+ type=float,
+ default=0.4,
+ help='weight for auxiliary loss')
+parser.add_argument(
+ '--cutout', action='store_true', default=False, help='use cutout')
+parser.add_argument(
+ '--cutout_length', type=int, default=16, help='cutout length')
+parser.add_argument(
+ '--drop_path_prob', type=float, default=0.2, help='drop path probability')
+parser.add_argument('--save', type=str, default='EXP', help='experiment name')
+parser.add_argument(
+ '--arch', type=str, default='DARTS', help='which architecture to use')
+parser.add_argument(
+ '--grad_clip', type=float, default=5, help='gradient clipping')
+parser.add_argument(
+ '--lr_exp_decay',
+ action='store_true',
+ default=False,
+ help='use exponential_decay learning_rate')
+parser.add_argument('--mix_alpha', type=float, default=0.5, help='mixup alpha')
+parser.add_argument(
+ '--lrc_loss_lambda', default=0, type=float, help='lrc_loss_lambda')
+parser.add_argument(
+ '--loss_type',
+ default=1,
+ type=float,
+ help='loss_type 0: cross entropy 1: multi margin loss 2: max margin loss')
+
+args = parser.parse_args()
+
+CIFAR_CLASSES = 10
+dataset_train_size = 50000
+image_size = 32
+
+
+def main():
+ image_shape = [3, image_size, image_size]
+ devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
+ devices_num = len(devices.split(","))
+ logging.info("args = %s", args)
+ genotype = eval("genotypes.%s" % args.arch)
+ model = Network(args.init_channels, CIFAR_CLASSES, args.layers,
+ args.auxiliary, genotype)
+ steps_one_epoch = dataset_train_size / (devices_num * args.batch_size)
+ train(model, args, image_shape, steps_one_epoch)
+
+
+def build_program(main_prog, startup_prog, args, is_train, model, im_shape,
+ steps_one_epoch):
+ out = []
+ with fluid.program_guard(main_prog, startup_prog):
+ py_reader = model.build_input(im_shape, args.batch_size, is_train)
+ if is_train:
+ with fluid.unique_name.guard():
+ loss = model.train_model(py_reader, args.init_channels,
+ args.auxiliary, args.auxiliary_weight,
+ args.batch_size, args.lrc_loss_lambda)
+ optimizer = fluid.optimizer.Momentum(
+ learning_rate=cosine_decay(args.learning_rate, \
+ args.epochs, steps_one_epoch),
+ regularization=fluid.regularizer.L2Decay(\
+ args.weight_decay),
+ momentum=args.momentum)
+ optimizer.minimize(loss)
+ out = [py_reader, loss]
+ else:
+ with fluid.unique_name.guard():
+ loss, acc_1, acc_5 = model.test_model(py_reader,
+ args.init_channels)
+ out = [py_reader, loss, acc_1, acc_5]
+ return out
+
+
+def train(model, args, im_shape, steps_one_epoch):
+ train_startup_prog = fluid.Program()
+ test_startup_prog = fluid.Program()
+ train_prog = fluid.Program()
+ test_prog = fluid.Program()
+
+ train_py_reader, loss_train = build_program(train_prog, train_startup_prog,
+ args, True, model, im_shape,
+ steps_one_epoch)
+
+ test_py_reader, loss_test, acc_1, acc_5 = build_program(
+ test_prog, test_startup_prog, args, False, model, im_shape,
+ steps_one_epoch)
+
+ test_prog = test_prog.clone(for_test=True)
+
+ place = fluid.CUDAPlace(0)
+ exe = fluid.Executor(place)
+ exe.run(train_startup_prog)
+ exe.run(test_startup_prog)
+
+ exec_strategy = fluid.ExecutionStrategy()
+ exec_strategy.num_threads = 1
+ train_exe = fluid.ParallelExecutor(
+ main_program=train_prog,
+ use_cuda=True,
+ loss_name=loss_train.name,
+ exec_strategy=exec_strategy)
+ train_reader = reader.train10(args)
+ test_reader = reader.test10(args)
+ train_py_reader.decorate_paddle_reader(train_reader)
+ test_py_reader.decorate_paddle_reader(test_reader)
+
+ fluid.clip.set_gradient_clip(fluid.clip.GradientClipByNorm(args.grad_clip))
+ fluid.memory_optimize(fluid.default_main_program())
+
+ def save_model(postfix, main_prog):
+ model_path = os.path.join(args.model_path, postfix)
+ if os.path.isdir(model_path):
+ shutil.rmtree(model_path)
+ fluid.io.save_persistables(exe, model_path, main_program=main_prog)
+
+ def test(epoch_id):
+ test_fetch_list = [loss_test, acc_1, acc_5]
+ objs = utils.AvgrageMeter()
+ top1 = utils.AvgrageMeter()
+ top5 = utils.AvgrageMeter()
+ test_py_reader.start()
+ test_start_time = time.time()
+ step_id = 0
+ try:
+ while True:
+ prev_test_start_time = test_start_time
+ test_start_time = time.time()
+ loss_test_v, acc_1_v, acc_5_v = exe.run(
+ test_prog, fetch_list=test_fetch_list)
+ objs.update(np.array(loss_test_v), args.batch_size)
+ top1.update(np.array(acc_1_v), args.batch_size)
+ top5.update(np.array(acc_5_v), args.batch_size)
+ if step_id % args.report_freq == 0:
+ print("Epoch {}, Step {}, acc_1 {}, acc_5 {}, time {}".
+ format(epoch_id, step_id,
+ np.array(acc_1_v),
+ np.array(acc_5_v), test_start_time -
+ prev_test_start_time))
+ step_id += 1
+ except fluid.core.EOFException:
+ test_py_reader.reset()
+ print("Epoch {0}, top1 {1}, top5 {2}".format(epoch_id, top1.avg,
+ top5.avg))
+
+ train_fetch_list = [loss_train]
+ epoch_start_time = time.time()
+ for epoch_id in range(args.epochs):
+ model.drop_path_prob = args.drop_path_prob * epoch_id / args.epochs
+ train_py_reader.start()
+ epoch_end_time = time.time()
+ if epoch_id > 0:
+ print("Epoch {}, total time {}".format(epoch_id - 1, epoch_end_time
+ - epoch_start_time))
+ epoch_start_time = epoch_end_time
+ epoch_end_time
+ start_time = time.time()
+ step_id = 0
+ try:
+ while True:
+ prev_start_time = start_time
+ start_time = time.time()
+ loss_v, = train_exe.run(
+ fetch_list=[v.name for v in train_fetch_list])
+ print("Epoch {}, Step {}, loss {}, time {}".format(epoch_id, step_id, \
+ np.array(loss_v).mean(), start_time-prev_start_time))
+ step_id += 1
+ sys.stdout.flush()
+ except fluid.core.EOFException:
+ train_py_reader.reset()
+ if epoch_id % 50 == 0 or epoch_id == args.epochs - 1:
+ save_model(str(epoch_id), train_prog)
+ test(epoch_id)
+
+
+if __name__ == '__main__':
+ main()
diff --git a/fluid/AutoDL/LRC/utils.py b/fluid/AutoDL/LRC/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4002b57c6e91f9a4f7992156c4fa07f9e55d628c
--- /dev/null
+++ b/fluid/AutoDL/LRC/utils.py
@@ -0,0 +1,55 @@
+# Copyright (c) 2019 PaddlePaddle 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.
+#
+# Based on:
+# --------------------------------------------------------
+# DARTS
+# Copyright (c) 2018, Hanxiao Liu.
+# Licensed under the Apache License, Version 2.0;
+# --------------------------------------------------------
+
+import os
+import sys
+import time
+import math
+import numpy as np
+
+
+def mixup_data(x, y, batch_size, alpha=1.0):
+ '''Compute the mixup data. Return mixed inputs, pairs of targets, and lambda'''
+ if alpha > 0.:
+ lam = np.random.beta(alpha, alpha)
+ else:
+ lam = 1.
+ index = np.random.permutation(batch_size)
+
+ mixed_x = lam * x + (1 - lam) * x[index, :]
+ y_a, y_b = y, y[index]
+ return mixed_x.astype('float32'), y_a.astype('int64'),\
+ y_b.astype('int64'), np.array(lam, dtype='float32')
+
+
+class AvgrageMeter(object):
+ def __init__(self):
+ self.reset()
+
+ def reset(self):
+ self.avg = 0
+ self.sum = 0
+ self.cnt = 0
+
+ def update(self, val, n=1):
+ self.sum += val * n
+ self.cnt += n
+ self.avg = self.sum / self.cnt
diff --git a/fluid/DeepQNetwork/DQN_agent.py b/fluid/DeepQNetwork/DQN_agent.py
index 67eb3ce6a29bb723b481d6b1c2f517f037d52942..5b474325f656533b91965fd59d70c2d421e16fc3 100644
--- a/fluid/DeepQNetwork/DQN_agent.py
+++ b/fluid/DeepQNetwork/DQN_agent.py
@@ -1,11 +1,10 @@
#-*- coding: utf-8 -*-
+import math
+import numpy as np
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
-import numpy as np
-import math
from tqdm import tqdm
-from utils import fluid_flatten
class DQNModel(object):
@@ -39,34 +38,51 @@ class DQNModel(object):
name='isOver', shape=[], dtype='bool')
def _build_net(self):
- state, action, reward, next_s, isOver = self._get_inputs()
- self.pred_value = self.get_DQN_prediction(state)
- self.predict_program = fluid.default_main_program().clone()
+ self.predict_program = fluid.Program()
+ self.train_program = fluid.Program()
+ self._sync_program = fluid.Program()
- reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
+ with fluid.program_guard(self.predict_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ self.pred_value = self.get_DQN_prediction(state)
- action_onehot = fluid.layers.one_hot(action, self.action_dim)
- action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
+ with fluid.program_guard(self.train_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ pred_value = self.get_DQN_prediction(state)
- pred_action_value = fluid.layers.reduce_sum(
- fluid.layers.elementwise_mul(action_onehot, self.pred_value), dim=1)
+ reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
- targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
- best_v = fluid.layers.reduce_max(targetQ_predict_value, dim=1)
- best_v.stop_gradient = True
+ action_onehot = fluid.layers.one_hot(action, self.action_dim)
+ action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
- target = reward + (1.0 - fluid.layers.cast(
- isOver, dtype='float32')) * self.gamma * best_v
- cost = fluid.layers.square_error_cost(pred_action_value, target)
- cost = fluid.layers.reduce_mean(cost)
+ pred_action_value = fluid.layers.reduce_sum(
+ fluid.layers.elementwise_mul(action_onehot, pred_value), dim=1)
- self._sync_program = self._build_sync_target_network()
+ targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
+ best_v = fluid.layers.reduce_max(targetQ_predict_value, dim=1)
+ best_v.stop_gradient = True
- optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
- optimizer.minimize(cost)
+ target = reward + (1.0 - fluid.layers.cast(
+ isOver, dtype='float32')) * self.gamma * best_v
+ cost = fluid.layers.square_error_cost(pred_action_value, target)
+ cost = fluid.layers.reduce_mean(cost)
- # define program
- self.train_program = fluid.default_main_program()
+ optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
+ optimizer.minimize(cost)
+
+ vars = list(self.train_program.list_vars())
+ policy_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
+ target_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
+ policy_vars.sort(key=lambda x: x.name)
+ target_vars.sort(key=lambda x: x.name)
+
+ with fluid.program_guard(self._sync_program):
+ sync_ops = []
+ for i, var in enumerate(policy_vars):
+ sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
+ sync_ops.append(sync_op)
# fluid exe
place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
@@ -81,50 +97,50 @@ class DQNModel(object):
conv1 = fluid.layers.conv2d(
input=image,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv1'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv1_b'.format(variable_field)))
max_pool1 = fluid.layers.pool2d(
- input=conv1, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv1, pool_size=2, pool_stride=2, pool_type='max')
conv2 = fluid.layers.conv2d(
input=max_pool1,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv2'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv2_b'.format(variable_field)))
max_pool2 = fluid.layers.pool2d(
- input=conv2, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv2, pool_size=2, pool_stride=2, pool_type='max')
conv3 = fluid.layers.conv2d(
input=max_pool2,
num_filters=64,
- filter_size=[4, 4],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=4,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv3'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv3_b'.format(variable_field)))
max_pool3 = fluid.layers.pool2d(
- input=conv3, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv3, pool_size=2, pool_stride=2, pool_type='max')
conv4 = fluid.layers.conv2d(
input=max_pool3,
num_filters=64,
- filter_size=[3, 3],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=3,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv4'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv4_b'.format(variable_field)))
- flatten = fluid_flatten(conv4)
+ flatten = fluid.layers.flatten(conv4, axis=1)
out = fluid.layers.fc(
input=flatten,
@@ -133,23 +149,6 @@ class DQNModel(object):
bias_attr=ParamAttr(name='{}_fc1_b'.format(variable_field)))
return out
- def _build_sync_target_network(self):
- vars = list(fluid.default_main_program().list_vars())
- policy_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
- target_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
- policy_vars.sort(key=lambda x: x.name)
- target_vars.sort(key=lambda x: x.name)
-
- sync_program = fluid.default_main_program().clone()
- with fluid.program_guard(sync_program):
- sync_ops = []
- for i, var in enumerate(policy_vars):
- sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
- sync_ops.append(sync_op)
- sync_program = sync_program.prune(sync_ops)
- return sync_program
def act(self, state, train_or_test):
sample = np.random.random()
diff --git a/fluid/DeepQNetwork/DoubleDQN_agent.py b/fluid/DeepQNetwork/DoubleDQN_agent.py
index 09b4b2119bab3fbdfa9bb9cfb8fae40fa34f87e1..c95ae5632fd2e904a625f680f4a9147d5615b765 100644
--- a/fluid/DeepQNetwork/DoubleDQN_agent.py
+++ b/fluid/DeepQNetwork/DoubleDQN_agent.py
@@ -1,11 +1,10 @@
#-*- coding: utf-8 -*-
+import math
+import numpy as np
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
-import numpy as np
from tqdm import tqdm
-import math
-from utils import fluid_argmax, fluid_flatten
class DoubleDQNModel(object):
@@ -39,41 +38,59 @@ class DoubleDQNModel(object):
name='isOver', shape=[], dtype='bool')
def _build_net(self):
- state, action, reward, next_s, isOver = self._get_inputs()
- self.pred_value = self.get_DQN_prediction(state)
- self.predict_program = fluid.default_main_program().clone()
+ self.predict_program = fluid.Program()
+ self.train_program = fluid.Program()
+ self._sync_program = fluid.Program()
- reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
+ with fluid.program_guard(self.predict_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ self.pred_value = self.get_DQN_prediction(state)
- action_onehot = fluid.layers.one_hot(action, self.action_dim)
- action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
+ with fluid.program_guard(self.train_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ pred_value = self.get_DQN_prediction(state)
- pred_action_value = fluid.layers.reduce_sum(
- fluid.layers.elementwise_mul(action_onehot, self.pred_value), dim=1)
+ reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
- targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
+ action_onehot = fluid.layers.one_hot(action, self.action_dim)
+ action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
- next_s_predcit_value = self.get_DQN_prediction(next_s)
- greedy_action = fluid_argmax(next_s_predcit_value)
+ pred_action_value = fluid.layers.reduce_sum(
+ fluid.layers.elementwise_mul(action_onehot, pred_value), dim=1)
- predict_onehot = fluid.layers.one_hot(greedy_action, self.action_dim)
- best_v = fluid.layers.reduce_sum(
- fluid.layers.elementwise_mul(predict_onehot, targetQ_predict_value),
- dim=1)
- best_v.stop_gradient = True
+ targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
- target = reward + (1.0 - fluid.layers.cast(
- isOver, dtype='float32')) * self.gamma * best_v
- cost = fluid.layers.square_error_cost(pred_action_value, target)
- cost = fluid.layers.reduce_mean(cost)
+ next_s_predcit_value = self.get_DQN_prediction(next_s)
+ greedy_action = fluid.layers.argmax(next_s_predcit_value, axis=1)
+ greedy_action = fluid.layers.unsqueeze(greedy_action, axes=[1])
- self._sync_program = self._build_sync_target_network()
+ predict_onehot = fluid.layers.one_hot(greedy_action, self.action_dim)
+ best_v = fluid.layers.reduce_sum(
+ fluid.layers.elementwise_mul(predict_onehot, targetQ_predict_value),
+ dim=1)
+ best_v.stop_gradient = True
- optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
- optimizer.minimize(cost)
+ target = reward + (1.0 - fluid.layers.cast(
+ isOver, dtype='float32')) * self.gamma * best_v
+ cost = fluid.layers.square_error_cost(pred_action_value, target)
+ cost = fluid.layers.reduce_mean(cost)
- # define program
- self.train_program = fluid.default_main_program()
+ optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
+ optimizer.minimize(cost)
+
+ vars = list(self.train_program.list_vars())
+ policy_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
+ target_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
+ policy_vars.sort(key=lambda x: x.name)
+ target_vars.sort(key=lambda x: x.name)
+
+ with fluid.program_guard(self._sync_program):
+ sync_ops = []
+ for i, var in enumerate(policy_vars):
+ sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
+ sync_ops.append(sync_op)
# fluid exe
place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
@@ -88,50 +105,50 @@ class DoubleDQNModel(object):
conv1 = fluid.layers.conv2d(
input=image,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv1'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv1_b'.format(variable_field)))
max_pool1 = fluid.layers.pool2d(
- input=conv1, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv1, pool_size=2, pool_stride=2, pool_type='max')
conv2 = fluid.layers.conv2d(
input=max_pool1,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv2'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv2_b'.format(variable_field)))
max_pool2 = fluid.layers.pool2d(
- input=conv2, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv2, pool_size=2, pool_stride=2, pool_type='max')
conv3 = fluid.layers.conv2d(
input=max_pool2,
num_filters=64,
- filter_size=[4, 4],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=4,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv3'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv3_b'.format(variable_field)))
max_pool3 = fluid.layers.pool2d(
- input=conv3, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv3, pool_size=2, pool_stride=2, pool_type='max')
conv4 = fluid.layers.conv2d(
input=max_pool3,
num_filters=64,
- filter_size=[3, 3],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=3,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv4'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv4_b'.format(variable_field)))
- flatten = fluid_flatten(conv4)
+ flatten = fluid.layers.flatten(conv4, axis=1)
out = fluid.layers.fc(
input=flatten,
@@ -140,23 +157,6 @@ class DoubleDQNModel(object):
bias_attr=ParamAttr(name='{}_fc1_b'.format(variable_field)))
return out
- def _build_sync_target_network(self):
- vars = list(fluid.default_main_program().list_vars())
- policy_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
- target_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
- policy_vars.sort(key=lambda x: x.name)
- target_vars.sort(key=lambda x: x.name)
-
- sync_program = fluid.default_main_program().clone()
- with fluid.program_guard(sync_program):
- sync_ops = []
- for i, var in enumerate(policy_vars):
- sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
- sync_ops.append(sync_op)
- sync_program = sync_program.prune(sync_ops)
- return sync_program
def act(self, state, train_or_test):
sample = np.random.random()
diff --git a/fluid/DeepQNetwork/DuelingDQN_agent.py b/fluid/DeepQNetwork/DuelingDQN_agent.py
index 271a767b7b5841cf1abe213fc477859e3cf5dd05..cf2ff71bb811e5dce62be78beab1f0afb05d31f9 100644
--- a/fluid/DeepQNetwork/DuelingDQN_agent.py
+++ b/fluid/DeepQNetwork/DuelingDQN_agent.py
@@ -1,11 +1,10 @@
#-*- coding: utf-8 -*-
+import math
+import numpy as np
import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
-import numpy as np
from tqdm import tqdm
-import math
-from utils import fluid_flatten
class DuelingDQNModel(object):
@@ -39,34 +38,51 @@ class DuelingDQNModel(object):
name='isOver', shape=[], dtype='bool')
def _build_net(self):
- state, action, reward, next_s, isOver = self._get_inputs()
- self.pred_value = self.get_DQN_prediction(state)
- self.predict_program = fluid.default_main_program().clone()
+ self.predict_program = fluid.Program()
+ self.train_program = fluid.Program()
+ self._sync_program = fluid.Program()
- reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
+ with fluid.program_guard(self.predict_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ self.pred_value = self.get_DQN_prediction(state)
- action_onehot = fluid.layers.one_hot(action, self.action_dim)
- action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
+ with fluid.program_guard(self.train_program):
+ state, action, reward, next_s, isOver = self._get_inputs()
+ pred_value = self.get_DQN_prediction(state)
- pred_action_value = fluid.layers.reduce_sum(
- fluid.layers.elementwise_mul(action_onehot, self.pred_value), dim=1)
+ reward = fluid.layers.clip(reward, min=-1.0, max=1.0)
- targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
- best_v = fluid.layers.reduce_max(targetQ_predict_value, dim=1)
- best_v.stop_gradient = True
+ action_onehot = fluid.layers.one_hot(action, self.action_dim)
+ action_onehot = fluid.layers.cast(action_onehot, dtype='float32')
- target = reward + (1.0 - fluid.layers.cast(
- isOver, dtype='float32')) * self.gamma * best_v
- cost = fluid.layers.square_error_cost(pred_action_value, target)
- cost = fluid.layers.reduce_mean(cost)
+ pred_action_value = fluid.layers.reduce_sum(
+ fluid.layers.elementwise_mul(action_onehot, pred_value), dim=1)
- self._sync_program = self._build_sync_target_network()
+ targetQ_predict_value = self.get_DQN_prediction(next_s, target=True)
+ best_v = fluid.layers.reduce_max(targetQ_predict_value, dim=1)
+ best_v.stop_gradient = True
- optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
- optimizer.minimize(cost)
+ target = reward + (1.0 - fluid.layers.cast(
+ isOver, dtype='float32')) * self.gamma * best_v
+ cost = fluid.layers.square_error_cost(pred_action_value, target)
+ cost = fluid.layers.reduce_mean(cost)
- # define program
- self.train_program = fluid.default_main_program()
+ optimizer = fluid.optimizer.Adam(1e-3 * 0.5, epsilon=1e-3)
+ optimizer.minimize(cost)
+
+ vars = list(self.train_program.list_vars())
+ policy_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
+ target_vars = list(filter(
+ lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
+ policy_vars.sort(key=lambda x: x.name)
+ target_vars.sort(key=lambda x: x.name)
+
+ with fluid.program_guard(self._sync_program):
+ sync_ops = []
+ for i, var in enumerate(policy_vars):
+ sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
+ sync_ops.append(sync_op)
# fluid exe
place = fluid.CUDAPlace(0) if self.use_cuda else fluid.CPUPlace()
@@ -81,50 +97,50 @@ class DuelingDQNModel(object):
conv1 = fluid.layers.conv2d(
input=image,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv1'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv1_b'.format(variable_field)))
max_pool1 = fluid.layers.pool2d(
- input=conv1, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv1, pool_size=2, pool_stride=2, pool_type='max')
conv2 = fluid.layers.conv2d(
input=max_pool1,
num_filters=32,
- filter_size=[5, 5],
- stride=[1, 1],
- padding=[2, 2],
+ filter_size=5,
+ stride=1,
+ padding=2,
act='relu',
param_attr=ParamAttr(name='{}_conv2'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv2_b'.format(variable_field)))
max_pool2 = fluid.layers.pool2d(
- input=conv2, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv2, pool_size=2, pool_stride=2, pool_type='max')
conv3 = fluid.layers.conv2d(
input=max_pool2,
num_filters=64,
- filter_size=[4, 4],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=4,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv3'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv3_b'.format(variable_field)))
max_pool3 = fluid.layers.pool2d(
- input=conv3, pool_size=[2, 2], pool_stride=[2, 2], pool_type='max')
+ input=conv3, pool_size=2, pool_stride=2, pool_type='max')
conv4 = fluid.layers.conv2d(
input=max_pool3,
num_filters=64,
- filter_size=[3, 3],
- stride=[1, 1],
- padding=[1, 1],
+ filter_size=3,
+ stride=1,
+ padding=1,
act='relu',
param_attr=ParamAttr(name='{}_conv4'.format(variable_field)),
bias_attr=ParamAttr(name='{}_conv4_b'.format(variable_field)))
- flatten = fluid_flatten(conv4)
+ flatten = fluid.layers.flatten(conv4, axis=1)
value = fluid.layers.fc(
input=flatten,
@@ -143,24 +159,6 @@ class DuelingDQNModel(object):
advantage, dim=1, keep_dim=True))
return Q
- def _build_sync_target_network(self):
- vars = list(fluid.default_main_program().list_vars())
- policy_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'policy' in x.name, vars))
- target_vars = list(filter(
- lambda x: 'GRAD' not in x.name and 'target' in x.name, vars))
- policy_vars.sort(key=lambda x: x.name)
- target_vars.sort(key=lambda x: x.name)
-
- sync_program = fluid.default_main_program().clone()
- with fluid.program_guard(sync_program):
- sync_ops = []
- for i, var in enumerate(policy_vars):
- sync_op = fluid.layers.assign(policy_vars[i], target_vars[i])
- sync_ops.append(sync_op)
- # The prune API is deprecated, please don't use it any more.
- sync_program = sync_program._prune(sync_ops)
- return sync_program
def act(self, state, train_or_test):
sample = np.random.random()
@@ -186,12 +184,14 @@ class DuelingDQNModel(object):
self.global_step += 1
action = np.expand_dims(action, -1)
- self.exe.run(self.train_program, \
- feed={'state': state.astype('float32'), \
- 'action': action.astype('int32'), \
- 'reward': reward, \
- 'next_s': next_state.astype('float32'), \
- 'isOver': isOver})
+ self.exe.run(self.train_program,
+ feed={
+ 'state': state.astype('float32'),
+ 'action': action.astype('int32'),
+ 'reward': reward,
+ 'next_s': next_state.astype('float32'),
+ 'isOver': isOver
+ })
def sync_target_network(self):
self.exe.run(self._sync_program)
diff --git a/fluid/DeepQNetwork/README.md b/fluid/DeepQNetwork/README.md
index e72920bcad29ce7ffd78bfb90a1406654298248d..1edeaaa884318ec3a530ec4fdb7d031d07411b56 100644
--- a/fluid/DeepQNetwork/README.md
+++ b/fluid/DeepQNetwork/README.md
@@ -29,7 +29,7 @@ The average game rewards that can be obtained for the three models as the number
+ gym
+ tqdm
+ opencv-python
-+ paddlepaddle-gpu>=0.12.0
++ paddlepaddle-gpu>=1.0.0
+ ale_python_interface
### Install Dependencies:
diff --git a/fluid/DeepQNetwork/README_cn.md b/fluid/DeepQNetwork/README_cn.md
index 68a65bffe8fab79ce563fefc894dd035c1572065..640d775ad8fed2be360d308b6c5df41c86d77c04 100644
--- a/fluid/DeepQNetwork/README_cn.md
+++ b/fluid/DeepQNetwork/README_cn.md
@@ -28,7 +28,7 @@
+ gym
+ tqdm
+ opencv-python
-+ paddlepaddle-gpu>=0.12.0
++ paddlepaddle-gpu>=1.0.0
+ ale_python_interface
### 下载依赖：
diff --git a/fluid/DeepQNetwork/utils.py b/fluid/DeepQNetwork/utils.py
deleted file mode 100644
index 26ed7fbdb54494c3cf9a983f8ecafdfbcd4d2719..0000000000000000000000000000000000000000
--- a/fluid/DeepQNetwork/utils.py
+++ /dev/null
@@ -1,20 +0,0 @@
-#-*- coding: utf-8 -*-
-#File: utils.py
-
-import paddle.fluid as fluid
-import numpy as np
-
-
-def fluid_argmax(x):
- """
- Get index of max value for the last dimension
- """
- _, max_index = fluid.layers.topk(x, k=1)
- return max_index
-
-
-def fluid_flatten(x):
- """
- Flatten fluid variable along the first dimension
- """
- return fluid.layers.reshape(x, shape=[-1, np.prod(x.shape[1:])])
diff --git a/fluid/PaddleCV/deeplabv3+/README.md b/fluid/PaddleCV/deeplabv3+/README.md
index 97e1600db9ff6e2f9de2a254681a2eacb2f9359b..b9990a20845f637eb3611f4875213a657b4491d8 100644
--- a/fluid/PaddleCV/deeplabv3+/README.md
+++ b/fluid/PaddleCV/deeplabv3+/README.md
@@ -1,4 +1,4 @@
-DeepLab运行本目录下的程序示例需要使用PaddlePaddle Fluid v1.0.0版本或以上。如果您的PaddlePaddle安装版本低于此要求，请按照安装文档中的说明更新PaddlePaddle安装版本，如果使用GPU，该程序需要使用cuDNN v7版本。
+DeepLab运行本目录下的程序示例需要使用PaddlePaddle Fluid v1.3.0版本或以上。如果您的PaddlePaddle安装版本低于此要求，请按照安装文档中的说明更新PaddlePaddle安装版本，如果使用GPU，该程序需要使用cuDNN v7版本。
## 代码结构
@@ -38,15 +38,16 @@ data/cityscape/
# 预训练模型准备
+我们为了节约更多的显存，在这里我们使用Group Norm作为我们的归一化手段。
如果需要从头开始训练模型，用户需要下载我们的初始化模型
```
-wget http://paddlemodels.cdn.bcebos.com/deeplab/deeplabv3plus_xception65_initialize.tar.gz
-tar -xf deeplabv3plus_xception65_initialize.tar.gz && rm deeplabv3plus_xception65_initialize.tar.gz
+wget https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn_init.tgz
+tar -xf deeplabv3plus_gn_init.tgz && rm deeplabv3plus_gn_init.tgz
```
如果需要最终训练模型进行fine tune或者直接用于预测，请下载我们的最终模型
```
-wget http://paddlemodels.cdn.bcebos.com/deeplab/deeplabv3plus.tar.gz
-tar -xf deeplabv3plus.tar.gz && rm deeplabv3plus.tar.gz
+wget https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn.tgz
+tar -xf deeplabv3plus_gn.tgz && rm deeplabv3plus_gn.tgz
```
@@ -59,6 +60,7 @@ python ./train.py \
--batch_size=1 \
--train_crop_size=769 \
--total_step=50 \
+ --norm_type=gn \
--init_weights_path=$INIT_WEIGHTS_PATH \
--save_weights_path=$SAVE_WEIGHTS_PATH \
--dataset_path=$DATASET_PATH
@@ -72,19 +74,25 @@ python train.py --help
```
python ./train.py \
--batch_size=8 \
- --parallel=true \
+ --parallel=True \
+ --norm_type=gn \
--train_crop_size=769 \
--total_step=90000 \
- --init_weights_path=deeplabv3plus_xception65_initialize.params \
- --save_weights_path=output/ \
+ --base_lr=0.001 \
+ --init_weights_path=deeplabv3plus_gn_init \
+ --save_weights_path=output \
--dataset_path=$DATASET_PATH
```
+如果您的显存不足，可以尝试减小`batch_size`，同时等比例放大`total_step`, 保证相乘的值不变，这得益于Group Norm的特性，改变 `batch_size` 并不会显著影响结果，而且能够节约更多显存, 比如您可以设置`--batch_size=4 --total_step=180000`。
+
+如果您希望使用多卡进行训练，可以同比增加`batch_size`，减小`total_step`, 比如原来单卡训练是`--batch_size=4 --total_step=180000`，使用4卡训练则是`--batch_size=16 --total_step=45000`
### 测试
执行以下命令在`Cityscape`测试数据集上进行测试：
```
python ./eval.py \
- --init_weights=deeplabv3plus.params \
+ --init_weights=deeplabv3plus_gn \
+ --norm_type=gn \
--dataset_path=$DATASET_PATH
```
需要通过选项`--model_path`指定模型文件。测试脚本的输出的评估指标为mean IoU。
@@ -93,15 +101,17 @@ python ./eval.py \
## 实验结果
训练完成以后，使用`eval.py`在验证集上进行测试，得到以下结果：
```
-load from: ../models/deeplabv3p
+load from: ../models/deeplabv3plus_gn
total number 500
-step: 500, mIoU: 0.7873
+step: 500, mIoU: 0.7881
```
## 其他信息
-|数据集 | pretrained model | trained model | mean IoU
-|---|---|---|---|
-|CityScape | [deeplabv3plus_xception65_initialize.tar.gz](http://paddlemodels.cdn.bcebos.com/deeplab/deeplabv3plus_xception65_initialize.tar.gz) | [deeplabv3plus.tar.gz](http://paddlemodels.cdn.bcebos.com/deeplab/deeplabv3plus.tar.gz) | 0.7873 |
+
+|数据集 | norm type | pretrained model | trained model | mean IoU
+|---|---|---|---|---|
+|CityScape | batch norm | [deeplabv3plus_xception65_initialize.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_xception65_initialize.tgz) | [deeplabv3plus.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus.tgz) | 0.7873 |
+|CityScape | group norm | [deeplabv3plus_gn_init.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn_init.tgz) | [deeplabv3plus_gn.tgz](https://paddle-deeplab.bj.bcebos.com/deeplabv3plus_gn.tgz) | 0.7881 |
## 参考
diff --git a/fluid/PaddleCV/deeplabv3+/eval.py b/fluid/PaddleCV/deeplabv3+/eval.py
index 5699f2fac3ff52e39932eba71e8d25a189bf8fc6..4620dd5d7cb9c91735619478a550c3a4cc747aa6 100644
--- a/fluid/PaddleCV/deeplabv3+/eval.py
+++ b/fluid/PaddleCV/deeplabv3+/eval.py
@@ -2,7 +2,9 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
-os.environ['FLAGS_fraction_of_gpu_memory_to_use'] = '0.98'
+if 'FLAGS_fraction_of_gpu_memory_to_use' not in os.environ:
+ os.environ['FLAGS_fraction_of_gpu_memory_to_use'] = '0.98'
+os.environ['FLAGS_enable_parallel_graph'] = '1'
import paddle
import paddle.fluid as fluid
@@ -12,21 +14,21 @@ from reader import CityscapeDataset
import reader
import models
import sys
+import utility
+parser = argparse.ArgumentParser()
+add_arg = lambda *args: utility.add_arguments(*args, argparser=parser)
-def add_argument(name, type, default, help):
- parser.add_argument('--' + name, default=default, type=type, help=help)
-
-
-def add_arguments():
- add_argument('total_step', int, -1,
- "Number of the step to be evaluated, -1 for full evaluation.")
- add_argument('init_weights_path', str, None,
- "Path of the weights to evaluate.")
- add_argument('dataset_path', str, None, "Cityscape dataset path.")
- add_argument('verbose', bool, False, "Print mIoU for each step if verbose.")
- add_argument('use_gpu', bool, True, "Whether use GPU or CPU.")
- add_argument('num_classes', int, 19, "Number of classes.")
+# yapf: disable
+add_arg('total_step', int, -1, "Number of the step to be evaluated, -1 for full evaluation.")
+add_arg('init_weights_path', str, None, "Path of the weights to evaluate.")
+add_arg('dataset_path', str, None, "Cityscape dataset path.")
+add_arg('verbose', bool, False, "Print mIoU for each step if verbose.")
+add_arg('use_gpu', bool, True, "Whether use GPU or CPU.")
+add_arg('num_classes', int, 19, "Number of classes.")
+add_arg('use_py_reader', bool, True, "Use py_reader.")
+add_arg('norm_type', str, 'bn', "Normalization type, should be 'bn' or 'gn'.")
+#yapf: enable
def mean_iou(pred, label):
@@ -43,7 +45,7 @@ def mean_iou(pred, label):
def load_model():
- if args.init_weights_path.endswith('/'):
+ if os.path.isdir(args.init_weights_path):
fluid.io.load_params(
exe, dirname=args.init_weights_path, main_program=tp)
else:
@@ -53,13 +55,11 @@ def load_model():
CityscapeDataset = reader.CityscapeDataset
-parser = argparse.ArgumentParser()
-add_arguments()
-
args = parser.parse_args()
models.clean()
models.is_train = False
+models.default_norm_type = args.norm_type
deeplabv3p = models.deeplabv3p
image_shape = [1025, 2049]
@@ -73,8 +73,15 @@ reader.default_config['shuffle'] = False
num_classes = args.num_classes
with fluid.program_guard(tp, sp):
- img = fluid.layers.data(name='img', shape=[3, 0, 0], dtype='float32')
- label = fluid.layers.data(name='label', shape=eval_shape, dtype='int32')
+ if args.use_py_reader:
+ py_reader = fluid.layers.py_reader(capacity=64,
+ shapes=[[1, 3, 0, 0], [1] + eval_shape],
+ dtypes=['float32', 'int32'])
+ img, label = fluid.layers.read_file(py_reader)
+ else:
+ img = fluid.layers.data(name='img', shape=[3, 0, 0], dtype='float32')
+ label = fluid.layers.data(name='label', shape=eval_shape, dtype='int32')
+
img = fluid.layers.resize_bilinear(img, image_shape)
logit = deeplabv3p(img)
logit = fluid.layers.resize_bilinear(logit, eval_shape)
@@ -105,16 +112,25 @@ else:
total_step = args.total_step
batches = dataset.get_batch_generator(batch_size, total_step)
+if args.use_py_reader:
+ py_reader.decorate_tensor_provider(lambda :[ (yield b[1],b[2]) for b in batches])
+ py_reader.start()
sum_iou = 0
all_correct = np.array([0], dtype=np.int64)
all_wrong = np.array([0], dtype=np.int64)
-for i, imgs, labels, names in batches:
- result = exe.run(tp,
- feed={'img': imgs,
- 'label': labels},
- fetch_list=[pred, miou, out_wrong, out_correct])
+for i in range(total_step):
+ if not args.use_py_reader:
+ _, imgs, labels, names = next(batches)
+ result = exe.run(tp,
+ feed={'img': imgs,
+ 'label': labels},
+ fetch_list=[pred, miou, out_wrong, out_correct])
+ else:
+ result = exe.run(tp,
+ fetch_list=[pred, miou, out_wrong, out_correct])
+
wrong = result[2][:-1] + all_wrong
right = result[3][:-1] + all_correct
all_wrong = wrong.copy()
@@ -122,7 +138,6 @@ for i, imgs, labels, names in batches:
mp = (wrong + right) != 0
miou2 = np.mean((right[mp] * 1.0 / (right[mp] + wrong[mp])))
if args.verbose:
- print('step: %s, mIoU: %s' % (i + 1, miou2))
+ print('step: %s, mIoU: %s' % (i + 1, miou2), flush=True)
else:
- print('\rstep: %s, mIoU: %s' % (i + 1, miou2))
- sys.stdout.flush()
+ print('\rstep: %s, mIoU: %s' % (i + 1, miou2), end='\r', flush=True)
diff --git a/fluid/PaddleCV/deeplabv3+/models.py b/fluid/PaddleCV/deeplabv3+/models.py
index c1ea12296af3e9b6e0bb783cfa10efe5adfa15aa..117ab5da539da1a403fb99d8642b3f7b4f864355 100644
--- a/fluid/PaddleCV/deeplabv3+/models.py
+++ b/fluid/PaddleCV/deeplabv3+/models.py
@@ -5,6 +5,7 @@ import paddle
import paddle.fluid as fluid
import contextlib
+import os
name_scope = ""
decode_channel = 48
@@ -146,10 +147,12 @@ def bn_relu(data):
def relu(data):
- return append_op_result(fluid.layers.relu(data), 'relu')
+ return append_op_result(
+ fluid.layers.relu(
+ data, name=name_scope + 'relu'), 'relu')
-def seq_conv(input, channel, stride, filter, dilation=1, act=None):
+def seperate_conv(input, channel, stride, filter, dilation=1, act=None):
with scope('depthwise'):
input = conv(
input,
@@ -187,14 +190,14 @@ def xception_block(input,
with scope('separable_conv' + str(i + 1)):
if not activation_fn_in_separable_conv:
data = relu(data)
- data = seq_conv(
+ data = seperate_conv(
data,
channels[i],
strides[i],
filters[i],
dilation=dilation)
else:
- data = seq_conv(
+ data = seperate_conv(
data,
channels[i],
strides[i],
@@ -273,11 +276,11 @@ def encoder(input):
with scope("aspp0"):
aspp0 = bn_relu(conv(input, channel, 1, 1, groups=1, padding=0))
with scope("aspp1"):
- aspp1 = seq_conv(input, channel, 1, 3, dilation=6, act=relu)
+ aspp1 = seperate_conv(input, channel, 1, 3, dilation=6, act=relu)
with scope("aspp2"):
- aspp2 = seq_conv(input, channel, 1, 3, dilation=12, act=relu)
+ aspp2 = seperate_conv(input, channel, 1, 3, dilation=12, act=relu)
with scope("aspp3"):
- aspp3 = seq_conv(input, channel, 1, 3, dilation=18, act=relu)
+ aspp3 = seperate_conv(input, channel, 1, 3, dilation=18, act=relu)
with scope("concat"):
data = append_op_result(
fluid.layers.concat(
@@ -300,10 +303,10 @@ def decoder(encode_data, decode_shortcut):
[encode_data, decode_shortcut], axis=1)
append_op_result(encode_data, 'concat')
with scope("separable_conv1"):
- encode_data = seq_conv(
+ encode_data = seperate_conv(
encode_data, encode_channel, 1, 3, dilation=1, act=relu)
with scope("separable_conv2"):
- encode_data = seq_conv(
+ encode_data = seperate_conv(
encode_data, encode_channel, 1, 3, dilation=1, act=relu)
return encode_data
diff --git a/fluid/PaddleCV/deeplabv3+/train.py b/fluid/PaddleCV/deeplabv3+/train.py
index e009f76e0e16be9e4a5db532615cefac258fada1..799a1900f5c975fdb6040852b868aaebd4565404 100755
--- a/fluid/PaddleCV/deeplabv3+/train.py
+++ b/fluid/PaddleCV/deeplabv3+/train.py
@@ -2,7 +2,8 @@ from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
-os.environ['FLAGS_fraction_of_gpu_memory_to_use'] = '0.98'
+if 'FLAGS_fraction_of_gpu_memory_to_use' not in os.environ:
+ os.environ['FLAGS_fraction_of_gpu_memory_to_use'] = '0.98'
import paddle
import paddle.fluid as fluid
@@ -12,105 +13,94 @@ from reader import CityscapeDataset
import reader
import models
import time
+import contextlib
+import paddle.fluid.profiler as profiler
+import utility
-
-def add_argument(name, type, default, help):
- parser.add_argument('--' + name, default=default, type=type, help=help)
-
-
-def add_arguments():
- add_argument('batch_size', int, 2,
- "The number of images in each batch during training.")
- add_argument('train_crop_size', int, 769,
- "'Image crop size during training.")
- add_argument('base_lr', float, 0.0001,
- "The base learning rate for model training.")
- add_argument('total_step', int, 90000, "Number of the training step.")
- add_argument('init_weights_path', str, None,
- "Path of the initial weights in paddlepaddle format.")
- add_argument('save_weights_path', str, None,
- "Path of the saved weights during training.")
- add_argument('dataset_path', str, None, "Cityscape dataset path.")
- add_argument('parallel', bool, False, "using ParallelExecutor.")
- add_argument('use_gpu', bool, True, "Whether use GPU or CPU.")
- add_argument('num_classes', int, 19, "Number of classes.")
- parser.add_argument(
- '--enable_ce',
- action='store_true',
- help='If set, run the task with continuous evaluation logs.')
-
+parser = argparse.ArgumentParser()
+add_arg = lambda *args: utility.add_arguments(*args, argparser=parser)
+
+# yapf: disable
+add_arg('batch_size', int, 2, "The number of images in each batch during training.")
+add_arg('train_crop_size', int, 769, "Image crop size during training.")
+add_arg('base_lr', float, 0.0001, "The base learning rate for model training.")
+add_arg('total_step', int, 90000, "Number of the training step.")
+add_arg('init_weights_path', str, None, "Path of the initial weights in paddlepaddle format.")
+add_arg('save_weights_path', str, None, "Path of the saved weights during training.")
+add_arg('dataset_path', str, None, "Cityscape dataset path.")
+add_arg('parallel', bool, True, "using ParallelExecutor.")
+add_arg('use_gpu', bool, True, "Whether use GPU or CPU.")
+add_arg('num_classes', int, 19, "Number of classes.")
+add_arg('load_logit_layer', bool, True, "Load last logit fc layer or not. If you are training with different number of classes, you should set to False.")
+add_arg('memory_optimize', bool, True, "Using memory optimizer.")
+add_arg('norm_type', str, 'bn', "Normalization type, should be 'bn' or 'gn'.")
+add_arg('profile', bool, False, "Enable profiler.")
+add_arg('use_py_reader', bool, True, "Use py reader.")
+parser.add_argument(
+ '--enable_ce',
+ action='store_true',
+ help='If set, run the task with continuous evaluation logs.')
+#yapf: enable
+
+@contextlib.contextmanager
+def profile_context(profile=True):
+ if profile:
+ with profiler.profiler('All', 'total', '/tmp/profile_file2'):
+ yield
+ else:
+ yield
def load_model():
- myvars = [
- x for x in tp.list_vars()
- if isinstance(x, fluid.framework.Parameter) and x.name.find('logit') ==
- -1
- ]
- if args.init_weights_path.endswith('/'):
- if args.num_classes == 19:
+ if os.path.isdir(args.init_weights_path):
+ load_vars = [
+ x for x in tp.list_vars()
+ if isinstance(x, fluid.framework.Parameter) and x.name.find('logit') ==
+ -1
+ ]
+ if args.load_logit_layer:
fluid.io.load_params(
exe, dirname=args.init_weights_path, main_program=tp)
else:
- fluid.io.load_vars(exe, dirname=args.init_weights_path, vars=myvars)
+ fluid.io.load_vars(exe, dirname=args.init_weights_path, vars=load_vars)
else:
- if args.num_classes == 19:
- fluid.io.load_params(
- exe,
- dirname="",
- filename=args.init_weights_path,
- main_program=tp)
- else:
- fluid.io.load_vars(
- exe, dirname="", filename=args.init_weights_path, vars=myvars)
+ fluid.io.load_params(
+ exe,
+ dirname="",
+ filename=args.init_weights_path,
+ main_program=tp)
+
def save_model():
- if args.save_weights_path.endswith('/'):
- fluid.io.save_params(
- exe, dirname=args.save_weights_path, main_program=tp)
- else:
- fluid.io.save_params(
- exe, dirname="", filename=args.save_weights_path, main_program=tp)
+ assert not os.path.isfile(args.save_weights_path)
+ fluid.io.save_params(
+ exe, dirname=args.save_weights_path, main_program=tp)
def loss(logit, label):
- label_nignore = (label < num_classes).astype('float32')
- label = fluid.layers.elementwise_min(
- label,
- fluid.layers.assign(np.array(
- [num_classes - 1], dtype=np.int32)))
+ label_nignore = fluid.layers.less_than(
+ label.astype('float32'),
+ fluid.layers.assign(np.array([num_classes], 'float32')),
+ force_cpu=False).astype('float32')
logit = fluid.layers.transpose(logit, [0, 2, 3, 1])
logit = fluid.layers.reshape(logit, [-1, num_classes])
label = fluid.layers.reshape(label, [-1, 1])
label = fluid.layers.cast(label, 'int64')
label_nignore = fluid.layers.reshape(label_nignore, [-1, 1])
- loss = fluid.layers.softmax_with_cross_entropy(logit, label)
- loss = loss * label_nignore
- no_grad_set.add(label_nignore.name)
- no_grad_set.add(label.name)
+ loss = fluid.layers.softmax_with_cross_entropy(logit, label, ignore_index=255, numeric_stable_mode=True)
+ label_nignore.stop_gradient = True
+ label.stop_gradient = True
return loss, label_nignore
-def get_cards(args):
- if args.enable_ce:
- cards = os.environ.get('CUDA_VISIBLE_DEVICES')
- num = len(cards.split(","))
- return num
- else:
- return args.num_devices
-
-
-CityscapeDataset = reader.CityscapeDataset
-parser = argparse.ArgumentParser()
-
-add_arguments()
-
args = parser.parse_args()
+utility.print_arguments(args)
models.clean()
models.bn_momentum = 0.9997
models.dropout_keep_prop = 0.9
models.label_number = args.num_classes
+models.default_norm_type = args.norm_type
deeplabv3p = models.deeplabv3p
sp = fluid.Program()
@@ -133,12 +123,17 @@ weight_decay = 0.00004
base_lr = args.base_lr
total_step = args.total_step
-no_grad_set = set()
-
with fluid.program_guard(tp, sp):
- img = fluid.layers.data(
- name='img', shape=[3] + image_shape, dtype='float32')
- label = fluid.layers.data(name='label', shape=image_shape, dtype='int32')
+ if args.use_py_reader:
+ batch_size_each = batch_size // fluid.core.get_cuda_device_count()
+ py_reader = fluid.layers.py_reader(capacity=64,
+ shapes=[[batch_size_each, 3] + image_shape, [batch_size_each] + image_shape],
+ dtypes=['float32', 'int32'])
+ img, label = fluid.layers.read_file(py_reader)
+ else:
+ img = fluid.layers.data(
+ name='img', shape=[3] + image_shape, dtype='float32')
+ label = fluid.layers.data(name='label', shape=image_shape, dtype='int32')
logit = deeplabv3p(img)
pred = fluid.layers.argmax(logit, axis=1).astype('int32')
loss, mask = loss(logit, label)
@@ -154,11 +149,21 @@ with fluid.program_guard(tp, sp):
lr,
momentum=0.9,
regularization=fluid.regularizer.L2DecayRegularizer(
- regularization_coeff=weight_decay), )
- retv = opt.minimize(loss_mean, startup_program=sp, no_grad_set=no_grad_set)
-
-fluid.memory_optimize(
- tp, print_log=False, skip_opt_set=set([pred.name, loss_mean.name]), level=1)
+ regularization_coeff=weight_decay))
+ optimize_ops, params_grads = opt.minimize(loss_mean, startup_program=sp)
+ # ir memory optimizer has some issues, we need to seed grad persistable to
+ # avoid this issue
+ for p,g in params_grads: g.persistable = True
+
+
+exec_strategy = fluid.ExecutionStrategy()
+exec_strategy.num_threads = fluid.core.get_cuda_device_count()
+exec_strategy.num_iteration_per_drop_scope = 100
+build_strategy = fluid.BuildStrategy()
+if args.memory_optimize:
+ build_strategy.fuse_relu_depthwise_conv = True
+ build_strategy.enable_inplace = True
+ build_strategy.memory_optimize = True
place = fluid.CPUPlace()
if args.use_gpu:
@@ -170,47 +175,58 @@ if args.init_weights_path:
print("load from:", args.init_weights_path)
load_model()
-dataset = CityscapeDataset(args.dataset_path, 'train')
+dataset = reader.CityscapeDataset(args.dataset_path, 'train')
if args.parallel:
- exe_p = fluid.ParallelExecutor(
- use_cuda=True, loss_name=loss_mean.name, main_program=tp)
-
-batches = dataset.get_batch_generator(batch_size, total_step)
-
+ binary = fluid.compiler.CompiledProgram(tp).with_data_parallel(
+ loss_name=loss_mean.name,
+ build_strategy=build_strategy,
+ exec_strategy=exec_strategy)
+else:
+ binary = fluid.compiler.CompiledProgram(main)
+
+if args.use_py_reader:
+ assert(batch_size % fluid.core.get_cuda_device_count() == 0)
+ def data_gen():
+ batches = dataset.get_batch_generator(
+ batch_size // fluid.core.get_cuda_device_count(),
+ total_step * fluid.core.get_cuda_device_count())
+ for b in batches:
+ yield b[1], b[2]
+ py_reader.decorate_tensor_provider(data_gen)
+ py_reader.start()
+else:
+ batches = dataset.get_batch_generator(batch_size, total_step)
total_time = 0.0
epoch_idx = 0
train_loss = 0
-for i, imgs, labels, names in batches:
- epoch_idx += 1
- begin_time = time.time()
- prev_start_time = time.time()
- if args.parallel:
- retv = exe_p.run(fetch_list=[pred.name, loss_mean.name],
- feed={'img': imgs,
- 'label': labels})
- else:
- retv = exe.run(tp,
- feed={'img': imgs,
- 'label': labels},
- fetch_list=[pred, loss_mean])
- end_time = time.time()
- total_time += end_time - begin_time
- if i % 100 == 0:
- print("Model is saved to", args.save_weights_path)
- save_model()
- print("step {:d}, loss: {:.6f}, step_time_cost: {:.3f}".format(
- i, np.mean(retv[1]), end_time - prev_start_time))
-
- # only for ce
- train_loss = np.mean(retv[1])
+with profile_context(args.profile):
+ for i in range(total_step):
+ epoch_idx += 1
+ begin_time = time.time()
+ prev_start_time = time.time()
+ if not args.use_py_reader:
+ _, imgs, labels, names = next(batches)
+ train_loss, = exe.run(binary,
+ feed={'img': imgs,
+ 'label': labels}, fetch_list=[loss_mean])
+ else:
+ train_loss, = exe.run(binary, fetch_list=[loss_mean])
+ train_loss = np.mean(train_loss)
+ end_time = time.time()
+ total_time += end_time - begin_time
+ if i % 100 == 0:
+ print("Model is saved to", args.save_weights_path)
+ save_model()
+ print("step {:d}, loss: {:.6f}, step_time_cost: {:.3f}".format(
+ i, train_loss, end_time - prev_start_time))
+
+print("Training done. Model is saved to", args.save_weights_path)
+save_model()
if args.enable_ce:
- gpu_num = get_cards(args)
+ gpu_num = fluid.core.get_cuda_device_count()
print("kpis\teach_pass_duration_card%s\t%s" %
(gpu_num, total_time / epoch_idx))
print("kpis\ttrain_loss_card%s\t%s" % (gpu_num, train_loss))
-
-print("Training done. Model is saved to", args.save_weights_path)
-save_model()
diff --git a/fluid/PaddleCV/deeplabv3+/utility.py b/fluid/PaddleCV/deeplabv3+/utility.py
new file mode 100644
index 0000000000000000000000000000000000000000..aebb9acbf4f450b50f020d96ccd3b13be5d7afaf
--- /dev/null
+++ b/fluid/PaddleCV/deeplabv3+/utility.py
@@ -0,0 +1,60 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 __future__ import division
+from __future__ import print_function
+import distutils.util
+import six
+
+
+def print_arguments(args):
+ """Print argparse's arguments.
+
+ Usage:
+
+ .. code-block:: python
+
+ parser = argparse.ArgumentParser()
+ parser.add_argument("name", default="Jonh", type=str, help="User name.")
+ args = parser.parse_args()
+ print_arguments(args)
+
+ :param args: Input argparse.Namespace for printing.
+ :type args: argparse.Namespace
+ """
+ print("----------- Configuration Arguments -----------")
+ for arg, value in sorted(six.iteritems(vars(args))):
+ print("%s: %s" % (arg, value))
+ print("------------------------------------------------")
+
+
+def add_arguments(argname, type, default, help, argparser, **kwargs):
+ """Add argparse's argument.
+
+ Usage:
+
+ .. code-block:: python
+
+ parser = argparse.ArgumentParser()
+ add_argument("name", str, "Jonh", "User name.", parser)
+ args = parser.parse_args()
+ """
+ type = distutils.util.strtobool if type == bool else type
+ argparser.add_argument(
+ "--" + argname,
+ default=default,
+ type=type,
+ help=help + ' Default: %(default)s.',
+ **kwargs)
diff --git a/fluid/PaddleCV/face_detection/widerface_eval.py b/fluid/PaddleCV/face_detection/widerface_eval.py
index 1544442c78c38bcbcb537cd81374f5c72c7bfc5a..46eed9be5d064d50c824dd0769b07b9b425dfeb4 100644
--- a/fluid/PaddleCV/face_detection/widerface_eval.py
+++ b/fluid/PaddleCV/face_detection/widerface_eval.py
@@ -121,7 +121,7 @@ def detect_face(image, shrink):
return_numpy=False)
detection = np.array(detection)
# layout: xmin, ymin, xmax. ymax, score
- if detection.shape == (1, ):
+ if np.prod(detection.shape) == 1:
print("No face detected")
return np.array([[0, 0, 0, 0, 0]])
det_conf = detection[:, 1]
diff --git a/fluid/PaddleCV/faster_rcnn/image/Faster_RCNN.jpg b/fluid/PaddleCV/faster_rcnn/image/Faster_RCNN.jpg
deleted file mode 100644
index c2ab8085c914979eb23a59734d54797b6580e956..0000000000000000000000000000000000000000
Binary files a/fluid/PaddleCV/faster_rcnn/image/Faster_RCNN.jpg and /dev/null differ
diff --git a/fluid/PaddleCV/icnet/README.md b/fluid/PaddleCV/icnet/README.md
index dc350ff5e66993b33b976018df36369b773a90c3..84e067ab081f648a4107ece906bad9a52ae13bbc 100644
--- a/fluid/PaddleCV/icnet/README.md
+++ b/fluid/PaddleCV/icnet/README.md
@@ -103,7 +103,7 @@ python infer.py \
## 其他信息
|数据集 | pretrained model |
|---|---|
-|CityScape | [Model]()[md: ] |
+|CityScape | [pretrained_model](https://paddle-icnet-models.bj.bcebos.com/model_1000.tar.gz) |
## 参考
diff --git a/fluid/PaddleCV/icnet/cityscape.py b/fluid/PaddleCV/icnet/cityscape.py
index c5c08afcf3a3c85b9f43c9110e8a8dedc5900d5b..281658fa1e00bf38a9e7f0fa1a7ed2e9b7559539 100644
--- a/fluid/PaddleCV/icnet/cityscape.py
+++ b/fluid/PaddleCV/icnet/cityscape.py
@@ -155,6 +155,17 @@ class DataGenerater:
else:
return np.pad(image, ((0, pad_h), (0, pad_w)), 'constant')
+ def random_crop(self, im, out_shape, is_color=True):
+ h, w = im.shape[:2]
+ h_start = np.random.randint(0, h - out_shape[0] + 1)
+ w_start = np.random.randint(0, w - out_shape[1] + 1)
+ h_end, w_end = h_start + out_shape[0], w_start + out_shape[1]
+ if is_color:
+ im = im[h_start:h_end, w_start:w_end, :]
+ else:
+ im = im[h_start:h_end, w_start:w_end]
+ return im
+
def resize(self, image, label, out_size):
"""
Resize image and label by padding or cropping.
@@ -166,8 +177,7 @@ class DataGenerater:
combined = np.concatenate((image, label), axis=2)
combined = self.padding_as(
combined, out_size[0], out_size[1], is_color=True)
- combined = dataset.image.random_crop(
- combined, out_size[0], is_color=True)
+ combined = self.random_crop(combined, out_size, is_color=True)
image = combined[:, :, 0:3]
label = combined[:, :, 3:4] + ignore_label
return image, label
diff --git a/fluid/PaddleCV/icnet/icnet.py b/fluid/PaddleCV/icnet/icnet.py
index d640621eb9def4bfb1411667ea68f5384fbd5489..3286ce74072f0fde2b215763d50156dcd152a99c 100644
--- a/fluid/PaddleCV/icnet/icnet.py
+++ b/fluid/PaddleCV/icnet/icnet.py
@@ -235,12 +235,12 @@ def proj_block(input, filter_num, padding=0, dilation=None, stride=1,
def sub_net_4(input, input_shape):
- tmp = interp(input, out_shape=np.ceil(input_shape // 32))
+ tmp = interp(input, out_shape=(input_shape // 32))
tmp = dilation_convs(tmp)
tmp = pyramis_pooling(tmp, input_shape)
tmp = conv(tmp, 1, 1, 256, 1, 1, name="conv5_4_k1")
tmp = bn(tmp, relu=True)
- tmp = interp(tmp, input_shape // 16)
+ tmp = interp(tmp, out_shape=np.ceil(input_shape / 16))
return tmp
diff --git a/fluid/PaddleCV/image_classification/README.md b/fluid/PaddleCV/image_classification/README.md
index 57dc26005334eff06528dcb22a99c17659a61d2c..3a20bdf4aa44e2752b8939bfa49886e7c0d5f6f0 100644
--- a/fluid/PaddleCV/image_classification/README.md
+++ b/fluid/PaddleCV/image_classification/README.md
@@ -81,7 +81,7 @@ python train.py \
* **lr**: initialized learning rate. Default: 0.1.
* **pretrained_model**: model path for pretraining. Default: None.
* **checkpoint**: the checkpoint path to resume. Default: None.
-* **model_category**: the category of models, ("models"|"models_name"). Default: "models".
+* **model_category**: the category of models, ("models"|"models_name"). Default: "models_name".
Or can start the training step by running the ```run.sh```.
@@ -209,6 +209,7 @@ Models are trained by starting with learning rate ```0.1``` and decaying it by `
|[VGG16](https://paddle-imagenet-models-name.bj.bcebos.com/VGG16_pretrained.zip) | 72.08%/90.63% | 71.65%/90.57% |
|[VGG19](https://paddle-imagenet-models-name.bj.bcebos.com/VGG19_pretrained.zip) | 72.56%/90.83% | 72.32%/90.98% |
|[MobileNetV1](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.zip) | 70.91%/89.54% | 70.51%/89.35% |
+|[MobileNetV2](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.zip) | 71.90%/90.55% | 71.53%/90.41% |
|[ResNet50](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.zip) | 76.35%/92.80% | 76.22%/92.92% |
|[ResNet101](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.zip) | 77.49%/93.57% | 77.56%/93.64% |
|[ResNet152](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet152_pretrained.zip) | 78.12%/93.93% | 77.92%/93.87% |
@@ -220,6 +221,8 @@ Models are trained by starting with learning rate ```0.1``` and decaying it by `
- Released models: not specify parameter names
+**NOTE: These are trained by using model_category=models**
+
|model | top-1/top-5 accuracy(PIL)| top-1/top-5 accuracy(CV2) |
|- |:-: |:-:|
|[ResNet152](http://paddle-imagenet-models.bj.bcebos.com/ResNet152_pretrained.zip) | 78.18%/93.93% | 78.11%/94.04% |
diff --git a/fluid/PaddleCV/image_classification/README_cn.md b/fluid/PaddleCV/image_classification/README_cn.md
index 7fc35a643e95dae8c2197a96e1fab44b60e458a4..c9f553e311ff07e76eab2274646270f5de9b8fa2 100644
--- a/fluid/PaddleCV/image_classification/README_cn.md
+++ b/fluid/PaddleCV/image_classification/README_cn.md
@@ -79,7 +79,7 @@ python train.py \
* **lr**: initialized learning rate. Default: 0.1.
* **pretrained_model**: model path for pretraining. Default: None.
* **checkpoint**: the checkpoint path to resume. Default: None.
-* **model_category**: the category of models, ("models"|"models_name"). Default:"models".
+* **model_category**: the category of models, ("models"|"models_name"). Default:"models_name".
**数据读取器说明：** 数据读取器定义在```reader.py```和```reader_cv2.py```中, 一般, CV2 reader可以提高数据读取速度, reader(PIL)可以得到相对更高的精度, 在[训练阶段](#training-a-model), 默认采用的增广方式是随机裁剪与水平翻转, 而在[评估](#inference)与[推断](#inference)阶段用的默认方式是中心裁剪。当前支持的数据增广方式有：
* 旋转
@@ -204,6 +204,7 @@ Models包括两种模型：带有参数名字的模型，和不带有参数名
|[VGG16](https://paddle-imagenet-models-name.bj.bcebos.com/VGG16_pretrained.zip) | 72.08%/90.63% | 71.65%/90.57% |
|[VGG19](https://paddle-imagenet-models-name.bj.bcebos.com/VGG19_pretrained.zip) | 72.56%/90.83% | 72.32%/90.98% |
|[MobileNetV1](http://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV1_pretrained.zip) | 70.91%/89.54% | 70.51%/89.35% |
+|[MobileNetV2](https://paddle-imagenet-models-name.bj.bcebos.com/MobileNetV2_pretrained.zip) | 71.90%/90.55% | 71.53%/90.41% |
|[ResNet50](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_pretrained.zip) | 76.35%/92.80% | 76.22%/92.92% |
|[ResNet101](http://paddle-imagenet-models-name.bj.bcebos.com/ResNet101_pretrained.zip) | 77.49%/93.57% | 77.56%/93.64% |
|[ResNet152](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet152_pretrained.zip) | 78.12%/93.93% | 77.92%/93.87% |
@@ -212,6 +213,8 @@ Models包括两种模型：带有参数名字的模型，和不带有参数名
- Released models: not specify parameter names
+**注意：这是model_category = models 的预训练模型**
+
|model | top-1/top-5 accuracy(PIL)| top-1/top-5 accuracy(CV2) |
|- |:-: |:-:|
|[ResNet152](http://paddle-imagenet-models.bj.bcebos.com/ResNet152_pretrained.zip) | 78.18%/93.93% | 78.11%/94.04% |
diff --git a/fluid/PaddleCV/image_classification/dist_train/README.md b/fluid/PaddleCV/image_classification/dist_train/README.md
index 0b2729cce4fa2e0780b8db5f87da49a8e221c665..8875429146045e60457246fbc9b0a34d874d8855 100644
--- a/fluid/PaddleCV/image_classification/dist_train/README.md
+++ b/fluid/PaddleCV/image_classification/dist_train/README.md
@@ -39,6 +39,8 @@ You can test if distributed training works on a single node before deploying to
***NOTE: for best performance, we recommend using multi-process mode, see No.3. And together with fp16.***
+***NOTE: for nccl2 distributed mode, you must ensure each node train same number of samples, or set skip_unbalanced_data to 1 to do sync training.***
+
1. simply run `python dist_train.py` to start local training with default configuratioins.
2. for pserver mode, run `bash run_ps_mode.sh` to start 2 pservers and 2 trainers, these 2 trainers
will use GPU 0 and 1 to simulate 2 workers.
@@ -90,4 +92,19 @@ The default resnet50 distributed training config is based on this paper: https:/
### Performance
-TBD
+The below figure shows fluid distributed training performances. We did these on a 4-node V100 GPU cluster,
+each has 8 V100 GPU card, with total of 32 GPUs. All modes can reach the "state of the art (choose loss scale carefully when using fp16 mode)" of ResNet50 model with imagenet dataset. The Y axis in the figure shows
+the images/s while the X-axis shows the number of GPUs.
+
+
+
+Performance of Multiple-GPU Training of Resnet50 on Imagenet
+
+
+The second figure shows speed-ups when using multiple GPUs according to the above figure.
+
+
+
+Speed-ups of Multiple-GPU Training of Resnet50 on Imagenet
+
+
diff --git a/fluid/PaddleCV/image_classification/eval.py b/fluid/PaddleCV/image_classification/eval.py
index ddce243fe1fcae81ee6064c7ff185fb8a045a402..0660efe13750467ad6bf964b484c9db0ab44b1ee 100644
--- a/fluid/PaddleCV/image_classification/eval.py
+++ b/fluid/PaddleCV/image_classification/eval.py
@@ -7,8 +7,6 @@ import time
import sys
import paddle
import paddle.fluid as fluid
-#import models
-import models_name as models
#import reader_cv2 as reader
import reader as reader
import argparse
@@ -26,10 +24,21 @@ add_arg('class_dim', int, 1000, "Class number.")
add_arg('image_shape', str, "3,224,224", "Input image size")
add_arg('with_mem_opt', bool, True, "Whether to use memory optimization or not.")
add_arg('pretrained_model', str, None, "Whether to use pretrained model.")
-add_arg('model', str, "SE_ResNeXt50_32x4d", "Set the network to use.")
+add_arg('model', str, "SE_ResNeXt50_32x4d", "Set the network to use.")
+add_arg('model_category', str, "models_name", "Whether to use models_name or not, valid value:'models','models_name'." )
+
# yapf: enable
-model_list = [m for m in dir(models) if "__" not in m]
+
+def set_models(model_category):
+ global models
+ assert model_category in ["models", "models_name"
+ ], "{} is not in lists: {}".format(
+ model_category, ["models", "models_name"])
+ if model_category == "models_name":
+ import models_name as models
+ else:
+ import models as models
def eval(args):
@@ -40,6 +49,7 @@ def eval(args):
with_memory_optimization = args.with_mem_opt
image_shape = [int(m) for m in args.image_shape.split(",")]
+ model_list = [m for m in dir(models) if "__" not in m]
assert model_name in model_list, "{} is not in lists: {}".format(args.model,
model_list)
@@ -63,11 +73,11 @@ def eval(args):
acc_top5 = fluid.layers.accuracy(input=out0, label=label, k=5)
else:
out = model.net(input=image, class_dim=class_dim)
- cost = fluid.layers.cross_entropy(input=out, label=label)
-
+ cost, pred = fluid.layers.softmax_with_cross_entropy(
+ out, label, return_softmax=True)
avg_cost = fluid.layers.mean(x=cost)
- acc_top1 = fluid.layers.accuracy(input=out, label=label, k=1)
- acc_top5 = fluid.layers.accuracy(input=out, label=label, k=5)
+ acc_top1 = fluid.layers.accuracy(input=pred, label=label, k=1)
+ acc_top5 = fluid.layers.accuracy(input=pred, label=label, k=5)
test_program = fluid.default_main_program().clone(for_test=True)
@@ -125,6 +135,7 @@ def eval(args):
def main():
args = parser.parse_args()
print_arguments(args)
+ set_models(args.model_category)
eval(args)
diff --git a/fluid/PaddleCV/image_classification/images/imagenet_dist_performance.png b/fluid/PaddleCV/image_classification/images/imagenet_dist_performance.png
new file mode 100644
index 0000000000000000000000000000000000000000..d6be7f5f3a9f79c495a684c0b5f4ed459d465c5e
Binary files /dev/null and b/fluid/PaddleCV/image_classification/images/imagenet_dist_performance.png differ
diff --git a/fluid/PaddleCV/image_classification/images/imagenet_dist_speedup.png b/fluid/PaddleCV/image_classification/images/imagenet_dist_speedup.png
new file mode 100644
index 0000000000000000000000000000000000000000..de834c95ff7e4f0f8aec780ebc7cafdb77872e4d
Binary files /dev/null and b/fluid/PaddleCV/image_classification/images/imagenet_dist_speedup.png differ
diff --git a/fluid/PaddleCV/image_classification/infer.py b/fluid/PaddleCV/image_classification/infer.py
index e89c08d923cdc37596c76dc7146a2666b719844d..88ccf42912b67035895cd81f5f982edca1bd0a3e 100644
--- a/fluid/PaddleCV/image_classification/infer.py
+++ b/fluid/PaddleCV/image_classification/infer.py
@@ -7,7 +7,6 @@ import time
import sys
import paddle
import paddle.fluid as fluid
-import models
import reader
import argparse
import functools
@@ -23,9 +22,19 @@ add_arg('image_shape', str, "3,224,224", "Input image size")
add_arg('with_mem_opt', bool, True, "Whether to use memory optimization or not.")
add_arg('pretrained_model', str, None, "Whether to use pretrained model.")
add_arg('model', str, "SE_ResNeXt50_32x4d", "Set the network to use.")
+add_arg('model_category', str, "models_name", "Whether to use models_name or not, valid value:'models','models_name'." )
# yapf: enable
-model_list = [m for m in dir(models) if "__" not in m]
+
+def set_models(model_category):
+ global models
+ assert model_category in ["models", "models_name"
+ ], "{} is not in lists: {}".format(
+ model_category, ["models", "models_name"])
+ if model_category == "models_name":
+ import models_name as models
+ else:
+ import models as models
def infer(args):
@@ -35,7 +44,7 @@ def infer(args):
pretrained_model = args.pretrained_model
with_memory_optimization = args.with_mem_opt
image_shape = [int(m) for m in args.image_shape.split(",")]
-
+ model_list = [m for m in dir(models) if "__" not in m]
assert model_name in model_list, "{} is not in lists: {}".format(args.model,
model_list)
@@ -85,6 +94,7 @@ def infer(args):
def main():
args = parser.parse_args()
print_arguments(args)
+ set_models(args.model_category)
infer(args)
diff --git a/fluid/PaddleCV/image_classification/run.sh b/fluid/PaddleCV/image_classification/run.sh
index fbdacb87633a70b60ecdedf9a6f74e7287d2b2d0..b0cc2255b03db82bc88397f625ed68023280d2f0 100644
--- a/fluid/PaddleCV/image_classification/run.sh
+++ b/fluid/PaddleCV/image_classification/run.sh
@@ -1,16 +1,19 @@
#Hyperparameters config
+#Example: SE_ResNext50_32x4d
python train.py \
--model=SE_ResNeXt50_32x4d \
- --batch_size=32 \
+ --batch_size=400 \
--total_images=1281167 \
--class_dim=1000 \
--image_shape=3,224,224 \
--model_save_dir=output/ \
- --with_mem_opt=False \
- --lr_strategy=piecewise_decay \
- --lr=0.1
+ --with_mem_opt=True \
+ --lr_strategy=cosine_decay \
+ --lr=0.1 \
+ --num_epochs=200 \
+ --l2_decay=1.2e-4 \
+ --model_category=models_name \
# >log_SE_ResNeXt50_32x4d.txt 2>&1 &
-
#AlexNet:
#python train.py \
# --model=AlexNet \
@@ -19,24 +22,12 @@ python train.py \
# --class_dim=1000 \
# --image_shape=3,224,224 \
# --model_save_dir=output/ \
-# --with_mem_opt=False \
+# --with_mem_opt=True \
+# --model_category=models_name \
# --lr_strategy=piecewise_decay \
# --num_epochs=120 \
-# --lr=0.01
-
-#VGG11:
-#python train.py \
-# --model=VGG11 \
-# --batch_size=512 \
-# --total_images=1281167 \
-# --class_dim=1000 \
-# --image_shape=3,224,224 \
-# --model_save_dir=output/ \
-# --with_mem_opt=False \
-# --lr_strategy=piecewise_decay \
-# --num_epochs=120 \
-# --lr=0.1
-
+# --lr=0.01 \
+# --l2_decay=1e-4
#MobileNet v1:
#python train.py \
@@ -46,12 +37,26 @@ python train.py \
# --class_dim=1000 \
# --image_shape=3,224,224 \
# --model_save_dir=output/ \
-# --with_mem_opt=False \
+# --with_mem_opt=True \
+# --model_category=models_name \
# --lr_strategy=piecewise_decay \
# --num_epochs=120 \
-# --lr=0.1
-
+# --lr=0.1 \
+# --l2_decay=3e-5
+#python train.py \
+# --model=MobileNetV2 \
+# --batch_size=500 \
+# --total_images=1281167 \
+# --class_dim=1000 \
+# --image_shape=3,224,224 \
+# --model_save_dir=output/ \
+# --model_category=models_name \
+# --with_mem_opt=True \
+# --lr_strategy=cosine_decay \
+# --num_epochs=240 \
+# --lr=0.1 \
+# --l2_decay=4e-5
#ResNet50:
#python train.py \
# --model=ResNet50 \
@@ -60,10 +65,12 @@ python train.py \
# --class_dim=1000 \
# --image_shape=3,224,224 \
# --model_save_dir=output/ \
-# --with_mem_opt=False \
+# --with_mem_opt=True \
+# --model_category=models_name \
# --lr_strategy=piecewise_decay \
# --num_epochs=120 \
-# --lr=0.1
+# --lr=0.1 \
+# --l2_decay=1e-4
#ResNet101:
#python train.py \
@@ -73,44 +80,58 @@ python train.py \
# --class_dim=1000 \
# --image_shape=3,224,224 \
# --model_save_dir=output/ \
-# --with_mem_opt=False \
+# --model_category=models_name \
+# --with_mem_opt=True \
# --lr_strategy=piecewise_decay \
# --num_epochs=120 \
-# --lr=0.1
+# --lr=0.1 \
+# --l2_decay=1e-4
#ResNet152:
#python train.py \
# --model=ResNet152 \
# --batch_size=256 \
# --total_images=1281167 \
+# --class_dim=1000 \
# --image_shape=3,224,224 \
+# --model_save_dir=output/ \
# --lr_strategy=piecewise_decay \
+# --model_category=models_name \
+# --with_mem_opt=True \
# --lr=0.1 \
# --num_epochs=120 \
-# --l2_decay=1e-4 \(TODO)
+# --l2_decay=1e-4
-#SE_ResNeXt50:
+#SE_ResNeXt50_32x4d:
#python train.py \
-# --model=SE_ResNeXt50 \
+# --model=SE_ResNeXt50_32x4d \
# --batch_size=400 \
# --total_images=1281167 \
+# --class_dim=1000 \
# --image_shape=3,224,224 \
# --lr_strategy=cosine_decay \
+# --model_category=models_name \
+# --model_save_dir=output/ \
# --lr=0.1 \
# --num_epochs=200 \
-# --l2_decay=12e-5 \(TODO)
+# --with_mem_opt=True \
+# --l2_decay=1.2e-4
-#SE_ResNeXt101:
+#SE_ResNeXt101_32x4d:
#python train.py \
-# --model=SE_ResNeXt101 \
+# --model=SE_ResNeXt101_32x4d \
# --batch_size=400 \
# --total_images=1281167 \
+# --class_dim=1000 \
# --image_shape=3,224,224 \
# --lr_strategy=cosine_decay \
+# --model_category=models_name \
+# --model_save_dir=output/ \
# --lr=0.1 \
# --num_epochs=200 \
-# --l2_decay=15e-5 \(TODO)
+# --with_mem_opt=True \
+# --l2_decay=1.5e-5
#VGG11:
#python train.py \
@@ -119,17 +140,55 @@ python train.py \
# --total_images=1281167 \
# --image_shape=3,224,224 \
# --lr_strategy=cosine_decay \
+# --class_dim=1000 \
+# --model_category=models_name \
+# --model_save_dir=output/ \
# --lr=0.1 \
# --num_epochs=90 \
-# --l2_decay=2e-4 \(TODO)
+# --with_mem_opt=True \
+# --l2_decay=2e-4
#VGG13:
#python train.py
# --model=VGG13 \
# --batch_size=256 \
# --total_images=1281167 \
+# --class_dim=1000 \
# --image_shape=3,224,224 \
# --lr_strategy=cosine_decay \
# --lr=0.01 \
# --num_epochs=90 \
-# --l2_decay=3e-4 \(TODO)
+# --model_category=models_name \
+# --model_save_dir=output/ \
+# --with_mem_opt=True \
+# --l2_decay=3e-4
+
+#VGG16:
+#python train.py
+# --model=VGG16 \
+# --batch_size=256 \
+# --total_images=1281167 \
+# --class_dim=1000 \
+# --lr_strategy=cosine_decay \
+# --image_shape=3,224,224 \
+# --model_category=models_name \
+# --model_save_dir=output/ \
+# --lr=0.01 \
+# --num_epochs=90 \
+# --with_mem_opt=True \
+# --l2_decay=3e-4
+
+#VGG19:
+#python train.py
+# --model=VGG19 \
+# --batch_size=256 \
+# --total_images=1281167 \
+# --class_dim=1000 \
+# --image_shape=3,224,224 \
+# --lr_strategy=cosine_decay \
+# --lr=0.01 \
+# --num_epochs=90 \
+# --with_mem_opt=True \
+# --model_category=models_name \
+# --model_save_dir=output/ \
+# --l2_decay=3e-4
diff --git a/fluid/PaddleCV/image_classification/train.py b/fluid/PaddleCV/image_classification/train.py
index 6830773b91f2fa07c2b6f530a6370cedde82ffd7..145b288620bbbb27693bacbb7145e4df8371a4c2 100644
--- a/fluid/PaddleCV/image_classification/train.py
+++ b/fluid/PaddleCV/image_classification/train.py
@@ -10,7 +10,6 @@ import math
import paddle
import paddle.fluid as fluid
import paddle.dataset.flowers as flowers
-import models
import reader
import argparse
import functools
@@ -19,8 +18,8 @@ import utils
from utils.learning_rate import cosine_decay
from utils.fp16_utils import create_master_params_grads, master_param_to_train_param
from utility import add_arguments, print_arguments
-import models
-import models_name
+
+IMAGENET1000 = 1281167
parser = argparse.ArgumentParser(description=__doc__)
add_arg = functools.partial(add_arguments, argparser=parser)
@@ -40,25 +39,32 @@ add_arg('lr_strategy', str, "piecewise_decay", "Set the learning rate
add_arg('model', str, "SE_ResNeXt50_32x4d", "Set the network to use.")
add_arg('enable_ce', bool, False, "If set True, enable continuous evaluation job.")
add_arg('data_dir', str, "./data/ILSVRC2012", "The ImageNet dataset root dir.")
-add_arg('model_category', str, "models", "Whether to use models_name or not, valid value:'models','models_name'" )
+add_arg('model_category', str, "models_name", "Whether to use models_name or not, valid value:'models','models_name'." )
add_arg('fp16', bool, False, "Enable half precision training with fp16." )
add_arg('scale_loss', float, 1.0, "Scale loss for fp16." )
+add_arg('l2_decay', float, 1e-4, "L2_decay parameter.")
+add_arg('momentum_rate', float, 0.9, "momentum_rate.")
# yapf: enable
-def set_models(model):
+def set_models(model_category):
global models
- if model == "models":
- models = models
+ assert model_category in ["models", "models_name"
+ ], "{} is not in lists: {}".format(
+ model_category, ["models", "models_name"])
+ if model_category == "models_name":
+ import models_name as models
else:
- models = models_name
+ import models as models
def optimizer_setting(params):
ls = params["learning_strategy"]
+ l2_decay = params["l2_decay"]
+ momentum_rate = params["momentum_rate"]
if ls["name"] == "piecewise_decay":
if "total_images" not in params:
- total_images = 1281167
+ total_images = IMAGENET1000
else:
total_images = params["total_images"]
batch_size = ls["batch_size"]
@@ -71,16 +77,17 @@ def optimizer_setting(params):
optimizer = fluid.optimizer.Momentum(
learning_rate=fluid.layers.piecewise_decay(
boundaries=bd, values=lr),
- momentum=0.9,
- regularization=fluid.regularizer.L2Decay(1e-4))
+ momentum=momentum_rate,
+ regularization=fluid.regularizer.L2Decay(l2_decay))
elif ls["name"] == "cosine_decay":
if "total_images" not in params:
- total_images = 1281167
+ total_images = IMAGENET1000
else:
total_images = params["total_images"]
-
batch_size = ls["batch_size"]
+ l2_decay = params["l2_decay"]
+ momentum_rate = params["momentum_rate"]
step = int(total_images / batch_size + 1)
lr = params["lr"]
@@ -89,43 +96,42 @@ def optimizer_setting(params):
optimizer = fluid.optimizer.Momentum(
learning_rate=cosine_decay(
learning_rate=lr, step_each_epoch=step, epochs=num_epochs),
- momentum=0.9,
- regularization=fluid.regularizer.L2Decay(4e-5))
- elif ls["name"] == "exponential_decay":
+ momentum=momentum_rate,
+ regularization=fluid.regularizer.L2Decay(l2_decay))
+ elif ls["name"] == "linear_decay":
if "total_images" not in params:
- total_images = 1281167
+ total_images = IMAGENET1000
else:
total_images = params["total_images"]
batch_size = ls["batch_size"]
- step = int(total_images / batch_size +1)
- lr = params["lr"]
num_epochs = params["num_epochs"]
- learning_decay_rate_factor=ls["learning_decay_rate_factor"]
- num_epochs_per_decay = ls["num_epochs_per_decay"]
- NUM_GPUS = 1
-
+ start_lr = params["lr"]
+ l2_decay = params["l2_decay"]
+ momentum_rate = params["momentum_rate"]
+ end_lr = 0
+ total_step = int((total_images / batch_size) * num_epochs)
+ lr = fluid.layers.polynomial_decay(
+ start_lr, total_step, end_lr, power=1)
optimizer = fluid.optimizer.Momentum(
- learning_rate=fluid.layers.exponential_decay(
- learning_rate = lr * NUM_GPUS,
- decay_steps = step * num_epochs_per_decay / NUM_GPUS,
- decay_rate = learning_decay_rate_factor),
- momentum=0.9,
-
- regularization = fluid.regularizer.L2Decay(4e-5))
-
+ learning_rate=lr,
+ momentum=momentum_rate,
+ regularization=fluid.regularizer.L2Decay(l2_decay))
else:
lr = params["lr"]
+ l2_decay = params["l2_decay"]
+ momentum_rate = params["momentum_rate"]
optimizer = fluid.optimizer.Momentum(
learning_rate=lr,
- momentum=0.9,
- regularization=fluid.regularizer.L2Decay(1e-4))
+ momentum=momentum_rate,
+ regularization=fluid.regularizer.L2Decay(l2_decay))
return optimizer
+
def net_config(image, label, model, args):
model_list = [m for m in dir(models) if "__" not in m]
- assert args.model in model_list,"{} is not lists: {}".format(
- args.model, model_list)
+ assert args.model in model_list, "{} is not lists: {}".format(args.model,
+ model_list)
class_dim = args.class_dim
model_name = args.model
@@ -148,8 +154,9 @@ def net_config(image, label, model, args):
acc_top1 = fluid.layers.accuracy(input=out0, label=label, k=1)
acc_top5 = fluid.layers.accuracy(input=out0, label=label, k=5)
else:
- out = model.net(input=image, class_dim=class_dim)
- cost, pred = fluid.layers.softmax_with_cross_entropy(out, label, return_softmax=True)
+ out = model.net(input=image, class_dim=class_dim)
+ cost, pred = fluid.layers.softmax_with_cross_entropy(
+ out, label, return_softmax=True)
if args.scale_loss > 1:
avg_cost = fluid.layers.mean(x=cost) * float(args.scale_loss)
else:
@@ -190,19 +197,25 @@ def build_program(is_train, main_prog, startup_prog, args):
params["num_epochs"] = args.num_epochs
params["learning_strategy"]["batch_size"] = args.batch_size
params["learning_strategy"]["name"] = args.lr_strategy
+ params["l2_decay"] = args.l2_decay
+ params["momentum_rate"] = args.momentum_rate
optimizer = optimizer_setting(params)
-
if args.fp16:
params_grads = optimizer.backward(avg_cost)
master_params_grads = create_master_params_grads(
params_grads, main_prog, startup_prog, args.scale_loss)
optimizer.apply_gradients(master_params_grads)
- master_param_to_train_param(master_params_grads, params_grads, main_prog)
+ master_param_to_train_param(master_params_grads,
+ params_grads, main_prog)
else:
optimizer.minimize(avg_cost)
+ global_lr = optimizer._global_learning_rate()
- return py_reader, avg_cost, acc_top1, acc_top5
+ if is_train:
+ return py_reader, avg_cost, acc_top1, acc_top5, global_lr
+ else:
+ return py_reader, avg_cost, acc_top1, acc_top5
def train(args):
@@ -220,7 +233,7 @@ def train(args):
startup_prog.random_seed = 1000
train_prog.random_seed = 1000
- train_py_reader, train_cost, train_acc1, train_acc5 = build_program(
+ train_py_reader, train_cost, train_acc1, train_acc5, global_lr = build_program(
is_train=True,
main_prog=train_prog,
startup_prog=startup_prog,
@@ -255,7 +268,8 @@ def train(args):
if visible_device:
device_num = len(visible_device.split(','))
else:
- device_num = subprocess.check_output(['nvidia-smi', '-L']).decode().count('\n')
+ device_num = subprocess.check_output(
+ ['nvidia-smi', '-L']).decode().count('\n')
train_batch_size = args.batch_size / device_num
test_batch_size = 16
@@ -283,11 +297,12 @@ def train(args):
use_cuda=bool(args.use_gpu),
loss_name=train_cost.name)
- train_fetch_list = [train_cost.name, train_acc1.name, train_acc5.name]
+ train_fetch_list = [
+ train_cost.name, train_acc1.name, train_acc5.name, global_lr.name
+ ]
test_fetch_list = [test_cost.name, test_acc1.name, test_acc5.name]
params = models.__dict__[args.model]().params
-
for pass_id in range(params["num_epochs"]):
train_py_reader.start()
@@ -299,7 +314,9 @@ def train(args):
try:
while True:
t1 = time.time()
- loss, acc1, acc5 = train_exe.run(fetch_list=train_fetch_list)
+ loss, acc1, acc5, lr = train_exe.run(
+ fetch_list=train_fetch_list)
+
t2 = time.time()
period = t2 - t1
loss = np.mean(np.array(loss))
@@ -308,12 +325,14 @@ def train(args):
train_info[0].append(loss)
train_info[1].append(acc1)
train_info[2].append(acc5)
+ lr = np.mean(np.array(lr))
train_time.append(period)
+
if batch_id % 10 == 0:
print("Pass {0}, trainbatch {1}, loss {2}, \
- acc1 {3}, acc5 {4} time {5}"
- .format(pass_id, batch_id, loss, acc1, acc5,
- "%2.2f sec" % period))
+ acc1 {3}, acc5 {4}, lr{5}, time {6}"
+ .format(pass_id, batch_id, loss, acc1, acc5, "%.5f" %
+ lr, "%2.2f sec" % period))
sys.stdout.flush()
batch_id += 1
except fluid.core.EOFException:
@@ -322,7 +341,8 @@ def train(args):
train_loss = np.array(train_info[0]).mean()
train_acc1 = np.array(train_info[1]).mean()
train_acc5 = np.array(train_info[2]).mean()
- train_speed = np.array(train_time).mean() / (train_batch_size * device_num)
+ train_speed = np.array(train_time).mean() / (train_batch_size *
+ device_num)
test_py_reader.start()
@@ -394,10 +414,7 @@ def train(args):
def main():
args = parser.parse_args()
- models_now = args.model_category
- assert models_now in ["models", "models_name"], "{} is not in lists: {}".format(
- models_now, ["models", "models_name"])
- set_models(models_now)
+ set_models(args.model_category)
print_arguments(args)
train(args)
diff --git a/fluid/PaddleCV/ocr_recognition/README.md b/fluid/PaddleCV/ocr_recognition/README.md
index 8b2d95694631e46d541d46c3f4950fd9a99ce0e3..1c9553993e84d10376441407704088ec4dd66c0c 100644
--- a/fluid/PaddleCV/ocr_recognition/README.md
+++ b/fluid/PaddleCV/ocr_recognition/README.md
@@ -202,5 +202,5 @@ env CUDA_VISIBLE_DEVICE=0 python infer.py \
|模型| 错误率|
|- |:-: |
-|[ocr_ctc_params](https://drive.google.com/open?id=1gsg2ODO2_F2pswXwW5MXpf8RY8-BMRyZ) | 22.3% |
-|[ocr_attention_params](https://drive.google.com/open?id=1Bx7-94mngyTaMA5kVjzYHDPAdXxOYbRm) | 15.8%|
+|[ocr_ctc_params](https://paddle-ocr-models.bj.bcebos.com/ocr_ctc.zip) | 22.3% |
+|[ocr_attention_params](https://paddle-ocr-models.bj.bcebos.com/ocr_attention.zip) | 15.8%|
diff --git a/fluid/PaddleCV/faster_rcnn/.gitignore b/fluid/PaddleCV/rcnn/.gitignore
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/.gitignore
rename to fluid/PaddleCV/rcnn/.gitignore
diff --git a/fluid/PaddleCV/faster_rcnn/.run_ce.sh b/fluid/PaddleCV/rcnn/.run_ce.sh
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/.run_ce.sh
rename to fluid/PaddleCV/rcnn/.run_ce.sh
diff --git a/fluid/PaddleCV/faster_rcnn/README.md b/fluid/PaddleCV/rcnn/README.md
similarity index 77%
rename from fluid/PaddleCV/faster_rcnn/README.md
rename to fluid/PaddleCV/rcnn/README.md
index 0a5f68c34adda54ba0e27f44f16c18cafe057830..824709be70af4e7628f8c92e0c7ee7a5b0edf0d8 100644
--- a/fluid/PaddleCV/faster_rcnn/README.md
+++ b/fluid/PaddleCV/rcnn/README.md
@@ -1,4 +1,4 @@
-# Faster RCNN Objective Detection
+# RCNN Objective Detection
---
## Table of Contents
@@ -9,7 +9,6 @@
- [Training](#training)
- [Evaluation](#evaluation)
- [Inference and Visualization](#inference-and-visualization)
-- [Appendix](#appendix)
## Installation
@@ -17,17 +16,20 @@ Running sample code in this directory requires PaddelPaddle Fluid v.1.0.0 and la
## Introduction
-[Faster Rcnn](https://arxiv.org/abs/1506.01497) is a typical two stage detector. The total framework of network can be divided into four parts, as shown below:
-
-
-Faster RCNN model
-
+Region Convolutional Neural Network (RCNN) models are two stages detector. According to proposals and feature extraction, obtain class and more precise proposals.
+Now RCNN model contains two typical models: Faster RCNN and Mask RCNN.
+
+[Faster RCNN](https://arxiv.org/abs/1506.01497), The total framework of network can be divided into four parts:
1. Base conv layer. As a CNN objective dection, Faster RCNN extract feature maps using a basic convolutional network. The feature maps then can be shared by RPN and fc layers. This sampel uses [ResNet-50](https://arxiv.org/abs/1512.03385) as base conv layer.
2. Region Proposal Network (RPN). RPN generates proposals for detection。This block generates anchors by a set of size and ratio and classifies anchors into fore-ground and back-ground by softmax. Then refine anchors to obtain more precise proposals using box regression.
3. RoI Align. This layer takes feature maps and proposals as input. The proposals are mapped to feature maps and pooled to the same size. The output are sent to fc layers for classification and regression. RoIPool and RoIAlign are used separately to this layer and it can be set in roi\_func in config.py.
4. Detection layer. Using the output of roi pooling to compute the class and locatoin of each proposal in two fc layers.
+[Mask RCNN](https://arxiv.org/abs/1703.06870) is a classical instance segmentation model and an extension of Faster RCNN
+
+Mask RCNN is a two stage model as well. At the first stage, it generates proposals from input images. At the second stage, it obtains class result, bbox and mask which is the result from segmentation branch on original Faster RCNN model. It decouples the relation between mask and classification.
+
## Data preparation
Train the model on [MS-COCO dataset](http://cocodataset.org/#download), download dataset as below:
@@ -62,12 +64,24 @@ To train the model, [cocoapi](https://github.com/cocodataset/cocoapi) is needed.
After data preparation, one can start the training step by:
+- Faster RCNN
+
python train.py \
--model_save_dir=output/ \
- --pretrained_model=${path_to_pretrain_model}
- --data_dir=${path_to_data}
+ --pretrained_model=${path_to_pretrain_model} \
+ --data_dir=${path_to_data} \
+ --MASK_ON=False
+
+- Mask RCNN
+
+ python train.py \
+ --model_save_dir=output/ \
+ --pretrained_model=${path_to_pretrain_model} \
+ --data_dir=${path_to_data} \
+ --MASK_ON=True
- Set ```export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7``` to specifiy 8 GPU to train.
+- Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
- For more help on arguments:
python train.py --help
@@ -93,7 +107,6 @@ After data preparation, one can start the training step by:
* In first 500 iteration, the learning rate increases linearly from 0.00333 to 0.01. Then lr is decayed at 120000, 160000 iteration with multiplier 0.1, 0.01. The maximum iteration is 180000. Also, we released a 2x model which has 360000 iterations and lr is decayed at 240000, 320000. These configuration can be set by max_iter and lr_steps in config.py.
* Set the learning rate of bias to two times as global lr in non basic convolutional layers.
* In basic convolutional layers, parameters of affine layers and res body do not update.
-* Use Nvidia Tesla V100 8GPU, total time for training is about 40 hours.
## Evaluation
@@ -101,14 +114,27 @@ Evaluation is to evaluate the performance of a trained model. This sample provid
`eval_coco_map.py` is the main executor for evalution, one can start evalution step by:
+- Faster RCNN
+
python eval_coco_map.py \
--dataset=coco2017 \
--pretrained_model=${path_to_pretrain_model} \
+ --MASK_ON=False
+
+- Mask RCNN
+
+ python eval_coco_map.py \
+ --dataset=coco2017 \
+ --pretrained_model=${path_to_pretrain_model} \
+ --MASK_ON=True
- Set ```export CUDA_VISIBLE_DEVICES=0``` to specifiy one GPU to eval.
+- Set ```MASK_ON``` to choose Faster RCNN or Mask RCNN model.
Evalutaion result is shown as below:
+Faster RCNN:
+
| Model | RoI function | Batch size | Max iteration | mAP |
| :--------------- | :--------: | :------------: | :------------------: |------: |
| [Fluid RoIPool minibatch padding](http://paddlemodels.bj.bcebos.com/faster_rcnn/model_pool_minibatch_padding.tar.gz) | RoIPool | 8 | 180000 | 0.316 |
@@ -121,6 +147,14 @@ Evalutaion result is shown as below:
* Fluid RoIAlign no padding: Images without padding.
* Fluid RoIAlign no padding 2x: Images without padding, train for 360000 iterations, learning rate is decayed at 240000, 320000.
+Mask RCNN:
+
+| Model | Batch size | Max iteration | box mAP | mask mAP |
+| :--------------- | :--------: | :------------: | :--------: |------: |
+| [Fluid mask no padding](https://paddlemodels.bj.bcebos.com/faster_rcnn/Fluid_mask_no_padding.tar.gz) | 8 | 180000 | 0.359 | 0.314 |
+
+* Fluid mask no padding: Use RoIAlign. Images without padding.
+
## Inference and Visualization
Inference is used to get prediction score or image features based on trained models. `infer.py` is the main executor for inference, one can start infer step by:
@@ -135,8 +169,12 @@ Inference is used to get prediction score or image features based on trained mod
Visualization of infer result is shown as below:
-
-
-
+
Faster RCNN Visualization Examples
+
+
+
+
+Mask RCNN Visualization Examples
+
diff --git a/fluid/PaddleCV/faster_rcnn/README_cn.md b/fluid/PaddleCV/rcnn/README_cn.md
similarity index 76%
rename from fluid/PaddleCV/faster_rcnn/README_cn.md
rename to fluid/PaddleCV/rcnn/README_cn.md
index 29adfcfd274b82f2ddaba1894be6ad1c7ece1e6a..57e622e60f049d2976038bddb4cb39bd8fbf9756 100644
--- a/fluid/PaddleCV/faster_rcnn/README_cn.md
+++ b/fluid/PaddleCV/rcnn/README_cn.md
@@ -1,4 +1,4 @@
-# Faster RCNN 目标检测
+# RCNN 系列目标检测
---
## 内容
@@ -9,25 +9,27 @@
- [模型训练](#模型训练)
- [模型评估](#模型评估)
- [模型推断及可视化](#模型推断及可视化)
-- [附录](#附录)
## 安装
在当前目录下运行样例代码需要PadddlePaddle Fluid的v.1.0.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/0.15.0/beginners_guide/install/install_doc.html#paddlepaddle)中的说明来更新PaddlePaddle。
## 简介
+区域卷积神经网络（RCNN）系列模型为两阶段目标检测器。通过对图像生成候选区域，提取特征，判别特征类别并修正候选框位置。
+RCNN系列目前包含两个代表模型：Faster RCNN，Mask RCNN
-[Faster Rcnn](https://arxiv.org/abs/1506.01497) 是典型的两阶段目标检测器。如下图所示，整体网络可以分为4个主要内容：
-
-
-Faster RCNN 目标检测模型
-
+[Faster RCNN](https://arxiv.org/abs/1506.01497) 整体网络可以分为4个主要内容：
1. 基础卷积层。作为一种卷积神经网络目标检测方法，Faster RCNN首先使用一组基础的卷积网络提取图像的特征图。特征图被后续RPN层和全连接层共享。本示例采用[ResNet-50](https://arxiv.org/abs/1512.03385)作为基础卷积层。
2. 区域生成网络(RPN)。RPN网络用于生成候选区域(proposals)。该层通过一组固定的尺寸和比例得到一组锚点(anchors), 通过softmax判断锚点属于前景或者背景，再利用区域回归修正锚点从而获得精确的候选区域。
3. RoI Align。该层收集输入的特征图和候选区域，将候选区域映射到特征图中并池化为统一大小的区域特征图，送入全连接层判定目标类别, 该层可选用RoIPool和RoIAlign两种方式，在config.py中设置roi\_func。
4. 检测层。利用区域特征图计算候选区域的类别，同时再次通过区域回归获得检测框最终的精确位置。
+[Mask RCNN](https://arxiv.org/abs/1703.06870) 扩展自Faster RCNN，是经典的实例分割模型。
+
+Mask RCNN同样为两阶段框架，第一阶段扫描图像生成候选框；第二阶段根据候选框得到分类结果，边界框，同时在原有Faster RCNN模型基础上添加分割分支，得到掩码结果，实现了掩码和类别预测关系的解藕。
+
+
## 数据准备
在[MS-COCO数据集](http://cocodataset.org/#download)上进行训练，通过如下方式下载数据集。
@@ -61,12 +63,24 @@ Faster RCNN 目标检测模型
数据准备完毕后，可以通过如下的方式启动训练：
+- Faster RCNN
+
python train.py \
--model_save_dir=output/ \
- --pretrained_model=${path_to_pretrain_model}
- --data_dir=${path_to_data}
+ --pretrained_model=${path_to_pretrain_model} \
+ --data_dir=${path_to_data} \
+ --MASK_ON=False
+
+- Mask RCNN
+
+ python train.py \
+ --model_save_dir=output/ \
+ --pretrained_model=${path_to_pretrain_model} \
+ --data_dir=${path_to_data} \
+ --MASK_ON=True
- 通过设置export CUDA\_VISIBLE\_DEVICES=0,1,2,3,4,5,6,7指定8卡GPU训练。
+- 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
- 可选参数见：
python train.py --help
@@ -83,11 +97,10 @@ Faster RCNN 目标检测模型
**训练策略：**
-* 采用momentum优化算法训练Faster RCNN，momentum=0.9。
+* 采用momentum优化算法训练，momentum=0.9。
* 权重衰减系数为0.0001，前500轮学习率从0.00333线性增加至0.01。在120000，160000轮时使用0.1,0.01乘子进行学习率衰减，最大训练180000轮。同时我们也提供了2x模型，该模型采用更多的迭代轮数进行训练，训练360000轮，学习率在240000，320000轮衰减，其他参数不变，训练最大轮数和学习率策略可以在config.py中对max_iter和lr_steps进行设置。
* 非基础卷积层卷积bias学习率为整体学习率2倍。
* 基础卷积层中，affine_layers参数不更新，res2层参数不更新。
-* 使用Nvidia Tesla V100 8卡并行，总共训练时长大约40小时。
## 模型评估
@@ -95,14 +108,27 @@ Faster RCNN 目标检测模型
`eval_coco_map.py`是评估模块的主要执行程序，调用示例如下：
+- Faster RCNN
+
python eval_coco_map.py \
--dataset=coco2017 \
--pretrained_model=${path_to_pretrain_model} \
+ --MASK_ON=False
+
+- Mask RCNN
+
+ python eval_coco_map.py \
+ --dataset=coco2017 \
+ --pretrained_model=${path_to_pretrain_model} \
+ --MASK_ON=True
- 通过设置export CUDA\_VISIBLE\_DEVICES=0指定单卡GPU评估。
+- 通过设置```MASK_ON```选择Faster RCNN和Mask RCNN模型。
下表为模型评估结果：
+Faster RCNN
+
| 模型 | RoI处理方式 | 批量大小 | 迭代次数 | mAP |
| :--------------- | :--------: | :------------: | :------------------: |------: |
| [Fluid RoIPool minibatch padding](http://paddlemodels.bj.bcebos.com/faster_rcnn/model_pool_minibatch_padding.tar.gz) | RoIPool | 8 | 180000 | 0.316 |
@@ -117,6 +143,14 @@ Faster RCNN 目标检测模型
* Fluid RoIAlign no padding: 使用RoIAlign，不对图像做填充处理。
* Fluid RoIAlign no padding 2x: 使用RoIAlign，不对图像做填充处理。训练360000轮，学习率在240000，320000轮衰减。
+Mask RCNN:
+
+| 模型 | 批量大小 | 迭代次数 | box mAP | mask mAP |
+| :--------------- | :--------: | :------------: | :--------: |------: |
+| [Fluid mask no padding](https://paddlemodels.bj.bcebos.com/faster_rcnn/Fluid_mask_no_padding.tar.gz) | 8 | 180000 | 0.359 | 0.314 |
+
+* Fluid mask no padding: 使用RoIAlign，不对图像做填充处理
+
## 模型推断及可视化
模型推断可以获取图像中的物体及其对应的类别，`infer.py`是主要执行程序，调用示例如下：
@@ -131,8 +165,12 @@ Faster RCNN 目标检测模型
下图为模型可视化预测结果：
-
-
-
+
Faster RCNN 预测可视化
+
+
+
+
+Mask RCNN 预测可视化
+
diff --git a/fluid/PaddleCV/faster_rcnn/__init__.py b/fluid/PaddleCV/rcnn/__init__.py
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/__init__.py
rename to fluid/PaddleCV/rcnn/__init__.py
diff --git a/fluid/PaddleCV/faster_rcnn/_ce.py b/fluid/PaddleCV/rcnn/_ce.py
similarity index 78%
rename from fluid/PaddleCV/faster_rcnn/_ce.py
rename to fluid/PaddleCV/rcnn/_ce.py
index 9d5850fd22c3d023eb866fa474b6f6f586ca326e..e331d1bb7cccce5ac914dfa3417fe9090bd9cf99 100644
--- a/fluid/PaddleCV/faster_rcnn/_ce.py
+++ b/fluid/PaddleCV/rcnn/_ce.py
@@ -6,18 +6,19 @@ sys.path.append(os.environ['ceroot'])
from kpi import CostKpi
from kpi import DurationKpi
-
-each_pass_duration_card1_kpi = DurationKpi('each_pass_duration_card1', 0.08, 0, actived=True)
+each_pass_duration_card1_kpi = DurationKpi(
+ 'each_pass_duration_card1', 0.08, 0, actived=True)
train_loss_card1_kpi = CostKpi('train_loss_card1', 0.08, 0)
-each_pass_duration_card4_kpi = DurationKpi('each_pass_duration_card4', 0.08, 0, actived=True)
+each_pass_duration_card4_kpi = DurationKpi(
+ 'each_pass_duration_card4', 0.08, 0, actived=True)
train_loss_card4_kpi = CostKpi('train_loss_card4', 0.08, 0)
tracking_kpis = [
- each_pass_duration_card1_kpi,
- train_loss_card1_kpi,
- each_pass_duration_card4_kpi,
- train_loss_card4_kpi,
- ]
+ each_pass_duration_card1_kpi,
+ train_loss_card1_kpi,
+ each_pass_duration_card4_kpi,
+ train_loss_card4_kpi,
+]
def parse_log(log):
diff --git a/fluid/PaddleCV/faster_rcnn/box_utils.py b/fluid/PaddleCV/rcnn/box_utils.py
similarity index 88%
rename from fluid/PaddleCV/faster_rcnn/box_utils.py
rename to fluid/PaddleCV/rcnn/box_utils.py
index 64d7d96948b856f4ae5c28594e9fb19a3a18480e..bb3fe9c8f0cb261004578abba651ad7210518a22 100644
--- a/fluid/PaddleCV/faster_rcnn/box_utils.py
+++ b/fluid/PaddleCV/rcnn/box_utils.py
@@ -69,6 +69,7 @@ def clip_xyxy_to_image(x1, y1, x2, y2, height, width):
y2 = np.minimum(height - 1., np.maximum(0., y2))
return x1, y1, x2, y2
+
def nms(dets, thresh):
"""Apply classic DPM-style greedy NMS."""
if dets.shape[0] == 0:
@@ -123,3 +124,21 @@ def nms(dets, thresh):
return np.where(suppressed == 0)[0]
+
+def expand_boxes(boxes, scale):
+ """Expand an array of boxes by a given scale."""
+ w_half = (boxes[:, 2] - boxes[:, 0]) * .5
+ h_half = (boxes[:, 3] - boxes[:, 1]) * .5
+ x_c = (boxes[:, 2] + boxes[:, 0]) * .5
+ y_c = (boxes[:, 3] + boxes[:, 1]) * .5
+
+ w_half *= scale
+ h_half *= scale
+
+ boxes_exp = np.zeros(boxes.shape)
+ boxes_exp[:, 0] = x_c - w_half
+ boxes_exp[:, 2] = x_c + w_half
+ boxes_exp[:, 1] = y_c - h_half
+ boxes_exp[:, 3] = y_c + h_half
+
+ return boxes_exp
diff --git a/fluid/PaddleCV/rcnn/colormap.py b/fluid/PaddleCV/rcnn/colormap.py
new file mode 100644
index 0000000000000000000000000000000000000000..8c2447794fc2e9841b30c2cdf11e8fc70d20d764
--- /dev/null
+++ b/fluid/PaddleCV/rcnn/colormap.py
@@ -0,0 +1,61 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+#
+# Based on:
+# --------------------------------------------------------
+# Detectron
+# Copyright (c) 2017-present, Facebook, Inc.
+# Licensed under the Apache License, Version 2.0;
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import numpy as np
+
+
+def colormap(rgb=False):
+ color_list = np.array([
+ 0.000, 0.447, 0.741, 0.850, 0.325, 0.098, 0.929, 0.694, 0.125, 0.494,
+ 0.184, 0.556, 0.466, 0.674, 0.188, 0.301, 0.745, 0.933, 0.635, 0.078,
+ 0.184, 0.300, 0.300, 0.300, 0.600, 0.600, 0.600, 1.000, 0.000, 0.000,
+ 1.000, 0.500, 0.000, 0.749, 0.749, 0.000, 0.000, 1.000, 0.000, 0.000,
+ 0.000, 1.000, 0.667, 0.000, 1.000, 0.333, 0.333, 0.000, 0.333, 0.667,
+ 0.000, 0.333, 1.000, 0.000, 0.667, 0.333, 0.000, 0.667, 0.667, 0.000,
+ 0.667, 1.000, 0.000, 1.000, 0.333, 0.000, 1.000, 0.667, 0.000, 1.000,
+ 1.000, 0.000, 0.000, 0.333, 0.500, 0.000, 0.667, 0.500, 0.000, 1.000,
+ 0.500, 0.333, 0.000, 0.500, 0.333, 0.333, 0.500, 0.333, 0.667, 0.500,
+ 0.333, 1.000, 0.500, 0.667, 0.000, 0.500, 0.667, 0.333, 0.500, 0.667,
+ 0.667, 0.500, 0.667, 1.000, 0.500, 1.000, 0.000, 0.500, 1.000, 0.333,
+ 0.500, 1.000, 0.667, 0.500, 1.000, 1.000, 0.500, 0.000, 0.333, 1.000,
+ 0.000, 0.667, 1.000, 0.000, 1.000, 1.000, 0.333, 0.000, 1.000, 0.333,
+ 0.333, 1.000, 0.333, 0.667, 1.000, 0.333, 1.000, 1.000, 0.667, 0.000,
+ 1.000, 0.667, 0.333, 1.000, 0.667, 0.667, 1.000, 0.667, 1.000, 1.000,
+ 1.000, 0.000, 1.000, 1.000, 0.333, 1.000, 1.000, 0.667, 1.000, 0.167,
+ 0.000, 0.000, 0.333, 0.000, 0.000, 0.500, 0.000, 0.000, 0.667, 0.000,
+ 0.000, 0.833, 0.000, 0.000, 1.000, 0.000, 0.000, 0.000, 0.167, 0.000,
+ 0.000, 0.333, 0.000, 0.000, 0.500, 0.000, 0.000, 0.667, 0.000, 0.000,
+ 0.833, 0.000, 0.000, 1.000, 0.000, 0.000, 0.000, 0.167, 0.000, 0.000,
+ 0.333, 0.000, 0.000, 0.500, 0.000, 0.000, 0.667, 0.000, 0.000, 0.833,
+ 0.000, 0.000, 1.000, 0.000, 0.000, 0.000, 0.143, 0.143, 0.143, 0.286,
+ 0.286, 0.286, 0.429, 0.429, 0.429, 0.571, 0.571, 0.571, 0.714, 0.714,
+ 0.714, 0.857, 0.857, 0.857, 1.000, 1.000, 1.000
+ ]).astype(np.float32)
+ color_list = color_list.reshape((-1, 3)) * 255
+ if not rgb:
+ color_list = color_list[:, ::-1]
+ return color_list
diff --git a/fluid/PaddleCV/faster_rcnn/config.py b/fluid/PaddleCV/rcnn/config.py
similarity index 93%
rename from fluid/PaddleCV/faster_rcnn/config.py
rename to fluid/PaddleCV/rcnn/config.py
index 44b35f7509eeb1adf316e3e725aef8a729bf6499..2a8ebdf7c1871f5863facd6e2138993ed4d7ffd1 100644
--- a/fluid/PaddleCV/faster_rcnn/config.py
+++ b/fluid/PaddleCV/rcnn/config.py
@@ -90,6 +90,9 @@ _C.TRAIN.freeze_at = 2
# min area of ground truth box
_C.TRAIN.gt_min_area = -1
+# Use horizontally-flipped images during training?
+_C.TRAIN.use_flipped = True
+
#
# Inference options
#
@@ -120,7 +123,7 @@ _C.TEST.rpn_post_nms_top_n = 1000
_C.TEST.rpn_min_size = 0.0
# max number of detections
-_C.TEST.detectiions_per_im = 100
+_C.TEST.detections_per_im = 100
# NMS threshold used on RPN proposals
_C.TEST.rpn_nms_thresh = 0.7
@@ -129,6 +132,9 @@ _C.TEST.rpn_nms_thresh = 0.7
# Model options
#
+# Whether use mask rcnn head
+_C.MASK_ON = True
+
# weight for bbox regression targets
_C.bbox_reg_weights = [0.1, 0.1, 0.2, 0.2]
@@ -156,6 +162,15 @@ _C.roi_resolution = 14
# spatial scale
_C.spatial_scale = 1. / 16.
+# resolution to represent mask labels
+_C.resolution = 14
+
+# Number of channels in the mask head
+_C.dim_reduced = 256
+
+# Threshold for converting soft masks to hard masks
+_C.mrcnn_thresh_binarize = 0.5
+
#
# SOLVER options
#
@@ -204,12 +219,6 @@ _C.pixel_means = [102.9801, 115.9465, 122.7717]
# clip box to prevent overflowing
_C.bbox_clip = np.log(1000. / 16.)
-# dataset path
-_C.train_file_list = 'annotations/instances_train2017.json'
-_C.train_data_dir = 'train2017'
-_C.val_file_list = 'annotations/instances_val2017.json'
-_C.val_data_dir = 'val2017'
-
def merge_cfg_from_args(args, mode):
"""Merge config keys, values in args into the global config."""
diff --git a/fluid/PaddleCV/faster_rcnn/data_utils.py b/fluid/PaddleCV/rcnn/data_utils.py
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/data_utils.py
rename to fluid/PaddleCV/rcnn/data_utils.py
diff --git a/fluid/PaddleCV/faster_rcnn/dataset/coco/download.sh b/fluid/PaddleCV/rcnn/dataset/coco/download.sh
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/dataset/coco/download.sh
rename to fluid/PaddleCV/rcnn/dataset/coco/download.sh
diff --git a/fluid/PaddleCV/faster_rcnn/edict.py b/fluid/PaddleCV/rcnn/edict.py
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/edict.py
rename to fluid/PaddleCV/rcnn/edict.py
diff --git a/fluid/PaddleCV/faster_rcnn/eval_coco_map.py b/fluid/PaddleCV/rcnn/eval_coco_map.py
similarity index 53%
rename from fluid/PaddleCV/faster_rcnn/eval_coco_map.py
rename to fluid/PaddleCV/rcnn/eval_coco_map.py
index f8c755a3d0f880a47791f1c43aa161cfa0e5ff98..f3671fa25124040fbb674354c3e6cbbf5231988c 100644
--- a/fluid/PaddleCV/faster_rcnn/eval_coco_map.py
+++ b/fluid/PaddleCV/rcnn/eval_coco_map.py
@@ -18,8 +18,7 @@ from __future__ import print_function
import os
import time
import numpy as np
-from eval_helper import get_nmsed_box
-from eval_helper import get_dt_res
+from eval_helper import *
import paddle
import paddle.fluid as fluid
import reader
@@ -30,21 +29,21 @@ import json
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval, Params
from config import cfg
+from roidbs import DatasetPath
def eval():
- if '2014' in cfg.dataset:
- test_list = 'annotations/instances_val2014.json'
- elif '2017' in cfg.dataset:
- test_list = 'annotations/instances_val2017.json'
+
+ data_path = DatasetPath('val')
+ test_list = data_path.get_file_list()
image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
class_nums = cfg.class_num
devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
devices_num = len(devices.split(","))
total_batch_size = devices_num * cfg.TRAIN.im_per_batch
- cocoGt = COCO(os.path.join(cfg.data_dir, test_list))
- numId_to_catId_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
+ cocoGt = COCO(test_list)
+ num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
category_ids = cocoGt.getCatIds()
label_list = {
item['id']: item['name']
@@ -52,51 +51,88 @@ def eval():
}
label_list[0] = ['background']
- model = model_builder.FasterRCNN(
+ model = model_builder.RCNN(
add_conv_body_func=resnet.add_ResNet50_conv4_body,
add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
use_pyreader=False,
- is_train=False)
+ mode='val')
model.build_model(image_shape)
- rpn_rois, confs, locs = model.eval_out()
+ pred_boxes = model.eval_bbox_out()
+ if cfg.MASK_ON:
+ masks = model.eval_mask_out()
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
+ exe.run(fluid.default_startup_program())
# yapf: disable
if cfg.pretrained_model:
def if_exist(var):
return os.path.exists(os.path.join(cfg.pretrained_model, var.name))
fluid.io.load_vars(exe, cfg.pretrained_model, predicate=if_exist)
+
# yapf: enable
test_reader = reader.test(total_batch_size)
feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
dts_res = []
- fetch_list = [rpn_rois, confs, locs]
+ segms_res = []
+ if cfg.MASK_ON:
+ fetch_list = [pred_boxes, masks]
+ else:
+ fetch_list = [pred_boxes]
+ eval_start = time.time()
for batch_id, batch_data in enumerate(test_reader()):
start = time.time()
im_info = []
for data in batch_data:
im_info.append(data[1])
- rpn_rois_v, confs_v, locs_v = exe.run(
- fetch_list=[v.name for v in fetch_list],
- feed=feeder.feed(batch_data),
- return_numpy=False)
- new_lod, nmsed_out = get_nmsed_box(rpn_rois_v, confs_v, locs_v,
- class_nums, im_info,
- numId_to_catId_map)
-
- dts_res += get_dt_res(total_batch_size, new_lod, nmsed_out, batch_data)
+ results = exe.run(fetch_list=[v.name for v in fetch_list],
+ feed=feeder.feed(batch_data),
+ return_numpy=False)
+
+ pred_boxes_v = results[0]
+ if cfg.MASK_ON:
+ masks_v = results[1]
+
+ new_lod = pred_boxes_v.lod()
+ nmsed_out = pred_boxes_v
+
+ dts_res += get_dt_res(total_batch_size, new_lod[0], nmsed_out,
+ batch_data, num_id_to_cat_id_map)
+
+ if cfg.MASK_ON and np.array(masks_v).shape != (1, 1):
+ segms_out = segm_results(nmsed_out, masks_v, im_info)
+ segms_res += get_segms_res(total_batch_size, new_lod[0], segms_out,
+ batch_data, num_id_to_cat_id_map)
end = time.time()
print('batch id: {}, time: {}'.format(batch_id, end - start))
- with open("detection_result.json", 'w') as outfile:
+ eval_end = time.time()
+ total_time = eval_end - eval_start
+ print('average time of eval is: {}'.format(total_time / (batch_id + 1)))
+ assert len(dts_res) > 0, "The number of valid bbox detected is zero.\n \
+ Please use reasonable model and check input data."
+
+ assert len(segms_res) > 0, "The number of valid mask detected is zero.\n \
+ Please use reasonable model and check input data.."
+
+ with open("detection_bbox_result.json", 'w') as outfile:
json.dump(dts_res, outfile)
- print("start evaluate using coco api")
- cocoDt = cocoGt.loadRes("detection_result.json")
+ print("start evaluate bbox using coco api")
+ cocoDt = cocoGt.loadRes("detection_bbox_result.json")
cocoEval = COCOeval(cocoGt, cocoDt, 'bbox')
cocoEval.evaluate()
cocoEval.accumulate()
cocoEval.summarize()
+ if cfg.MASK_ON:
+ with open("detection_segms_result.json", 'w') as outfile:
+ json.dump(segms_res, outfile)
+ print("start evaluate mask using coco api")
+ cocoDt = cocoGt.loadRes("detection_segms_result.json")
+ cocoEval = COCOeval(cocoGt, cocoDt, 'segm')
+ cocoEval.evaluate()
+ cocoEval.accumulate()
+ cocoEval.summarize()
+
if __name__ == '__main__':
args = parse_args()
diff --git a/fluid/PaddleCV/faster_rcnn/eval_helper.py b/fluid/PaddleCV/rcnn/eval_helper.py
similarity index 53%
rename from fluid/PaddleCV/faster_rcnn/eval_helper.py
rename to fluid/PaddleCV/rcnn/eval_helper.py
index 852b52955915bf268f930ce3b0fa35de5734b1ea..92f76e68ea697ac2fb987ddb10c815a82d746020 100644
--- a/fluid/PaddleCV/faster_rcnn/eval_helper.py
+++ b/fluid/PaddleCV/rcnn/eval_helper.py
@@ -21,6 +21,10 @@ from PIL import Image
from PIL import ImageDraw
from PIL import ImageFont
from config import cfg
+import pycocotools.mask as mask_util
+import six
+from colormap import colormap
+import cv2
def box_decoder(deltas, boxes, weights):
@@ -80,8 +84,7 @@ def clip_tiled_boxes(boxes, im_shape):
return boxes
-def get_nmsed_box(rpn_rois, confs, locs, class_nums, im_info,
- numId_to_catId_map):
+def get_nmsed_box(rpn_rois, confs, locs, class_nums, im_info):
lod = rpn_rois.lod()[0]
rpn_rois_v = np.array(rpn_rois)
variance_v = np.array(cfg.bbox_reg_weights)
@@ -106,38 +109,41 @@ def get_nmsed_box(rpn_rois, confs, locs, class_nums, im_info,
inds = np.where(scores_n[:, j] > cfg.TEST.score_thresh)[0]
scores_j = scores_n[inds, j]
rois_j = rois_n[inds, j * 4:(j + 1) * 4]
- dets_j = np.hstack((rois_j, scores_j[:, np.newaxis])).astype(
+ dets_j = np.hstack((scores_j[:, np.newaxis], rois_j)).astype(
np.float32, copy=False)
keep = box_utils.nms(dets_j, cfg.TEST.nms_thresh)
nms_dets = dets_j[keep, :]
#add labels
- cat_id = numId_to_catId_map[j]
- label = np.array([cat_id for _ in range(len(keep))])
+ label = np.array([j for _ in range(len(keep))])
nms_dets = np.hstack((nms_dets, label[:, np.newaxis])).astype(
np.float32, copy=False)
cls_boxes[j] = nms_dets
# Limit to max_per_image detections **over all classes**
image_scores = np.hstack(
- [cls_boxes[j][:, -2] for j in range(1, class_nums)])
- if len(image_scores) > cfg.TEST.detectiions_per_im:
- image_thresh = np.sort(image_scores)[-cfg.TEST.detectiions_per_im]
+ [cls_boxes[j][:, 1] for j in range(1, class_nums)])
+ if len(image_scores) > cfg.TEST.detections_per_im:
+ image_thresh = np.sort(image_scores)[-cfg.TEST.detections_per_im]
for j in range(1, class_nums):
- keep = np.where(cls_boxes[j][:, -2] >= image_thresh)[0]
+ keep = np.where(cls_boxes[j][:, 1] >= image_thresh)[0]
cls_boxes[j] = cls_boxes[j][keep, :]
im_results_n = np.vstack([cls_boxes[j] for j in range(1, class_nums)])
im_results[i] = im_results_n
new_lod.append(len(im_results_n) + new_lod[-1])
- boxes = im_results_n[:, :-2]
- scores = im_results_n[:, -2]
- labels = im_results_n[:, -1]
+ boxes = im_results_n[:, 2:]
+ scores = im_results_n[:, 1]
+ labels = im_results_n[:, 0]
im_results = np.vstack([im_results[k] for k in range(len(lod) - 1)])
return new_lod, im_results
-def get_dt_res(batch_size, lod, nmsed_out, data):
+def get_dt_res(batch_size, lod, nmsed_out, data, num_id_to_cat_id_map):
dts_res = []
nmsed_out_v = np.array(nmsed_out)
+ if nmsed_out_v.shape == (
+ 1,
+ 1, ):
+ return dts_res
assert (len(lod) == batch_size + 1), \
"Error Lod Tensor offset dimension. Lod({}) vs. batch_size({})"\
.format(len(lod), batch_size)
@@ -150,7 +156,8 @@ def get_dt_res(batch_size, lod, nmsed_out, data):
for j in range(dt_num_this_img):
dt = nmsed_out_v[k]
k = k + 1
- xmin, ymin, xmax, ymax, score, category_id = dt.tolist()
+ num_id, score, xmin, ymin, xmax, ymax = dt.tolist()
+ category_id = num_id_to_cat_id_map[num_id]
w = xmax - xmin + 1
h = ymax - ymin + 1
bbox = [xmin, ymin, w, h]
@@ -164,24 +171,131 @@ def get_dt_res(batch_size, lod, nmsed_out, data):
return dts_res
-def draw_bounding_box_on_image(image_path, nms_out, draw_threshold, label_list):
- image = Image.open(image_path)
+def get_segms_res(batch_size, lod, segms_out, data, num_id_to_cat_id_map):
+ segms_res = []
+ segms_out_v = np.array(segms_out)
+ k = 0
+ for i in range(batch_size):
+ dt_num_this_img = lod[i + 1] - lod[i]
+ image_id = int(data[i][-1])
+ for j in range(dt_num_this_img):
+ dt = segms_out_v[k]
+ k = k + 1
+ segm, num_id, score = dt.tolist()
+ cat_id = num_id_to_cat_id_map[num_id]
+ if six.PY3:
+ if 'counts' in segm:
+ segm['counts'] = segm['counts'].decode("utf8")
+ segm_res = {
+ 'image_id': image_id,
+ 'category_id': cat_id,
+ 'segmentation': segm,
+ 'score': score
+ }
+ segms_res.append(segm_res)
+ return segms_res
+
+
+def draw_bounding_box_on_image(image_path,
+ nms_out,
+ draw_threshold,
+ label_list,
+ num_id_to_cat_id_map,
+ image=None):
+ if image is None:
+ image = Image.open(image_path)
draw = ImageDraw.Draw(image)
im_width, im_height = image.size
- for dt in nms_out:
- xmin, ymin, xmax, ymax, score, category_id = dt.tolist()
+ for dt in np.array(nms_out):
+ num_id, score, xmin, ymin, xmax, ymax = dt.tolist()
+ category_id = num_id_to_cat_id_map[num_id]
if score < draw_threshold:
continue
- bbox = dt[:4]
- xmin, ymin, xmax, ymax = bbox
draw.line(
[(xmin, ymin), (xmin, ymax), (xmax, ymax), (xmax, ymin),
(xmin, ymin)],
- width=4,
+ width=2,
fill='red')
if image.mode == 'RGB':
draw.text((xmin, ymin), label_list[int(category_id)], (255, 255, 0))
image_name = image_path.split('/')[-1]
print("image with bbox drawed saved as {}".format(image_name))
image.save(image_name)
+
+
+def draw_mask_on_image(image_path, segms_out, draw_threshold, alpha=0.7):
+ image = Image.open(image_path)
+ draw = ImageDraw.Draw(image)
+ im_width, im_height = image.size
+ mask_color_id = 0
+ w_ratio = .4
+ image = np.array(image).astype('float32')
+ for dt in np.array(segms_out):
+ segm, num_id, score = dt.tolist()
+ if score < draw_threshold:
+ continue
+ mask = mask_util.decode(segm) * 255
+ color_list = colormap(rgb=True)
+ color_mask = color_list[mask_color_id % len(color_list), 0:3]
+ mask_color_id += 1
+ for c in range(3):
+ color_mask[c] = color_mask[c] * (1 - w_ratio) + w_ratio * 255
+ idx = np.nonzero(mask)
+ image[idx[0], idx[1], :] *= 1.0 - alpha
+ image[idx[0], idx[1], :] += alpha * color_mask
+ image = Image.fromarray(image.astype('uint8'))
+ return image
+
+
+def segm_results(im_results, masks, im_info):
+ im_results = np.array(im_results)
+ class_num = cfg.class_num
+ M = cfg.resolution
+ scale = (M + 2.0) / M
+ lod = masks.lod()[0]
+ masks_v = np.array(masks)
+ boxes = im_results[:, 2:]
+ labels = im_results[:, 0]
+ segms_results = [[] for _ in range(len(lod) - 1)]
+ sum = 0
+ for i in range(len(lod) - 1):
+ im_results_n = im_results[lod[i]:lod[i + 1]]
+ cls_segms = []
+ masks_n = masks_v[lod[i]:lod[i + 1]]
+ boxes_n = boxes[lod[i]:lod[i + 1]]
+ labels_n = labels[lod[i]:lod[i + 1]]
+ im_h = int(round(im_info[i][0] / im_info[i][2]))
+ im_w = int(round(im_info[i][1] / im_info[i][2]))
+ boxes_n = box_utils.expand_boxes(boxes_n, scale)
+ boxes_n = boxes_n.astype(np.int32)
+ padded_mask = np.zeros((M + 2, M + 2), dtype=np.float32)
+ for j in range(len(im_results_n)):
+ class_id = int(labels_n[j])
+ padded_mask[1:-1, 1:-1] = masks_n[j, class_id, :, :]
+
+ ref_box = boxes_n[j, :]
+ w = ref_box[2] - ref_box[0] + 1
+ h = ref_box[3] - ref_box[1] + 1
+ w = np.maximum(w, 1)
+ h = np.maximum(h, 1)
+
+ mask = cv2.resize(padded_mask, (w, h))
+ mask = np.array(mask > cfg.mrcnn_thresh_binarize, dtype=np.uint8)
+ im_mask = np.zeros((im_h, im_w), dtype=np.uint8)
+
+ x_0 = max(ref_box[0], 0)
+ x_1 = min(ref_box[2] + 1, im_w)
+ y_0 = max(ref_box[1], 0)
+ y_1 = min(ref_box[3] + 1, im_h)
+ im_mask[y_0:y_1, x_0:x_1] = mask[(y_0 - ref_box[1]):(y_1 - ref_box[
+ 1]), (x_0 - ref_box[0]):(x_1 - ref_box[0])]
+ sum += im_mask.sum()
+ rle = mask_util.encode(
+ np.array(
+ im_mask[:, :, np.newaxis], order='F'))[0]
+ cls_segms.append(rle)
+ segms_results[i] = np.array(cls_segms)[:, np.newaxis]
+ segms_results = np.vstack([segms_results[k] for k in range(len(lod) - 1)])
+ im_results = np.hstack([segms_results, im_results])
+ return im_results[:, :3]
diff --git a/fluid/PaddleCV/faster_rcnn/image/000000000139.jpg b/fluid/PaddleCV/rcnn/image/000000000139.jpg
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/image/000000000139.jpg
rename to fluid/PaddleCV/rcnn/image/000000000139.jpg
diff --git a/fluid/PaddleCV/rcnn/image/000000000139_mask.jpg b/fluid/PaddleCV/rcnn/image/000000000139_mask.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..47dfa9a435bf81c8585e8100413cfc0d6719754c
Binary files /dev/null and b/fluid/PaddleCV/rcnn/image/000000000139_mask.jpg differ
diff --git a/fluid/PaddleCV/faster_rcnn/image/000000127517.jpg b/fluid/PaddleCV/rcnn/image/000000127517.jpg
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/image/000000127517.jpg
rename to fluid/PaddleCV/rcnn/image/000000127517.jpg
diff --git a/fluid/PaddleCV/rcnn/image/000000127517_mask.jpg b/fluid/PaddleCV/rcnn/image/000000127517_mask.jpg
new file mode 100644
index 0000000000000000000000000000000000000000..c0284591deadf6010bf780acf16124231c42d677
Binary files /dev/null and b/fluid/PaddleCV/rcnn/image/000000127517_mask.jpg differ
diff --git a/fluid/PaddleCV/faster_rcnn/image/000000203864.jpg b/fluid/PaddleCV/rcnn/image/000000203864.jpg
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/image/000000203864.jpg
rename to fluid/PaddleCV/rcnn/image/000000203864.jpg
diff --git a/fluid/PaddleCV/faster_rcnn/image/000000515077.jpg b/fluid/PaddleCV/rcnn/image/000000515077.jpg
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/image/000000515077.jpg
rename to fluid/PaddleCV/rcnn/image/000000515077.jpg
diff --git a/fluid/PaddleCV/faster_rcnn/infer.py b/fluid/PaddleCV/rcnn/infer.py
similarity index 59%
rename from fluid/PaddleCV/faster_rcnn/infer.py
rename to fluid/PaddleCV/rcnn/infer.py
index 3c7200f9de57bbd8d42df9dcb7d72c8fdca7e253..c9e0598c54162f9418b76ff360f67f715f9dfed6 100644
--- a/fluid/PaddleCV/faster_rcnn/infer.py
+++ b/fluid/PaddleCV/rcnn/infer.py
@@ -1,9 +1,7 @@
import os
import time
import numpy as np
-from eval_helper import get_nmsed_box
-from eval_helper import get_dt_res
-from eval_helper import draw_bounding_box_on_image
+from eval_helper import *
import paddle
import paddle.fluid as fluid
import reader
@@ -14,17 +12,16 @@ import json
from pycocotools.coco import COCO
from pycocotools.cocoeval import COCOeval, Params
from config import cfg
+from roidbs import DatasetPath
def infer():
- if '2014' in cfg.dataset:
- test_list = 'annotations/instances_val2014.json'
- elif '2017' in cfg.dataset:
- test_list = 'annotations/instances_val2017.json'
+ data_path = DatasetPath('val')
+ test_list = data_path.get_file_list()
- cocoGt = COCO(os.path.join(cfg.data_dir, test_list))
- numId_to_catId_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
+ cocoGt = COCO(test_list)
+ num_id_to_cat_id_map = {i + 1: v for i, v in enumerate(cocoGt.getCatIds())}
category_ids = cocoGt.getCatIds()
label_list = {
item['id']: item['name']
@@ -34,13 +31,15 @@ def infer():
image_shape = [3, cfg.TEST.max_size, cfg.TEST.max_size]
class_nums = cfg.class_num
- model = model_builder.FasterRCNN(
+ model = model_builder.RCNN(
add_conv_body_func=resnet.add_ResNet50_conv4_body,
add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
use_pyreader=False,
- is_train=False)
+ mode='infer')
model.build_model(image_shape)
- rpn_rois, confs, locs = model.eval_out()
+ pred_boxes = model.eval_bbox_out()
+ if cfg.MASK_ON:
+ masks = model.eval_mask_out()
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
# yapf: disable
@@ -53,17 +52,29 @@ def infer():
feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
dts_res = []
- fetch_list = [rpn_rois, confs, locs]
+ segms_res = []
+ if cfg.MASK_ON:
+ fetch_list = [pred_boxes, masks]
+ else:
+ fetch_list = [pred_boxes]
data = next(infer_reader())
im_info = [data[0][1]]
- rpn_rois_v, confs_v, locs_v = exe.run(
- fetch_list=[v.name for v in fetch_list],
- feed=feeder.feed(data),
- return_numpy=False)
- new_lod, nmsed_out = get_nmsed_box(rpn_rois_v, confs_v, locs_v, class_nums,
- im_info, numId_to_catId_map)
+ result = exe.run(fetch_list=[v.name for v in fetch_list],
+ feed=feeder.feed(data),
+ return_numpy=False)
+ pred_boxes_v = result[0]
+ if cfg.MASK_ON:
+ masks_v = result[1]
+ new_lod = pred_boxes_v.lod()
+ nmsed_out = pred_boxes_v
path = os.path.join(cfg.image_path, cfg.image_name)
- draw_bounding_box_on_image(path, nmsed_out, cfg.draw_threshold, label_list)
+ image = None
+ if cfg.MASK_ON:
+ segms_out = segm_results(nmsed_out, masks_v, im_info)
+ image = draw_mask_on_image(path, segms_out, cfg.draw_threshold)
+
+ draw_bounding_box_on_image(path, nmsed_out, cfg.draw_threshold, label_list,
+ num_id_to_cat_id_map, image)
if __name__ == '__main__':
diff --git a/fluid/PaddleCV/faster_rcnn/learning_rate.py b/fluid/PaddleCV/rcnn/learning_rate.py
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/learning_rate.py
rename to fluid/PaddleCV/rcnn/learning_rate.py
diff --git a/fluid/PaddleCV/faster_rcnn/models/__init__.py b/fluid/PaddleCV/rcnn/models/__init__.py
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/models/__init__.py
rename to fluid/PaddleCV/rcnn/models/__init__.py
diff --git a/fluid/PaddleCV/faster_rcnn/models/model_builder.py b/fluid/PaddleCV/rcnn/models/model_builder.py
similarity index 57%
rename from fluid/PaddleCV/faster_rcnn/models/model_builder.py
rename to fluid/PaddleCV/rcnn/models/model_builder.py
index 9be2f330a62081107d57566962aadc32e1ac687a..1f82d8847155e09a74a3310fc41544759878c0ba 100644
--- a/fluid/PaddleCV/faster_rcnn/models/model_builder.py
+++ b/fluid/PaddleCV/rcnn/models/model_builder.py
@@ -16,23 +16,23 @@ import paddle.fluid as fluid
from paddle.fluid.param_attr import ParamAttr
from paddle.fluid.initializer import Constant
from paddle.fluid.initializer import Normal
+from paddle.fluid.initializer import MSRA
from paddle.fluid.regularizer import L2Decay
from config import cfg
-class FasterRCNN(object):
+class RCNN(object):
def __init__(self,
add_conv_body_func=None,
add_roi_box_head_func=None,
- is_train=True,
+ mode='train',
use_pyreader=True,
use_random=True):
self.add_conv_body_func = add_conv_body_func
self.add_roi_box_head_func = add_roi_box_head_func
- self.is_train = is_train
+ self.mode = mode
self.use_pyreader = use_pyreader
self.use_random = use_random
- #self.py_reader = None
def build_model(self, image_shape):
self.build_input(image_shape)
@@ -41,31 +41,62 @@ class FasterRCNN(object):
self.rpn_heads(body_conv)
# Fast RCNN
self.fast_rcnn_heads(body_conv)
+ if self.mode != 'train':
+ self.eval_bbox()
+ # Mask RCNN
+ if cfg.MASK_ON:
+ self.mask_rcnn_heads(body_conv)
def loss(self):
+ losses = []
# Fast RCNN loss
loss_cls, loss_bbox = self.fast_rcnn_loss()
# RPN loss
rpn_cls_loss, rpn_reg_loss = self.rpn_loss()
- return loss_cls, loss_bbox, rpn_cls_loss, rpn_reg_loss,
+ losses = [loss_cls, loss_bbox, rpn_cls_loss, rpn_reg_loss]
+ rkeys = ['loss', 'loss_cls', 'loss_bbox', \
+ 'loss_rpn_cls', 'loss_rpn_bbox',]
+ if cfg.MASK_ON:
+ loss_mask = self.mask_rcnn_loss()
+ losses = losses + [loss_mask]
+ rkeys = rkeys + ["loss_mask"]
+ loss = fluid.layers.sum(losses)
+ rloss = [loss] + losses
+ return rloss, rkeys
- def eval_out(self):
- cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
- return [self.rpn_rois, cls_prob, self.bbox_pred]
+ def eval_mask_out(self):
+ return self.mask_fcn_logits
+
+ def eval_bbox_out(self):
+ return self.pred_result
def build_input(self, image_shape):
if self.use_pyreader:
+ in_shapes = [[-1] + image_shape, [-1, 4], [-1, 1], [-1, 1],
+ [-1, 3], [-1, 1]]
+ lod_levels = [0, 1, 1, 1, 0, 0]
+ dtypes = [
+ 'float32', 'float32', 'int32', 'int32', 'float32', 'int32'
+ ]
+ if cfg.MASK_ON:
+ in_shapes.append([-1, 2])
+ lod_levels.append(3)
+ dtypes.append('float32')
self.py_reader = fluid.layers.py_reader(
capacity=64,
- shapes=[[-1] + image_shape, [-1, 4], [-1, 1], [-1, 1], [-1, 3],
- [-1, 1]],
- lod_levels=[0, 1, 1, 1, 0, 0],
- dtypes=[
- "float32", "float32", "int32", "int32", "float32", "int32"
- ],
+ shapes=in_shapes,
+ lod_levels=lod_levels,
+ dtypes=dtypes,
use_double_buffer=True)
- self.image, self.gt_box, self.gt_label, self.is_crowd, \
- self.im_info, self.im_id = fluid.layers.read_file(self.py_reader)
+ ins = fluid.layers.read_file(self.py_reader)
+ self.image = ins[0]
+ self.gt_box = ins[1]
+ self.gt_label = ins[2]
+ self.is_crowd = ins[3]
+ self.im_info = ins[4]
+ self.im_id = ins[5]
+ if cfg.MASK_ON:
+ self.gt_masks = ins[6]
else:
self.image = fluid.layers.data(
name='image', shape=image_shape, dtype='float32')
@@ -74,24 +105,57 @@ class FasterRCNN(object):
self.gt_label = fluid.layers.data(
name='gt_label', shape=[1], dtype='int32', lod_level=1)
self.is_crowd = fluid.layers.data(
- name='is_crowd',
- shape=[-1],
- dtype='int32',
- lod_level=1,
- append_batch_size=False)
+ name='is_crowd', shape=[1], dtype='int32', lod_level=1)
self.im_info = fluid.layers.data(
name='im_info', shape=[3], dtype='float32')
self.im_id = fluid.layers.data(
name='im_id', shape=[1], dtype='int32')
+ if cfg.MASK_ON:
+ self.gt_masks = fluid.layers.data(
+ name='gt_masks', shape=[2], dtype='float32', lod_level=3)
def feeds(self):
- if not self.is_train:
+ if self.mode == 'infer':
+ return [self.image, self.im_info]
+ if self.mode == 'val':
return [self.image, self.im_info, self.im_id]
+ if not cfg.MASK_ON:
+ return [
+ self.image, self.gt_box, self.gt_label, self.is_crowd,
+ self.im_info, self.im_id
+ ]
return [
self.image, self.gt_box, self.gt_label, self.is_crowd, self.im_info,
- self.im_id
+ self.im_id, self.gt_masks
]
+ def eval_bbox(self):
+ self.im_scale = fluid.layers.slice(
+ self.im_info, [1], starts=[2], ends=[3])
+ im_scale_lod = fluid.layers.sequence_expand(self.im_scale,
+ self.rpn_rois)
+ boxes = self.rpn_rois / im_scale_lod
+ cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
+ bbox_pred_reshape = fluid.layers.reshape(self.bbox_pred,
+ (-1, cfg.class_num, 4))
+ decoded_box = fluid.layers.box_coder(
+ prior_box=boxes,
+ prior_box_var=cfg.bbox_reg_weights,
+ target_box=bbox_pred_reshape,
+ code_type='decode_center_size',
+ box_normalized=False,
+ axis=1)
+ cliped_box = fluid.layers.box_clip(
+ input=decoded_box, im_info=self.im_info)
+ self.pred_result = fluid.layers.multiclass_nms(
+ bboxes=cliped_box,
+ scores=cls_prob,
+ score_threshold=cfg.TEST.score_thresh,
+ nms_top_k=-1,
+ nms_threshold=cfg.TEST.nms_thresh,
+ keep_top_k=cfg.TEST.detections_per_im,
+ normalized=False)
+
def rpn_heads(self, rpn_input):
# RPN hidden representation
dim_out = rpn_input.shape[1]
@@ -151,13 +215,13 @@ class FasterRCNN(object):
rpn_cls_score_prob = fluid.layers.sigmoid(
self.rpn_cls_score, name='rpn_cls_score_prob')
- param_obj = cfg.TRAIN if self.is_train else cfg.TEST
+ param_obj = cfg.TRAIN if self.mode == 'train' else cfg.TEST
pre_nms_top_n = param_obj.rpn_pre_nms_top_n
post_nms_top_n = param_obj.rpn_post_nms_top_n
nms_thresh = param_obj.rpn_nms_thresh
min_size = param_obj.rpn_min_size
eta = param_obj.rpn_eta
- rpn_rois, rpn_roi_probs = fluid.layers.generate_proposals(
+ self.rpn_rois, self.rpn_roi_probs = fluid.layers.generate_proposals(
scores=rpn_cls_score_prob,
bbox_deltas=self.rpn_bbox_pred,
im_info=self.im_info,
@@ -168,10 +232,9 @@ class FasterRCNN(object):
nms_thresh=nms_thresh,
min_size=min_size,
eta=eta)
- self.rpn_rois = rpn_rois
- if self.is_train:
+ if self.mode == 'train':
outs = fluid.layers.generate_proposal_labels(
- rpn_rois=rpn_rois,
+ rpn_rois=self.rpn_rois,
gt_classes=self.gt_label,
is_crowd=self.is_crowd,
gt_boxes=self.gt_box,
@@ -191,27 +254,28 @@ class FasterRCNN(object):
self.bbox_inside_weights = outs[3]
self.bbox_outside_weights = outs[4]
+ if cfg.MASK_ON:
+ mask_out = fluid.layers.generate_mask_labels(
+ im_info=self.im_info,
+ gt_classes=self.gt_label,
+ is_crowd=self.is_crowd,
+ gt_segms=self.gt_masks,
+ rois=self.rois,
+ labels_int32=self.labels_int32,
+ num_classes=cfg.class_num,
+ resolution=cfg.resolution)
+ self.mask_rois = mask_out[0]
+ self.roi_has_mask_int32 = mask_out[1]
+ self.mask_int32 = mask_out[2]
+
def fast_rcnn_heads(self, roi_input):
- if self.is_train:
+ if self.mode == 'train':
pool_rois = self.rois
else:
pool_rois = self.rpn_rois
- if cfg.roi_func == 'RoIPool':
- pool = fluid.layers.roi_pool(
- input=roi_input,
- rois=pool_rois,
- pooled_height=cfg.roi_resolution,
- pooled_width=cfg.roi_resolution,
- spatial_scale=cfg.spatial_scale)
- elif cfg.roi_func == 'RoIAlign':
- pool = fluid.layers.roi_align(
- input=roi_input,
- rois=pool_rois,
- pooled_height=cfg.roi_resolution,
- pooled_width=cfg.roi_resolution,
- spatial_scale=cfg.spatial_scale,
- sampling_ratio=cfg.sampling_ratio)
- rcnn_out = self.add_roi_box_head_func(pool)
+ self.res5_2_sum = self.add_roi_box_head_func(roi_input, pool_rois)
+ rcnn_out = fluid.layers.pool2d(
+ self.res5_2_sum, pool_type='avg', pool_size=7, name='res5_pool')
self.cls_score = fluid.layers.fc(input=rcnn_out,
size=cfg.class_num,
act=None,
@@ -237,15 +301,87 @@ class FasterRCNN(object):
learning_rate=2.,
regularizer=L2Decay(0.)))
+ def SuffixNet(self, conv5):
+ mask_out = fluid.layers.conv2d_transpose(
+ input=conv5,
+ num_filters=cfg.dim_reduced,
+ filter_size=2,
+ stride=2,
+ act='relu',
+ param_attr=ParamAttr(
+ name='conv5_mask_w', initializer=MSRA(uniform=False)),
+ bias_attr=ParamAttr(
+ name='conv5_mask_b', learning_rate=2., regularizer=L2Decay(0.)))
+ act_func = None
+ if self.mode != 'train':
+ act_func = 'sigmoid'
+ mask_fcn_logits = fluid.layers.conv2d(
+ input=mask_out,
+ num_filters=cfg.class_num,
+ filter_size=1,
+ act=act_func,
+ param_attr=ParamAttr(
+ name='mask_fcn_logits_w', initializer=MSRA(uniform=False)),
+ bias_attr=ParamAttr(
+ name="mask_fcn_logits_b",
+ learning_rate=2.,
+ regularizer=L2Decay(0.)))
+
+ if self.mode != 'train':
+ mask_fcn_logits = fluid.layers.lod_reset(mask_fcn_logits,
+ self.pred_result)
+ return mask_fcn_logits
+
+ def mask_rcnn_heads(self, mask_input):
+ if self.mode == 'train':
+ conv5 = fluid.layers.gather(self.res5_2_sum,
+ self.roi_has_mask_int32)
+ self.mask_fcn_logits = self.SuffixNet(conv5)
+ else:
+ self.eval_bbox()
+ pred_res_shape = fluid.layers.shape(self.pred_result)
+ shape = fluid.layers.reduce_prod(pred_res_shape)
+ shape = fluid.layers.reshape(shape, [1, 1])
+ ones = fluid.layers.fill_constant([1, 1], value=1, dtype='int32')
+ cond = fluid.layers.equal(x=shape, y=ones)
+ ie = fluid.layers.IfElse(cond)
+
+ with ie.true_block():
+ pred_res_null = ie.input(self.pred_result)
+ ie.output(pred_res_null)
+ with ie.false_block():
+ pred_res = ie.input(self.pred_result)
+ pred_boxes = fluid.layers.slice(
+ pred_res, [1], starts=[2], ends=[6])
+ im_scale_lod = fluid.layers.sequence_expand(self.im_scale,
+ pred_boxes)
+ mask_rois = pred_boxes * im_scale_lod
+ conv5 = self.add_roi_box_head_func(mask_input, mask_rois)
+ mask_fcn = self.SuffixNet(conv5)
+ ie.output(mask_fcn)
+ self.mask_fcn_logits = ie()[0]
+
+ def mask_rcnn_loss(self):
+ mask_label = fluid.layers.cast(x=self.mask_int32, dtype='float32')
+ reshape_dim = cfg.class_num * cfg.resolution * cfg.resolution
+ mask_fcn_logits_reshape = fluid.layers.reshape(self.mask_fcn_logits,
+ (-1, reshape_dim))
+
+ loss_mask = fluid.layers.sigmoid_cross_entropy_with_logits(
+ x=mask_fcn_logits_reshape,
+ label=mask_label,
+ ignore_index=-1,
+ normalize=True)
+ loss_mask = fluid.layers.reduce_sum(loss_mask, name='loss_mask')
+ return loss_mask
+
def fast_rcnn_loss(self):
labels_int64 = fluid.layers.cast(x=self.labels_int32, dtype='int64')
labels_int64.stop_gradient = True
- #loss_cls = fluid.layers.softmax_with_cross_entropy(
- # logits=cls_score,
- # label=labels_int64
- # )
- cls_prob = fluid.layers.softmax(self.cls_score, use_cudnn=False)
- loss_cls = fluid.layers.cross_entropy(cls_prob, labels_int64)
+ loss_cls = fluid.layers.softmax_with_cross_entropy(
+ logits=self.cls_score,
+ label=labels_int64,
+ numeric_stable_mode=True, )
loss_cls = fluid.layers.reduce_mean(loss_cls)
loss_bbox = fluid.layers.smooth_l1(
x=self.bbox_pred,
@@ -303,5 +439,4 @@ class FasterRCNN(object):
norm = fluid.layers.reduce_prod(score_shape)
norm.stop_gradient = True
rpn_reg_loss = rpn_reg_loss / norm
-
return rpn_cls_loss, rpn_reg_loss
diff --git a/fluid/PaddleCV/faster_rcnn/models/resnet.py b/fluid/PaddleCV/rcnn/models/resnet.py
similarity index 88%
rename from fluid/PaddleCV/faster_rcnn/models/resnet.py
rename to fluid/PaddleCV/rcnn/models/resnet.py
index e868a1506afe4124036d2ecef4acf83676ba02f9..8093470241b3297c44a2e42b5162e25cac1514be 100644
--- a/fluid/PaddleCV/faster_rcnn/models/resnet.py
+++ b/fluid/PaddleCV/rcnn/models/resnet.py
@@ -160,8 +160,22 @@ def add_ResNet50_conv4_body(body_input):
return res4
-def add_ResNet_roi_conv5_head(head_input):
- res5 = layer_warp(bottleneck, head_input, 512, 3, 2, name="res5")
- res5_pool = fluid.layers.pool2d(
- res5, pool_type='avg', pool_size=7, name='res5_pool')
- return res5_pool
+def add_ResNet_roi_conv5_head(head_input, rois):
+ if cfg.roi_func == 'RoIPool':
+ pool = fluid.layers.roi_pool(
+ input=head_input,
+ rois=rois,
+ pooled_height=cfg.roi_resolution,
+ pooled_width=cfg.roi_resolution,
+ spatial_scale=cfg.spatial_scale)
+ elif cfg.roi_func == 'RoIAlign':
+ pool = fluid.layers.roi_align(
+ input=head_input,
+ rois=rois,
+ pooled_height=cfg.roi_resolution,
+ pooled_width=cfg.roi_resolution,
+ spatial_scale=cfg.spatial_scale,
+ sampling_ratio=cfg.sampling_ratio)
+
+ res5 = layer_warp(bottleneck, pool, 512, 3, 2, name="res5")
+ return res5
diff --git a/fluid/PaddleCV/faster_rcnn/pretrained/download.sh b/fluid/PaddleCV/rcnn/pretrained/download.sh
similarity index 100%
rename from fluid/PaddleCV/faster_rcnn/pretrained/download.sh
rename to fluid/PaddleCV/rcnn/pretrained/download.sh
diff --git a/fluid/PaddleCV/faster_rcnn/profile.py b/fluid/PaddleCV/rcnn/profile.py
similarity index 77%
rename from fluid/PaddleCV/faster_rcnn/profile.py
rename to fluid/PaddleCV/rcnn/profile.py
index 73634bd6773ecb1606a43b297f0966e2d55506b3..92f089b4238a545595723bd8251c6a2e715a59d3 100644
--- a/fluid/PaddleCV/faster_rcnn/profile.py
+++ b/fluid/PaddleCV/rcnn/profile.py
@@ -37,18 +37,15 @@ def train():
devices = os.getenv("CUDA_VISIBLE_DEVICES") or ""
devices_num = len(devices.split(","))
total_batch_size = devices_num * cfg.TRAIN.im_per_batch
- model = model_builder.FasterRCNN(
+ model = model_builder.RCNN(
add_conv_body_func=resnet.add_ResNet50_conv4_body,
add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
use_pyreader=cfg.use_pyreader,
use_random=False)
model.build_model(image_shape)
- loss_cls, loss_bbox, rpn_cls_loss, rpn_reg_loss = model.loss()
- loss_cls.persistable = True
- loss_bbox.persistable = True
- rpn_cls_loss.persistable = True
- rpn_reg_loss.persistable = True
- loss = loss_cls + loss_bbox + rpn_cls_loss + rpn_reg_loss
+ losses, keys = model.loss()
+ loss = losses[0]
+ fetch_list = [loss]
boundaries = cfg.lr_steps
gamma = cfg.lr_gamma
@@ -95,8 +92,6 @@ def train():
train_reader = reader.train(batch_size=total_batch_size, shuffle=False)
feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
- fetch_list = [loss, loss_cls, loss_bbox, rpn_cls_loss, rpn_reg_loss]
-
def run(iterations):
reader_time = []
run_time = []
@@ -109,20 +104,16 @@ def train():
reader_time.append(end_time - start_time)
start_time = time.time()
if cfg.parallel:
- losses = train_exe.run(fetch_list=[v.name for v in fetch_list],
- feed=feeder.feed(data))
+ outs = train_exe.run(fetch_list=[v.name for v in fetch_list],
+ feed=feeder.feed(data))
else:
- losses = exe.run(fluid.default_main_program(),
- fetch_list=[v.name for v in fetch_list],
- feed=feeder.feed(data))
+ outs = exe.run(fluid.default_main_program(),
+ fetch_list=[v.name for v in fetch_list],
+ feed=feeder.feed(data))
end_time = time.time()
run_time.append(end_time - start_time)
total_images += len(data)
-
- lr = np.array(fluid.global_scope().find_var('learning_rate')
- .get_tensor())
- print("Batch {:d}, lr {:.6f}, loss {:.6f} ".format(batch_id, lr[0],
- losses[0][0]))
+ print("Batch {:d}, loss {:.6f} ".format(batch_id, np.mean(outs[0])))
return reader_time, run_time, total_images
def run_pyreader(iterations):
@@ -135,18 +126,16 @@ def train():
for batch_id in range(iterations):
start_time = time.time()
if cfg.parallel:
- losses = train_exe.run(
+ outs = train_exe.run(
fetch_list=[v.name for v in fetch_list])
else:
- losses = exe.run(fluid.default_main_program(),
- fetch_list=[v.name for v in fetch_list])
+ outs = exe.run(fluid.default_main_program(),
+ fetch_list=[v.name for v in fetch_list])
end_time = time.time()
run_time.append(end_time - start_time)
total_images += devices_num
- lr = np.array(fluid.global_scope().find_var('learning_rate')
- .get_tensor())
- print("Batch {:d}, lr {:.6f}, loss {:.6f} ".format(batch_id, lr[
- 0], losses[0][0]))
+ print("Batch {:d}, loss {:.6f} ".format(batch_id,
+ np.mean(outs[0])))
except fluid.core.EOFException:
py_reader.reset()
diff --git a/fluid/PaddleCV/faster_rcnn/reader.py b/fluid/PaddleCV/rcnn/reader.py
similarity index 59%
rename from fluid/PaddleCV/faster_rcnn/reader.py
rename to fluid/PaddleCV/rcnn/reader.py
index 50b3d88b3995442c49833e6f69c7d6f04ea84064..a1c7373bf1495b3f31973b3b7a5bbdfb6830d04c 100644
--- a/fluid/PaddleCV/faster_rcnn/reader.py
+++ b/fluid/PaddleCV/rcnn/reader.py
@@ -27,6 +27,48 @@ from collections import deque
from roidbs import JsonDataset
import data_utils
from config import cfg
+import segm_utils
+
+
+def roidb_reader(roidb, mode):
+ im, im_scales = data_utils.get_image_blob(roidb, mode)
+ im_id = roidb['id']
+ im_height = np.round(roidb['height'] * im_scales)
+ im_width = np.round(roidb['width'] * im_scales)
+ im_info = np.array([im_height, im_width, im_scales], dtype=np.float32)
+ if mode == 'infer':
+ return im, im_info
+ if mode == 'val':
+ return im, im_info, im_id
+
+ gt_boxes = roidb['gt_boxes'].astype('float32')
+ gt_classes = roidb['gt_classes'].astype('int32')
+ is_crowd = roidb['is_crowd'].astype('int32')
+ segms = roidb['segms']
+
+ outs = (im, gt_boxes, gt_classes, is_crowd, im_info, im_id)
+
+ if cfg.MASK_ON:
+ gt_masks = []
+ valid = True
+ segms = roidb['segms']
+ assert len(segms) == is_crowd.shape[0]
+ for i in range(len(roidb['segms'])):
+ segm, iscrowd = segms[i], is_crowd[i]
+ gt_segm = []
+ if iscrowd:
+ gt_segm.append([[0, 0]])
+ else:
+ for poly in segm:
+ if len(poly) == 0:
+ valid = False
+ break
+ gt_segm.append(np.array(poly).reshape(-1, 2))
+ if (not valid) or len(gt_segm) == 0:
+ break
+ gt_masks.append(gt_segm)
+ outs = outs + (gt_masks, )
+ return outs
def coco(mode,
@@ -34,48 +76,16 @@ def coco(mode,
total_batch_size=None,
padding_total=False,
shuffle=False):
- if 'coco2014' in cfg.dataset:
- cfg.train_file_list = 'annotations/instances_train2014.json'
- cfg.train_data_dir = 'train2014'
- cfg.val_file_list = 'annotations/instances_val2014.json'
- cfg.val_data_dir = 'val2014'
- elif 'coco2017' in cfg.dataset:
- cfg.train_file_list = 'annotations/instances_train2017.json'
- cfg.train_data_dir = 'train2017'
- cfg.val_file_list = 'annotations/instances_val2017.json'
- cfg.val_data_dir = 'val2017'
- else:
- raise NotImplementedError('Dataset {} not supported'.format(
- cfg.dataset))
cfg.mean_value = np.array(cfg.pixel_means)[np.newaxis,
np.newaxis, :].astype('float32')
total_batch_size = total_batch_size if total_batch_size else batch_size
if mode != 'infer':
assert total_batch_size % batch_size == 0
- if mode == 'train':
- cfg.train_file_list = os.path.join(cfg.data_dir, cfg.train_file_list)
- cfg.train_data_dir = os.path.join(cfg.data_dir, cfg.train_data_dir)
- elif mode == 'test' or mode == 'infer':
- cfg.val_file_list = os.path.join(cfg.data_dir, cfg.val_file_list)
- cfg.val_data_dir = os.path.join(cfg.data_dir, cfg.val_data_dir)
- json_dataset = JsonDataset(train=(mode == 'train'))
+ json_dataset = JsonDataset(mode)
roidbs = json_dataset.get_roidb()
print("{} on {} with {} roidbs".format(mode, cfg.dataset, len(roidbs)))
- def roidb_reader(roidb, mode):
- im, im_scales = data_utils.get_image_blob(roidb, mode)
- im_id = roidb['id']
- im_height = np.round(roidb['height'] * im_scales)
- im_width = np.round(roidb['width'] * im_scales)
- im_info = np.array([im_height, im_width, im_scales], dtype=np.float32)
- if mode == 'test' or mode == 'infer':
- return im, im_info, im_id
- gt_boxes = roidb['gt_boxes'].astype('float32')
- gt_classes = roidb['gt_classes'].astype('int32')
- is_crowd = roidb['is_crowd'].astype('int32')
- return im, gt_boxes, gt_classes, is_crowd, im_info, im_id
-
def padding_minibatch(batch_data):
if len(batch_data) == 1:
return batch_data
@@ -93,39 +103,53 @@ def coco(mode,
def reader():
if mode == "train":
- roidb_perm = deque(np.random.permutation(roidbs))
+ if shuffle:
+ roidb_perm = deque(np.random.permutation(roidbs))
+ else:
+ roidb_perm = deque(roidbs)
roidb_cur = 0
+ count = 0
batch_out = []
+ device_num = total_batch_size / batch_size
while True:
roidb = roidb_perm[0]
roidb_cur += 1
roidb_perm.rotate(-1)
if roidb_cur >= len(roidbs):
- roidb_perm = deque(np.random.permutation(roidbs))
+ if shuffle:
+ roidb_perm = deque(np.random.permutation(roidbs))
+ else:
+ roidb_perm = deque(roidbs)
roidb_cur = 0
- im, gt_boxes, gt_classes, is_crowd, im_info, im_id = roidb_reader(
- roidb, mode)
- if gt_boxes.shape[0] == 0:
+ # im, gt_boxes, gt_classes, is_crowd, im_info, im_id, gt_masks
+ datas = roidb_reader(roidb, mode)
+ if datas[1].shape[0] == 0:
continue
- batch_out.append(
- (im, gt_boxes, gt_classes, is_crowd, im_info, im_id))
+ if cfg.MASK_ON:
+ if len(datas[-1]) != datas[1].shape[0]:
+ continue
+ batch_out.append(datas)
if not padding_total:
if len(batch_out) == batch_size:
yield padding_minibatch(batch_out)
+ count += 1
batch_out = []
else:
if len(batch_out) == total_batch_size:
batch_out = padding_minibatch(batch_out)
- for i in range(total_batch_size / batch_size):
+ for i in range(device_num):
sub_batch_out = []
for j in range(batch_size):
sub_batch_out.append(batch_out[i * batch_size +
j])
yield sub_batch_out
+ count += 1
sub_batch_out = []
batch_out = []
-
- elif mode == "test":
+ iter_id = count // device_num
+ if iter_id >= cfg.max_iter:
+ return
+ elif mode == "val":
batch_out = []
for roidb in roidbs:
im, im_info, im_id = roidb_reader(roidb, mode)
@@ -140,8 +164,8 @@ def coco(mode,
for roidb in roidbs:
if cfg.image_name not in roidb['image']:
continue
- im, im_info, im_id = roidb_reader(roidb, mode)
- batch_out = [(im, im_info, im_id)]
+ im, im_info = roidb_reader(roidb, mode)
+ batch_out = [(im, im_info)]
yield batch_out
return reader
@@ -153,7 +177,7 @@ def train(batch_size, total_batch_size=None, padding_total=False, shuffle=True):
def test(batch_size, total_batch_size=None, padding_total=False):
- return coco('test', batch_size, total_batch_size, shuffle=False)
+ return coco('val', batch_size, total_batch_size, shuffle=False)
def infer():
diff --git a/fluid/PaddleCV/faster_rcnn/roidbs.py b/fluid/PaddleCV/rcnn/roidbs.py
similarity index 83%
rename from fluid/PaddleCV/faster_rcnn/roidbs.py
rename to fluid/PaddleCV/rcnn/roidbs.py
index b21dc9ed1fb01275aa57b158b0151a56ae297dc7..bd7e581999e6e7052238cadc7472b852ec88dab8 100644
--- a/fluid/PaddleCV/faster_rcnn/roidbs.py
+++ b/fluid/PaddleCV/rcnn/roidbs.py
@@ -36,24 +36,39 @@ import matplotlib
matplotlib.use('Agg')
from pycocotools.coco import COCO
import box_utils
+import segm_utils
from config import cfg
logger = logging.getLogger(__name__)
+class DatasetPath(object):
+ def __init__(self, mode):
+ self.mode = mode
+ mode_name = 'train' if mode == 'train' else 'val'
+ if cfg.dataset != 'coco2014' and cfg.dataset != 'coco2017':
+ raise NotImplementedError('Dataset {} not supported'.format(
+ cfg.dataset))
+ self.sub_name = mode_name + cfg.dataset[-4:]
+
+ def get_data_dir(self):
+ return os.path.join(cfg.data_dir, self.sub_name)
+
+ def get_file_list(self):
+ sfile_list = 'annotations/instances_' + self.sub_name + '.json'
+ return os.path.join(cfg.data_dir, sfile_list)
+
+
class JsonDataset(object):
"""A class representing a COCO json dataset."""
- def __init__(self, train=False):
+ def __init__(self, mode):
print('Creating: {}'.format(cfg.dataset))
self.name = cfg.dataset
- self.is_train = train
- if self.is_train:
- data_dir = cfg.train_data_dir
- file_list = cfg.train_file_list
- else:
- data_dir = cfg.val_data_dir
- file_list = cfg.val_file_list
+ self.is_train = mode == 'train'
+ data_path = DatasetPath(mode)
+ data_dir = data_path.get_data_dir()
+ file_list = data_path.get_file_list()
self.image_directory = data_dir
self.COCO = COCO(file_list)
# Set up dataset classes
@@ -91,8 +106,9 @@ class JsonDataset(object):
end_time = time.time()
print('_add_gt_annotations took {:.3f}s'.format(end_time -
start_time))
- print('Appending horizontally-flipped training examples...')
- self._extend_with_flipped_entries(roidb)
+ if cfg.TRAIN.use_flipped:
+ print('Appending horizontally-flipped training examples...')
+ self._extend_with_flipped_entries(roidb)
print('Loaded dataset: {:s}'.format(self.name))
print('{:d} roidb entries'.format(len(roidb)))
if self.is_train:
@@ -111,6 +127,7 @@ class JsonDataset(object):
entry['gt_classes'] = np.empty((0), dtype=np.int32)
entry['gt_id'] = np.empty((0), dtype=np.int32)
entry['is_crowd'] = np.empty((0), dtype=np.bool)
+ entry['segms'] = []
# Remove unwanted fields that come from the json file (if they exist)
for k in ['date_captured', 'url', 'license', 'file_name']:
if k in entry:
@@ -126,9 +143,15 @@ class JsonDataset(object):
objs = self.COCO.loadAnns(ann_ids)
# Sanitize bboxes -- some are invalid
valid_objs = []
+ valid_segms = []
width = entry['width']
height = entry['height']
for obj in objs:
+ if isinstance(obj['segmentation'], list):
+ # Valid polygons have >= 3 points, so require >= 6 coordinates
+ obj['segmentation'] = [
+ p for p in obj['segmentation'] if len(p) >= 6
+ ]
if obj['area'] < cfg.TRAIN.gt_min_area:
continue
if 'ignore' in obj and obj['ignore'] == 1:
@@ -141,6 +164,8 @@ class JsonDataset(object):
if obj['area'] > 0 and x2 > x1 and y2 > y1:
obj['clean_bbox'] = [x1, y1, x2, y2]
valid_objs.append(obj)
+ valid_segms.append(obj['segmentation'])
+
num_valid_objs = len(valid_objs)
gt_boxes = np.zeros((num_valid_objs, 4), dtype=entry['gt_boxes'].dtype)
@@ -158,6 +183,7 @@ class JsonDataset(object):
entry['gt_classes'] = np.append(entry['gt_classes'], gt_classes)
entry['gt_id'] = np.append(entry['gt_id'], gt_id)
entry['is_crowd'] = np.append(entry['is_crowd'], is_crowd)
+ entry['segms'].extend(valid_segms)
def _extend_with_flipped_entries(self, roidb):
"""Flip each entry in the given roidb and return a new roidb that is the
@@ -175,11 +201,13 @@ class JsonDataset(object):
gt_boxes[:, 2] = width - oldx1 - 1
assert (gt_boxes[:, 2] >= gt_boxes[:, 0]).all()
flipped_entry = {}
- dont_copy = ('gt_boxes', 'flipped')
+ dont_copy = ('gt_boxes', 'flipped', 'segms')
for k, v in entry.items():
if k not in dont_copy:
flipped_entry[k] = v
flipped_entry['gt_boxes'] = gt_boxes
+ flipped_entry['segms'] = segm_utils.flip_segms(
+ entry['segms'], entry['height'], entry['width'])
flipped_entry['flipped'] = True
flipped_roidb.append(flipped_entry)
roidb.extend(flipped_roidb)
diff --git a/fluid/PaddleCV/rcnn/scripts/eval.sh b/fluid/PaddleCV/rcnn/scripts/eval.sh
new file mode 100644
index 0000000000000000000000000000000000000000..922380acf52e594931506e791990319d152d9260
--- /dev/null
+++ b/fluid/PaddleCV/rcnn/scripts/eval.sh
@@ -0,0 +1,17 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+model=$1 # faster_rcnn, mask_rcnn
+if [ "$model" = "faster_rcnn" ]; then
+ mask_on="--MASK_ON False"
+elif [ "$model" = "mask_rcnn" ]; then
+ mask_on="--MASK_ON True"
+else
+ echo "Invalid model provided. Please use one of {faster_rcnn, mask_rcnn}"
+ exit 1
+fi
+
+python -u ../eval_coco_map.py \
+ $mask_on \
+ --pretrained_model=../output/model_iter179999 \
+ --data_dir=../dataset/coco/ \
diff --git a/fluid/PaddleCV/rcnn/scripts/infer.sh b/fluid/PaddleCV/rcnn/scripts/infer.sh
new file mode 100644
index 0000000000000000000000000000000000000000..6f0e02730b9db07568c31a280825f75e321eab64
--- /dev/null
+++ b/fluid/PaddleCV/rcnn/scripts/infer.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+model=$1 # faster_rcnn, mask_rcnn
+if [ "$model" = "faster_rcnn" ]; then
+ mask_on="--MASK_ON False"
+elif [ "$model" = "mask_rcnn" ]; then
+ mask_on="--MASK_ON True"
+else
+ echo "Invalid model provided. Please use one of {faster_rcnn, mask_rcnn}"
+ exit 1
+fi
+
+python -u ../infer.py \
+ $mask_on \
+ --pretrained_model=../output/model_iter179999 \
+ --image_path=../dataset/coco/val2017/ \
+ --image_name=000000000139.jpg \
+ --draw_threshold=0.6
diff --git a/fluid/PaddleCV/rcnn/scripts/train.sh b/fluid/PaddleCV/rcnn/scripts/train.sh
new file mode 100755
index 0000000000000000000000000000000000000000..83c67e6c39121c0fecec5cd7c037d14ab53c619d
--- /dev/null
+++ b/fluid/PaddleCV/rcnn/scripts/train.sh
@@ -0,0 +1,19 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+
+model=$1 # faster_rcnn, mask_rcnn
+if [ "$model" = "faster_rcnn" ]; then
+ mask_on="--MASK_ON False"
+elif [ "$model" = "mask_rcnn" ]; then
+ mask_on="--MASK_ON True"
+else
+ echo "Invalid model provided. Please use one of {faster_rcnn, mask_rcnn}"
+ exit 1
+fi
+
+python -u ../train.py \
+ $mask_on \
+ --model_save_dir=../output/ \
+ --pretrained_model=../imagenet_resnet50_fusebn/ \
+ --data_dir=../dataset/coco/ \
+
diff --git a/fluid/PaddleCV/rcnn/segm_utils.py b/fluid/PaddleCV/rcnn/segm_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..17b72228bc4284dc5936d4a3fda5c2422c4aa958
--- /dev/null
+++ b/fluid/PaddleCV/rcnn/segm_utils.py
@@ -0,0 +1,88 @@
+# Copyright (c) 2018 PaddlePaddle 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://w_idxw.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.
+#
+# Based on:
+# --------------------------------------------------------
+# Detectron
+# Copyright (c) 2017-present, Facebook, Inc.
+# Licensed under the Apache License, Version 2.0;
+# Written by Ross Girshick
+# --------------------------------------------------------
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+from __future__ import unicode_literals
+
+import numpy as np
+import pycocotools.mask as mask_util
+import cv2
+
+
+def is_poly(segm):
+ """Determine if segm is a polygon. Valid segm expected (polygon or RLE)."""
+ assert isinstance(segm, (list, dict)), \
+ 'Invalid segm type: {}'.format(type(segm))
+ return isinstance(segm, list)
+
+
+def segms_to_rle(segms, height, width):
+ rle = segms
+ if isinstance(segms, list):
+ # polygon -- a single object might consist of multiple parts
+ # we merge all parts into one mask rle code
+ rles = mask_util.frPyObjects(segms, height, width)
+ rle = mask_util.merge(rles)
+ elif isinstance(segms['counts'], list):
+ # uncompressed RLE
+ rle = mask_util.frPyObjects(segms, height, width)
+ return rle
+
+
+def segms_to_mask(segms, iscrowd, height, width):
+ print('segms: ', segms)
+ if iscrowd:
+ return [[0 for i in range(width)] for j in range(height)]
+ rle = segms_to_rle(segms, height, width)
+ mask = mask_util.decode(rle)
+ return mask
+
+
+def flip_segms(segms, height, width):
+ """Left/right flip each mask in a list of masks."""
+
+ def _flip_poly(poly, width):
+ flipped_poly = np.array(poly)
+ flipped_poly[0::2] = width - np.array(poly[0::2]) - 1
+ return flipped_poly.tolist()
+
+ def _flip_rle(rle, height, width):
+ if 'counts' in rle and type(rle['counts']) == list:
+ # Magic RLE format handling painfully discovered by looking at the
+ # COCO API showAnns function.
+ rle = mask_util.frPyObjects([rle], height, width)
+ mask = mask_util.decode(rle)
+ mask = mask[:, ::-1, :]
+ rle = mask_util.encode(np.array(mask, order='F', dtype=np.uint8))
+ return rle
+
+ flipped_segms = []
+ for segm in segms:
+ if is_poly(segm):
+ # Polygon format
+ flipped_segms.append([_flip_poly(poly, width) for poly in segm])
+ else:
+ # RLE format
+ flipped_segms.append(_flip_rle(segm, height, width))
+ return flipped_segms
diff --git a/fluid/PaddleCV/faster_rcnn/train.py b/fluid/PaddleCV/rcnn/train.py
similarity index 62%
rename from fluid/PaddleCV/faster_rcnn/train.py
rename to fluid/PaddleCV/rcnn/train.py
index b840d2855c09e1df91601d30df1503a6003aeef5..8404de31d0be066fb41e0cbd44166bd53787c7ee 100644
--- a/fluid/PaddleCV/faster_rcnn/train.py
+++ b/fluid/PaddleCV/rcnn/train.py
@@ -20,7 +20,8 @@ import sys
import numpy as np
import time
import shutil
-from utility import parse_args, print_arguments, SmoothedValue
+from utility import parse_args, print_arguments, SmoothedValue, TrainingStats, now_time
+import collections
import paddle
import paddle.fluid as fluid
@@ -35,7 +36,7 @@ def train():
learning_rate = cfg.learning_rate
image_shape = [3, cfg.TRAIN.max_size, cfg.TRAIN.max_size]
- if cfg.debug or cfg.enable_ce:
+ if cfg.enable_ce:
fluid.default_startup_program().random_seed = 1000
fluid.default_main_program().random_seed = 1000
import random
@@ -49,36 +50,36 @@ def train():
use_random = True
if cfg.enable_ce:
use_random = False
- model = model_builder.FasterRCNN(
+ model = model_builder.RCNN(
add_conv_body_func=resnet.add_ResNet50_conv4_body,
add_roi_box_head_func=resnet.add_ResNet_roi_conv5_head,
use_pyreader=cfg.use_pyreader,
use_random=use_random)
model.build_model(image_shape)
- loss_cls, loss_bbox, rpn_cls_loss, rpn_reg_loss = model.loss()
- loss_cls.persistable = True
- loss_bbox.persistable = True
- rpn_cls_loss.persistable = True
- rpn_reg_loss.persistable = True
- loss = loss_cls + loss_bbox + rpn_cls_loss + rpn_reg_loss
+ losses, keys = model.loss()
+ loss = losses[0]
+ fetch_list = losses
boundaries = cfg.lr_steps
gamma = cfg.lr_gamma
step_num = len(cfg.lr_steps)
values = [learning_rate * (gamma**i) for i in range(step_num + 1)]
+ lr = exponential_with_warmup_decay(
+ learning_rate=learning_rate,
+ boundaries=boundaries,
+ values=values,
+ warmup_iter=cfg.warm_up_iter,
+ warmup_factor=cfg.warm_up_factor)
optimizer = fluid.optimizer.Momentum(
- learning_rate=exponential_with_warmup_decay(
- learning_rate=learning_rate,
- boundaries=boundaries,
- values=values,
- warmup_iter=cfg.warm_up_iter,
- warmup_factor=cfg.warm_up_factor),
+ learning_rate=lr,
regularization=fluid.regularizer.L2Decay(cfg.weight_decay),
momentum=cfg.momentum)
optimizer.minimize(loss)
+ fetch_list = fetch_list + [lr]
- fluid.memory_optimize(fluid.default_main_program())
+ fluid.memory_optimize(
+ fluid.default_main_program(), skip_opt_set=set(fetch_list))
place = fluid.CUDAPlace(0) if cfg.use_gpu else fluid.CPUPlace()
exe = fluid.Executor(place)
@@ -107,7 +108,8 @@ def train():
py_reader = model.py_reader
py_reader.decorate_paddle_reader(train_reader)
else:
- train_reader = reader.train(batch_size=total_batch_size, shuffle=shuffle)
+ train_reader = reader.train(
+ batch_size=total_batch_size, shuffle=shuffle)
feeder = fluid.DataFeeder(place=place, feed_list=model.feeds())
def save_model(postfix):
@@ -116,88 +118,72 @@ def train():
shutil.rmtree(model_path)
fluid.io.save_persistables(exe, model_path)
- fetch_list = [loss, rpn_cls_loss, rpn_reg_loss, loss_cls, loss_bbox]
-
def train_loop_pyreader():
py_reader.start()
- smoothed_loss = SmoothedValue(cfg.log_window)
+ train_stats = TrainingStats(cfg.log_window, keys)
try:
start_time = time.time()
prev_start_time = start_time
- total_time = 0
- last_loss = 0
- every_pass_loss = []
for iter_id in range(cfg.max_iter):
prev_start_time = start_time
start_time = time.time()
- losses = train_exe.run(fetch_list=[v.name for v in fetch_list])
- every_pass_loss.append(np.mean(np.array(losses[0])))
- smoothed_loss.add_value(np.mean(np.array(losses[0])))
- lr = np.array(fluid.global_scope().find_var('learning_rate')
- .get_tensor())
- print("Iter {:d}, lr {:.6f}, loss {:.6f}, time {:.5f}".format(
- iter_id, lr[0],
- smoothed_loss.get_median_value(
- ), start_time - prev_start_time))
- end_time = time.time()
- total_time += end_time - start_time
- last_loss = np.mean(np.array(losses[0]))
-
+ outs = train_exe.run(fetch_list=[v.name for v in fetch_list])
+ stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
+ train_stats.update(stats)
+ logs = train_stats.log()
+ strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
+ now_time(), iter_id,
+ np.mean(outs[-1]), logs, start_time - prev_start_time)
+ print(strs)
sys.stdout.flush()
if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0:
save_model("model_iter{}".format(iter_id))
- # only for ce
+ end_time = time.time()
+ total_time = end_time - start_time
+ last_loss = np.array(outs[0]).mean()
if cfg.enable_ce:
gpu_num = devices_num
epoch_idx = iter_id + 1
loss = last_loss
print("kpis\teach_pass_duration_card%s\t%s" %
- (gpu_num, total_time / epoch_idx))
- print("kpis\ttrain_loss_card%s\t%s" %
- (gpu_num, loss))
-
- except fluid.core.EOFException:
+ (gpu_num, total_time / epoch_idx))
+ print("kpis\ttrain_loss_card%s\t%s" % (gpu_num, loss))
+ except (StopIteration, fluid.core.EOFException):
py_reader.reset()
- return np.mean(every_pass_loss)
def train_loop():
start_time = time.time()
prev_start_time = start_time
start = start_time
- total_time = 0
- last_loss = 0
- every_pass_loss = []
- smoothed_loss = SmoothedValue(cfg.log_window)
+ train_stats = TrainingStats(cfg.log_window, keys)
for iter_id, data in enumerate(train_reader()):
prev_start_time = start_time
start_time = time.time()
- losses = train_exe.run(fetch_list=[v.name for v in fetch_list],
- feed=feeder.feed(data))
- loss_v = np.mean(np.array(losses[0]))
- every_pass_loss.append(loss_v)
- smoothed_loss.add_value(loss_v)
- lr = np.array(fluid.global_scope().find_var('learning_rate')
- .get_tensor())
- end_time = time.time()
- total_time += end_time - start_time
- last_loss = loss_v
- print("Iter {:d}, lr {:.6f}, loss {:.6f}, time {:.5f}".format(
- iter_id, lr[0],
- smoothed_loss.get_median_value(), start_time - prev_start_time))
+ outs = train_exe.run(fetch_list=[v.name for v in fetch_list],
+ feed=feeder.feed(data))
+ stats = {k: np.array(v).mean() for k, v in zip(keys, outs[:-1])}
+ train_stats.update(stats)
+ logs = train_stats.log()
+ strs = '{}, iter: {}, lr: {:.5f}, {}, time: {:.3f}'.format(
+ now_time(), iter_id,
+ np.mean(outs[-1]), logs, start_time - prev_start_time)
+ print(strs)
sys.stdout.flush()
if (iter_id + 1) % cfg.TRAIN.snapshot_iter == 0:
save_model("model_iter{}".format(iter_id))
if (iter_id + 1) == cfg.max_iter:
break
+ end_time = time.time()
+ total_time = end_time - start_time
+ last_loss = np.array(outs[0]).mean()
# only for ce
if cfg.enable_ce:
gpu_num = devices_num
epoch_idx = iter_id + 1
loss = last_loss
print("kpis\teach_pass_duration_card%s\t%s" %
- (gpu_num, total_time / epoch_idx))
- print("kpis\ttrain_loss_card%s\t%s" %
- (gpu_num, loss))
+ (gpu_num, total_time / epoch_idx))
+ print("kpis\ttrain_loss_card%s\t%s" % (gpu_num, loss))
return np.mean(every_pass_loss)
diff --git a/fluid/PaddleCV/faster_rcnn/utility.py b/fluid/PaddleCV/rcnn/utility.py
similarity index 83%
rename from fluid/PaddleCV/faster_rcnn/utility.py
rename to fluid/PaddleCV/rcnn/utility.py
index f428de4c17ac9a6bd1600f52267d6718426adc78..7948bc13fb9c540a92603bca8f423d02aecf81c6 100644
--- a/fluid/PaddleCV/faster_rcnn/utility.py
+++ b/fluid/PaddleCV/rcnn/utility.py
@@ -22,7 +22,9 @@ import sys
import distutils.util
import numpy as np
import six
+import collections
from collections import deque
+import datetime
from paddle.fluid import core
import argparse
import functools
@@ -85,6 +87,37 @@ class SmoothedValue(object):
return np.median(self.deque)
+def now_time():
+ return datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S.%f')
+
+
+class TrainingStats(object):
+ def __init__(self, window_size, stats_keys):
+ self.smoothed_losses_and_metrics = {
+ key: SmoothedValue(window_size)
+ for key in stats_keys
+ }
+
+ def update(self, stats):
+ for k, v in self.smoothed_losses_and_metrics.items():
+ v.add_value(stats[k])
+
+ def get(self, extras=None):
+ stats = collections.OrderedDict()
+ if extras:
+ for k, v in extras.items():
+ stats[k] = v
+ for k, v in self.smoothed_losses_and_metrics.items():
+ stats[k] = round(v.get_median_value(), 3)
+
+ return stats
+
+ def log(self, extras=None):
+ d = self.get(extras)
+ strs = ', '.join(str(dict({x: y})).strip('{}') for x, y in d.items())
+ return strs
+
+
def parse_args():
"""return all args
"""
@@ -108,14 +141,15 @@ def parse_args():
add_arg('learning_rate', float, 0.01, "Learning rate.")
add_arg('max_iter', int, 180000, "Iter number.")
add_arg('log_window', int, 20, "Log smooth window, set 1 for debug, set 20 for train.")
- # FAST RCNN
+ # RCNN
# RPN
add_arg('anchor_sizes', int, [32,64,128,256,512], "The size of anchors.")
add_arg('aspect_ratios', float, [0.5,1.0,2.0], "The ratio of anchors.")
add_arg('variance', float, [1.,1.,1.,1.], "The variance of anchors.")
add_arg('rpn_stride', float, [16.,16.], "Stride of the feature map that RPN is attached.")
add_arg('rpn_nms_thresh', float, 0.7, "NMS threshold used on RPN proposals")
- # TRAIN TEST INFER
+ # TRAIN VAL INFER
+ add_arg('MASK_ON', bool, False, "Option for different models. If False, choose faster_rcnn. If True, choose mask_rcnn")
add_arg('im_per_batch', int, 1, "Minibatch size.")
add_arg('max_size', int, 1333, "The resized image height.")
add_arg('scales', int, [800], "The resized image height.")
@@ -124,7 +158,6 @@ def parse_args():
add_arg('nms_thresh', float, 0.5, "NMS threshold.")
add_arg('score_thresh', float, 0.05, "score threshold for NMS.")
add_arg('snapshot_stride', int, 10000, "save model every snapshot stride.")
- add_arg('debug', bool, False, "Debug mode")
# SINGLE EVAL AND DRAW
add_arg('draw_threshold', float, 0.8, "Confidence threshold to draw bbox.")
add_arg('image_path', str, 'dataset/coco/val2017', "The image path used to inference and visualize.")
@@ -138,5 +171,5 @@ def parse_args():
if 'train' in file_name or 'profile' in file_name:
merge_cfg_from_args(args, 'train')
else:
- merge_cfg_from_args(args, 'test')
+ merge_cfg_from_args(args, 'val')
return args
diff --git a/fluid/PaddleCV/video/.gitignore b/fluid/PaddleCV/video/.gitignore
new file mode 100644
index 0000000000000000000000000000000000000000..7052bdda1c76c2ab1adebd204bdef9ebf1a39755
--- /dev/null
+++ b/fluid/PaddleCV/video/.gitignore
@@ -0,0 +1,5 @@
+checkpoints
+output*
+*.pyc
+*.swp
+*_result
diff --git a/fluid/PaddleCV/video/README.md b/fluid/PaddleCV/video/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..b6b6cdd2dd817268b2fe42f79da8e9e952f96f74
--- /dev/null
+++ b/fluid/PaddleCV/video/README.md
@@ -0,0 +1,130 @@
+
+## 简介
+本教程期望给开发者提供基于PaddlePaddle的便捷、高效的使用深度学习算法解决视频理解、视频编辑、视频生成等一系列模型。目前包含视频分类模型，后续会不断的扩展到其他更多场景。
+
+目前视频分类模型包括:
+
+| 模型 | 类别 | 描述 |
+| :--------------- | :--------: | :------------: |
+| [Attention Cluster](./models/attention_cluster/README.md) | 视频分类| CVPR'18提出的视频多模态特征注意力聚簇融合方法 |
+| [Attention LSTM](./models/attention_lstm/README.md) | 视频分类| 常用模型，速度快精度高 |
+| [NeXtVLAD](./models/nextvlad/README.md) | 视频分类| 2nd-Youtube-8M最优单模型 |
+| [StNet](./models/stnet/README.md) | 视频分类| AAAI'19提出的视频联合时空建模方法 |
+| [TSN](./models/tsn/README.md) | 视频分类| ECCV'16提出的基于2D-CNN经典解决方案 |
+
+### 主要特点
+
+- 包含视频分类方向的多个主流领先模型，其中Attention LSTM，Attention Cluster和NeXtVLAD是比较流行的特征序列模型，TSN和StNet是两个End-to-End的视频分类模型。Attention LSTM模型速度快精度高，NeXtVLAD是2nd-Youtube-8M比赛中最好的单模型, TSN是基于2D-CNN的经典解决方案。Attention Cluster和StNet是百度自研模型，分别发表于CVPR2018和AAAI2019，是Kinetics600比赛第一名中使用到的模型。
+
+- 提供了适合视频分类任务的通用骨架代码，用户可一键式高效配置模型完成训练和评测。
+
+## 安装
+
+在当前模型库运行样例代码需要PadddlePaddle Fluid v.1.2.0或以上的版本。如果你的运行环境中的PaddlePaddle低于此版本，请根据[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)中的说明来更新PaddlePaddle。
+
+## 数据准备
+
+视频模型库使用Youtube-8M和Kinetics数据集, 具体使用方法请参考[数据说明](./dataset/README.md)
+
+## 快速使用
+
+视频模型库提供通用的train/test/infer框架，通过`train.py/test.py/infer.py`指定模型名、模型配置参数等可一键式进行训练和预测。
+
+以StNet模型为例：
+
+单卡训练：
+
+``` bash
+export CUDA_VISIBLE_DEVICES=0
+python train.py --model-name=STNET
+ --config=./configs/stnet.txt
+ --save-dir=checkpoints
+```
+
+多卡训练：
+
+``` bash
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+python train.py --model-name=STNET
+ --config=./configs/stnet.txt
+ --save-dir=checkpoints
+```
+
+视频模型库同时提供了快速训练脚本，脚本位于`scripts/train`目录下，可通过如下命令启动训练:
+
+``` bash
+bash scripts/train/train_stnet.sh
+```
+
+- 请根据`CUDA_VISIBLE_DEVICES`指定卡数修改`config`文件中的`num_gpus`和`batch_size`配置。
+
+## 模型库结构
+
+### 代码结构
+
+```
+configs/
+ stnet.txt
+ tsn.txt
+ ...
+dataset/
+ youtube/
+ kinetics/
+datareader/
+ feature_readeer.py
+ kinetics_reader.py
+ ...
+metrics/
+ kinetics/
+ youtube8m/
+ ...
+models/
+ stnet/
+ tsn/
+ ...
+scripts/
+ train/
+ test/
+train.py
+test.py
+infer.py
+```
+
+- `configs`: 各模型配置文件模板
+- `datareader`: 提供Youtube-8M，Kinetics数据集reader
+- `metrics`: Youtube-8，Kinetics数据集评估脚本
+- `models`: 各模型网络结构构建脚本
+- `scripts`: 各模型快速训练评估脚本
+- `train.py`: 一键式训练脚本，可通过指定模型名，配置文件等一键式启动训练
+- `test.py`: 一键式评估脚本，可通过指定模型名，配置文件，模型权重等一键式启动评估
+- `infer.py`: 一键式推断脚本，可通过指定模型名，配置文件，模型权重，待推断文件列表等一键式启动推断
+
+## Model Zoo
+
+- 基于Youtube-8M数据集模型：
+
+| 模型 | Batch Size | 环境配置 | cuDNN版本 | GAP | 下载链接 |
+| :-------: | :---: | :---------: | :-----: | :----: | :----------: |
+| Attention Cluster | 2048 | 8卡P40 | 7.1 | 0.84 | [model](https://paddlemodels.bj.bcebos.com/video_classification/attention_cluster_youtube8m.tar.gz) |
+| Attention LSTM | 1024 | 8卡P40 | 7.1 | 0.86 | [model](https://paddlemodels.bj.bcebos.com/video_classification/attention_lstm_youtube8m.tar.gz) |
+| NeXtVLAD | 160 | 4卡P40 | 7.1 | 0.87 | [model](https://paddlemodels.bj.bcebos.com/video_classification/nextvlad_youtube8m.tar.gz) |
+
+- 基于Kinetics数据集模型：
+
+| 模型 | Batch Size | 环境配置 | cuDNN版本 | Top-1 | 下载链接 |
+| :-------: | :---: | :---------: | :----: | :----: | :----------: |
+| StNet | 128 | 8卡P40 | 5.1 | 0.69 | [model](https://paddlemodels.bj.bcebos.com/video_classification/stnet_kinetics.tar.gz) |
+| TSN | 256 | 8卡P40 | 7.1 | 0.67 | [model](https://paddlemodels.bj.bcebos.com/video_classification/tsn_kinetics.tar.gz) |
+
+## 参考文献
+
+- [Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification](https://arxiv.org/abs/1711.09550), Xiang Long, Chuang Gan, Gerard de Melo, Jiajun Wu, Xiao Liu, Shilei Wen
+- [Beyond Short Snippets: Deep Networks for Video Classification](https://arxiv.org/abs/1503.08909) Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, George Toderici
+- [NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification](https://arxiv.org/abs/1811.05014), Rongcheng Lin, Jing Xiao, Jianping Fan
+- [StNet:Local and Global Spatial-Temporal Modeling for Human Action Recognition](https://arxiv.org/abs/1811.01549), Dongliang He, Zhichao Zhou, Chuang Gan, Fu Li, Xiao Liu, Yandong Li, Limin Wang, Shilei Wen
+- [Temporal Segment Networks: Towards Good Practices for Deep Action Recognition](https://arxiv.org/abs/1608.00859), Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, Luc Van Gool
+
+## 版本更新
+
+- 3/2019: 新增模型库，发布Attention Cluster，Attention LSTM，NeXtVLAD，StNet，TSN五个视频分类模型。
+
diff --git a/fluid/PaddleCV/video/config.py b/fluid/PaddleCV/video/config.py
new file mode 100755
index 0000000000000000000000000000000000000000..a534536c35c9446ed7dd4139c831757654e02222
--- /dev/null
+++ b/fluid/PaddleCV/video/config.py
@@ -0,0 +1,58 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+
+try:
+ from configparser import ConfigParser
+except:
+ from ConfigParser import ConfigParser
+
+from utils import AttrDict
+
+CONFIG_SECS = [
+ 'train',
+ 'valid',
+ 'test',
+ 'infer',
+ ]
+
+
+def parse_config(cfg_file):
+ parser = ConfigParser()
+ cfg = AttrDict()
+ parser.read(cfg_file)
+ for sec in parser.sections():
+ sec_dict = AttrDict()
+ for k, v in parser.items(sec):
+ try:
+ v = eval(v)
+ except:
+ pass
+ setattr(sec_dict, k, v)
+ setattr(cfg, sec.upper(), sec_dict)
+
+ return cfg
+
+def merge_configs(cfg, sec, args_dict):
+ assert sec in CONFIG_SECS, "invalid config section {}".format(sec)
+ sec_dict = getattr(cfg, sec.upper())
+ for k, v in args_dict.items():
+ if v is None:
+ continue
+ try:
+ if hasattr(sec_dict, k):
+ setattr(sec_dict, k, v)
+ except:
+ pass
+ return cfg
+
diff --git a/fluid/PaddleCV/video/configs/attention_cluster.txt b/fluid/PaddleCV/video/configs/attention_cluster.txt
new file mode 100755
index 0000000000000000000000000000000000000000..0ce7c4b213fe4008e9f02beeb124d11a0ec1f785
--- /dev/null
+++ b/fluid/PaddleCV/video/configs/attention_cluster.txt
@@ -0,0 +1,33 @@
+[MODEL]
+name = "AttentionCluster"
+dataset = "YouTube-8M"
+bone_network = None
+drop_rate = 0.5
+feature_num = 2
+feature_names = ['rgb', 'audio']
+feature_dims = [1024, 128]
+seg_num = 100
+cluster_nums = [32, 32]
+num_classes = 3862
+topk = 20
+
+[TRAIN]
+epoch = 5
+learning_rate = 0.001
+pretrain_base = None
+batch_size = 2048
+use_gpu = True
+num_gpus = 8
+filelist = "dataset/youtube8m/train.list"
+
+[VALID]
+batch_size = 2048
+filelist = "dataset/youtube8m/val.list"
+
+[TEST]
+batch_size = 256
+filelist = "dataset/youtube8m/test.list"
+
+[INFER]
+batch_size = 1
+filelist = "dataset/youtube8m/infer.list"
diff --git a/fluid/PaddleCV/video/configs/attention_lstm.txt b/fluid/PaddleCV/video/configs/attention_lstm.txt
new file mode 100755
index 0000000000000000000000000000000000000000..9154fe2c17282e1066f248a797b50ece080994e7
--- /dev/null
+++ b/fluid/PaddleCV/video/configs/attention_lstm.txt
@@ -0,0 +1,37 @@
+[MODEL]
+name = "AttentionLSTM"
+dataset = "YouTube-8M"
+bone_nework = None
+drop_rate = 0.5
+feature_num = 2
+feature_names = ['rgb', 'audio']
+feature_dims = [1024, 128]
+embedding_size = 512
+lstm_size = 1024
+num_classes = 3862
+topk = 20
+
+[TRAIN]
+epoch = 10
+learning_rate = 0.001
+decay_epochs = [5]
+decay_gamma = 0.1
+weight_decay = 0.0008
+num_samples = 5000000
+pretrain_base = None
+batch_size = 1024
+use_gpu = True
+num_gpus = 8
+filelist = "dataset/youtube8m/train.list"
+
+[VALID]
+batch_size = 1024
+filelist = "dataset/youtube8m/val.list"
+
+[TEST]
+batch_size = 128
+filelist = "dataset/youtube8m/test.list"
+
+[INFER]
+batch_size = 1
+filelist = "dataset/youtube8m/infer.list"
diff --git a/fluid/PaddleCV/video/configs/nextvlad.txt b/fluid/PaddleCV/video/configs/nextvlad.txt
new file mode 100755
index 0000000000000000000000000000000000000000..18779b1f2eaf78cf9db3c25d5fbd991e16e2ed54
--- /dev/null
+++ b/fluid/PaddleCV/video/configs/nextvlad.txt
@@ -0,0 +1,39 @@
+[MODEL]
+name = "NEXTVLAD"
+num_classes = 3862
+topk = 20
+video_feature_size = 1024
+audio_feature_size = 128
+cluster_size = 128
+hidden_size = 2048
+groups = 8
+expansion = 2
+drop_rate = 0.5
+gating_reduction = 8
+eigen_file = "./dataset/youtube8m/yt8m_pca/eigenvals.npy"
+
+[TRAIN]
+epoch = 6
+learning_rate = 0.0002
+lr_boundary_examples = 2000000
+max_iter = 700000
+learning_rate_decay = 0.8
+l2_penalty = 1e-5
+gradient_clip_norm = 1.0
+use_gpu = True
+num_gpus = 4
+batch_size = 160
+filelist = "./dataset/youtube8m/train.list"
+
+[VALID]
+batch_size = 160
+filelist = "./dataset/youtube8m/val.list"
+
+[TEST]
+batch_size = 40
+filelist = "./dataset/youtube8m/test.list"
+
+[INFER]
+batch_size = 1
+filelist = "./dataset/youtube8m/infer.list"
+
diff --git a/fluid/PaddleCV/video/configs/stnet.txt b/fluid/PaddleCV/video/configs/stnet.txt
new file mode 100755
index 0000000000000000000000000000000000000000..ff3e4ddd25202b0d75c4fb53425dfe41a8f4222a
--- /dev/null
+++ b/fluid/PaddleCV/video/configs/stnet.txt
@@ -0,0 +1,51 @@
+[MODEL]
+name = "STNET"
+format = "pkl"
+num_classes = 400
+seg_num = 7
+seglen = 5
+image_mean = [0.485, 0.456, 0.406]
+image_std = [0.229, 0.224, 0.225]
+num_layers = 50
+
+[TRAIN]
+epoch = 60
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 128
+num_gpus = 8
+use_gpu = True
+filelist = "./dataset/kinetics/train.list"
+learning_rate = 0.01
+learning_rate_decay = 0.1
+l2_weight_decay = 1e-4
+momentum = 0.9
+total_videos = 224684
+pretrain_base = "./dataset/pretrained/ResNet50_pretrained"
+
+[VALID]
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 128
+filelist = "./dataset/kinetics/val.list"
+
+[TEST]
+short_size = 256
+target_size = 256
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 16
+filelist = "./dataset/kinetics/test.list"
+
+[INFER]
+short_size = 256
+target_size = 256
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 1
+filelist = "./dataset/kinetics/infer.list"
+
diff --git a/fluid/PaddleCV/video/configs/tsn.txt b/fluid/PaddleCV/video/configs/tsn.txt
new file mode 100755
index 0000000000000000000000000000000000000000..bca5ff349a9792bb07b18c815d7f994419cb82f5
--- /dev/null
+++ b/fluid/PaddleCV/video/configs/tsn.txt
@@ -0,0 +1,50 @@
+[MODEL]
+name = "TSN"
+format = "pkl"
+num_classes = 400
+seg_num = 3
+seglen = 1
+image_mean = [0.485, 0.456, 0.406]
+image_std = [0.229, 0.224, 0.225]
+num_layers = 50
+
+[TRAIN]
+epoch = 45
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 256
+use_gpu = True
+num_gpus = 8
+filelist = "./dataset/kinetics/train.list"
+learning_rate = 0.01
+learning_rate_decay = 0.1
+l2_weight_decay = 1e-4
+momentum = 0.9
+total_videos = 224684
+
+[VALID]
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 256
+filelist = "./dataset/kinetics/val.list"
+
+[TEST]
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 32
+filelist = "./dataset/kinetics/test.list"
+
+[INFER]
+short_size = 256
+target_size = 224
+num_reader_threads = 12
+buf_size = 1024
+batch_size = 1
+filelist = "./dataset/kinetics/infer.list"
+
diff --git a/fluid/PaddleCV/video/datareader/__init__.py b/fluid/PaddleCV/video/datareader/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..8f94515902de7768ea7d9a4c67a5f3c7595facaf
--- /dev/null
+++ b/fluid/PaddleCV/video/datareader/__init__.py
@@ -0,0 +1,12 @@
+from .reader_utils import regist_reader, get_reader
+from .feature_reader import FeatureReader
+from .kinetics_reader import KineticsReader
+from .nonlocal_reader import NonlocalReader
+
+regist_reader("ATTENTIONCLUSTER", FeatureReader)
+regist_reader("NEXTVLAD", FeatureReader)
+regist_reader("ATTENTIONLSTM", FeatureReader)
+regist_reader("TSN", KineticsReader)
+regist_reader("TSM", KineticsReader)
+regist_reader("STNET", KineticsReader)
+regist_reader("NONLOCAL", NonlocalReader)
diff --git a/fluid/PaddleCV/video/datareader/feature_reader.py b/fluid/PaddleCV/video/datareader/feature_reader.py
new file mode 100644
index 0000000000000000000000000000000000000000..9f465c09474446529dd6804a26d3c71204b2fcfa
--- /dev/null
+++ b/fluid/PaddleCV/video/datareader/feature_reader.py
@@ -0,0 +1,135 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 sys
+from .reader_utils import DataReader
+try:
+ import cPickle as pickle
+ from cStringIO import StringIO
+except ImportError:
+ import pickle
+ from io import BytesIO
+import numpy as np
+import random
+
+python_ver = sys.version_info
+
+
+class FeatureReader(DataReader):
+ """
+ Data reader for youtube-8M dataset, which was stored as features extracted by prior networks
+ This is for the three models: lstm, attention cluster, nextvlad
+
+ dataset cfg: num_classes
+ batch_size
+ list
+ NextVlad only: eigen_file
+ """
+
+ def __init__(self, name, mode, cfg):
+ self.name = name
+ self.mode = mode
+ self.num_classes = cfg.MODEL.num_classes
+
+ # set batch size and file list
+ self.batch_size = cfg[mode.upper()]['batch_size']
+ self.filelist = cfg[mode.upper()]['filelist']
+ self.eigen_file = cfg.MODEL.get('eigen_file', None)
+ self.seg_num = cfg.MODEL.get('seg_num', None)
+
+ def create_reader(self):
+ fl = open(self.filelist).readlines()
+ fl = [line.strip() for line in fl if line.strip() != '']
+ if self.mode == 'train':
+ random.shuffle(fl)
+
+ def reader():
+ batch_out = []
+ for filepath in fl:
+ if python_ver < (3, 0):
+ data = pickle.load(open(filepath, 'rb'))
+ else:
+ data = pickle.load(open(filepath, 'rb'), encoding='bytes')
+ indexes = list(range(len(data)))
+ if self.mode == 'train':
+ random.shuffle(indexes)
+ for i in indexes:
+ record = data[i]
+ nframes = record[b'nframes']
+ rgb = record[b'feature'].astype(float)
+ audio = record[b'audio'].astype(float)
+ if self.mode != 'infer':
+ label = record[b'label']
+ one_hot_label = make_one_hot(label, self.num_classes)
+ video = record[b'video']
+
+ rgb = rgb[0:nframes, :]
+ audio = audio[0:nframes, :]
+
+ rgb = dequantize(
+ rgb, max_quantized_value=2., min_quantized_value=-2.)
+ audio = dequantize(
+ audio, max_quantized_value=2, min_quantized_value=-2)
+
+ if self.name == 'NEXTVLAD':
+ # add the effect of eigen values
+ eigen_file = self.eigen_file
+ eigen_val = np.sqrt(np.load(eigen_file)
+ [:1024, 0]).astype(np.float32)
+ eigen_val = eigen_val + 1e-4
+ rgb = (rgb - 4. / 512) * eigen_val
+ if self.name == 'ATTENTIONCLUSTER':
+ sample_inds = generate_random_idx(rgb.shape[0],
+ self.seg_num)
+ rgb = rgb[sample_inds]
+ audio = audio[sample_inds]
+ if self.mode != 'infer':
+ batch_out.append((rgb, audio, one_hot_label))
+ else:
+ batch_out.append((rgb, audio, video))
+ if len(batch_out) == self.batch_size:
+ yield batch_out
+ batch_out = []
+
+ return reader
+
+
+def dequantize(feat_vector, max_quantized_value=2., min_quantized_value=-2.):
+ """
+ Dequantize the feature from the byte format to the float format
+ """
+
+ assert max_quantized_value > min_quantized_value
+ quantized_range = max_quantized_value - min_quantized_value
+ scalar = quantized_range / 255.0
+ bias = (quantized_range / 512.0) + min_quantized_value
+
+ return feat_vector * scalar + bias
+
+
+def make_one_hot(label, dim=3862):
+ one_hot_label = np.zeros(dim)
+ one_hot_label = one_hot_label.astype(float)
+ for ind in label:
+ one_hot_label[int(ind)] = 1
+ return one_hot_label
+
+
+def generate_random_idx(feature_len, seg_num):
+ idxs = []
+ stride = float(feature_len) / seg_num
+ for i in range(seg_num):
+ pos = (i + np.random.random()) * stride
+ idxs.append(min(feature_len - 1, int(pos)))
+ return idxs
diff --git a/fluid/PaddleCV/video/datareader/kinetics_reader.py b/fluid/PaddleCV/video/datareader/kinetics_reader.py
new file mode 100644
index 0000000000000000000000000000000000000000..c7bbf17241383ffc32778330db9ac78308683b46
--- /dev/null
+++ b/fluid/PaddleCV/video/datareader/kinetics_reader.py
@@ -0,0 +1,353 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import sys
+import math
+import random
+import functools
+try:
+ import cPickle as pickle
+ from cStringIO import StringIO
+except ImportError:
+ import pickle
+ from io import BytesIO
+import numpy as np
+import paddle
+from PIL import Image, ImageEnhance
+import logging
+
+from .reader_utils import DataReader
+
+logger = logging.getLogger(__name__)
+python_ver = sys.version_info
+
+
+class KineticsReader(DataReader):
+ """
+ Data reader for kinetics dataset of two format mp4 and pkl.
+ 1. mp4, the original format of kinetics400
+ 2. pkl, the mp4 was decoded previously and stored as pkl
+ In both case, load the data, and then get the frame data in the form of numpy and label as an integer.
+ dataset cfg: format
+ num_classes
+ seg_num
+ short_size
+ target_size
+ num_reader_threads
+ buf_size
+ image_mean
+ image_std
+ batch_size
+ list
+ """
+
+ def __init__(self, name, mode, cfg):
+ self.name = name
+ self.mode = mode
+ self.format = cfg.MODEL.format
+ self.num_classes = cfg.MODEL.num_classes
+ self.seg_num = cfg.MODEL.seg_num
+ self.seglen = cfg.MODEL.seglen
+ self.short_size = cfg[mode.upper()]['short_size']
+ self.target_size = cfg[mode.upper()]['target_size']
+ self.num_reader_threads = cfg[mode.upper()]['num_reader_threads']
+ self.buf_size = cfg[mode.upper()]['buf_size']
+
+ self.img_mean = np.array(cfg.MODEL.image_mean).reshape(
+ [3, 1, 1]).astype(np.float32)
+ self.img_std = np.array(cfg.MODEL.image_std).reshape(
+ [3, 1, 1]).astype(np.float32)
+ # set batch size and file list
+ self.batch_size = cfg[mode.upper()]['batch_size']
+ self.filelist = cfg[mode.upper()]['filelist']
+
+ def create_reader(self):
+ _reader = _reader_creator(self.filelist, self.mode, seg_num=self.seg_num, seglen = self.seglen, \
+ short_size = self.short_size, target_size = self.target_size, \
+ img_mean = self.img_mean, img_std = self.img_std, \
+ shuffle = (self.mode == 'train'), \
+ num_threads = self.num_reader_threads, \
+ buf_size = self.buf_size, format = self.format)
+
+ def _batch_reader():
+ batch_out = []
+ for imgs, label in _reader():
+ if imgs is None:
+ continue
+ batch_out.append((imgs, label))
+ if len(batch_out) == self.batch_size:
+ yield batch_out
+ batch_out = []
+
+ return _batch_reader
+
+
+def _reader_creator(pickle_list,
+ mode,
+ seg_num,
+ seglen,
+ short_size,
+ target_size,
+ img_mean,
+ img_std,
+ shuffle=False,
+ num_threads=1,
+ buf_size=1024,
+ format='pkl'):
+ def reader():
+ with open(pickle_list) as flist:
+ lines = [line.strip() for line in flist]
+ if shuffle:
+ random.shuffle(lines)
+ for line in lines:
+ pickle_path = line.strip()
+ yield [pickle_path]
+
+ if format == 'pkl':
+ decode_func = decode_pickle
+ elif format == 'mp4':
+ decode_func = decode_mp4
+ else:
+ raise "Not implemented format {}".format(format)
+
+ mapper = functools.partial(
+ decode_func,
+ mode=mode,
+ seg_num=seg_num,
+ seglen=seglen,
+ short_size=short_size,
+ target_size=target_size,
+ img_mean=img_mean,
+ img_std=img_std)
+
+ return paddle.reader.xmap_readers(mapper, reader, num_threads, buf_size)
+
+
+def decode_mp4(sample, mode, seg_num, seglen, short_size, target_size, img_mean,
+ img_std):
+ sample = sample[0].split(' ')
+ mp4_path = sample[0]
+ # when infer, we store vid as label
+ label = int(sample[1])
+ try:
+ imgs = mp4_loader(mp4_path, seg_num, seglen, mode)
+ if len(imgs) < 1:
+ logger.error('{} frame length {} less than 1.'.format(mp4_path,
+ len(imgs)))
+ return None, None
+ except:
+ logger.error('Error when loading {}'.format(mp4_path))
+ return None, None
+
+ return imgs_transform(imgs, label, mode, seg_num, seglen, \
+ short_size, target_size, img_mean, img_std)
+
+
+def decode_pickle(sample, mode, seg_num, seglen, short_size, target_size,
+ img_mean, img_std):
+ pickle_path = sample[0]
+ try:
+ if python_ver < (3, 0):
+ data_loaded = pickle.load(open(pickle_path, 'rb'))
+ else:
+ data_loaded = pickle.load(open(pickle_path, 'rb'), encoding='bytes')
+
+ vid, label, frames = data_loaded
+ if len(frames) < 1:
+ logger.error('{} frame length {} less than 1.'.format(pickle_path,
+ len(frames)))
+ return None, None
+ except:
+ logger.info('Error when loading {}'.format(pickle_path))
+ return None, None
+
+ if mode == 'train' or mode == 'valid' or mode == 'test':
+ ret_label = label
+ elif mode == 'infer':
+ ret_label = vid
+
+ imgs = video_loader(frames, seg_num, seglen, mode)
+ return imgs_transform(imgs, ret_label, mode, seg_num, seglen, \
+ short_size, target_size, img_mean, img_std)
+
+
+def imgs_transform(imgs, label, mode, seg_num, seglen, short_size, target_size,
+ img_mean, img_std):
+ imgs = group_scale(imgs, short_size)
+
+ if mode == 'train':
+ imgs = group_random_crop(imgs, target_size)
+ imgs = group_random_flip(imgs)
+ else:
+ imgs = group_center_crop(imgs, target_size)
+
+ np_imgs = (np.array(imgs[0]).astype('float32').transpose(
+ (2, 0, 1))).reshape(1, 3, target_size, target_size) / 255
+ for i in range(len(imgs) - 1):
+ img = (np.array(imgs[i + 1]).astype('float32').transpose(
+ (2, 0, 1))).reshape(1, 3, target_size, target_size) / 255
+ np_imgs = np.concatenate((np_imgs, img))
+ imgs = np_imgs
+ imgs -= img_mean
+ imgs /= img_std
+ imgs = np.reshape(imgs, (seg_num, seglen * 3, target_size, target_size))
+
+ return imgs, label
+
+
+def group_random_crop(img_group, target_size):
+ w, h = img_group[0].size
+ th, tw = target_size, target_size
+
+ assert (w >= target_size) and (h >= target_size), \
+ "image width({}) and height({}) should be larger than crop size".format(w, h, target_size)
+
+ out_images = []
+ x1 = random.randint(0, w - tw)
+ y1 = random.randint(0, h - th)
+
+ for img in img_group:
+ if w == tw and h == th:
+ out_images.append(img)
+ else:
+ out_images.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+ return out_images
+
+
+def group_random_flip(img_group):
+ v = random.random()
+ if v < 0.5:
+ ret = [img.transpose(Image.FLIP_LEFT_RIGHT) for img in img_group]
+ return ret
+ else:
+ return img_group
+
+
+def group_center_crop(img_group, target_size):
+ img_crop = []
+ for img in img_group:
+ w, h = img.size
+ th, tw = target_size, target_size
+ assert (w >= target_size) and (h >= target_size), \
+ "image width({}) and height({}) should be larger than crop size".format(w, h, target_size)
+ x1 = int(round((w - tw) / 2.))
+ y1 = int(round((h - th) / 2.))
+ img_crop.append(img.crop((x1, y1, x1 + tw, y1 + th)))
+
+ return img_crop
+
+
+def group_scale(imgs, target_size):
+ resized_imgs = []
+ for i in range(len(imgs)):
+ img = imgs[i]
+ w, h = img.size
+ if (w <= h and w == target_size) or (h <= w and h == target_size):
+ resized_imgs.append(img)
+ continue
+
+ if w < h:
+ ow = target_size
+ oh = int(target_size * 4.0 / 3.0)
+ resized_imgs.append(img.resize((ow, oh), Image.BILINEAR))
+ else:
+ oh = target_size
+ ow = int(target_size * 4.0 / 3.0)
+ resized_imgs.append(img.resize((ow, oh), Image.BILINEAR))
+
+ return resized_imgs
+
+
+def imageloader(buf):
+ if isinstance(buf, str):
+ img = Image.open(StringIO(buf))
+ else:
+ img = Image.open(BytesIO(buf))
+
+ return img.convert('RGB')
+
+
+def video_loader(frames, nsample, seglen, mode):
+ videolen = len(frames)
+ average_dur = int(videolen / nsample)
+
+ imgs = []
+ for i in range(nsample):
+ idx = 0
+ if mode == 'train':
+ if average_dur >= seglen:
+ idx = random.randint(0, average_dur - seglen)
+ idx += i * average_dur
+ elif average_dur >= 1:
+ idx += i * average_dur
+ else:
+ idx = i
+ else:
+ if average_dur >= seglen:
+ idx = (average_dur - seglen) // 2
+ idx += i * average_dur
+ elif average_dur >= 1:
+ idx += i * average_dur
+ else:
+ idx = i
+
+ for jj in range(idx, idx + seglen):
+ imgbuf = frames[int(jj % videolen)]
+ img = imageloader(imgbuf)
+ imgs.append(img)
+
+ return imgs
+
+
+def mp4_loader(filepath, nsample, seglen, mode):
+ cap = cv2.VideoCapture(filepath)
+ videolen = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+ average_dur = int(videolen / nsample)
+ sampledFrames = []
+ for i in range(videolen):
+ ret, frame = cap.read()
+ # maybe first frame is empty
+ if ret == False:
+ continue
+ img = frame[:, :, ::-1]
+ sampledFrames.append(img)
+
+ imgs = []
+ for i in range(nsample):
+ idx = 0
+ if mode == 'train':
+ if average_dur >= seglen:
+ idx = random.randint(0, average_dur - seglen)
+ idx += i * average_dur
+ elif average_dur >= 1:
+ idx += i * average_dur
+ else:
+ idx = i
+ else:
+ if average_dur >= seglen:
+ idx = (average_dur - 1) // 2
+ idx += i * average_dur
+ elif average_dur >= 1:
+ idx += i * average_dur
+ else:
+ idx = i
+
+ for jj in range(idx, idx + seglen):
+ imgbuf = sampledFrames[int(jj % videolen)]
+ img = Image.fromarray(imgbuf, mode='RGB')
+ imgs.append(img)
+
+ return imgs
diff --git a/fluid/PaddleCV/video/datareader/nonlocal_reader.py b/fluid/PaddleCV/video/datareader/nonlocal_reader.py
new file mode 100644
index 0000000000000000000000000000000000000000..8edb5ac2ab7fa279e6f04830c8543b9846f23b38
--- /dev/null
+++ b/fluid/PaddleCV/video/datareader/nonlocal_reader.py
@@ -0,0 +1,338 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import random
+import time
+import multiprocessing
+import numpy as np
+import cv2
+import logging
+
+from .reader_utils import DataReader
+
+logger = logging.getLogger(__name__)
+
+
+class NonlocalReader(DataReader):
+ """
+ Data reader for kinetics dataset, which read mp4 file and decode into numpy.
+ This is for nonlocal neural network model.
+ cfg: num_classes
+ num_reader_threads
+ image_mean
+ image_std
+ batch_size
+ list
+ crop_size
+ sample_rate
+ video_length
+ jitter_scales
+ Test only cfg: num_test_clips
+ use_multi_crop
+ """
+
+ def __init__(self, name, mode, cfg):
+ self.name = name
+ self.mode = mode
+ self.cfg = cfg
+
+ def create_reader(self):
+ cfg = self.cfg
+ mode = self.mode
+ num_reader_threads = cfg[mode.upper()]['num_reader_threads']
+ assert num_reader_threads >=1, \
+ "number of reader threads({}) should be a positive integer".format(num_reader_threads)
+ if num_reader_threads == 1:
+ reader_func = make_reader
+ else:
+ reader_func = make_multi_reader
+
+ dataset_args = {}
+ dataset_args['image_mean'] = cfg.MODEL.image_mean
+ dataset_args['image_std'] = cfg.MODEL.image_std
+ dataset_args['crop_size'] = cfg[mode.upper()]['crop_size']
+ dataset_args['sample_rate'] = cfg[mode.upper()]['sample_rate']
+ dataset_args['video_length'] = cfg[mode.upper()]['video_length']
+ dataset_args['min_size'] = cfg[mode.upper()]['jitter_scales'][0]
+ dataset_args['max_size'] = cfg[mode.upper()]['jitter_scales'][1]
+ dataset_args['num_reader_threads'] = num_reader_threads
+ filelist = cfg[mode.upper()]['list']
+ batch_size = cfg[mode.upper()]['batch_size']
+
+ if self.mode == 'train':
+ sample_times = 1
+ return reader_func(filelist, batch_size, sample_times, True, True,
+ **dataset_args)
+ elif self.mode == 'valid':
+ sample_times = 1
+ return reader_func(filelist, batch_size, sample_times, False, False,
+ **dataset_args)
+ elif self.mode == 'test':
+ sample_times = cfg['TEST']['num_test_clips']
+ if cfg['TEST']['use_multi_crop'] == 1:
+ sample_times = int(sample_times / 3)
+ if cfg['TEST']['use_multi_crop'] == 2:
+ sample_times = int(sample_times / 6)
+ return reader_func(filelist, batch_size, sample_times, False, False,
+ **dataset_args)
+ else:
+ logger.info('Not implemented')
+ raise NotImplementedError
+
+
+def video_fast_get_frame(video_path,
+ sampling_rate=1,
+ length=64,
+ start_frm=-1,
+ sample_times=1):
+ cap = cv2.VideoCapture(video_path)
+ frame_cnt = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
+ width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
+ height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+
+ sampledFrames = []
+
+ video_output = np.ndarray(shape=[length, height, width, 3], dtype=np.uint8)
+
+ use_start_frm = start_frm
+ if start_frm < 0:
+ if (frame_cnt - length * sampling_rate > 0):
+ use_start_frm = random.randint(0,
+ frame_cnt - length * sampling_rate)
+ else:
+ use_start_frm = 0
+ else:
+ frame_gaps = float(frame_cnt) / float(sample_times)
+ use_start_frm = int(frame_gaps * start_frm) % frame_cnt
+
+ for i in range(frame_cnt):
+ ret, frame = cap.read()
+ # maybe first frame is empty
+ if ret == False:
+ continue
+ img = frame[:, :, ::-1]
+ sampledFrames.append(img)
+
+ for idx in range(length):
+ i = use_start_frm + idx * sampling_rate
+ i = i % len(sampledFrames)
+ video_output[idx] = sampledFrames[i]
+
+ cap.release()
+ return video_output
+
+
+def apply_resize(rgbdata, min_size, max_size):
+ length, height, width, channel = rgbdata.shape
+ ratio = 1.0
+ # generate random scale between [min_size, max_size]
+ if min_size == max_size:
+ side_length = min_size
+ else:
+ side_length = np.random.randint(min_size, max_size)
+ if height > width:
+ ratio = float(side_length) / float(width)
+ else:
+ ratio = float(side_length) / float(height)
+ out_height = int(height * ratio)
+ out_width = int(width * ratio)
+ outdata = np.zeros(
+ (length, out_height, out_width, channel), dtype=rgbdata.dtype)
+ for i in range(length):
+ outdata[i] = cv2.resize(rgbdata[i], (out_width, out_height))
+ return outdata
+
+
+def crop_mirror_transform(rgbdata,
+ mean,
+ std,
+ cropsize=224,
+ use_mirror=True,
+ center_crop=False,
+ spatial_pos=-1):
+ channel, length, height, width = rgbdata.shape
+ assert height >= cropsize, "crop size should not be larger than video height"
+ assert width >= cropsize, "crop size should not be larger than video width"
+ # crop to specific scale
+ if center_crop:
+ h_off = int((height - cropsize) / 2)
+ w_off = int((width - cropsize) / 2)
+ if spatial_pos >= 0:
+ now_pos = spatial_pos % 3
+ if h_off > 0:
+ h_off = h_off * now_pos
+ else:
+ w_off = w_off * now_pos
+ else:
+ h_off = np.random.randint(0, height - cropsize)
+ w_off = np.random.randint(0, width - cropsize)
+ outdata = np.zeros(
+ (channel, length, cropsize, cropsize), dtype=rgbdata.dtype)
+ outdata[:, :, :, :] = rgbdata[:, :, h_off:h_off + cropsize, w_off:w_off +
+ cropsize]
+ # apply mirror
+ mirror_indicator = (np.random.rand() > 0.5)
+ mirror_me = use_mirror and mirror_indicator
+ if spatial_pos > 0:
+ mirror_me = (int(spatial_pos / 3) > 0)
+ if mirror_me:
+ outdata = outdata[:, :, :, ::-1]
+ # substract mean and divide std
+ outdata = outdata.astype(np.float32)
+ outdata = (outdata - mean) / std
+ return outdata
+
+
+def make_reader(filelist, batch_size, sample_times, is_training, shuffle,
+ **dataset_args):
+ # should add smaple_times param
+ fl = open(filelist).readlines()
+ fl = [line.strip() for line in fl if line.strip() != '']
+
+ if shuffle:
+ random.shuffle(fl)
+
+ def reader():
+ batch_out = []
+ for line in fl:
+ # start_time = time.time()
+ line_items = line.split(' ')
+ fn = line_items[0]
+ label = int(line_items[1])
+ if len(line_items) > 2:
+ start_frm = int(line_items[2])
+ spatial_pos = int(line_items[3])
+ in_sample_times = sample_times
+ else:
+ start_frm = -1
+ spatial_pos = -1
+ in_sample_times = 1
+ label = np.array([label]).astype(np.int64)
+ # 1, get rgb data for fixed length of frames
+ try:
+ rgbdata = video_fast_get_frame(fn, \
+ sampling_rate = dataset_args['sample_rate'], length = dataset_args['video_length'], \
+ start_frm = start_frm, sample_times = in_sample_times)
+ except:
+ logger.info('Error when loading {}, just skip this file'.format(
+ fn))
+ continue
+ # add prepocessing
+ # 2, reszie to randomly scale between [min_size, max_size] when training, or cgf.TEST.SCALE when inference
+ min_size = dataset_args['min_size']
+ max_size = dataset_args['max_size']
+ rgbdata = apply_resize(rgbdata, min_size, max_size)
+ # transform [length, height, width, channel] to [channel, length, height, width]
+ rgbdata = np.transpose(rgbdata, [3, 0, 1, 2])
+
+ # 3 crop, mirror and transform
+ rgbdata = crop_mirror_transform(rgbdata, mean = dataset_args['image_mean'], \
+ std = dataset_args['image_std'], cropsize = dataset_args['crop_size'], \
+ use_mirror = is_training, center_crop = (not is_training), \
+ spatial_pos = spatial_pos)
+
+ batch_out.append((rgbdata, label))
+ if len(batch_out) == batch_size:
+ yield batch_out
+ batch_out = []
+
+ return reader
+
+
+def make_multi_reader(filelist, batch_size, sample_times, is_training, shuffle,
+ **dataset_args):
+ fl = open(filelist).readlines()
+ fl = [line.strip() for line in fl if line.strip() != '']
+
+ if shuffle:
+ random.shuffle(fl)
+
+ n = dataset_args['num_reader_threads']
+ queue_size = 20
+ reader_lists = [None] * n
+ file_num = int(len(fl) // n)
+ for i in range(n):
+ if i < len(reader_lists) - 1:
+ tmp_list = fl[i * file_num:(i + 1) * file_num]
+ else:
+ tmp_list = fl[i * file_num:]
+ reader_lists[i] = tmp_list
+
+ def read_into_queue(flq, queue):
+ batch_out = []
+ for line in flq:
+ line_items = line.split(' ')
+ fn = line_items[0]
+ label = int(line_items[1])
+ if len(line_items) > 2:
+ start_frm = int(line_items[2])
+ spatial_pos = int(line_items[3])
+ in_sample_times = sample_times
+ else:
+ start_frm = -1
+ spatial_pos = -1
+ in_sample_times = 1
+ label = np.array([label]).astype(np.int64)
+ # 1, get rgb data for fixed length of frames
+ try:
+ rgbdata = video_fast_get_frame(fn, \
+ sampling_rate = dataset_args['sample_rate'], length = dataset_args['video_length'], \
+ start_frm = start_frm, sample_times = in_sample_times)
+ except:
+ logger.info('Error when loading {}, just skip this file'.format(
+ fn))
+ continue
+ # add prepocessing
+ # 2, reszie to randomly scale between [min_size, max_size] when training, or cgf.TEST.SCALE when inference
+ min_size = dataset_args['min_size']
+ max_size = dataset_args['max_size']
+ rgbdata = apply_resize(rgbdata, min_size, max_size)
+ # transform [length, height, width, channel] to [channel, length, height, width]
+ rgbdata = np.transpose(rgbdata, [3, 0, 1, 2])
+
+ # 3 crop, mirror and transform
+ rgbdata = crop_mirror_transform(rgbdata, mean = dataset_args['image_mean'], \
+ std = dataset_args['image_std'], cropsize = dataset_args['crop_size'], \
+ use_mirror = is_training, center_crop = (not is_training), \
+ spatial_pos = spatial_pos)
+
+ batch_out.append((rgbdata, label))
+ if len(batch_out) == batch_size:
+ queue.put(batch_out)
+ batch_out = []
+ queue.put(None)
+
+ def queue_reader():
+ queue = multiprocessing.Queue(queue_size)
+ p_list = [None] * len(reader_lists)
+ # for reader_list in reader_lists:
+ for i in range(len(reader_lists)):
+ reader_list = reader_lists[i]
+ p_list[i] = multiprocessing.Process(
+ target=read_into_queue, args=(reader_list, queue))
+ p_list[i].start()
+ reader_num = len(reader_lists)
+ finish_num = 0
+ while finish_num < reader_num:
+ sample = queue.get()
+ if sample is None:
+ finish_num += 1
+ else:
+ yield sample
+ for i in range(len(p_list)):
+ p_list[i].terminate()
+ p_list[i].join()
+
+ return queue_reader
diff --git a/fluid/PaddleCV/video/datareader/reader_utils.py b/fluid/PaddleCV/video/datareader/reader_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c8b436a74335c9b3b3361947123b1a3bb3d43dd
--- /dev/null
+++ b/fluid/PaddleCV/video/datareader/reader_utils.py
@@ -0,0 +1,75 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 pickle
+import cv2
+import numpy as np
+import random
+
+
+class ReaderNotFoundError(Exception):
+ "Error: reader not found"
+
+ def __init__(self, reader_name, avail_readers):
+ super(ReaderNotFoundError, self).__init__()
+ self.reader_name = reader_name
+ self.avail_readers = avail_readers
+
+ def __str__(self):
+ msg = "Reader {} Not Found.\nAvailiable readers:\n".format(
+ self.reader_name)
+ for reader in self.avail_readers:
+ msg += " {}\n".format(reader)
+ return msg
+
+
+class DataReader(object):
+ """data reader for video input"""
+
+ def __init__(self, model_name, mode, cfg):
+ """Not implemented"""
+ pass
+
+ def create_reader(self):
+ """Not implemented"""
+ pass
+
+
+class ReaderZoo(object):
+ def __init__(self):
+ self.reader_zoo = {}
+
+ def regist(self, name, reader):
+ assert reader.__base__ == DataReader, "Unknow model type {}".format(
+ type(reader))
+ self.reader_zoo[name] = reader
+
+ def get(self, name, mode, cfg):
+ for k, v in self.reader_zoo.items():
+ if k == name:
+ return v(name, mode, cfg)
+ raise ReaderNotFoundError(name, self.reader_zoo.keys())
+
+
+# singleton reader_zoo
+reader_zoo = ReaderZoo()
+
+
+def regist_reader(name, reader):
+ reader_zoo.regist(name, reader)
+
+
+def get_reader(name, mode, cfg):
+ reader_model = reader_zoo.get(name, mode, cfg)
+ return reader_model.create_reader()
diff --git a/fluid/PaddleCV/video/dataset/README.md b/fluid/PaddleCV/video/dataset/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..3d4c78f9734a95f694b1d3b79816ed527ca4799e
--- /dev/null
+++ b/fluid/PaddleCV/video/dataset/README.md
@@ -0,0 +1,119 @@
+# 数据使用说明
+
+- [Youtube-8M](#Youtube-8M数据集)
+- [Kinetics](#Kinetics数据集)
+
+## Youtube-8M数据集
+这里用到的是YouTube-8M 2018年更新之后的数据集。使用官方数据集，并将TFRecord文件转化为pickle文件以便PaddlePaddle使用。Youtube-8M数据集官方提供了frame-level和video-level的特征，这里只需使用到frame-level的特征。
+
+### 数据下载
+请使用Youtube-8M官方链接分别下载[训练集](http://us.data.yt8m.org/2/frame/train/index.html)和[验证集](http://us.data.yt8m.org/2/frame/validate/index.html)。每个链接里各提供了3844个文件的下载地址，用户也可以使用官方提供的[下载脚本](https://research.google.com/youtube8m/download.html)下载数据。数据下载完成后，将会得到3844个训练数据文件和3844个验证数据文件（TFRecord格式）。
+假设存放视频模型代码库的主目录为: Code\_Root，进入dataset/youtube8m目录
+
+ cd dataset/youtube8m
+
+在youtube8m下新建目录tf/train和tf/val
+
+ mkdir tf && cd tf
+
+ mkdir train && mkdir val
+
+并分别将下载的train和validate数据存放在其中。
+
+### 数据格式转化
+
+为了适用于PaddlePaddle训练，需要离线将下载好的TFRecord文件格式转成了pickle格式，转换脚本请使用[dataset/youtube8m/tf2pkl.py](./youtube8m/tf2pkl.py)。
+
+在dataset/youtube8m 目录下新建目录pkl/train和pkl/val
+
+ cd dataset/youtube8m
+
+ mkdir pkl && cd pkl
+
+ mkdir train && mkdir val
+
+
+转化文件格式(TFRecord -> pkl)，进入dataset/youtube8m目录，运行脚本
+
+ python tf2pkl.py ./tf/train ./pkl/train
+
+和
+
+ python tf2pkl.py ./tf/val ./pkl/val
+
+分别将train和validate数据集转化为pkl文件。tf2pkl.py文件运行时需要两个参数，分别是数据源tf文件存放路径和转化后的pkl文件存放路径。
+
+备注：由于TFRecord文件的读取需要用到Tensorflow，用户要先安装Tensorflow，或者在安装有Tensorflow的环境中转化完数据，再拷贝到dataset/youtube8m/pkl目录下。为了避免和PaddlePaddle环境冲突，建议先在其他地方转化完成再将数据拷贝过来。
+
+### 生成文件列表
+
+进入dataset/youtube8m目录
+
+ ls $Code_Root/dataset/youtube8m/pkl/train/* > train.list
+
+ ls $Code_Root/dataset/youtube8m/pkl/val/* > val.list
+
+在dataset/youtube8m目录下将生成两个文件，train.list和val.list，每一行分别保存了一个pkl文件的绝对路径。
+
+## Kinetics数据集
+
+Kinetics数据集是DeepMind公开的大规模视频动作识别数据集，有Kinetics400与Kinetics600两个版本。这里使用Kinetics400数据集，具体的数据预处理过程如下。
+
+### mp4视频下载
+在Code\_Root目录下创建文件夹
+
+ cd $Code_Root/dataset && mkdir kinetics
+
+ cd kinetics && mkdir data_k400 && cd data_k400
+
+ mkdir train_mp4 && mkdir val_mp4
+
+ActivityNet官方提供了Kinetics的下载工具，具体参考其[官方repo ](https://github.com/activitynet/ActivityNet/tree/master/Crawler/Kinetics)即可下载Kinetics400的mp4视频集合。将kinetics400的训练与验证集合分别下载到dataset/kinetics/data\_k400/train\_mp4与dataset/kinetics/data\_k400/val\_mp4。
+
+### mp4文件预处理
+
+为提高数据读取速度，提前将mp4文件解帧并打pickle包，dataloader从视频的pkl文件中读取数据（该方法耗费更多存储空间）。pkl文件里打包的内容为(video-id,[frame1, frame2,...,frameN],label)。
+
+在 dataset/kinetics/data\_k400目录下创建目录train\_pkl和val\_pkl
+
+ cd $Code_Root/dataset/kinetics/data_k400
+
+ mkdir train_pkl && mkdir val_pkl
+
+进入$Code\_Root/dataset/kinetics目录，使用video2pkl.py脚本进行数据转化。首先需要下载[train](https://github.com/activitynet/ActivityNet/tree/master/Crawler/Kinetics/data/kinetics-400_train.csv)和[validation](https://github.com/activitynet/ActivityNet/tree/master/Crawler/Kinetics/data/kinetics-400_val.csv)数据集的文件列表。
+
+首先生成预处理需要的数据集标签文件
+
+ python generate_label.py kinetics-400_train.csv kinetics400_label.txt
+
+然后执行如下程序：
+
+ python video2pkl.py kinetics-400_train.csv $Source_dir $Target_dir 8 #以8个进程为例
+
+- 该脚本依赖`ffmpeg`库，请预先安装`ffmpeg`
+
+对于train数据，
+
+ Source_dir = $Code_Root/dataset/kinetics/data_k400/train_mp4
+
+ Target_dir = $Code_Root/dataset/kinetics/data_k400/train_pkl
+
+对于val数据，
+
+ Source_dir = $Code_Root/dataset/kinetics/data_k400/val_mp4
+
+ Target_dir = $Code_Root/dataset/kinetics/data_k400/val_pkl
+
+这样即可将mp4文件解码并保存为pkl文件。
+
+### 生成训练和验证集list
+
+ cd $Code_Root/dataset/kinetics
+
+ ls $Code_Root/dataset/kinetics/data_k400/train_pkl /* > train.list
+
+ ls $Code_Root/dataset/kinetics/data_k400/val_pkl /* > val.list
+
+
+即可生成相应的文件列表，train.list和val.list的每一行表示一个pkl文件的绝对路径。
+
diff --git a/fluid/PaddleCV/video/dataset/kinetics/generate_label.py b/fluid/PaddleCV/video/dataset/kinetics/generate_label.py
new file mode 100644
index 0000000000000000000000000000000000000000..4f7c504c56821527cde57bacf7e9a2d07c666c8f
--- /dev/null
+++ b/fluid/PaddleCV/video/dataset/kinetics/generate_label.py
@@ -0,0 +1,31 @@
+import sys
+
+# kinetics-400_train.csv should be down loaded first and set as sys.argv[1]
+# sys.argv[2] can be set as kinetics400_label.txt
+# python generate_label.py kinetics-400_train.csv kinetics400_label.txt
+
+num_classes = 400
+
+fname = sys.argv[1]
+outname = sys.argv[2]
+fl = open(fname).readlines()
+fl = fl[1:]
+outf = open(outname, 'w')
+
+label_list = []
+for line in fl:
+ label = line.strip().split(',')[0].strip('"')
+ if label in label_list:
+ continue
+ else:
+ label_list.append(label)
+
+assert len(label_list
+ ) == num_classes, "there should be {} labels in list, but ".format(
+ num_classes, len(label_list))
+
+label_list.sort()
+for i in range(num_classes):
+ outf.write('{} {}'.format(label_list[i], i) + '\n')
+
+outf.close()
diff --git a/fluid/PaddleCV/video/dataset/kinetics/video2pkl.py b/fluid/PaddleCV/video/dataset/kinetics/video2pkl.py
new file mode 100644
index 0000000000000000000000000000000000000000..881857c40c4ece2f192e681526e2622ef1ce2f81
--- /dev/null
+++ b/fluid/PaddleCV/video/dataset/kinetics/video2pkl.py
@@ -0,0 +1,84 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import sys
+import glob
+import cPickle
+from multiprocessing import Pool
+
+# example command line: python generate_k400_pkl.py kinetics-400_train.csv 8
+#
+# kinetics-400_train.csv is the training set file of K400 official release
+# each line contains laebl,youtube_id,time_start,time_end,split,is_cc
+
+assert (len(sys.argv) == 5)
+
+f = open(sys.argv[1])
+source_dir = sys.argv[2]
+target_dir = sys.argv[3]
+num_threads = sys.argv[4]
+all_video_entries = [x.strip().split(',') for x in f.readlines()]
+all_video_entries = all_video_entries[1:]
+f.close()
+
+category_label_map = {}
+f = open('kinetics400_label.txt')
+for line in f:
+ ens = line.strip().split(' ')
+ category = " ".join(ens[0:-1])
+ label = int(ens[-1])
+ category_label_map[category] = label
+f.close()
+
+
+def generate_pkl(entry):
+ mode = entry[4]
+ category = entry[0].strip('"')
+ category_dir = category
+ video_path = os.path.join(
+ './',
+ entry[1] + "_%06d" % int(entry[2]) + "_%06d" % int(entry[3]) + ".mp4")
+ video_path = os.path.join(source_dir, category_dir, video_path)
+ label = category_label_map[category]
+
+ vid = './' + video_path.split('/')[-1].split('.')[0]
+ if os.path.exists(video_path):
+ if not os.path.exists(vid):
+ os.makedirs(vid)
+ os.system('ffmpeg -i ' + video_path + ' -q 0 ' + vid + '/%06d.jpg')
+ else:
+ print("File not exists {}".format(video_path))
+ return
+
+ images = sorted(glob.glob(vid + '/*.jpg'))
+ ims = []
+ for img in images:
+ f = open(img)
+ ims.append(f.read())
+ f.close()
+
+ output_pkl = vid + ".pkl"
+ output_pkl = os.path.join(target_dir, output_pkl)
+ f = open(output_pkl, 'w')
+ cPickle.dump((vid, label, ims), f, -1)
+ f.close()
+
+ os.system('rm -rf %s' % vid)
+
+
+pool = Pool(processes=int(sys.argv[4]))
+pool.map(generate_pkl, all_video_entries)
+pool.close()
+pool.join()
diff --git a/fluid/PaddleCV/video/dataset/youtube8m/tf2pkl.py b/fluid/PaddleCV/video/dataset/youtube8m/tf2pkl.py
new file mode 100644
index 0000000000000000000000000000000000000000..3b32e3b41a705d6e294581ca3b92c911d238798f
--- /dev/null
+++ b/fluid/PaddleCV/video/dataset/youtube8m/tf2pkl.py
@@ -0,0 +1,278 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+"""Provides readers configured for different datasets."""
+import os, sys
+import numpy as np
+import tensorflow as tf
+from tensorflow import logging
+import cPickle
+
+from tensorflow.python.platform import gfile
+
+assert (len(sys.argv) == 3)
+source_dir = sys.argv[1]
+target_dir = sys.argv[2]
+
+
+def Dequantize(feat_vector, max_quantized_value=2, min_quantized_value=-2):
+ """Dequantize the feature from the byte format to the float format.
+
+ Args:
+ feat_vector: the input 1-d vector.
+ max_quantized_value: the maximum of the quantized value.
+ min_quantized_value: the minimum of the quantized value.
+
+ Returns:
+ A float vector which has the same shape as feat_vector.
+ """
+ assert max_quantized_value > min_quantized_value
+ quantized_range = max_quantized_value - min_quantized_value
+ scalar = quantized_range / 255.0
+ bias = (quantized_range / 512.0) + min_quantized_value
+ return feat_vector * scalar + bias
+
+
+def resize_axis(tensor, axis, new_size, fill_value=0):
+ """Truncates or pads a tensor to new_size on on a given axis.
+
+ Truncate or extend tensor such that tensor.shape[axis] == new_size. If the
+ size increases, the padding will be performed at the end, using fill_value.
+
+ Args:
+ tensor: The tensor to be resized.
+ axis: An integer representing the dimension to be sliced.
+ new_size: An integer or 0d tensor representing the new value for
+ tensor.shape[axis].
+ fill_value: Value to use to fill any new entries in the tensor. Will be
+ cast to the type of tensor.
+
+ Returns:
+ The resized tensor.
+ """
+ tensor = tf.convert_to_tensor(tensor)
+ shape = tf.unstack(tf.shape(tensor))
+
+ pad_shape = shape[:]
+ pad_shape[axis] = tf.maximum(0, new_size - shape[axis])
+
+ shape[axis] = tf.minimum(shape[axis], new_size)
+ shape = tf.stack(shape)
+
+ resized = tf.concat([
+ tf.slice(tensor, tf.zeros_like(shape), shape),
+ tf.fill(tf.stack(pad_shape), tf.cast(fill_value, tensor.dtype))
+ ], axis)
+
+ # Update shape.
+ new_shape = tensor.get_shape().as_list() # A copy is being made.
+ new_shape[axis] = new_size
+ resized.set_shape(new_shape)
+ return resized
+
+
+class BaseReader(object):
+ """Inherit from this class when implementing new readers."""
+
+ def prepare_reader(self, unused_filename_queue):
+ """Create a thread for generating prediction and label tensors."""
+ raise NotImplementedError()
+
+
+class YT8MFrameFeatureReader(BaseReader):
+ """Reads TFRecords of SequenceExamples.
+
+ The TFRecords must contain SequenceExamples with the sparse in64 'labels'
+ context feature and a fixed length byte-quantized feature vector, obtained
+ from the features in 'feature_names'. The quantized features will be mapped
+ back into a range between min_quantized_value and max_quantized_value.
+ """
+
+ def __init__(self,
+ num_classes=3862,
+ feature_sizes=[1024],
+ feature_names=["inc3"],
+ max_frames=300):
+ """Construct a YT8MFrameFeatureReader.
+
+ Args:
+ num_classes: a positive integer for the number of classes.
+ feature_sizes: positive integer(s) for the feature dimensions as a list.
+ feature_names: the feature name(s) in the tensorflow record as a list.
+ max_frames: the maximum number of frames to process.
+ """
+
+ assert len(feature_names) == len(feature_sizes), \
+ "length of feature_names (={}) != length of feature_sizes (={})".format( \
+ len(feature_names), len(feature_sizes))
+
+ self.num_classes = num_classes
+ self.feature_sizes = feature_sizes
+ self.feature_names = feature_names
+ self.max_frames = max_frames
+
+ def get_video_matrix(self, features, feature_size, max_frames,
+ max_quantized_value, min_quantized_value):
+ """Decodes features from an input string and quantizes it.
+
+ Args:
+ features: raw feature values
+ feature_size: length of each frame feature vector
+ max_frames: number of frames (rows) in the output feature_matrix
+ max_quantized_value: the maximum of the quantized value.
+ min_quantized_value: the minimum of the quantized value.
+
+ Returns:
+ feature_matrix: matrix of all frame-features
+ num_frames: number of frames in the sequence
+ """
+ decoded_features = tf.reshape(
+ tf.cast(tf.decode_raw(features, tf.uint8), tf.float32),
+ [-1, feature_size])
+
+ num_frames = tf.minimum(tf.shape(decoded_features)[0], max_frames)
+
+ feature_matrix = decoded_features
+
+ return feature_matrix, num_frames
+
+ def prepare_reader(self,
+ filename_queue,
+ max_quantized_value=2,
+ min_quantized_value=-2):
+ """Creates a single reader thread for YouTube8M SequenceExamples.
+
+ Args:
+ filename_queue: A tensorflow queue of filename locations.
+ max_quantized_value: the maximum of the quantized value.
+ min_quantized_value: the minimum of the quantized value.
+
+ Returns:
+ A tuple of video indexes, video features, labels, and padding data.
+ """
+ reader = tf.TFRecordReader()
+ _, serialized_example = reader.read(filename_queue)
+
+ contexts, features = tf.parse_single_sequence_example(
+ serialized_example,
+ context_features={
+ "id": tf.FixedLenFeature([], tf.string),
+ "labels": tf.VarLenFeature(tf.int64)
+ },
+ sequence_features={
+ feature_name: tf.FixedLenSequenceFeature(
+ [], dtype=tf.string)
+ for feature_name in self.feature_names
+ })
+
+ # read ground truth labels
+ labels = (tf.cast(
+ tf.sparse_to_dense(
+ contexts["labels"].values, (self.num_classes, ),
+ 1,
+ validate_indices=False),
+ tf.bool))
+
+ # loads (potentially) different types of features and concatenates them
+ num_features = len(self.feature_names)
+ assert num_features > 0, "No feature selected: feature_names is empty!"
+
+ assert len(self.feature_names) == len(self.feature_sizes), \
+ "length of feature_names (={}) != length of feature_sizes (={})".format( \
+ len(self.feature_names), len(self.feature_sizes))
+
+ num_frames = -1 # the number of frames in the video
+ feature_matrices = [None
+ ] * num_features # an array of different features
+
+ for feature_index in range(num_features):
+ feature_matrix, num_frames_in_this_feature = self.get_video_matrix(
+ features[self.feature_names[feature_index]],
+ self.feature_sizes[feature_index], self.max_frames,
+ max_quantized_value, min_quantized_value)
+ if num_frames == -1:
+ num_frames = num_frames_in_this_feature
+ #else:
+ # tf.assert_equal(num_frames, num_frames_in_this_feature)
+
+ feature_matrices[feature_index] = feature_matrix
+
+ # cap the number of frames at self.max_frames
+ num_frames = tf.minimum(num_frames, self.max_frames)
+
+ # concatenate different features
+ video_matrix = feature_matrices[0]
+ audio_matrix = feature_matrices[1]
+
+ return contexts["id"], video_matrix, audio_matrix, labels, num_frames
+
+
+def main(files_pattern):
+ data_files = gfile.Glob(files_pattern)
+ filename_queue = tf.train.string_input_producer(
+ data_files, num_epochs=1, shuffle=False)
+
+ reader = YT8MFrameFeatureReader(
+ feature_sizes=[1024, 128], feature_names=["rgb", "audio"])
+ vals = reader.prepare_reader(filename_queue)
+
+ with tf.Session() as sess:
+ sess.run(tf.initialize_local_variables())
+ sess.run(tf.initialize_all_variables())
+ coord = tf.train.Coordinator()
+ threads = tf.train.start_queue_runners(sess=sess, coord=coord)
+
+ vid_num = 0
+ all_data = []
+ try:
+ while not coord.should_stop():
+ vid, features, audios, labels, nframes = sess.run(vals)
+ label_index = np.where(labels == True)[0].tolist()
+ vid_num += 1
+
+ #print vid, features.shape, audios.shape, label_index, nframes
+
+ features_int = features.astype(np.uint8)
+ audios_int = audios.astype(np.uint8)
+
+ value_dict = {}
+ value_dict['video'] = vid
+ value_dict['feature'] = features_int
+ value_dict['audio'] = audios_int
+ value_dict['label'] = label_index
+ value_dict['nframes'] = nframes
+ all_data.append(value_dict)
+
+ except tf.errors.OutOfRangeError:
+ print('Finished extracting.')
+
+ finally:
+ coord.request_stop()
+ coord.join(threads)
+
+ print vid_num
+
+ record_name = files_pattern.split('/')[-1].split('.')[0]
+ outputdir = target_dir
+ fn = '%s.pkl' % record_name
+ outp = open(os.path.join(outputdir, fn), 'wb')
+ cPickle.dump(all_data, outp, protocol=cPickle.HIGHEST_PROTOCOL)
+ outp.close()
+
+
+if __name__ == '__main__':
+ record_dir = source_dir
+ record_files = os.listdir(record_dir)
+ for f in record_files:
+ record_path = os.path.join(record_dir, f)
+ main(record_path)
diff --git a/fluid/PaddleCV/video/dataset/youtube8m/yt8m_pca/eigenvals.npy b/fluid/PaddleCV/video/dataset/youtube8m/yt8m_pca/eigenvals.npy
new file mode 100644
index 0000000000000000000000000000000000000000..632506b9ad68f030d64643cc8100868b21c3eb98
Binary files /dev/null and b/fluid/PaddleCV/video/dataset/youtube8m/yt8m_pca/eigenvals.npy differ
diff --git a/fluid/PaddleCV/video/images/StNet.png b/fluid/PaddleCV/video/images/StNet.png
new file mode 100644
index 0000000000000000000000000000000000000000..fde8d77f8e76a5ede7c7cb9f9af96850fea137a1
Binary files /dev/null and b/fluid/PaddleCV/video/images/StNet.png differ
diff --git a/fluid/PaddleCV/video/images/attention_cluster.png b/fluid/PaddleCV/video/images/attention_cluster.png
new file mode 100644
index 0000000000000000000000000000000000000000..f4c1dd6e9a233de68f66b937a24765a8420f7e4b
Binary files /dev/null and b/fluid/PaddleCV/video/images/attention_cluster.png differ
diff --git a/fluid/PaddleCV/video/infer.py b/fluid/PaddleCV/video/infer.py
new file mode 100755
index 0000000000000000000000000000000000000000..43470cede76a39f7b7ffdcb43c0481e25aeca11f
--- /dev/null
+++ b/fluid/PaddleCV/video/infer.py
@@ -0,0 +1,152 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import sys
+import time
+import logging
+import argparse
+import numpy as np
+try:
+ import cPickle as pickle
+except:
+ import pickle
+import paddle.fluid as fluid
+
+from config import *
+import models
+from datareader import get_reader
+
+logging.root.handlers = []
+FORMAT = '[%(levelname)s: %(filename)s: %(lineno)4d]: %(message)s'
+logging.basicConfig(level=logging.DEBUG, format=FORMAT, stream=sys.stdout)
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ '--model-name',
+ type=str,
+ default='AttentionCluster',
+ help='name of model to train.')
+ parser.add_argument(
+ '--config',
+ type=str,
+ default='configs/attention_cluster.txt',
+ help='path to config file of model')
+ parser.add_argument(
+ '--use-gpu', type=bool, default=True, help='default use gpu.')
+ parser.add_argument(
+ '--weights',
+ type=str,
+ default=None,
+ help='weight path, None to use weights from Paddle.')
+ parser.add_argument(
+ '--batch-size',
+ type=int,
+ default=1,
+ help='sample number in a batch for inference.')
+ parser.add_argument(
+ '--filelist',
+ type=str,
+ default=None,
+ help='path to inferenece data file lists file.')
+ parser.add_argument(
+ '--log-interval',
+ type=int,
+ default=1,
+ help='mini-batch interval to log.')
+ parser.add_argument(
+ '--infer-topk',
+ type=int,
+ default=20,
+ help='topk predictions to restore.')
+ parser.add_argument(
+ '--save-dir', type=str, default='./', help='directory to store results')
+ args = parser.parse_args()
+ return args
+
+
+def infer(args):
+ # parse config
+ config = parse_config(args.config)
+ infer_config = merge_configs(config, 'infer', vars(args))
+ infer_model = models.get_model(args.model_name, infer_config, mode='infer')
+
+ infer_model.build_input(use_pyreader=False)
+ infer_model.build_model()
+ infer_feeds = infer_model.feeds()
+ infer_outputs = infer_model.outputs()
+
+ place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
+ exe = fluid.Executor(place)
+
+ filelist = args.filelist or infer_config.INFER.filelist
+ assert os.path.exists(filelist), "{} not exist.".format(args.filelist)
+
+ # get infer reader
+ infer_reader = get_reader(args.model_name.upper(), 'infer', infer_config)
+
+ if args.weights:
+ assert os.path.exists(
+ args.weights), "Given weight dir {} not exist.".format(args.weights)
+ # if no weight files specified, download weights from paddle
+ weights = args.weights or infer_model.get_weights()
+
+ def if_exist(var):
+ return os.path.exists(os.path.join(weights, var.name))
+
+ fluid.io.load_vars(exe, weights, predicate=if_exist)
+
+ infer_feeder = fluid.DataFeeder(place=place, feed_list=infer_feeds)
+ fetch_list = [x.name for x in infer_outputs]
+
+ periods = []
+ results = []
+ cur_time = time.time()
+ for infer_iter, data in enumerate(infer_reader()):
+ data_feed_in = [items[:-1] for items in data]
+ video_id = [items[-1] for items in data]
+ infer_outs = exe.run(fetch_list=fetch_list,
+ feed=infer_feeder.feed(data_feed_in))
+ predictions = np.array(infer_outs[0])
+ for i in range(len(predictions)):
+ topk_inds = predictions[i].argsort()[0 - args.infer_topk:]
+ topk_inds = topk_inds[::-1]
+ preds = predictions[i][topk_inds]
+ results.append(
+ (video_id[i], preds.tolist(), topk_inds.tolist()))
+ prev_time = cur_time
+ cur_time = time.time()
+ period = cur_time - prev_time
+ periods.append(period)
+ if args.log_interval > 0 and infer_iter % args.log_interval == 0:
+ logger.info('Processed {} samples'.format((infer_iter) * len(
+ predictions)))
+
+ logger.info('[INFER] infer finished. average time: {}'.format(
+ np.mean(periods)))
+
+ if not os.path.isdir(args.save_dir):
+ os.mkdir(args.save_dir)
+ result_file_name = os.path.join(args.save_dir,
+ "{}_infer_result".format(args.model_name))
+ pickle.dump(results, open(result_file_name, 'wb'))
+
+if __name__ == "__main__":
+ args = parse_args()
+ logger.info(args)
+
+ infer(args)
diff --git a/fluid/PaddleCV/video/metrics/__init__.py b/fluid/PaddleCV/video/metrics/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..0d1df762bdf3d3b920fc1e00d15a3a2ecdcdbe55
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/__init__.py
@@ -0,0 +1 @@
+from .metrics_util import get_metrics
diff --git a/fluid/PaddleCV/video/metrics/kinetics/__init__.py b/fluid/PaddleCV/video/metrics/kinetics/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fluid/PaddleCV/video/metrics/kinetics/accuracy_metrics.py b/fluid/PaddleCV/video/metrics/kinetics/accuracy_metrics.py
new file mode 100644
index 0000000000000000000000000000000000000000..d79bf2ee18ca8de7d5219dd5d1ab6452aec3fe5f
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/kinetics/accuracy_metrics.py
@@ -0,0 +1,107 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 __future__ import unicode_literals
+from __future__ import print_function
+from __future__ import division
+
+import numpy as np
+import datetime
+import logging
+
+logger = logging.getLogger(__name__)
+
+
+class MetricsCalculator():
+ def __init__(self, name, mode):
+ self.name = name
+ self.mode = mode # 'train', 'val', 'test'
+ self.reset()
+
+ def reset(self):
+ logger.info('Resetting {} metrics...'.format(self.mode))
+ self.aggr_acc1 = 0.0
+ self.aggr_acc5 = 0.0
+ self.aggr_loss = 0.0
+ self.aggr_batch_size = 0
+
+ def finalize_metrics(self):
+ self.avg_acc1 = self.aggr_acc1 / self.aggr_batch_size
+ self.avg_acc5 = self.aggr_acc5 / self.aggr_batch_size
+ self.avg_loss = self.aggr_loss / self.aggr_batch_size
+
+ def get_computed_metrics(self):
+ json_stats = {}
+ json_stats['avg_loss'] = self.avg_loss
+ json_stats['avg_acc1'] = self.avg_acc1
+ json_stats['avg_acc5'] = self.avg_acc5
+ return json_stats
+
+ def calculate_metrics(self, loss, softmax, labels):
+ accuracy1 = compute_topk_accuracy(softmax, labels, top_k=1) * 100.
+ accuracy5 = compute_topk_accuracy(softmax, labels, top_k=5) * 100.
+ return accuracy1, accuracy5
+
+ def accumulate(self, loss, softmax, labels):
+ cur_batch_size = softmax.shape[0]
+ # if returned loss is None for e.g. test, just set loss to be 0.
+ if loss is None:
+ cur_loss = 0.
+ else:
+ cur_loss = np.mean(np.array(loss)) #
+ self.aggr_batch_size += cur_batch_size
+ self.aggr_loss += cur_loss * cur_batch_size
+
+ accuracy1 = compute_topk_accuracy(softmax, labels, top_k=1) * 100.
+ accuracy5 = compute_topk_accuracy(softmax, labels, top_k=5) * 100.
+ self.aggr_acc1 += accuracy1 * cur_batch_size
+ self.aggr_acc5 += accuracy5 * cur_batch_size
+
+ return
+
+
+# ----------------------------------------------
+# other utils
+# ----------------------------------------------
+def compute_topk_correct_hits(top_k, preds, labels):
+ '''Compute the number of corret hits'''
+ batch_size = preds.shape[0]
+
+ top_k_preds = np.zeros((batch_size, top_k), dtype=np.float32)
+ for i in range(batch_size):
+ top_k_preds[i, :] = np.argsort(-preds[i, :])[:top_k]
+
+ correctness = np.zeros(batch_size, dtype=np.int32)
+ for i in range(batch_size):
+ if labels[i] in top_k_preds[i, :].astype(np.int32).tolist():
+ correctness[i] = 1
+ correct_hits = sum(correctness)
+
+ return correct_hits
+
+
+def compute_topk_accuracy(softmax, labels, top_k):
+
+ computed_metrics = {}
+
+ assert labels.shape[0] == softmax.shape[0], "Batch size mismatch."
+ aggr_batch_size = labels.shape[0]
+ aggr_top_k_correct_hits = compute_topk_correct_hits(top_k, softmax, labels)
+
+ # normalize results
+ computed_metrics = \
+ float(aggr_top_k_correct_hits) / aggr_batch_size
+
+ return computed_metrics
diff --git a/fluid/PaddleCV/video/metrics/metrics_util.py b/fluid/PaddleCV/video/metrics/metrics_util.py
new file mode 100644
index 0000000000000000000000000000000000000000..f769349136cc1c0b3dc5be9e44d4a5e186f2a39f
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/metrics_util.py
@@ -0,0 +1,196 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 __future__ import unicode_literals
+from __future__ import print_function
+from __future__ import division
+
+import logging
+
+import numpy as np
+from metrics.youtube8m import eval_util as youtube8m_metrics
+from metrics.kinetics import accuracy_metrics as kinetics_metrics
+from metrics.multicrop_test import multicrop_test_metrics as multicrop_test_metrics
+
+logger = logging.getLogger(__name__)
+
+
+class Metrics(object):
+ def __init__(self, name, mode, metrics_args):
+ """Not implemented"""
+ pass
+
+ def calculate_and_log_out(self, loss, pred, label, info=''):
+ """Not implemented"""
+ pass
+
+ def accumulate(self, loss, pred, label, info=''):
+ """Not implemented"""
+ pass
+
+ def finalize_and_log_out(self, info=''):
+ """Not implemented"""
+ pass
+
+ def reset(self):
+ """Not implemented"""
+ pass
+
+
+class Youtube8mMetrics(Metrics):
+ def __init__(self, name, mode, metrics_args):
+ self.name = name
+ self.mode = mode
+ self.num_classes = metrics_args['MODEL']['num_classes']
+ self.topk = metrics_args['MODEL']['topk']
+ self.calculator = youtube8m_metrics.EvaluationMetrics(self.num_classes,
+ self.topk)
+
+ def calculate_and_log_out(self, loss, pred, label, info=''):
+ loss = np.mean(np.array(loss))
+ hit_at_one = youtube8m_metrics.calculate_hit_at_one(pred, label)
+ perr = youtube8m_metrics.calculate_precision_at_equal_recall_rate(pred,
+ label)
+ gap = youtube8m_metrics.calculate_gap(pred, label)
+ logger.info(info + ' , loss = {0}, Hit@1 = {1}, PERR = {2}, GAP = {3}'.format(\
+ '%.6f' % loss, '%.2f' % hit_at_one, '%.2f' % perr, '%.2f' % gap))
+
+ def accumulate(self, loss, pred, label, info=''):
+ self.calculator.accumulate(loss, pred, label)
+
+ def finalize_and_log_out(self, info=''):
+ epoch_info_dict = self.calculator.get()
+ logger.info(info + '\tavg_hit_at_one: {0},\tavg_perr: {1},\tavg_loss :{2},\taps: {3},\tgap:{4}'\
+ .format(epoch_info_dict['avg_hit_at_one'], epoch_info_dict['avg_perr'], \
+ epoch_info_dict['avg_loss'], epoch_info_dict['aps'], epoch_info_dict['gap']))
+
+ def reset(self):
+ self.calculator.clear()
+
+
+class Kinetics400Metrics(Metrics):
+ def __init__(self, name, mode, metrics_args):
+ self.name = name
+ self.mode = mode
+ self.calculator = kinetics_metrics.MetricsCalculator(name, mode.lower())
+
+ def calculate_and_log_out(self, loss, pred, label, info=''):
+ if loss is not None:
+ loss = np.mean(np.array(loss))
+ else:
+ loss = 0.
+ acc1, acc5 = self.calculator.calculate_metrics(loss, pred, label)
+ logger.info(info + '\tLoss: {},\ttop1_acc: {}, \ttop5_acc: {}'.format('%.6f' % loss, \
+ '%.2f' % acc1, '%.2f' % acc5))
+
+ def accumulate(self, loss, pred, label, info=''):
+ self.calculator.accumulate(loss, pred, label)
+
+ def finalize_and_log_out(self, info=''):
+ self.calculator.finalize_metrics()
+ metrics_dict = self.calculator.get_computed_metrics()
+ loss = metrics_dict['avg_loss']
+ acc1 = metrics_dict['avg_acc1']
+ acc5 = metrics_dict['avg_acc5']
+ logger.info(info + '\tLoss: {},\ttop1_acc: {}, \ttop5_acc: {}'.format('%.6f' % loss, \
+ '%.2f' % acc1, '%.2f' % acc5))
+
+ def reset(self):
+ self.calculator.reset()
+
+
+class MulticropMetrics(Metrics):
+ def __init__(self, name, mode, metrics_args):
+ self.name = name
+ self.mode = mode
+ if mode == 'test':
+ args = {}
+ args['num_test_clips'] = metrics_args.TEST.num_test_clips
+ args['dataset_size'] = metrics_args.TEST.dataset_size
+ args['filename_gt'] = metrics_args.TEST.filename_gt
+ args['checkpoint_dir'] = metrics_args.TEST.checkpoint_dir
+ args['num_classes'] = metrics_args.MODEL.num_classes
+ self.calculator = multicrop_test_metrics.MetricsCalculator(
+ name, mode.lower(), **args)
+ else:
+ self.calculator = kinetics_metrics.MetricsCalculator(name,
+ mode.lower())
+
+ def calculate_and_log_out(self, loss, pred, label, info=''):
+ if self.mode == 'test':
+ pass
+ else:
+ if loss is not None:
+ loss = np.mean(np.array(loss))
+ else:
+ loss = 0.
+ acc1, acc5 = self.calculator.calculate_metrics(loss, pred, label)
+ logger.info(info + '\tLoss: {},\ttop1_acc: {}, \ttop5_acc: {}'.format('%.6f' % loss, \
+ '%.2f' % acc1, '%.2f' % acc5))
+
+ def accumulate(self, loss, pred, label):
+ self.calculator.accumulate(loss, pred, label)
+
+ def finalize_and_log_out(self, info=''):
+ if self.mode == 'test':
+ self.calculator.finalize_metrics()
+ else:
+ self.calculator.finalize_metrics()
+ metrics_dict = self.calculator.get_computed_metrics()
+ loss = metrics_dict['avg_loss']
+ acc1 = metrics_dict['avg_acc1']
+ acc5 = metrics_dict['avg_acc5']
+ logger.info(info + '\tLoss: {},\ttop1_acc: {}, \ttop5_acc: {}'.format('%.6f' % loss, \
+ '%.2f' % acc1, '%.2f' % acc5))
+
+ def reset(self):
+ self.calculator.reset()
+
+
+class MetricsZoo(object):
+ def __init__(self):
+ self.metrics_zoo = {}
+
+ def regist(self, name, metrics):
+ assert metrics.__base__ == Metrics, "Unknow model type {}".format(
+ type(metrics))
+ self.metrics_zoo[name] = metrics
+
+ def get(self, name, mode, cfg):
+ for k, v in self.metrics_zoo.items():
+ if k == name:
+ return v(name, mode, cfg)
+ raise MetricsNotFoundError(name, self.metrics_zoo.keys())
+
+
+# singleton metrics_zoo
+metrics_zoo = MetricsZoo()
+
+
+def regist_metrics(name, metrics):
+ metrics_zoo.regist(name, metrics)
+
+
+def get_metrics(name, mode, cfg):
+ return metrics_zoo.get(name, mode, cfg)
+
+
+regist_metrics("NEXTVLAD", Youtube8mMetrics)
+regist_metrics("ATTENTIONLSTM", Youtube8mMetrics)
+regist_metrics("ATTENTIONCLUSTER", Youtube8mMetrics)
+regist_metrics("TSN", Kinetics400Metrics)
+regist_metrics("TSM", Kinetics400Metrics)
+regist_metrics("STNET", Kinetics400Metrics)
+regist_metrics("NONLOCAL", MulticropMetrics)
diff --git a/fluid/PaddleCV/video/metrics/multicrop_test/__init__.py b/fluid/PaddleCV/video/metrics/multicrop_test/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fluid/PaddleCV/video/metrics/multicrop_test/multicrop_test_metrics.py b/fluid/PaddleCV/video/metrics/multicrop_test/multicrop_test_metrics.py
new file mode 100644
index 0000000000000000000000000000000000000000..9da8826cef1be209fce0bfac8d2c7f0b6d70d4a4
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/multicrop_test/multicrop_test_metrics.py
@@ -0,0 +1,213 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 __future__ import unicode_literals
+from __future__ import print_function
+from __future__ import division
+
+import sys
+import os
+import numpy as np
+import datetime
+import logging
+from collections import defaultdict
+import pickle
+
+logger = logging.getLogger(__name__)
+
+
+class MetricsCalculator():
+ def __init__(self, name, mode, **metrics_args):
+ """
+ metrics args:
+ num_test_clips, number of clips of each video when test
+ dataset_size, total number of videos in the dataset
+ filename_gt, a file with each line stores the groud truth of each video
+ checkpoint_dir, dir where to store the test results
+ num_classes, number of classes of the dataset
+ """
+ self.name = name
+ self.mode = mode # 'train', 'val', 'test'
+ self.metrics_args = metrics_args
+
+ self.num_test_clips = metrics_args['num_test_clips']
+ self.dataset_size = metrics_args['dataset_size']
+ self.filename_gt = metrics_args['filename_gt']
+ self.checkpoint_dir = metrics_args['checkpoint_dir']
+ self.num_classes = metrics_args['num_classes']
+ self.reset()
+
+ def reset(self):
+ logger.info('Resetting {} metrics...'.format(self.mode))
+ self.aggr_acc1 = 0.0
+ self.aggr_acc5 = 0.0
+ self.aggr_loss = 0.0
+ self.aggr_batch_size = 0
+ self.seen_inds = defaultdict(int)
+ self.results = []
+
+ def calculate_metrics(self, loss, pred, labels):
+ pass
+
+ def accumulate(self, loss, pred, labels):
+ labels = labels.astype(int)
+ for i in range(pred.shape[0]):
+ probs = pred[i, :].tolist()
+ vid = labels[i]
+ self.seen_inds[vid] += 1
+ if self.seen_inds[vid] > self.num_test_clips:
+ logger.warning('Video id {} have been seen. Skip.'.format(vid,
+ ))
+ continue
+ save_pairs = [vid, probs]
+ self.results.append(save_pairs)
+ logger.info("({0} / {1}) videos".format(\
+ len(self.seen_inds), self.dataset_size))
+
+ def finalize_metrics(self):
+ if self.filename_gt is not None:
+ evaluate_results(self.results, self.filename_gt, self.dataset_size, \
+ self.num_classes, self.num_test_clips)
+ # save temporary file
+ pkl_path = os.path.join(self.checkpoint_dir, "results_probs.pkl")
+
+ with open(pkl_path, 'w') as f:
+ pickle.dump(self.results, f)
+ logger.info('Temporary file saved to: {}'.format(pkl_path))
+
+
+def read_groundtruth(filename_gt):
+ f = open(filename_gt, 'r')
+ labels = []
+ for line in f:
+ rows = line.split()
+ labels.append(int(rows[1]))
+ f.close()
+ return labels
+
+
+def evaluate_results(results, filename_gt, test_dataset_size, num_classes,
+ num_test_clips):
+ gt_labels = read_groundtruth(filename_gt)
+ sample_num = test_dataset_size
+ class_num = num_classes
+ sample_video_times = num_test_clips
+ counts = np.zeros(sample_num, dtype=np.int32)
+ probs = np.zeros((sample_num, class_num))
+
+ assert (len(gt_labels) == sample_num)
+ """
+ clip_accuracy: the (e.g.) 10*19761 clips' average accuracy
+ clip1_accuracy: the 1st clip's accuracy (starting from frame 0)
+ """
+ clip_accuracy = 0
+ clip1_accuracy = 0
+ clip1_count = 0
+ seen_inds = defaultdict(int)
+
+ # evaluate
+ for entry in results:
+ vid = entry[0]
+ prob = np.array(entry[1])
+ probs[vid] += prob[0:class_num]
+ counts[vid] += 1
+
+ idx = prob.argmax()
+ if idx == gt_labels[vid]:
+ # clip accuracy
+ clip_accuracy += 1
+
+ # clip1 accuracy
+ seen_inds[vid] += 1
+ if seen_inds[vid] == 1:
+ clip1_count += 1
+ if idx == gt_labels[vid]:
+ clip1_accuracy += 1
+
+ # sanity checkcnt = 0
+ max_clips = 0
+ min_clips = sys.maxsize
+ count_empty = 0
+ count_corrupted = 0
+ for i in range(sample_num):
+ max_clips = max(max_clips, counts[i])
+ min_clips = min(min_clips, counts[i])
+ if counts[i] != sample_video_times:
+ count_corrupted += 1
+ logger.warning('Id: {} count: {}'.format(i, counts[i]))
+ if counts[i] == 0:
+ count_empty += 1
+
+ logger.info('Num of empty videos: {}'.format(count_empty))
+ logger.info('Num of corrupted videos: {}'.format(count_corrupted))
+ logger.info('Max num of clips in a video: {}'.format(max_clips))
+ logger.info('Min num of clips in a video: {}'.format(min_clips))
+
+ # clip1 accuracy for sanity (# print clip1 first as it is lowest)
+ logger.info('Clip1 accuracy: {:.2f} percent ({}/{})'.format(
+ 100. * clip1_accuracy / clip1_count, clip1_accuracy, clip1_count))
+
+ # clip accuracy for sanity
+ logger.info('Clip accuracy: {:.2f} percent ({}/{})'.format(
+ 100. * clip_accuracy / len(results), clip_accuracy, len(results)))
+
+ # compute accuracy
+ accuracy = 0
+ accuracy_top5 = 0
+ for i in range(sample_num):
+ prob = probs[i]
+
+ # top-1
+ idx = prob.argmax()
+ if idx == gt_labels[i] and counts[i] > 0:
+ accuracy = accuracy + 1
+
+ ids = np.argsort(prob)[::-1]
+ for j in range(5):
+ if ids[j] == gt_labels[i] and counts[i] > 0:
+ accuracy_top5 = accuracy_top5 + 1
+ break
+
+ accuracy = float(accuracy) / float(sample_num)
+ accuracy_top5 = float(accuracy_top5) / float(sample_num)
+
+ logger.info('-' * 80)
+ logger.info('top-1 accuracy: {:.2f} percent'.format(accuracy * 100))
+ logger.info('top-5 accuracy: {:.2f} percent'.format(accuracy_top5 * 100))
+ logger.info('-' * 80)
+
+ for i in range(sample_num):
+ prob = probs[i]
+
+ # top-1
+ idx = prob.argmax()
+ if idx == gt_labels[i] and counts[i] > 0:
+ accuracy = accuracy + 1
+
+ ids = np.argsort(prob)[::-1]
+ for j in range(5):
+ if ids[j] == gt_labels[i] and counts[i] > 0:
+ accuracy_top5 = accuracy_top5 + 1
+ break
+
+ accuracy = float(accuracy) / float(sample_num)
+ accuracy_top5 = float(accuracy_top5) / float(sample_num)
+
+ logger.info('-' * 80)
+ logger.info('top-1 accuracy: {:.2f} percent'.format(accuracy * 100))
+ logger.info('top-5 accuracy: {:.2f} percent'.format(accuracy_top5 * 100))
+ logger.info('-' * 80)
+
+ return
diff --git a/fluid/PaddleCV/video/metrics/youtube8m/__init__.py b/fluid/PaddleCV/video/metrics/youtube8m/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fluid/PaddleCV/video/metrics/youtube8m/average_precision_calculator.py b/fluid/PaddleCV/video/metrics/youtube8m/average_precision_calculator.py
new file mode 100644
index 0000000000000000000000000000000000000000..9bad69dd0aff1906e3548fb0322203f0bc5b408d
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/youtube8m/average_precision_calculator.py
@@ -0,0 +1,275 @@
+# Copyright 2016 Google Inc. 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.
+"""Calculate or keep track of the interpolated average precision.
+
+It provides an interface for calculating interpolated average precision for an
+entire list or the top-n ranked items. For the definition of the
+(non-)interpolated average precision:
+http://trec.nist.gov/pubs/trec15/appendices/CE.MEASURES06.pdf
+
+Example usages:
+1) Use it as a static function call to directly calculate average precision for
+a short ranked list in the memory.
+
+```
+import random
+
+p = np.array([random.random() for _ in xrange(10)])
+a = np.array([random.choice([0, 1]) for _ in xrange(10)])
+
+ap = average_precision_calculator.AveragePrecisionCalculator.ap(p, a)
+```
+
+2) Use it as an object for long ranked list that cannot be stored in memory or
+the case where partial predictions can be observed at a time (Tensorflow
+predictions). In this case, we first call the function accumulate many times
+to process parts of the ranked list. After processing all the parts, we call
+peek_interpolated_ap_at_n.
+```
+p1 = np.array([random.random() for _ in xrange(5)])
+a1 = np.array([random.choice([0, 1]) for _ in xrange(5)])
+p2 = np.array([random.random() for _ in xrange(5)])
+a2 = np.array([random.choice([0, 1]) for _ in xrange(5)])
+
+# interpolated average precision at 10 using 1000 break points
+calculator = average_precision_calculator.AveragePrecisionCalculator(10)
+calculator.accumulate(p1, a1)
+calculator.accumulate(p2, a2)
+ap3 = calculator.peek_ap_at_n()
+```
+"""
+
+import heapq
+import random
+import numbers
+
+import numpy
+
+
+class AveragePrecisionCalculator(object):
+ """Calculate the average precision and average precision at n."""
+
+ def __init__(self, top_n=None):
+ """Construct an AveragePrecisionCalculator to calculate average precision.
+
+ This class is used to calculate the average precision for a single label.
+
+ Args:
+ top_n: A positive Integer specifying the average precision at n, or
+ None to use all provided data points.
+
+ Raises:
+ ValueError: An error occurred when the top_n is not a positive integer.
+ """
+ if not ((isinstance(top_n, int) and top_n >= 0) or top_n is None):
+ raise ValueError("top_n must be a positive integer or None.")
+
+ self._top_n = top_n # average precision at n
+ self._total_positives = 0 # total number of positives have seen
+ self._heap = [] # max heap of (prediction, actual)
+
+ @property
+ def heap_size(self):
+ """Gets the heap size maintained in the class."""
+ return len(self._heap)
+
+ @property
+ def num_accumulated_positives(self):
+ """Gets the number of positive samples that have been accumulated."""
+ return self._total_positives
+
+ def accumulate(self, predictions, actuals, num_positives=None):
+ """Accumulate the predictions and their ground truth labels.
+
+ After the function call, we may call peek_ap_at_n to actually calculate
+ the average precision.
+ Note predictions and actuals must have the same shape.
+
+ Args:
+ predictions: a list storing the prediction scores.
+ actuals: a list storing the ground truth labels. Any value
+ larger than 0 will be treated as positives, otherwise as negatives.
+ num_positives = If the 'predictions' and 'actuals' inputs aren't complete,
+ then it's possible some true positives were missed in them. In that case,
+ you can provide 'num_positives' in order to accurately track recall.
+
+ Raises:
+ ValueError: An error occurred when the format of the input is not the
+ numpy 1-D array or the shape of predictions and actuals does not match.
+ """
+ if len(predictions) != len(actuals):
+ raise ValueError(
+ "the shape of predictions and actuals does not match.")
+
+ if not num_positives is None:
+ if not isinstance(num_positives,
+ numbers.Number) or num_positives < 0:
+ raise ValueError(
+ "'num_positives' was provided but it wan't a nonzero number."
+ )
+
+ if not num_positives is None:
+ self._total_positives += num_positives
+ else:
+ self._total_positives += numpy.size(numpy.where(actuals > 0))
+ topk = self._top_n
+ heap = self._heap
+
+ for i in range(numpy.size(predictions)):
+ if topk is None or len(heap) < topk:
+ heapq.heappush(heap, (predictions[i], actuals[i]))
+ else:
+ if predictions[i] > heap[0][0]: # heap[0] is the smallest
+ heapq.heappop(heap)
+ heapq.heappush(heap, (predictions[i], actuals[i]))
+
+ def clear(self):
+ """Clear the accumulated predictions."""
+ self._heap = []
+ self._total_positives = 0
+
+ def peek_ap_at_n(self):
+ """Peek the non-interpolated average precision at n.
+
+ Returns:
+ The non-interpolated average precision at n (default 0).
+ If n is larger than the length of the ranked list,
+ the average precision will be returned.
+ """
+ if self.heap_size <= 0:
+ return 0
+ predlists = numpy.array(list(zip(*self._heap)))
+
+ ap = self.ap_at_n(
+ predlists[0],
+ predlists[1],
+ n=self._top_n,
+ total_num_positives=self._total_positives)
+ return ap
+
+ @staticmethod
+ def ap(predictions, actuals):
+ """Calculate the non-interpolated average precision.
+
+ Args:
+ predictions: a numpy 1-D array storing the sparse prediction scores.
+ actuals: a numpy 1-D array storing the ground truth labels. Any value
+ larger than 0 will be treated as positives, otherwise as negatives.
+
+ Returns:
+ The non-interpolated average precision at n.
+ If n is larger than the length of the ranked list,
+ the average precision will be returned.
+
+ Raises:
+ ValueError: An error occurred when the format of the input is not the
+ numpy 1-D array or the shape of predictions and actuals does not match.
+ """
+ return AveragePrecisionCalculator.ap_at_n(predictions, actuals, n=None)
+
+ @staticmethod
+ def ap_at_n(predictions, actuals, n=20, total_num_positives=None):
+ """Calculate the non-interpolated average precision.
+
+ Args:
+ predictions: a numpy 1-D array storing the sparse prediction scores.
+ actuals: a numpy 1-D array storing the ground truth labels. Any value
+ larger than 0 will be treated as positives, otherwise as negatives.
+ n: the top n items to be considered in ap@n.
+ total_num_positives : (optionally) you can specify the number of total
+ positive
+ in the list. If specified, it will be used in calculation.
+
+ Returns:
+ The non-interpolated average precision at n.
+ If n is larger than the length of the ranked list,
+ the average precision will be returned.
+
+ Raises:
+ ValueError: An error occurred when
+ 1) the format of the input is not the numpy 1-D array;
+ 2) the shape of predictions and actuals does not match;
+ 3) the input n is not a positive integer.
+ """
+ if len(predictions) != len(actuals):
+ raise ValueError(
+ "the shape of predictions and actuals does not match.")
+
+ if n is not None:
+ if not isinstance(n, int) or n <= 0:
+ raise ValueError("n must be 'None' or a positive integer."
+ " It was '%s'." % n)
+
+ ap = 0.0
+
+ predictions = numpy.array(predictions)
+ actuals = numpy.array(actuals)
+
+ # add a shuffler to avoid overestimating the ap
+ predictions, actuals = AveragePrecisionCalculator._shuffle(predictions,
+ actuals)
+ sortidx = sorted(
+ range(len(predictions)), key=lambda k: predictions[k], reverse=True)
+
+ if total_num_positives is None:
+ numpos = numpy.size(numpy.where(actuals > 0))
+ else:
+ numpos = total_num_positives
+
+ if numpos == 0:
+ return 0
+
+ if n is not None:
+ numpos = min(numpos, n)
+ delta_recall = 1.0 / numpos
+ poscount = 0.0
+
+ # calculate the ap
+ r = len(sortidx)
+ if n is not None:
+ r = min(r, n)
+ for i in range(r):
+ if actuals[sortidx[i]] > 0:
+ poscount += 1
+ ap += poscount / (i + 1) * delta_recall
+ return ap
+
+ @staticmethod
+ def _shuffle(predictions, actuals):
+ random.seed(0)
+ suffidx = random.sample(range(len(predictions)), len(predictions))
+ predictions = predictions[suffidx]
+ actuals = actuals[suffidx]
+ return predictions, actuals
+
+ @staticmethod
+ def _zero_one_normalize(predictions, epsilon=1e-7):
+ """Normalize the predictions to the range between 0.0 and 1.0.
+
+ For some predictions like SVM predictions, we need to normalize them before
+ calculate the interpolated average precision. The normalization will not
+ change the rank in the original list and thus won't change the average
+ precision.
+
+ Args:
+ predictions: a numpy 1-D array storing the sparse prediction scores.
+ epsilon: a small constant to avoid denominator being zero.
+
+ Returns:
+ The normalized prediction.
+ """
+ denominator = numpy.max(predictions) - numpy.min(predictions)
+ ret = (predictions - numpy.min(predictions)) / numpy.max(denominator,
+ epsilon)
+ return ret
diff --git a/fluid/PaddleCV/video/metrics/youtube8m/eval_util.py b/fluid/PaddleCV/video/metrics/youtube8m/eval_util.py
new file mode 100644
index 0000000000000000000000000000000000000000..f7742236f1176073eae84fdc7c3a3a1a2e294fe0
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/youtube8m/eval_util.py
@@ -0,0 +1,245 @@
+# Copyright 2016 Google Inc. 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.
+"""Provides functions to help with evaluating models."""
+import datetime
+import numpy
+
+from . import mean_average_precision_calculator as map_calculator
+from . import average_precision_calculator as ap_calculator
+
+
+def flatten(l):
+ """ Merges a list of lists into a single list. """
+ return [item for sublist in l for item in sublist]
+
+
+def calculate_hit_at_one(predictions, actuals):
+ """Performs a local (numpy) calculation of the hit at one.
+
+ Args:
+ predictions: Matrix containing the outputs of the model.
+ Dimensions are 'batch' x 'num_classes'.
+ actuals: Matrix containing the ground truth labels.
+ Dimensions are 'batch' x 'num_classes'.
+
+ Returns:
+ float: The average hit at one across the entire batch.
+ """
+ top_prediction = numpy.argmax(predictions, 1)
+ hits = actuals[numpy.arange(actuals.shape[0]), top_prediction]
+ return numpy.average(hits)
+
+
+def calculate_precision_at_equal_recall_rate(predictions, actuals):
+ """Performs a local (numpy) calculation of the PERR.
+
+ Args:
+ predictions: Matrix containing the outputs of the model.
+ Dimensions are 'batch' x 'num_classes'.
+ actuals: Matrix containing the ground truth labels.
+ Dimensions are 'batch' x 'num_classes'.
+
+ Returns:
+ float: The average precision at equal recall rate across the entire batch.
+ """
+ aggregated_precision = 0.0
+ num_videos = actuals.shape[0]
+ for row in numpy.arange(num_videos):
+ num_labels = int(numpy.sum(actuals[row]))
+ top_indices = numpy.argpartition(predictions[row],
+ -num_labels)[-num_labels:]
+ item_precision = 0.0
+ for label_index in top_indices:
+ if predictions[row][label_index] > 0:
+ item_precision += actuals[row][label_index]
+ item_precision /= top_indices.size
+ aggregated_precision += item_precision
+ aggregated_precision /= num_videos
+ return aggregated_precision
+
+
+def calculate_gap(predictions, actuals, top_k=20):
+ """Performs a local (numpy) calculation of the global average precision.
+
+ Only the top_k predictions are taken for each of the videos.
+
+ Args:
+ predictions: Matrix containing the outputs of the model.
+ Dimensions are 'batch' x 'num_classes'.
+ actuals: Matrix containing the ground truth labels.
+ Dimensions are 'batch' x 'num_classes'.
+ top_k: How many predictions to use per video.
+
+ Returns:
+ float: The global average precision.
+ """
+ gap_calculator = ap_calculator.AveragePrecisionCalculator()
+ sparse_predictions, sparse_labels, num_positives = top_k_by_class(
+ predictions, actuals, top_k)
+ gap_calculator.accumulate(
+ flatten(sparse_predictions), flatten(sparse_labels), sum(num_positives))
+ return gap_calculator.peek_ap_at_n()
+
+
+def top_k_by_class(predictions, labels, k=20):
+ """Extracts the top k predictions for each video, sorted by class.
+
+ Args:
+ predictions: A numpy matrix containing the outputs of the model.
+ Dimensions are 'batch' x 'num_classes'.
+ k: the top k non-zero entries to preserve in each prediction.
+
+ Returns:
+ A tuple (predictions,labels, true_positives). 'predictions' and 'labels'
+ are lists of lists of floats. 'true_positives' is a list of scalars. The
+ length of the lists are equal to the number of classes. The entries in the
+ predictions variable are probability predictions, and
+ the corresponding entries in the labels variable are the ground truth for
+ those predictions. The entries in 'true_positives' are the number of true
+ positives for each class in the ground truth.
+
+ Raises:
+ ValueError: An error occurred when the k is not a positive integer.
+ """
+ if k <= 0:
+ raise ValueError("k must be a positive integer.")
+ k = min(k, predictions.shape[1])
+ num_classes = predictions.shape[1]
+ prediction_triplets = []
+ for video_index in range(predictions.shape[0]):
+ prediction_triplets.extend(
+ top_k_triplets(predictions[video_index], labels[video_index], k))
+ out_predictions = [[] for v in range(num_classes)]
+ out_labels = [[] for v in range(num_classes)]
+ for triplet in prediction_triplets:
+ out_predictions[triplet[0]].append(triplet[1])
+ out_labels[triplet[0]].append(triplet[2])
+ out_true_positives = [numpy.sum(labels[:, i]) for i in range(num_classes)]
+
+ return out_predictions, out_labels, out_true_positives
+
+
+def top_k_triplets(predictions, labels, k=20):
+ """Get the top_k for a 1-d numpy array. Returns a sparse list of tuples in
+ (prediction, class) format"""
+ m = len(predictions)
+ k = min(k, m)
+ indices = numpy.argpartition(predictions, -k)[-k:]
+ return [(index, predictions[index], labels[index]) for index in indices]
+
+
+class EvaluationMetrics(object):
+ """A class to store the evaluation metrics."""
+
+ def __init__(self, num_class, top_k):
+ """Construct an EvaluationMetrics object to store the evaluation metrics.
+
+ Args:
+ num_class: A positive integer specifying the number of classes.
+ top_k: A positive integer specifying how many predictions are considered per video.
+
+ Raises:
+ ValueError: An error occurred when MeanAveragePrecisionCalculator cannot
+ not be constructed.
+ """
+ self.sum_hit_at_one = 0.0
+ self.sum_perr = 0.0
+ self.sum_loss = 0.0
+ self.map_calculator = map_calculator.MeanAveragePrecisionCalculator(
+ num_class)
+ self.global_ap_calculator = ap_calculator.AveragePrecisionCalculator()
+ self.top_k = top_k
+ self.num_examples = 0
+
+ #def accumulate(self, predictions, labels, loss):
+ def accumulate(self, loss, predictions, labels):
+ """Accumulate the metrics calculated locally for this mini-batch.
+
+ Args:
+ predictions: A numpy matrix containing the outputs of the model.
+ Dimensions are 'batch' x 'num_classes'.
+ labels: A numpy matrix containing the ground truth labels.
+ Dimensions are 'batch' x 'num_classes'.
+ loss: A numpy array containing the loss for each sample.
+
+ Returns:
+ dictionary: A dictionary storing the metrics for the mini-batch.
+
+ Raises:
+ ValueError: An error occurred when the shape of predictions and actuals
+ does not match.
+ """
+ batch_size = labels.shape[0]
+ mean_hit_at_one = calculate_hit_at_one(predictions, labels)
+ mean_perr = calculate_precision_at_equal_recall_rate(predictions,
+ labels)
+ mean_loss = numpy.mean(loss)
+
+ # Take the top 20 predictions.
+ sparse_predictions, sparse_labels, num_positives = top_k_by_class(
+ predictions, labels, self.top_k)
+ self.map_calculator.accumulate(sparse_predictions, sparse_labels,
+ num_positives)
+ self.global_ap_calculator.accumulate(
+ flatten(sparse_predictions),
+ flatten(sparse_labels), sum(num_positives))
+
+ self.num_examples += batch_size
+ self.sum_hit_at_one += mean_hit_at_one * batch_size
+ self.sum_perr += mean_perr * batch_size
+ self.sum_loss += mean_loss * batch_size
+
+ return {
+ "hit_at_one": mean_hit_at_one,
+ "perr": mean_perr,
+ "loss": mean_loss
+ }
+
+ def get(self):
+ """Calculate the evaluation metrics for the whole epoch.
+
+ Raises:
+ ValueError: If no examples were accumulated.
+
+ Returns:
+ dictionary: a dictionary storing the evaluation metrics for the epoch. The
+ dictionary has the fields: avg_hit_at_one, avg_perr, avg_loss, and
+ aps (default nan).
+ """
+ if self.num_examples <= 0:
+ raise ValueError("total_sample must be positive.")
+ avg_hit_at_one = self.sum_hit_at_one / self.num_examples
+ avg_perr = self.sum_perr / self.num_examples
+ avg_loss = self.sum_loss / self.num_examples
+
+ aps = self.map_calculator.peek_map_at_n()
+ gap = self.global_ap_calculator.peek_ap_at_n()
+
+ epoch_info_dict = {}
+ return {
+ "avg_hit_at_one": avg_hit_at_one,
+ "avg_perr": avg_perr,
+ "avg_loss": avg_loss,
+ "aps": aps,
+ "gap": gap
+ }
+
+ def clear(self):
+ """Clear the evaluation metrics and reset the EvaluationMetrics object."""
+ self.sum_hit_at_one = 0.0
+ self.sum_perr = 0.0
+ self.sum_loss = 0.0
+ self.map_calculator.clear()
+ self.global_ap_calculator.clear()
+ self.num_examples = 0
diff --git a/fluid/PaddleCV/video/metrics/youtube8m/mean_average_precision_calculator.py b/fluid/PaddleCV/video/metrics/youtube8m/mean_average_precision_calculator.py
new file mode 100644
index 0000000000000000000000000000000000000000..0ae8b0ed3717aba13b7ed35b4af025be40423967
--- /dev/null
+++ b/fluid/PaddleCV/video/metrics/youtube8m/mean_average_precision_calculator.py
@@ -0,0 +1,114 @@
+# Copyright 2016 Google Inc. 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.
+"""Calculate the mean average precision.
+
+It provides an interface for calculating mean average precision
+for an entire list or the top-n ranked items.
+
+Example usages:
+We first call the function accumulate many times to process parts of the ranked
+list. After processing all the parts, we call peek_map_at_n
+to calculate the mean average precision.
+
+```
+import random
+
+p = np.array([[random.random() for _ in xrange(50)] for _ in xrange(1000)])
+a = np.array([[random.choice([0, 1]) for _ in xrange(50)]
+ for _ in xrange(1000)])
+
+# mean average precision for 50 classes.
+calculator = mean_average_precision_calculator.MeanAveragePrecisionCalculator(
+ num_class=50)
+calculator.accumulate(p, a)
+aps = calculator.peek_map_at_n()
+```
+"""
+
+import numpy
+from . import average_precision_calculator
+
+
+class MeanAveragePrecisionCalculator(object):
+ """This class is to calculate mean average precision.
+ """
+
+ def __init__(self, num_class):
+ """Construct a calculator to calculate the (macro) average precision.
+
+ Args:
+ num_class: A positive Integer specifying the number of classes.
+ top_n_array: A list of positive integers specifying the top n for each
+ class. The top n in each class will be used to calculate its average
+ precision at n.
+ The size of the array must be num_class.
+
+ Raises:
+ ValueError: An error occurred when num_class is not a positive integer;
+ or the top_n_array is not a list of positive integers.
+ """
+ if not isinstance(num_class, int) or num_class <= 1:
+ raise ValueError("num_class must be a positive integer.")
+
+ self._ap_calculators = [] # member of AveragePrecisionCalculator
+ self._num_class = num_class # total number of classes
+ for i in range(num_class):
+ self._ap_calculators.append(
+ average_precision_calculator.AveragePrecisionCalculator())
+
+ def accumulate(self, predictions, actuals, num_positives=None):
+ """Accumulate the predictions and their ground truth labels.
+
+ Args:
+ predictions: A list of lists storing the prediction scores. The outer
+ dimension corresponds to classes.
+ actuals: A list of lists storing the ground truth labels. The dimensions
+ should correspond to the predictions input. Any value
+ larger than 0 will be treated as positives, otherwise as negatives.
+ num_positives: If provided, it is a list of numbers representing the
+ number of true positives for each class. If not provided, the number of
+ true positives will be inferred from the 'actuals' array.
+
+ Raises:
+ ValueError: An error occurred when the shape of predictions and actuals
+ does not match.
+ """
+ if not num_positives:
+ num_positives = [None for i in predictions.shape[1]]
+
+ calculators = self._ap_calculators
+ for i in range(len(predictions)):
+ calculators[i].accumulate(predictions[i], actuals[i],
+ num_positives[i])
+
+ def clear(self):
+ for calculator in self._ap_calculators:
+ calculator.clear()
+
+ def is_empty(self):
+ return ([calculator.heap_size for calculator in self._ap_calculators] ==
+ [0 for _ in range(self._num_class)])
+
+ def peek_map_at_n(self):
+ """Peek the non-interpolated mean average precision at n.
+
+ Returns:
+ An array of non-interpolated average precision at n (default 0) for each
+ class.
+ """
+ aps = [
+ self._ap_calculators[i].peek_ap_at_n()
+ for i in range(self._num_class)
+ ]
+ return aps
diff --git a/fluid/PaddleCV/video/models/__init__.py b/fluid/PaddleCV/video/models/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..ae3da375a60279edd0b8be104ef399d9b0c7d725
--- /dev/null
+++ b/fluid/PaddleCV/video/models/__init__.py
@@ -0,0 +1,13 @@
+from .model import regist_model, get_model
+from .attention_cluster import AttentionCluster
+from .nextvlad import NEXTVLAD
+from .tsn import TSN
+from .stnet import STNET
+from .attention_lstm import AttentionLSTM
+
+# regist models
+regist_model("AttentionCluster", AttentionCluster)
+regist_model("NEXTVLAD", NEXTVLAD)
+regist_model("TSN", TSN)
+regist_model("STNET", STNET)
+regist_model("AttentionLSTM", AttentionLSTM)
diff --git a/fluid/PaddleCV/video/models/attention_cluster/README.md b/fluid/PaddleCV/video/models/attention_cluster/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..9f538f86dc02be3a1f2795f36a1d55da2355dae6
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_cluster/README.md
@@ -0,0 +1,103 @@
+# Attention Cluster 视频分类模型
+
+---
+## 目录
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+- [模型评估](#模型评估)
+- [模型推断](#模型推断)
+- [参考论文](#参考论文)
+
+
+## 模型简介
+
+Attention Cluster模型为ActivityNet Kinetics Challenge 2017中最佳序列模型。该模型通过带Shifting Opeation的Attention Clusters处理已抽取好的RGB、Flow、Audio数据，Attention Cluster结构如下图所示。
+
+
+
+Multimodal Attention Cluster with Shifting Operation
+
+
+详细内容请参考[Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification](https://arxiv.org/abs/1711.09550)
+
+## 数据准备
+
+Attention Cluster模型使用2nd-Youtube-8M数据集, 数据下载及准备请参考[数据说明](../../dataset/README.md)
+
+## 模型训练
+
+数据准备完毕后，可以通过如下两种方式启动训练：
+
+ python train.py --model-name=AttentionCluster
+ --config=./configs/attention_cluster.txt
+ --save-dir=checkpoints
+ --log-interval=10
+ --valid-interval=1
+
+ bash scripts/train/train_attention_cluster.sh
+
+- 可下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_cluster_youtube8m.tar.gz)通过`--resume`指定权重存放路径进行finetune等开发
+
+**数据读取器说明：** 模型读取Youtube-8M数据集中已抽取好的`rgb`和`audio`数据，对于每个视频的数据，均匀采样100帧，该值由配置文件中的`seg_num`参数指定。
+
+**模型设置：** 模型主要可配置参数为`cluster_nums`和`seg_num`参数，当配置`cluster_nums`为32, `seg_num`为100时，在Nvidia Tesla P40上单卡可跑`batch_size=256`。
+
+**训练策略：**
+
+* 采用Adam优化器，初始learning\_rate=0.001。
+* 训练过程中不使用权重衰减。
+* 参数主要使用MSRA初始化
+
+## 模型评估
+
+可通过如下两种方式进行模型评估:
+
+ python test.py --model-name=AttentionCluster
+ --config=configs/attention_cluster.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+
+ bash scripts/test/test_attention_cluster.sh
+
+- 使用`scripts/test/test_attention_cluster.sh`进行评估时，需要修改脚本中的`--weights`参数指定需要评估的权重。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_cluster_youtube8m.tar.gz)进行评估
+
+当取如下参数时:
+
+| 参数 | 取值 |
+| :---------: | :----: |
+| cluster\_nums | 32 |
+| seg\_num | 100 |
+| batch\_size | 2048 |
+| nums\_gpu | 7 |
+
+在2nd-YouTube-8M数据集下评估精度如下:
+
+
+| 精度指标 | 模型精度 |
+| :---------: | :----: |
+| Hit@1 | 0.87 |
+| PERR | 0.78 |
+| GAP | 0.84 |
+
+## 模型推断
+
+可通过如下命令进行模型推断：
+
+ python infer.py --model-name=attention_cluster
+ --config=configs/attention_cluster.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+ --filelist=$FILELIST
+
+- 模型推断结果存储于`AttentionCluster_infer_result`中，通过`pickle`格式存储。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_cluster_youtube8m.tar.gz)进行推断
+
+## 参考论文
+
+- [Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification](https://arxiv.org/abs/1711.09550), Xiang Long, Chuang Gan, Gerard de Melo, Jiajun Wu, Xiao Liu, Shilei Wen
+
diff --git a/fluid/PaddleCV/video/models/attention_cluster/__init__.py b/fluid/PaddleCV/video/models/attention_cluster/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..bd7ef3d595d3b3d286d765a24a0c44c4b29dc6c6
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_cluster/__init__.py
@@ -0,0 +1,3 @@
+from __future__ import absolute_import
+
+from .attention_cluster import *
diff --git a/fluid/PaddleCV/video/models/attention_cluster/attention_cluster.py b/fluid/PaddleCV/video/models/attention_cluster/attention_cluster.py
new file mode 100755
index 0000000000000000000000000000000000000000..84282544c95b21231a043e22f2c8dadd25579e8f
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_cluster/attention_cluster.py
@@ -0,0 +1,139 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+
+from ..model import ModelBase
+from .shifting_attention import ShiftingAttentionModel
+from .logistic_model import LogisticModel
+
+__all__ = ["AttentionCluster"]
+
+
+class AttentionCluster(ModelBase):
+ def __init__(self, name, cfg, mode='train'):
+ super(AttentionCluster, self).__init__(name, cfg, mode)
+ self.get_config()
+
+ def get_config(self):
+ # get model configs
+ self.feature_num = self.cfg.MODEL.feature_num
+ self.feature_names = self.cfg.MODEL.feature_names
+ self.feature_dims = self.cfg.MODEL.feature_dims
+ self.cluster_nums = self.cfg.MODEL.cluster_nums
+ self.seg_num = self.cfg.MODEL.seg_num
+ self.class_num = self.cfg.MODEL.num_classes
+ self.drop_rate = self.cfg.MODEL.drop_rate
+
+ if self.mode == 'train':
+ self.learning_rate = self.get_config_from_sec('train',
+ 'learning_rate', 1e-3)
+
+ def build_input(self, use_pyreader):
+ if use_pyreader:
+ assert self.mode != 'infer', \
+ 'pyreader is not recommendated when infer, please set use_pyreader to be false.'
+ shapes = []
+ for dim in self.feature_dims:
+ shapes.append([-1, self.seg_num, dim])
+ shapes.append([-1, self.class_num]) # label
+ self.py_reader = fluid.layers.py_reader(
+ capacity=1024,
+ shapes=shapes,
+ lod_levels=[0] * (self.feature_num + 1),
+ dtypes=['float32'] * (self.feature_num + 1),
+ name='train_py_reader'
+ if self.is_training else 'test_py_reader',
+ use_double_buffer=True)
+ inputs = fluid.layers.read_file(self.py_reader)
+ self.feature_input = inputs[:self.feature_num]
+ self.label_input = inputs[-1]
+ else:
+ self.feature_input = []
+ for name, dim in zip(self.feature_names, self.feature_dims):
+ self.feature_input.append(
+ fluid.layers.data(
+ shape=[self.seg_num, dim], dtype='float32', name=name))
+ if self.mode == 'infer':
+ self.label_input = None
+ else:
+ self.label_input = fluid.layers.data(
+ shape=[self.class_num], dtype='float32', name='label')
+
+ def build_model(self):
+ att_outs = []
+ for i, (input_dim, cluster_num, feature) in enumerate(
+ zip(self.feature_dims, self.cluster_nums, self.feature_input)):
+ att = ShiftingAttentionModel(input_dim, self.seg_num, cluster_num,
+ "satt{}".format(i))
+ att_out = att.forward(feature)
+ att_outs.append(att_out)
+ out = fluid.layers.concat(att_outs, axis=1)
+
+ if self.drop_rate > 0.:
+ out = fluid.layers.dropout(
+ out, self.drop_rate, is_test=(not self.is_training))
+
+ fc1 = fluid.layers.fc(
+ out,
+ size=1024,
+ act='tanh',
+ param_attr=ParamAttr(
+ name="fc1.weights",
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=ParamAttr(
+ name="fc1.bias", initializer=fluid.initializer.MSRA()))
+ fc2 = fluid.layers.fc(
+ fc1,
+ size=4096,
+ act='tanh',
+ param_attr=ParamAttr(
+ name="fc2.weights",
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=ParamAttr(
+ name="fc2.bias", initializer=fluid.initializer.MSRA()))
+
+ aggregate_model = LogisticModel()
+
+ self.output, self.logit = aggregate_model.build_model(
+ model_input=fc2,
+ vocab_size=self.class_num,
+ is_training=self.is_training)
+
+ def optimizer(self):
+ assert self.mode == 'train', "optimizer only can be get in train mode"
+ return fluid.optimizer.AdamOptimizer(self.learning_rate)
+
+ def loss(self):
+ assert self.mode != 'infer', "invalid loss calculationg in infer mode"
+ cost = fluid.layers.sigmoid_cross_entropy_with_logits(
+ x=self.logit, label=self.label_input)
+ cost = fluid.layers.reduce_sum(cost, dim=-1)
+ self.loss_ = fluid.layers.mean(x=cost)
+ return self.loss_
+
+ def outputs(self):
+ return [self.output, self.logit]
+
+ def feeds(self):
+ return self.feature_input if self.mode == 'infer' else self.feature_input + [
+ self.label_input
+ ]
+
+ def weights_info(self):
+ return (
+ "attention_cluster_youtube8m",
+ "https://paddlemodels.bj.bcebos.com/video_classification/attention_cluster_youtube8m.tar.gz"
+ )
diff --git a/fluid/PaddleCV/video/models/attention_cluster/logistic_model.py b/fluid/PaddleCV/video/models/attention_cluster/logistic_model.py
new file mode 100755
index 0000000000000000000000000000000000000000..6fad2a44ffc7df2049eeb04341a88b9c342c70ce
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_cluster/logistic_model.py
@@ -0,0 +1,47 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle
+import paddle.fluid as fluid
+
+
+class LogisticModel(object):
+ """Logistic model."""
+ def build_model(self,
+ model_input,
+ vocab_size,
+ **unused_params):
+ """Creates a logistic model.
+
+ Args:
+ model_input: 'batch' x 'num_features' matrix of input features.
+ vocab_size: The number of classes in the dataset.
+
+ Returns:
+ A dictionary with a tensor containing the probability predictions of the
+ model in the 'predictions' key. The dimensions of the tensor are
+ batch_size x num_classes."""
+ logit = fluid.layers.fc(
+ input=model_input,
+ size=vocab_size,
+ act=None,
+ name='logits_clf',
+ param_attr=fluid.ParamAttr(
+ name='logistic.weights',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=fluid.ParamAttr(
+ name='logistic.bias',
+ initializer=fluid.initializer.MSRA(uniform=False)))
+ output = fluid.layers.sigmoid(logit)
+ return output, logit
diff --git a/fluid/PaddleCV/video/models/attention_cluster/shifting_attention.py b/fluid/PaddleCV/video/models/attention_cluster/shifting_attention.py
new file mode 100755
index 0000000000000000000000000000000000000000..e27ad8dd58b882eb96fbb9763eecccc36ddfe28a
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_cluster/shifting_attention.py
@@ -0,0 +1,95 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+import numpy as np
+
+
+class ShiftingAttentionModel(object):
+ """Shifting Attention Model"""
+
+ def __init__(self, input_dim, seg_num, n_att, name):
+ self.n_att = n_att
+ self.input_dim = input_dim
+ self.seg_num = seg_num
+ self.name = name
+ self.gnorm = np.sqrt(n_att)
+
+ def softmax_m1(self, x):
+ x_shape = fluid.layers.shape(x)
+ x_shape.stop_gradient = True
+ flat_x = fluid.layers.reshape(x, shape=(-1, self.seg_num))
+ flat_softmax = fluid.layers.softmax(flat_x)
+ return fluid.layers.reshape(
+ flat_softmax, shape=x.shape, actual_shape=x_shape)
+
+ def glorot(self, n):
+ return np.sqrt(1.0 / np.sqrt(n))
+
+ def forward(self, x):
+ """Forward shifting attention model.
+
+ Args:
+ x: input features in shape of [N, L, F].
+
+ Returns:
+ out: output features in shape of [N, F * C]
+ """
+
+ trans_x = fluid.layers.transpose(x, perm=[0, 2, 1])
+ # scores and weight in shape [N, C, L], sum(weights, -1) = 1
+ trans_x = fluid.layers.unsqueeze(trans_x, [-1])
+ scores = fluid.layers.conv2d(
+ trans_x,
+ self.n_att,
+ filter_size=1,
+ param_attr=ParamAttr(
+ name=self.name + ".conv.weight",
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=ParamAttr(
+ name=self.name + ".conv.bias",
+ initializer=fluid.initializer.MSRA()))
+ scores = fluid.layers.squeeze(scores, [-1])
+ weights = self.softmax_m1(scores)
+
+ glrt = self.glorot(self.n_att)
+ self.w = fluid.layers.create_parameter(
+ shape=(self.n_att, ),
+ dtype=x.dtype,
+ attr=ParamAttr(self.name + ".shift_w"),
+ default_initializer=fluid.initializer.Normal(0.0, glrt))
+ self.b = fluid.layers.create_parameter(
+ shape=(self.n_att, ),
+ dtype=x.dtype,
+ attr=ParamAttr(name=self.name + ".shift_b"),
+ default_initializer=fluid.initializer.Normal(0.0, glrt))
+
+ outs = []
+ for i in range(self.n_att):
+ # slice weight and expand to shape [N, L, C]
+ weight = fluid.layers.slice(
+ weights, axes=[1], starts=[i], ends=[i + 1])
+ weight = fluid.layers.transpose(weight, perm=[0, 2, 1])
+ weight = fluid.layers.expand(weight, [1, 1, self.input_dim])
+
+ w_i = fluid.layers.slice(self.w, axes=[0], starts=[i], ends=[i + 1])
+ b_i = fluid.layers.slice(self.b, axes=[0], starts=[i], ends=[i + 1])
+ shift = fluid.layers.reduce_sum(x * weight, dim=1) * w_i + b_i
+
+ l2_norm = fluid.layers.l2_normalize(shift, axis=-1)
+ outs.append(l2_norm / self.gnorm)
+
+ out = fluid.layers.concat(outs, axis=1)
+ return out
diff --git a/fluid/PaddleCV/video/models/attention_lstm/README.md b/fluid/PaddleCV/video/models/attention_lstm/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..6458edb7d674ea2329a1f5d1b596c45e023a3f51
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_lstm/README.md
@@ -0,0 +1,93 @@
+# AttentionLSTM视频分类模型
+
+---
+## 内容
+
+- [模型简介](#简介)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+- [模型评估](#模型评估)
+- [模型推断](#模型推断)
+- [参考论文](#参考论文)
+
+## 模型简介
+
+递归神经网络（RNN）常用于序列数据的处理，可建模视频连续多帧的时序信息，在视频分类领域为基础常用方法。该模型采用了双向长短记忆网络（LSTM），将视频的所有帧特征依次编码。与传统方法直接采用LSTM最后一个时刻的输出不同，该模型增加了一个Attention层，每个时刻的隐状态输出都有一个自适应权重，然后线性加权得到最终特征向量。论文中实现的是两层LSTM结构，而本代码实现的是带Attention的双向LSTM，Attention层可参考论文[AttentionCluster](https://arxiv.org/abs/1711.09550)。
+
+详细内容请参考[Beyond Short Snippets: Deep Networks for Video Classification](https://arxiv.org/abs/1503.08909)。
+
+## 数据准备
+
+AttentionLSTM模型使用2nd-Youtube-8M数据集，关于数据部分请参考[数据说明](../../dataset/README.md)
+
+## 模型训练
+
+### 随机初始化开始训练
+
+数据准备完毕后，可以通过如下两种方式启动训练：
+
+ python train.py --model-name=AttentionLSTM
+ --config=./configs/attention_lstm.txt
+ --save-dir=checkpoints
+ --log-interval=10
+ --valid-interval=1
+
+ bash scripts/train/train_attention_lstm.sh
+
+- AttentionLSTM模型使用8卡Nvidia Tesla P40来训练的，总的batch size数是1024。
+
+### 使用预训练模型做finetune
+请先将提供的[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_lstm_youtube8m.tar.gz)下载到本地，并在上述脚本文件中添加`--resume`为所保存的预模型存放路径。
+
+## 模型评估
+可通过如下两种方式进行模型评估:
+
+ python test.py --model-name=AttentionLSTM
+ --config=configs/attention_lstm.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+
+ bash scripts/test/test_attention_lstm.sh
+
+- 使用`scripts/test/test_attention_LSTM.sh`进行评估时，需要修改脚本中的`--weights`参数指定需要评估的权重。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_lstm_youtube8m.tar.gz)进行评估
+
+模型参数列表如下：
+
+| 参数 | 取值 |
+| :---------: | :----: |
+| embedding\_size | 512 |
+| lstm\_size | 1024 |
+| drop\_rate | 0.5 |
+
+
+计算指标列表如下：
+
+| 精度指标 | 模型精度 |
+| :---------: | :----: |
+| Hit@1 | 0.8885 |
+| PERR | 0.8012 |
+| GAP | 0.8594 |
+
+
+## 模型推断
+
+可通过如下命令进行模型推断：
+
+ python infer.py --model-name=attention_lstm
+ --config=configs/attention_lstm.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+ --filelist=$FILELIST
+
+- 模型推断结果存储于`AttentionLSTM_infer_result`中，通过`pickle`格式存储。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/attention_lstm_youtube8m.tar.gz)进行推断
+
+## 参考论文
+
+- [Beyond Short Snippets: Deep Networks for Video Classification](https://arxiv.org/abs/1503.08909) Joe Yue-Hei Ng, Matthew Hausknecht, Sudheendra Vijayanarasimhan, Oriol Vinyals, Rajat Monga, George Toderici
+
+- [Attention Clusters: Purely Attention Based Local Feature Integration for Video Classification](https://arxiv.org/abs/1711.09550), Xiang Long, Chuang Gan, Gerard de Melo, Jiajun Wu, Xiao Liu, Shilei Wen
+
diff --git a/fluid/PaddleCV/video/models/attention_lstm/__init__.py b/fluid/PaddleCV/video/models/attention_lstm/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..cb872f0e43ab52054b42970896e5791a0eeb691d
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_lstm/__init__.py
@@ -0,0 +1 @@
+from .attention_lstm import *
diff --git a/fluid/PaddleCV/video/models/attention_lstm/attention_lstm.py b/fluid/PaddleCV/video/models/attention_lstm/attention_lstm.py
new file mode 100755
index 0000000000000000000000000000000000000000..5d28dc47e297ba76e2b07fd69de2efa3ae6ccb0f
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_lstm/attention_lstm.py
@@ -0,0 +1,151 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+
+from ..model import ModelBase
+from .lstm_attention import LSTMAttentionModel
+
+__all__ = ["AttentionLSTM"]
+
+
+class AttentionLSTM(ModelBase):
+ def __init__(self, name, cfg, mode='train'):
+ super(AttentionLSTM, self).__init__(name, cfg, mode)
+ self.get_config()
+
+ def get_config(self):
+ # get model configs
+ self.feature_num = self.cfg.MODEL.feature_num
+ self.feature_names = self.cfg.MODEL.feature_names
+ self.feature_dims = self.cfg.MODEL.feature_dims
+ self.num_classes = self.cfg.MODEL.num_classes
+ self.embedding_size = self.cfg.MODEL.embedding_size
+ self.lstm_size = self.cfg.MODEL.lstm_size
+ self.drop_rate = self.cfg.MODEL.drop_rate
+
+ # get mode configs
+ self.batch_size = self.get_config_from_sec(self.mode, 'batch_size', 1)
+ self.num_gpus = self.get_config_from_sec(self.mode, 'num_gpus', 1)
+
+ if self.mode == 'train':
+ self.learning_rate = self.get_config_from_sec('train',
+ 'learning_rate', 1e-3)
+ self.weight_decay = self.get_config_from_sec('train',
+ 'weight_decay', 8e-4)
+ self.num_samples = self.get_config_from_sec('train', 'num_samples',
+ 5000000)
+ self.decay_epochs = self.get_config_from_sec('train',
+ 'decay_epochs', [5])
+ self.decay_gamma = self.get_config_from_sec('train', 'decay_gamma',
+ 0.1)
+
+ def build_input(self, use_pyreader):
+ if use_pyreader:
+ assert self.mode != 'infer', \
+ 'pyreader is not recommendated when infer, please set use_pyreader to be false.'
+ shapes = []
+ for dim in self.feature_dims:
+ shapes.append([-1, dim])
+ shapes.append([-1, self.num_classes]) # label
+ self.py_reader = fluid.layers.py_reader(
+ capacity=1024,
+ shapes=shapes,
+ lod_levels=[1] * self.feature_num + [0],
+ dtypes=['float32'] * (self.feature_num + 1),
+ name='train_py_reader'
+ if self.is_training else 'test_py_reader',
+ use_double_buffer=True)
+ inputs = fluid.layers.read_file(self.py_reader)
+ self.feature_input = inputs[:self.feature_num]
+ self.label_input = inputs[-1]
+ else:
+ self.feature_input = []
+ for name, dim in zip(self.feature_names, self.feature_dims):
+ self.feature_input.append(
+ fluid.layers.data(
+ shape=[dim], lod_level=1, dtype='float32', name=name))
+ if self.mode == 'infer':
+ self.label_input = None
+ else:
+ self.label_input = fluid.layers.data(
+ shape=[self.num_classes], dtype='float32', name='label')
+
+ def build_model(self):
+ att_outs = []
+ for i, (input_dim, feature
+ ) in enumerate(zip(self.feature_dims, self.feature_input)):
+ att = LSTMAttentionModel(input_dim, self.embedding_size,
+ self.lstm_size, self.drop_rate)
+ att_out = att.forward(feature, is_training=(self.mode == 'train'))
+ att_outs.append(att_out)
+ out = fluid.layers.concat(att_outs, axis=1)
+
+ fc1 = fluid.layers.fc(
+ input=out,
+ size=8192,
+ act='relu',
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+ fc2 = fluid.layers.fc(
+ input=fc1,
+ size=4096,
+ act='tanh',
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+
+ self.logit = fluid.layers.fc(input=fc2, size=self.num_classes, act=None, \
+ bias_attr=ParamAttr(regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+
+ self.output = fluid.layers.sigmoid(self.logit)
+
+ def optimizer(self):
+ assert self.mode == 'train', "optimizer only can be get in train mode"
+ values = [
+ self.learning_rate * (self.decay_gamma**i)
+ for i in range(len(self.decay_epochs) + 1)
+ ]
+ iter_per_epoch = self.num_samples / self.batch_size
+ boundaries = [e * iter_per_epoch for e in self.decay_epochs]
+ return fluid.optimizer.RMSProp(
+ learning_rate=fluid.layers.piecewise_decay(
+ values=values, boundaries=boundaries),
+ centered=True,
+ regularization=fluid.regularizer.L2Decay(self.weight_decay))
+
+ def loss(self):
+ assert self.mode != 'infer', "invalid loss calculationg in infer mode"
+ cost = fluid.layers.sigmoid_cross_entropy_with_logits(
+ x=self.logit, label=self.label_input)
+ cost = fluid.layers.reduce_sum(cost, dim=-1)
+ sum_cost = fluid.layers.reduce_sum(cost)
+ self.loss_ = fluid.layers.scale(
+ sum_cost, scale=self.num_gpus, bias_after_scale=False)
+ return self.loss_
+
+ def outputs(self):
+ return [self.output, self.logit]
+
+ def feeds(self):
+ return self.feature_input if self.mode == 'infer' else self.feature_input + [
+ self.label_input
+ ]
+
+ def weights_info(self):
+ return ('attention_lstm_youtube8m',
+ 'https://paddlemodels.bj.bcebos.com/video_classification/attention_lstm_youtube8m.tar.gz')
diff --git a/fluid/PaddleCV/video/models/attention_lstm/lstm_attention.py b/fluid/PaddleCV/video/models/attention_lstm/lstm_attention.py
new file mode 100755
index 0000000000000000000000000000000000000000..5fce85c5c75ea176a3bf371de5f4eea5f02f25b3
--- /dev/null
+++ b/fluid/PaddleCV/video/models/attention_lstm/lstm_attention.py
@@ -0,0 +1,78 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+import numpy as np
+
+
+class LSTMAttentionModel(object):
+ """LSTM Attention Model"""
+
+ def __init__(self,
+ bias_attr,
+ embedding_size=512,
+ lstm_size=1024,
+ drop_rate=0.5):
+ self.lstm_size = lstm_size
+ self.embedding_size = embedding_size
+ self.drop_rate = drop_rate
+
+ def forward(self, input, is_training):
+ input_fc = fluid.layers.fc(
+ input=input,
+ size=self.embedding_size,
+ act='tanh',
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+
+ lstm_forward_fc = fluid.layers.fc(
+ input=input_fc,
+ size=self.lstm_size * 4,
+ act=None,
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+ lstm_forward, _ = fluid.layers.dynamic_lstm(
+ input=lstm_forward_fc, size=self.lstm_size * 4, is_reverse=False)
+
+ lsmt_backward_fc = fluid.layers.fc(
+ input=input_fc,
+ size=self.lstm_size * 4,
+ act=None,
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+ lstm_backward, _ = fluid.layers.dynamic_lstm(
+ input=lsmt_backward_fc, size=self.lstm_size * 4, is_reverse=True)
+
+ lstm_concat = fluid.layers.concat(
+ input=[lstm_forward, lstm_backward], axis=1)
+
+ lstm_dropout = fluid.layers.dropout(
+ x=lstm_concat, dropout_prob=self.drop_rate, is_test=(not is_training))
+
+ lstm_weight = fluid.layers.fc(
+ input=lstm_dropout,
+ size=1,
+ act='sequence_softmax',
+ bias_attr=ParamAttr(
+ regularizer=fluid.regularizer.L2Decay(0.0),
+ initializer=fluid.initializer.NormalInitializer(scale=0.0)))
+ scaled = fluid.layers.elementwise_mul(
+ x=lstm_dropout, y=lstm_weight, axis=0)
+ lstm_pool = fluid.layers.sequence_pool(input=scaled, pool_type='sum')
+
+ return lstm_pool
diff --git a/fluid/PaddleCV/video/models/model.py b/fluid/PaddleCV/video/models/model.py
new file mode 100755
index 0000000000000000000000000000000000000000..44f888ef39ef1445fae0a6f0e3622002bf6cb66a
--- /dev/null
+++ b/fluid/PaddleCV/video/models/model.py
@@ -0,0 +1,181 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import logging
+try:
+ from configparser import ConfigParser
+except:
+ from ConfigParser import ConfigParser
+
+import paddle.fluid as fluid
+from datareader import get_reader
+from metrics import get_metrics
+from .utils import download, AttrDict
+
+WEIGHT_DIR = os.path.expanduser("~/.paddle/weights")
+
+logger = logging.getLogger(__name__)
+
+
+class NotImplementError(Exception):
+ "Error: model function not implement"
+
+ def __init__(self, model, function):
+ super(NotImplementError, self).__init__()
+ self.model = model.__class__.__name__
+ self.function = function.__name__
+
+ def __str__(self):
+ return "Function {}() is not implemented in model {}".format(
+ self.function, self.model)
+
+
+class ModelNotFoundError(Exception):
+ "Error: model not found"
+
+ def __init__(self, model_name, avail_models):
+ super(ModelNotFoundError, self).__init__()
+ self.model_name = model_name
+ self.avail_models = avail_models
+
+ def __str__(self):
+ msg = "Model {} Not Found.\nAvailiable models:\n".format(
+ self.model_name)
+ for model in self.avail_models:
+ msg += " {}\n".format(model)
+ return msg
+
+
+class ModelBase(object):
+ def __init__(self, name, cfg, mode='train'):
+ assert mode in ['train', 'valid', 'test', 'infer'], \
+ "Unknown mode type {}".format(mode)
+ self.name = name
+ self.is_training = (mode == 'train')
+ self.mode = mode
+ self.py_reader = None
+
+ # parse config
+ # assert os.path.exists(cfg), \
+ # "Config file {} not exists".format(cfg)
+ # self._config = ModelConfig(cfg)
+ # self._config.parse()
+ # if args and isinstance(args, dict):
+ # self._config.merge_configs(mode, args)
+ # self.cfg = self._config.get_configs()
+ self.cfg = cfg
+
+ def build_model(self):
+ "build model struct"
+ raise NotImplementError(self, self.build_model)
+
+ def build_input(self, use_pyreader):
+ "build input Variable"
+ raise NotImplementError(self, self.build_input)
+
+ def optimizer(self):
+ "get model optimizer"
+ raise NotImplementError(self, self.optimizer)
+
+ def outputs():
+ "get output variable"
+ raise notimplementerror(self, self.outputs)
+
+ def loss(self):
+ "get loss variable"
+ raise notimplementerror(self, self.loss)
+
+ def feeds(self):
+ "get feed inputs list"
+ raise NotImplementError(self, self.feeds)
+
+ def weights_info(self):
+ "get model weight default path and download url"
+ raise NotImplementError(self, self.weights_info)
+
+ def get_weights(self):
+ "get model weight file path, download weight from Paddle if not exist"
+ path, url = self.weights_info()
+ path = os.path.join(WEIGHT_DIR, path)
+ if os.path.exists(path):
+ return path
+
+ logger.info("Download weights of {} from {}".format(self.name, url))
+ download(url, path)
+ return path
+
+ def pyreader(self):
+ return self.py_reader
+
+ def epoch_num(self):
+ "get train epoch num"
+ return self.cfg.TRAIN.epoch
+
+ def pretrain_info(self):
+ "get pretrain base model directory"
+ return (None, None)
+
+ def get_pretrain_weights(self):
+ "get model weight file path, download weight from Paddle if not exist"
+ path, url = self.pretrain_info()
+ if not path:
+ return None
+
+ path = os.path.join(WEIGHT_DIR, path)
+ if os.path.exists(path):
+ return path
+
+ logger.info("Download pretrain weights of {} from {}".format(
+ self.name, url))
+ download(url, path)
+ return path
+
+ def load_pretrain_params(self, exe, pretrain, prog, place):
+ logger.info("Load pretrain weights from {}".format(pretrain))
+ fluid.io.load_params(exe, pretrain, main_program=prog)
+
+ def get_config_from_sec(self, sec, item, default=None):
+ if sec.upper() not in self.cfg:
+ return default
+ return self.cfg[sec.upper()].get(item, default)
+
+
+class ModelZoo(object):
+ def __init__(self):
+ self.model_zoo = {}
+
+ def regist(self, name, model):
+ assert model.__base__ == ModelBase, "Unknow model type {}".format(
+ type(model))
+ self.model_zoo[name] = model
+
+ def get(self, name, cfg, mode='train'):
+ for k, v in self.model_zoo.items():
+ if k == name:
+ return v(name, cfg, mode)
+ raise ModelNotFoundError(name, self.model_zoo.keys())
+
+
+# singleton model_zoo
+model_zoo = ModelZoo()
+
+
+def regist_model(name, model):
+ model_zoo.regist(name, model)
+
+
+def get_model(name, cfg, mode='train'):
+ return model_zoo.get(name, cfg, mode)
+
diff --git a/fluid/PaddleCV/video/models/nextvlad/README.md b/fluid/PaddleCV/video/models/nextvlad/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..61b634e7fd52e2910913a5ca31c0413ce43e5021
--- /dev/null
+++ b/fluid/PaddleCV/video/models/nextvlad/README.md
@@ -0,0 +1,82 @@
+# NeXtVLAD视频分类模型
+
+---
+## 目录
+
+- [算法介绍](#模型简介)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+- [模型评估](#模型评估)
+- [模型推断](#模型推断)
+- [参考论文](#参考论文)
+
+
+## 算法介绍
+NeXtVLAD模型是第二届Youtube-8M视频理解竞赛中效果最好的单模型，在参数量小于80M的情况下，能得到高于0.87的GAP指标。该模型提供了一种将桢级别的视频特征转化并压缩成特征向量，以适用于大尺寸视频文件的分类的方法。其基本出发点是在NetVLAD模型的基础上，将高维度的特征先进行分组，通过引入attention机制聚合提取时间维度的信息，这样既可以获得较高的准确率，又可以使用更少的参数量。详细内容请参考[NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification](https://arxiv.org/abs/1811.05014)。
+
+这里实现了论文中的单模型结构，使用2nd-Youtube-8M的train数据集作为训练集，在val数据集上做测试。
+
+## 数据准备
+
+NeXtVLAD模型使用2nd-Youtube-8M数据集, 数据下载及准备请参考[数据说明](../../dataset/README.md)
+
+## 模型训练
+
+### 随机初始化开始训练
+在video目录下运行如下脚本即可
+
+ bash ./scripts/train/train_nextvlad.sh
+
+### 使用预训练模型做finetune
+
+请先将提供的预训练模型[model](https://paddlemodels.bj.bcebos.com/video_classification/nextvlad_youtube8m.tar.gz)下载到本地，并在上述脚本文件中添加--resume为所保存的预模型存放路径。
+
+使用4卡Nvidia Tesla P40，总的batch size数是160。
+
+### 训练策略
+
+* 使用Adam优化器，初始learning\_rate=0.0002
+* 每2,000,000个样本做一次学习率衰减，learning\_rate\_decay = 0.8
+* 正则化使用l2\_weight\_decay = 1e-5
+
+## 模型评估
+
+用户可以下载的预训练模型参数，或者使用自己训练好的模型参数，请在./scripts/test/test\_nextvald.sh
+文件中修改--weights参数为保存模型参数的目录。运行
+
+ bash ./scripts/test/test_nextvlad.sh
+
+由于youtube-8m提供的数据中test数据集是没有ground truth标签的，所以这里使用validation数据集来做测试。
+
+模型参数列表如下：
+
+| 参数 | 取值 |
+| :---------: | :----: |
+| cluster\_size | 128 |
+| hidden\_size | 2048 |
+| groups | 8 |
+| expansion | 2 |
+| drop\_rate | 0.5 |
+| gating\_reduction | 8 |
+
+计算指标列表如下：
+
+| 精度指标 | 模型精度 |
+| :---------: | :----: |
+| Hit@1 | 0.8960 |
+| PERR | 0.8132 |
+| GAP | 0.8709 |
+
+## 模型推断
+
+用户可以下载的预训练模型参数，或者使用自己训练好的模型参数，请在./scripts/infer/infer\_nextvald.sh
+文件中修改--weights参数为保存模型参数的目录，运行如下脚本
+
+ bash ./scripts/infer/infer_nextvald.sh
+
+推断结果会保存在NEXTVLAD\_infer\_result文件中，通过pickle格式存储。
+
+## 参考论文
+
+- [NeXtVLAD: An Efficient Neural Network to Aggregate Frame-level Features for Large-scale Video Classification](https://arxiv.org/abs/1811.05014), Rongcheng Lin, Jing Xiao, Jianping Fan
+
diff --git a/fluid/PaddleCV/video/models/nextvlad/__init__.py b/fluid/PaddleCV/video/models/nextvlad/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..d9a233374a1ac7069280801413872e9227f820a8
--- /dev/null
+++ b/fluid/PaddleCV/video/models/nextvlad/__init__.py
@@ -0,0 +1,3 @@
+from __future__ import absolute_import
+
+from .nextvlad import *
diff --git a/fluid/PaddleCV/video/models/nextvlad/clf_model.py b/fluid/PaddleCV/video/models/nextvlad/clf_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..70728dfb1139a1d32e6a5d921629ba018ed6cea9
--- /dev/null
+++ b/fluid/PaddleCV/video/models/nextvlad/clf_model.py
@@ -0,0 +1,50 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle
+import paddle.fluid as fluid
+
+
+class LogisticModel(object):
+ """Logistic model with L2 regularization."""
+
+ def create_model(self,
+ model_input,
+ vocab_size,
+ l2_penalty=None,
+ **unused_params):
+ """Creates a logistic model.
+
+ Args:
+ model_input: 'batch' x 'num_features' matrix of input features.
+ vocab_size: The number of classes in the dataset.
+
+ Returns:
+ A dictionary with a tensor containing the probability predictions of the
+ model in the 'predictions' key. The dimensions of the tensor are
+ batch_size x num_classes."""
+ logits = fluid.layers.fc(
+ input=model_input,
+ size=vocab_size,
+ act=None,
+ name='logits_clf',
+ param_attr=fluid.ParamAttr(
+ name='logits_clf_weights',
+ initializer=fluid.initializer.MSRA(uniform=False),
+ regularizer=fluid.regularizer.L2DecayRegularizer(l2_penalty)),
+ bias_attr=fluid.ParamAttr(
+ name='logits_clf_bias',
+ regularizer=fluid.regularizer.L2DecayRegularizer(l2_penalty)))
+ output = fluid.layers.sigmoid(logits)
+ return {'predictions': output, 'logits': logits}
diff --git a/fluid/PaddleCV/video/models/nextvlad/nextvlad.py b/fluid/PaddleCV/video/models/nextvlad/nextvlad.py
new file mode 100755
index 0000000000000000000000000000000000000000..62a96b5d1d61e3447699b5ec974662566c2e45f0
--- /dev/null
+++ b/fluid/PaddleCV/video/models/nextvlad/nextvlad.py
@@ -0,0 +1,167 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+
+from ..model import ModelBase
+from .clf_model import LogisticModel
+from . import nextvlad_model
+
+__all__ = ["NEXTVLAD"]
+
+
+class NEXTVLAD(ModelBase):
+ def __init__(self, name, cfg, mode='train'):
+ super(NEXTVLAD, self).__init__(name, cfg, mode=mode)
+ self.get_config()
+
+ def get_config(self):
+ # model params
+ self.num_classes = self.get_config_from_sec('model', 'num_classes')
+ self.video_feature_size = self.get_config_from_sec('model',
+ 'video_feature_size')
+ self.audio_feature_size = self.get_config_from_sec('model',
+ 'audio_feature_size')
+ self.cluster_size = self.get_config_from_sec('model', 'cluster_size')
+ self.hidden_size = self.get_config_from_sec('model', 'hidden_size')
+ self.groups = self.get_config_from_sec('model', 'groups')
+ self.expansion = self.get_config_from_sec('model', 'expansion')
+ self.drop_rate = self.get_config_from_sec('model', 'drop_rate')
+ self.gating_reduction = self.get_config_from_sec('model',
+ 'gating_reduction')
+ self.eigen_file = self.get_config_from_sec('model', 'eigen_file')
+ # training params
+ self.base_learning_rate = self.get_config_from_sec('train',
+ 'learning_rate')
+ self.lr_boundary_examples = self.get_config_from_sec(
+ 'train', 'lr_boundary_examples')
+ self.max_iter = self.get_config_from_sec('train', 'max_iter')
+ self.learning_rate_decay = self.get_config_from_sec(
+ 'train', 'learning_rate_decay')
+ self.l2_penalty = self.get_config_from_sec('train', 'l2_penalty')
+ self.gradient_clip_norm = self.get_config_from_sec('train',
+ 'gradient_clip_norm')
+ self.use_gpu = self.get_config_from_sec('train', 'use_gpu')
+ self.num_gpus = self.get_config_from_sec('train', 'num_gpus')
+
+ # other params
+ self.batch_size = self.get_config_from_sec(self.mode, 'batch_size')
+
+ def build_input(self, use_pyreader=True):
+ rgb_shape = [self.video_feature_size]
+ audio_shape = [self.audio_feature_size]
+ label_shape = [self.num_classes]
+ if use_pyreader:
+ assert self.mode != 'infer', \
+ 'pyreader is not recommendated when infer, please set use_pyreader to be false.'
+ py_reader = fluid.layers.py_reader(
+ capacity=100,
+ shapes=[[-1] + rgb_shape, [-1] + audio_shape,
+ [-1] + label_shape],
+ lod_levels=[1, 1, 0],
+ dtypes=['float32', 'float32', 'float32'],
+ name='train_py_reader'
+ if self.is_training else 'test_py_reader',
+ use_double_buffer=True)
+ rgb, audio, label = fluid.layers.read_file(py_reader)
+ self.py_reader = py_reader
+ else:
+ rgb = fluid.layers.data(
+ name='train_rgb' if self.is_training else 'test_rgb',
+ shape=rgb_shape,
+ dtype='float32',
+ lod_level=1)
+ audio = fluid.layers.data(
+ name='train_audio' if self.is_training else 'test_audio',
+ shape=audio_shape,
+ dtype='float32',
+ lod_level=1)
+ if self.mode == 'infer':
+ label = None
+ else:
+ label = fluid.layers.data(
+ name='train_label' if self.is_training else 'test_label',
+ shape=label_shape,
+ dtype='float32')
+ self.feature_input = [rgb, audio]
+ self.label_input = label
+
+ def create_model_args(self):
+ model_args = {}
+ model_args['class_dim'] = self.num_classes
+ model_args['cluster_size'] = self.cluster_size
+ model_args['hidden_size'] = self.hidden_size
+ model_args['groups'] = self.groups
+ model_args['expansion'] = self.expansion
+ model_args['drop_rate'] = self.drop_rate
+ model_args['gating_reduction'] = self.gating_reduction
+ model_args['l2_penalty'] = self.l2_penalty
+ return model_args
+
+ def build_model(self):
+ model_args = self.create_model_args()
+ videomodel = nextvlad_model.NeXtVLADModel()
+ rgb = self.feature_input[0]
+ audio = self.feature_input[1]
+ out = videomodel.create_model(
+ rgb, audio, is_training=(self.mode == 'train'), **model_args)
+ self.logits = out['logits']
+ self.predictions = out['predictions']
+ self.network_outputs = [out['predictions']]
+
+ def optimizer(self):
+ assert self.mode == 'train', "optimizer only can be get in train mode"
+ im_per_batch = self.batch_size
+ lr_bounds, lr_values = get_learning_rate_decay_list(
+ self.base_learning_rate, self.learning_rate_decay, self.max_iter,
+ self.lr_boundary_examples, im_per_batch)
+ return fluid.optimizer.AdamOptimizer(
+ learning_rate=fluid.layers.piecewise_decay(
+ boundaries=lr_bounds, values=lr_values))
+
+ def loss(self):
+ assert self.mode != 'infer', "invalid loss calculationg in infer mode"
+ cost = fluid.layers.sigmoid_cross_entropy_with_logits(
+ x=self.logits, label=self.label_input)
+ cost = fluid.layers.reduce_sum(cost, dim=-1)
+ self.loss_ = fluid.layers.mean(x=cost)
+ return self.loss_
+
+ def outputs(self):
+ return self.network_outputs
+
+ def feeds(self):
+ return self.feature_input if self.mode == 'infer' else self.feature_input + [
+ self.label_input
+ ]
+
+ def weights_info(self):
+ return ('nextvlad_youtube8m',
+ 'https://paddlemodels.bj.bcebos.com/video_classification/nextvlad_youtube8m.tar.gz')
+
+
+def get_learning_rate_decay_list(base_learning_rate, decay, max_iter,
+ decay_examples, total_batch_size):
+ decay_step = decay_examples // total_batch_size
+ lr_bounds = []
+ lr_values = [base_learning_rate]
+ i = 1
+ while True:
+ if i * decay_step >= max_iter:
+ break
+ lr_bounds.append(i * decay_step)
+ lr_values.append(base_learning_rate * (decay**i))
+ i += 1
+ return lr_bounds, lr_values
diff --git a/fluid/PaddleCV/video/models/nextvlad/nextvlad_model.py b/fluid/PaddleCV/video/models/nextvlad/nextvlad_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..9e9efe83fd936989b0b94ff2fadaf487c37c86b7
--- /dev/null
+++ b/fluid/PaddleCV/video/models/nextvlad/nextvlad_model.py
@@ -0,0 +1,231 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle
+import paddle.fluid as fluid
+from . import clf_model
+
+
+class NeXtVLAD(object):
+ """
+ This is a paddlepaddle implementation of the NeXtVLAD model. For more
+ information, please refer to the paper,
+ https://static.googleusercontent.com/media/research.google.com/zh-CN//youtube8m/workshop2018/p_c03.pdf
+ """
+
+ def __init__(self,
+ feature_size,
+ cluster_size,
+ is_training=True,
+ expansion=2,
+ groups=None,
+ inputname='video'):
+ self.feature_size = feature_size
+ self.cluster_size = cluster_size
+ self.is_training = is_training
+ self.expansion = expansion
+ self.groups = groups
+ self.name = inputname + '_'
+
+ def forward(self, input):
+ input = fluid.layers.fc(
+ input=input,
+ size=self.expansion * self.feature_size,
+ act=None,
+ name=self.name + 'fc_expansion',
+ param_attr=fluid.ParamAttr(
+ name=self.name + 'fc_expansion_w',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=fluid.ParamAttr(
+ name=self.name + 'fc_expansion_b',
+ initializer=fluid.initializer.Constant(value=0.)))
+
+ # attention factor of per group
+ attention = fluid.layers.fc(
+ input=input,
+ size=self.groups,
+ act='sigmoid',
+ name=self.name + 'fc_group_attention',
+ param_attr=fluid.ParamAttr(
+ name=self.name + 'fc_group_attention_w',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=fluid.ParamAttr(
+ name=self.name + 'fc_group_attention_b',
+ initializer=fluid.initializer.Constant(value=0.)))
+
+ # calculate activation factor of per group per cluster
+ feature_size = self.feature_size * self.expansion // self.groups
+ cluster_weights = fluid.layers.create_parameter(
+ shape=[
+ self.expansion * self.feature_size,
+ self.groups * self.cluster_size
+ ],
+ dtype=input.dtype,
+ attr=fluid.ParamAttr(name=self.name + 'cluster_weights'),
+ default_initializer=fluid.initializer.MSRA(uniform=False))
+
+ activation = fluid.layers.matmul(input, cluster_weights)
+ activation = fluid.layers.batch_norm(
+ activation, is_test=(not self.is_training))
+
+ # reshape of activation
+ activation = fluid.layers.reshape(activation,
+ [-1, self.groups, self.cluster_size])
+ # softmax on per cluster
+ activation = fluid.layers.softmax(activation)
+ activation = fluid.layers.elementwise_mul(activation, attention, axis=0)
+ a_sum = fluid.layers.sequence_pool(activation, 'sum')
+ a_sum = fluid.layers.reduce_sum(a_sum, dim=1)
+
+ # create cluster_weights2
+ cluster_weights2 = fluid.layers.create_parameter(
+ shape=[self.cluster_size, feature_size],
+ dtype=input.dtype,
+ attr=fluid.ParamAttr(name=self.name + 'cluster_weights2'),
+ default_initializer=fluid.initializer.MSRA(uniform=False))
+
+ # expand a_sum dimension from [-1, self.cluster_size] to be [-1, self.cluster_size, feature_size]
+ a_sum = fluid.layers.reshape(a_sum, [-1, self.cluster_size, 1])
+ a_sum = fluid.layers.expand(a_sum, [1, 1, feature_size])
+
+ # element wise multiply a_sum and cluster_weights2
+ a = fluid.layers.elementwise_mul(
+ a_sum, cluster_weights2,
+ axis=1) # output shape [-1, self.cluster_size, feature_size]
+
+ # transpose activation from [-1, self.groups, self.cluster_size] to [-1, self.cluster_size, self.groups]
+ activation2 = fluid.layers.transpose(activation, perm=[0, 2, 1])
+ # transpose op will clear the lod infomation, so it should be reset
+ activation = fluid.layers.lod_reset(activation2, activation)
+
+ # reshape input from [-1, self.expansion * self.feature_size] to [-1, self.groups, feature_size]
+ reshaped_input = fluid.layers.reshape(input,
+ [-1, self.groups, feature_size])
+ # mat multiply activation and reshaped_input
+ vlad = fluid.layers.matmul(
+ activation,
+ reshaped_input) # output shape [-1, self.cluster_size, feature_size]
+ vlad = fluid.layers.sequence_pool(vlad, 'sum')
+ vlad = fluid.layers.elementwise_sub(vlad, a)
+
+ # l2_normalization
+ vlad = fluid.layers.transpose(vlad, [0, 2, 1])
+ vlad = fluid.layers.l2_normalize(vlad, axis=1)
+
+ # reshape and batch norm
+ vlad = fluid.layers.reshape(vlad,
+ [-1, self.cluster_size * feature_size])
+ vlad = fluid.layers.batch_norm(vlad, is_test=(not self.is_training))
+
+ return vlad
+
+
+class NeXtVLADModel(object):
+ """
+ Creates a NeXtVLAD based model.
+ Args:
+ model_input: A LoDTensor of [-1, N] for the input video frames.
+ vocab_size: The number of classes in the dataset.
+ """
+
+ def __init__(self):
+ pass
+
+ def create_model(self,
+ video_input,
+ audio_input,
+ is_training=True,
+ class_dim=None,
+ cluster_size=None,
+ hidden_size=None,
+ groups=None,
+ expansion=None,
+ drop_rate=None,
+ gating_reduction=None,
+ l2_penalty=None,
+ **unused_params):
+
+ # calcluate vlad of video and audio
+ video_nextvlad = NeXtVLAD(
+ 1024,
+ cluster_size,
+ is_training,
+ expansion=expansion,
+ groups=groups,
+ inputname='video')
+ audio_nextvlad = NeXtVLAD(
+ 128,
+ cluster_size,
+ is_training,
+ expansion=expansion,
+ groups=groups,
+ inputname='audio')
+ vlad_video = video_nextvlad.forward(video_input)
+ vlad_audio = audio_nextvlad.forward(audio_input)
+
+ # concat video and audio
+ vlad = fluid.layers.concat([vlad_video, vlad_audio], axis=1)
+
+ # drop out
+ if drop_rate > 0.:
+ vlad = fluid.layers.dropout(
+ vlad, drop_rate, is_test=(not is_training))
+
+ # add fc
+ activation = fluid.layers.fc(
+ input=vlad,
+ size=hidden_size,
+ act=None,
+ name='hidden1_fc',
+ param_attr=fluid.ParamAttr(
+ name='hidden1_fc_weights',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=False)
+ activation = fluid.layers.batch_norm(
+ activation, is_test=(not is_training))
+
+ # add fc, gate 1
+ gates = fluid.layers.fc(
+ input=activation,
+ size=hidden_size // gating_reduction,
+ act=None,
+ name='gating_fc1',
+ param_attr=fluid.ParamAttr(
+ name='gating_fc1_weights',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=False)
+ gates = fluid.layers.batch_norm(
+ gates, is_test=(not is_training), act='relu')
+
+ # add fc, gate 2
+ gates = fluid.layers.fc(
+ input=gates,
+ size=hidden_size,
+ act='sigmoid',
+ name='gating_fc2',
+ param_attr=fluid.ParamAttr(
+ name='gating_fc2_weights',
+ initializer=fluid.initializer.MSRA(uniform=False)),
+ bias_attr=False)
+
+ activation = fluid.layers.elementwise_mul(activation, gates)
+ aggregate_model = clf_model.LogisticModel # set classification model
+
+ return aggregate_model().create_model(
+ model_input=activation,
+ vocab_size=class_dim,
+ l2_penalty=l2_penalty,
+ is_training=is_training,
+ **unused_params)
diff --git a/fluid/PaddleCV/video/models/stnet/README.md b/fluid/PaddleCV/video/models/stnet/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..2b849ae7ce68f218309b028204269940d702043d
--- /dev/null
+++ b/fluid/PaddleCV/video/models/stnet/README.md
@@ -0,0 +1,109 @@
+# StNet 视频分类模型
+
+---
+## 目录
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+- [模型评估](#模型评估)
+- [模型推断](#模型推断)
+- [参考论文](#参考论文)
+
+
+## 模型简介
+
+StNet模型框架为ActivityNet Kinetics Challenge 2018中夺冠的基础网络框架，本次开源的是基于ResNet50实现的StNet模型，基于其他backbone网络的框架用户可以依样配置。该模型提出“super-image"的概念，在super-image上进行2D卷积，建模视频中局部时空相关性。另外通过temporal modeling block建模视频的全局时空依赖，最后用一个temporal Xception block对抽取的特征序列进行长时序建模。StNet主体网络结构如下图所示：
+
+
+
+StNet Framework Overview
+
+
+详细内容请参考AAAI'2019年论文[StNet:Local and Global Spatial-Temporal Modeling for Human Action Recognition](https://arxiv.org/abs/1811.01549)
+
+## 数据准备
+
+StNet的训练数据采用由DeepMind公布的Kinetics-400动作识别数据集。数据下载及准备请参考[数据说明](../../dataset/README.md)
+
+## 模型训练
+
+数据准备完毕后，可以通过如下两种方式启动训练：
+
+ python train.py --model-name=STNET
+ --config=./configs/stnet.txt
+ --save-dir=checkpoints
+ --log-interval=10
+ --valid-interval=1
+
+ bash scripts/train/train_stnet.sh
+
+- 可下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/stnet_kinetics.tar.gz)通过`--resume`指定权重存放路径进行finetune等开发
+
+**数据读取器说明：** 模型读取Kinetics-400数据集中的`mp4`数据，每条数据抽取`seg_num`段，每段抽取`seg_len`帧图像，对每帧图像做随机增强后，缩放至`target_size`。
+
+**训练策略：**
+
+* 采用Momentum优化算法训练，momentum=0.9
+* 权重衰减系数为1e-4
+* 学习率在训练的总epoch数的1/3和2/3时分别做0.1的衰减
+
+**备注：**
+
+* 在训练StNet模型时使用PaddlePaddle Fluid 1.3 + cudnn5.1。使用cudnn7.0以上版本时batchnorm计算moving mean和moving average会出现异常，此问题还在修复中。建议用户安装PaddlePaddle时指定cudnn版本,
+
+ pip install paddlepaddle\_gpu==1.3.0.post85
+
+或者在PaddlePaddle的whl包[下载页面](http://paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/Tables.html/#permalink-4--whl-release)选择下载cuda8.0\_cudnn5\_avx\_mkl对应的whl包安装。
+关于安装PaddlePaddle的详细操作请参考[安装文档](http://www.paddlepaddle.org/documentation/docs/zh/1.3/beginners_guide/install/index_cn.html)。
+
+
+## 模型评估
+
+可通过如下两种方式进行模型评估:
+
+ python test.py --model-name=STNET
+ --config=configs/stnet.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+
+ bash scripts/test/test__stnet.sh
+
+- 使用`scripts/test/test_stnet.sh`进行评估时，需要修改脚本中的`--weights`参数指定需要评估的权重。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/stnet_kinetics.tar.gz)进行评估
+
+当取如下参数时:
+
+| 参数 | 取值 |
+| :---------: | :----: |
+| seg\_num | 25 |
+| seglen | 5 |
+| target\_size | 256 |
+
+在Kinetics400的validation数据集下评估精度如下:
+
+| 精度指标 | 模型精度 |
+| :---------: | :----: |
+| TOP\_1 | 0.69 |
+
+
+## 模型推断
+
+可通过如下命令进行模型推断：
+
+ python infer.py --model-name=stnet
+ --config=configs/stnet.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+ --filelist=$FILELIST
+
+- 模型推断结果存储于`STNET_infer_result`中，通过`pickle`格式存储。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/stnet_kinetics.tar.gz)进行推断
+
+
+## 参考论文
+
+- [StNet:Local and Global Spatial-Temporal Modeling for Human Action Recognition](https://arxiv.org/abs/1811.01549), Dongliang He, Zhichao Zhou, Chuang Gan, Fu Li, Xiao Liu, Yandong Li, Limin Wang, Shilei Wen
+
diff --git a/fluid/PaddleCV/video/models/stnet/__init__.py b/fluid/PaddleCV/video/models/stnet/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..db952550a12b34853556fa42bba04c823bc7cbe4
--- /dev/null
+++ b/fluid/PaddleCV/video/models/stnet/__init__.py
@@ -0,0 +1 @@
+from .stnet import *
diff --git a/fluid/PaddleCV/video/models/stnet/stnet.py b/fluid/PaddleCV/video/models/stnet/stnet.py
new file mode 100644
index 0000000000000000000000000000000000000000..c408aa0894265b2e7b8eccb8b49cae4c799018cc
--- /dev/null
+++ b/fluid/PaddleCV/video/models/stnet/stnet.py
@@ -0,0 +1,150 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+
+from ..model import ModelBase
+from .stnet_res_model import StNet_ResNet
+
+import logging
+logger = logging.getLogger(__name__)
+
+__all__ = ["STNET"]
+
+
+class STNET(ModelBase):
+ def __init__(self, name, cfg, mode='train'):
+ super(STNET, self).__init__(name, cfg, mode=mode)
+ self.get_config()
+
+ def get_config(self):
+ self.num_classes = self.get_config_from_sec('model', 'num_classes')
+ self.seg_num = self.get_config_from_sec('model', 'seg_num')
+ self.seglen = self.get_config_from_sec('model', 'seglen')
+ self.image_mean = self.get_config_from_sec('model', 'image_mean')
+ self.image_std = self.get_config_from_sec('model', 'image_std')
+ self.num_layers = self.get_config_from_sec('model', 'num_layers')
+
+ self.num_epochs = self.get_config_from_sec('train', 'epoch')
+ self.total_videos = self.get_config_from_sec('train', 'total_videos')
+ self.base_learning_rate = self.get_config_from_sec('train',
+ 'learning_rate')
+ self.learning_rate_decay = self.get_config_from_sec(
+ 'train', 'learning_rate_decay')
+ self.l2_weight_decay = self.get_config_from_sec('train',
+ 'l2_weight_decay')
+ self.momentum = self.get_config_from_sec('train', 'momentum')
+
+ self.target_size = self.get_config_from_sec(self.mode, 'target_size')
+ self.batch_size = self.get_config_from_sec(self.mode, 'batch_size')
+
+ def build_input(self, use_pyreader=True):
+ image_shape = [3, self.target_size, self.target_size]
+ image_shape[0] = image_shape[0] * self.seglen
+ image_shape = [self.seg_num] + image_shape
+ self.use_pyreader = use_pyreader
+ if use_pyreader:
+ assert self.mode != 'infer', \
+ 'pyreader is not recommendated when infer, please set use_pyreader to be false.'
+ py_reader = fluid.layers.py_reader(
+ capacity=100,
+ shapes=[[-1] + image_shape, [-1] + [1]],
+ dtypes=['float32', 'int64'],
+ name='train_py_reader'
+ if self.is_training else 'test_py_reader',
+ use_double_buffer=True)
+ image, label = fluid.layers.read_file(py_reader)
+ self.py_reader = py_reader
+ else:
+ image = fluid.layers.data(
+ name='image', shape=image_shape, dtype='float32')
+ if self.mode != 'infer':
+ label = fluid.layers.data(
+ name='label', shape=[1], dtype='int64')
+ else:
+ label = None
+ self.feature_input = [image]
+ self.label_input = label
+
+ def create_model_args(self):
+ cfg = {}
+ cfg['layers'] = self.num_layers
+ cfg['class_dim'] = self.num_classes
+ cfg['seg_num'] = self.seg_num
+ cfg['seglen'] = self.seglen
+ return cfg
+
+ def build_model(self):
+ cfg = self.create_model_args()
+ videomodel = StNet_ResNet(layers = cfg['layers'], seg_num = cfg['seg_num'], \
+ seglen = cfg['seglen'], is_training = (self.mode == 'train'))
+ out = videomodel.net(input=self.feature_input[0],
+ class_dim=cfg['class_dim'])
+ self.network_outputs = [out]
+
+ def optimizer(self):
+ epoch_points = [self.num_epochs / 3, self.num_epochs * 2 / 3]
+ total_videos = self.total_videos
+ step = int(total_videos / self.batch_size + 1)
+ bd = [e * step for e in epoch_points]
+ base_lr = self.base_learning_rate
+ lr_decay = self.learning_rate_decay
+ lr = [base_lr, base_lr * lr_decay, base_lr * lr_decay * lr_decay]
+ l2_weight_decay = self.l2_weight_decay
+ momentum = self.momentum
+ optimizer = fluid.optimizer.Momentum(
+ learning_rate=fluid.layers.piecewise_decay(
+ boundaries=bd, values=lr),
+ momentum=momentum,
+ regularization=fluid.regularizer.L2Decay(l2_weight_decay))
+
+ return optimizer
+
+ def loss(self):
+ cost = fluid.layers.cross_entropy(input=self.network_outputs[0], \
+ label=self.label_input, ignore_index=-1)
+ self.loss_ = fluid.layers.mean(x=cost)
+ return self.loss_
+
+ def outputs(self):
+ return self.network_outputs
+
+ def feeds(self):
+ return self.feature_input if self.mode == 'infer' else self.feature_input + [
+ self.label_input
+ ]
+
+ def pretrain_info(self):
+ return ('ResNet50_pretrained', 'https://paddlemodels.bj.bcebos.com/video_classification/ResNet50_pretrained.tar.gz')
+
+ def weights_info(self):
+ return ('stnet_kinetics',
+ 'https://paddlemodels.bj.bcebos.com/video_classification/stnet_kinetics.tar.gz')
+
+ def load_pretrain_params(self, exe, pretrain, prog, place):
+ def is_parameter(var):
+ if isinstance(var, fluid.framework.Parameter):
+ return isinstance(var, fluid.framework.Parameter) and (not ("fc_0" in var.name)) \
+ and (not ("batch_norm" in var.name)) and (not ("xception" in var.name)) and (not ("conv3d" in var.name))
+
+ logger.info("Load pretrain weights from {}, exclude fc, batch_norm, xception, conv3d layers.".format(pretrain))
+ vars = filter(is_parameter, prog.list_vars())
+ fluid.io.load_vars(exe, pretrain, vars=vars, main_program=prog)
+
+ param_tensor = fluid.global_scope().find_var(
+ "conv1_weights").get_tensor()
+ param_numpy = np.array(param_tensor)
+ param_numpy = np.mean(param_numpy, axis=1, keepdims=True) / self.seglen
+ param_numpy = np.repeat(param_numpy, 3 * self.seglen, axis=1)
+ param_tensor.set(param_numpy.astype(np.float32), place)
diff --git a/fluid/PaddleCV/video/models/stnet/stnet_res_model.py b/fluid/PaddleCV/video/models/stnet/stnet_res_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..71a22c4f869161a8b92ec5c79a69b85bc68d4c86
--- /dev/null
+++ b/fluid/PaddleCV/video/models/stnet/stnet_res_model.py
@@ -0,0 +1,312 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import time
+import sys
+import paddle.fluid as fluid
+import math
+
+
+class StNet_ResNet():
+ def __init__(self, layers=50, seg_num=7, seglen=5, is_training=True):
+ self.layers = layers
+ self.seglen = seglen
+ self.seg_num = seg_num
+ self.is_training = is_training
+
+ def temporal_conv_bn(
+ self,
+ input, #(B*seg_num, c, h, w)
+ num_filters,
+ filter_size=(3, 1, 1),
+ padding=(1, 0, 0)):
+ #(B, seg_num, c, h, w)
+ in_reshape = fluid.layers.reshape(
+ x=input,
+ shape=[
+ -1, self.seg_num, input.shape[-3], input.shape[-2],
+ input.shape[-1]
+ ])
+ in_transpose = fluid.layers.transpose(in_reshape, perm=[0, 2, 1, 3, 4])
+
+ conv = fluid.layers.conv3d(
+ input=in_transpose,
+ num_filters=num_filters,
+ filter_size=filter_size,
+ stride=1,
+ groups=1,
+ padding=padding,
+ act='relu',
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0.0)))
+
+ out = fluid.layers.batch_norm(
+ input=conv,
+ act=None,
+ is_test=(not self.is_training),
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=1.0)),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0.0)))
+ out = out + in_transpose
+ out = fluid.layers.transpose(out, perm=[0, 2, 1, 3, 4])
+ out = fluid.layers.reshape(x=out, shape=input.shape)
+ return out
+
+ def xception(self, input): #(B, C, seg_num,1)
+ bn = fluid.layers.batch_norm(
+ input=input,
+ act=None,
+ name="xception_bn",
+ is_test=(not self.is_training),
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=1.0)),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0.0)))
+ att_conv = fluid.layers.conv2d(
+ input=bn,
+ num_filters=2048,
+ filter_size=[3, 1],
+ stride=[1, 1],
+ padding=[1, 0],
+ groups=2048,
+ name="xception_att_conv",
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0)))
+ att_2 = fluid.layers.conv2d(
+ input=att_conv,
+ num_filters=1024,
+ filter_size=[1, 1],
+ stride=[1, 1],
+ name="xception_att_2",
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0)))
+ bndw = fluid.layers.batch_norm(
+ input=att_2,
+ act="relu",
+ name="xception_bndw",
+ is_test=(not self.is_training),
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=1.0)),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0.0)))
+ att1 = fluid.layers.conv2d(
+ input=bndw,
+ num_filters=1024,
+ filter_size=[3, 1],
+ stride=[1, 1],
+ padding=[1, 0],
+ groups=1024,
+ name="xception_att1",
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0)))
+ att1_2 = fluid.layers.conv2d(
+ input=att1,
+ num_filters=1024,
+ filter_size=[1, 1],
+ stride=[1, 1],
+ name="xception_att1_2",
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0)))
+ dw = fluid.layers.conv2d(
+ input=bn,
+ num_filters=1024,
+ filter_size=[1, 1],
+ stride=[1, 1],
+ name="xception_dw",
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.MSRAInitializer()),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0)))
+ add_to = dw + att1_2
+ bn2 = fluid.layers.batch_norm(
+ input=add_to,
+ act=None,
+ name='xception_bn2',
+ is_test=(not self.is_training),
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=1.0)),
+ bias_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.ConstantInitializer(value=0.0)))
+ return fluid.layers.relu(bn2)
+
+ def conv_bn_layer(self,
+ input,
+ num_filters,
+ filter_size,
+ stride=1,
+ groups=1,
+ act=None,
+ name=None):
+ conv = fluid.layers.conv2d(
+ input=input,
+ num_filters=num_filters,
+ filter_size=filter_size,
+ stride=stride,
+ padding=(filter_size - 1) // 2,
+ groups=groups,
+ act=None,
+ param_attr=fluid.param_attr.ParamAttr(name=name + "_weights"),
+ bias_attr=False,
+ #name = name+".conv2d.output.1"
+ )
+ if name == "conv1":
+ bn_name = "bn_" + name
+ else:
+ bn_name = "bn" + name[3:]
+ return fluid.layers.batch_norm(
+ input=conv,
+ act=act,
+ is_test=(not self.is_training),
+ #name=bn_name+'.output.1',
+ param_attr=fluid.param_attr.ParamAttr(name=bn_name + "_scale"),
+ bias_attr=fluid.param_attr.ParamAttr(bn_name + '_offset'),
+ moving_mean_name=bn_name + "_mean",
+ moving_variance_name=bn_name + '_variance')
+
+ def shortcut(self, input, ch_out, stride, name):
+ ch_in = input.shape[1]
+ if ch_in != ch_out or stride != 1:
+ return self.conv_bn_layer(input, ch_out, 1, stride, name=name)
+ else:
+ return input
+
+ def bottleneck_block(self, input, num_filters, stride, name):
+ conv0 = self.conv_bn_layer(
+ input=input,
+ num_filters=num_filters,
+ filter_size=1,
+ act='relu',
+ name=name + "_branch2a")
+ conv1 = self.conv_bn_layer(
+ input=conv0,
+ num_filters=num_filters,
+ filter_size=3,
+ stride=stride,
+ act='relu',
+ name=name + "_branch2b")
+ conv2 = self.conv_bn_layer(
+ input=conv1,
+ num_filters=num_filters * 4,
+ filter_size=1,
+ act=None,
+ name=name + "_branch2c")
+
+ short = self.shortcut(
+ input, num_filters * 4, stride, name=name + "_branch1")
+
+ return fluid.layers.elementwise_add(
+ x=short,
+ y=conv2,
+ act='relu',
+ #name=".add.output.5"
+ )
+
+ def net(self, input, class_dim=101):
+ layers = self.layers
+ seg_num = self.seg_num
+ seglen = self.seglen
+
+ supported_layers = [50, 101, 152]
+ if layers not in supported_layers:
+ print("supported layers are", supported_layers, \
+ "but input layer is ", layers)
+ exit()
+
+ # reshape input
+ # [B, seg_num, seglen*c, H, W] --> [B*seg_num, seglen*c, H, W]
+ channels = input.shape[2]
+ short_size = input.shape[3]
+ input = fluid.layers.reshape(
+ x=input, shape=[-1, channels, short_size, short_size])
+
+ if layers == 50:
+ depth = [3, 4, 6, 3]
+ elif layers == 101:
+ depth = [3, 4, 23, 3]
+ elif layers == 152:
+ depth = [3, 8, 36, 3]
+ num_filters = [64, 128, 256, 512]
+
+ conv = self.conv_bn_layer(
+ input=input,
+ num_filters=64,
+ filter_size=7,
+ stride=2,
+ act='relu',
+ name='conv1')
+ conv = fluid.layers.pool2d(
+ input=conv,
+ pool_size=3,
+ pool_stride=2,
+ pool_padding=1,
+ pool_type='max')
+
+ for block in range(len(depth)):
+ for i in range(depth[block]):
+ if layers in [101, 152] and block == 2:
+ if i == 0:
+ conv_name = "res" + str(block + 2) + "a"
+ else:
+ conv_name = "res" + str(block + 2) + "b" + str(i)
+ else:
+ conv_name = "res" + str(block + 2) + chr(97 + i)
+
+ conv = self.bottleneck_block(
+ input=conv,
+ num_filters=num_filters[block],
+ stride=2 if i == 0 and block != 0 else 1,
+ name=conv_name)
+ if block == 1:
+ #insert the first temporal modeling block
+ conv = self.temporal_conv_bn(input=conv, num_filters=512)
+ if block == 2:
+ #insert the second temporal modeling block
+ conv = self.temporal_conv_bn(input=conv, num_filters=1024)
+
+ pool = fluid.layers.pool2d(
+ input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+
+ feature = fluid.layers.reshape(
+ x=pool, shape=[-1, seg_num, pool.shape[1], 1])
+ feature = fluid.layers.transpose(feature, perm=[0, 2, 1, 3])
+
+ #append the temporal Xception block
+ xfeat = self.xception(feature) #(B, 1024, seg_num, 1)
+ out = fluid.layers.pool2d(
+ input=xfeat,
+ pool_size=(seg_num, 1),
+ pool_type='max',
+ global_pooling=True)
+ out = fluid.layers.reshape(x=out, shape=[-1, 1024])
+
+ stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
+ out = fluid.layers.fc(input=out,
+ size=class_dim,
+ act='softmax',
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.Uniform(-stdv,
+ stdv)))
+ return out
diff --git a/fluid/PaddleCV/video/models/tsn/README.md b/fluid/PaddleCV/video/models/tsn/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..6b030d9b7008fa320347a75cb2d897df48c50a83
--- /dev/null
+++ b/fluid/PaddleCV/video/models/tsn/README.md
@@ -0,0 +1,85 @@
+# TSN 视频分类模型
+
+---
+## 内容
+
+- [模型简介](#模型简介)
+- [数据准备](#数据准备)
+- [模型训练](#模型训练)
+- [模型评估](#模型评估)
+- [模型推断](#模型推断)
+- [参考论文](#参考论文)
+
+
+## 模型简介
+
+Temporal Segment Network (TSN) 是视频分类领域经典的基于2D-CNN的解决方案。该方法主要解决视频的长时间行为判断问题，通过稀疏采样视频帧的方式代替稠密采样，既能捕获视频全局信息，也能去除冗余，降低计算量。最终将每帧特征平均融合后得到视频的整体特征，并用于分类。本代码实现的模型为基于单路RGB图像的TSN网络结构，Backbone采用ResNet-50结构。
+
+详细内容请参考ECCV 2016年论文[StNet:Local and Global Spatial-Temporal Modeling for Human Action Recognition](https://arxiv.org/abs/1608.00859)
+
+## 数据准备
+
+TSN的训练数据采用由DeepMind公布的Kinetics-400动作识别数据集。数据下载及准备请参考[数据说明](../../dataset/README.md)
+
+## 模型训练
+
+数据准备完毕后，可以通过如下两种方式启动训练：
+
+ python train.py --model-name=TSN
+ --config=./configs/tsn.txt
+ --save-dir=checkpoints
+ --log-interval=10
+ --valid-interval=1
+
+ bash scripts/train/train_tsn.sh
+
+- 可下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/tsn_kinetics.tar.gz)通过`--resume`指定权重存放路径进行finetune等开发
+
+**数据读取器说明：** 模型读取Kinetics-400数据集中的`mp4`数据，每条数据抽取`seg_num`段，每段抽取1帧图像，对每帧图像做随机增强后，缩放至`target_size`。
+
+**训练策略：**
+
+* 采用Momentum优化算法训练，momentum=0.9
+* 权重衰减系数为1e-4
+* 学习率在训练的总epoch数的1/3和2/3时分别做0.1的衰减
+
+## 模型评估
+
+可通过如下两种方式进行模型评估:
+
+ python test.py --model-name=TSN
+ --config=configs/tsn.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+
+ bash scripts/test/test_tsn.sh
+
+- 使用`scripts/test/test_tsn.sh`进行评估时，需要修改脚本中的`--weights`参数指定需要评估的权重。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/tsn_kinetics.tar.gz)进行评估
+
+当取如下参数时，在Kinetics400的validation数据集下评估精度如下:
+
+| seg\_num | target\_size | Top-1 |
+| :------: | :----------: | :----: |
+| 3 | 224 | 0.66 |
+| 7 | 224 | 0.67 |
+
+## 模型推断
+
+可通过如下命令进行模型推断：
+
+ python infer.py --model-name=TSN
+ --config=configs/tsn.txt
+ --log-interval=1
+ --weights=$PATH_TO_WEIGHTS
+ --filelist=$FILELIST
+
+- 模型推断结果存储于`TSN_infer_result`中，通过`pickle`格式存储。
+
+- 若未指定`--weights`参数，脚本会下载已发布模型[model](https://paddlemodels.bj.bcebos.com/video_classification/tsn_kinetics.tar.gz)进行推断
+
+## 参考论文
+
+- [Temporal Segment Networks: Towards Good Practices for Deep Action Recognition](https://arxiv.org/abs/1608.00859), Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaoou Tang, Luc Van Gool
+
diff --git a/fluid/PaddleCV/video/models/tsn/__init__.py b/fluid/PaddleCV/video/models/tsn/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..bd57d2687bc948e63dd88306e9d435bbbb5a7978
--- /dev/null
+++ b/fluid/PaddleCV/video/models/tsn/__init__.py
@@ -0,0 +1 @@
+from .tsn import *
diff --git a/fluid/PaddleCV/video/models/tsn/tsn.py b/fluid/PaddleCV/video/models/tsn/tsn.py
new file mode 100644
index 0000000000000000000000000000000000000000..82fdb3279376a15e796df09818c343db24f048b4
--- /dev/null
+++ b/fluid/PaddleCV/video/models/tsn/tsn.py
@@ -0,0 +1,146 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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 paddle.fluid as fluid
+from paddle.fluid import ParamAttr
+
+from ..model import ModelBase
+from .tsn_res_model import TSN_ResNet
+
+import logging
+logger = logging.getLogger(__name__)
+
+__all__ = ["TSN"]
+
+
+class TSN(ModelBase):
+ def __init__(self, name, cfg, mode='train'):
+ super(TSN, self).__init__(name, cfg, mode=mode)
+ self.get_config()
+
+ def get_config(self):
+ self.num_classes = self.get_config_from_sec('model', 'num_classes')
+ self.seg_num = self.get_config_from_sec('model', 'seg_num')
+ self.seglen = self.get_config_from_sec('model', 'seglen')
+ self.image_mean = self.get_config_from_sec('model', 'image_mean')
+ self.image_std = self.get_config_from_sec('model', 'image_std')
+ self.num_layers = self.get_config_from_sec('model', 'num_layers')
+
+ self.num_epochs = self.get_config_from_sec('train', 'epoch')
+ self.total_videos = self.get_config_from_sec('train', 'total_videos')
+ self.base_learning_rate = self.get_config_from_sec('train',
+ 'learning_rate')
+ self.learning_rate_decay = self.get_config_from_sec(
+ 'train', 'learning_rate_decay')
+ self.l2_weight_decay = self.get_config_from_sec('train',
+ 'l2_weight_decay')
+ self.momentum = self.get_config_from_sec('train', 'momentum')
+
+ self.target_size = self.get_config_from_sec(self.mode, 'target_size')
+ self.batch_size = self.get_config_from_sec(self.mode, 'batch_size')
+
+ def build_input(self, use_pyreader=True):
+ image_shape = [3, self.target_size, self.target_size]
+ image_shape[0] = image_shape[0] * self.seglen
+ image_shape = [self.seg_num] + image_shape
+ self.use_pyreader = use_pyreader
+ if use_pyreader:
+ assert self.mode != 'infer', \
+ 'pyreader is not recommendated when infer, please set use_pyreader to be false.'
+ py_reader = fluid.layers.py_reader(
+ capacity=100,
+ shapes=[[-1] + image_shape, [-1] + [1]],
+ dtypes=['float32', 'int64'],
+ name='train_py_reader'
+ if self.is_training else 'test_py_reader',
+ use_double_buffer=True)
+ image, label = fluid.layers.read_file(py_reader)
+ self.py_reader = py_reader
+ else:
+ image = fluid.layers.data(
+ name='image', shape=image_shape, dtype='float32')
+ if self.mode != 'infer':
+ label = fluid.layers.data(
+ name='label', shape=[1], dtype='int64')
+ else:
+ label = None
+ self.feature_input = [image]
+ self.label_input = label
+
+ def create_model_args(self):
+ cfg = {}
+ cfg['layers'] = self.num_layers
+ cfg['class_dim'] = self.num_classes
+ cfg['seg_num'] = self.seg_num
+ return cfg
+
+ def build_model(self):
+ cfg = self.create_model_args()
+ videomodel = TSN_ResNet(
+ layers=cfg['layers'],
+ seg_num=cfg['seg_num'],
+ is_training=(self.mode == 'train'))
+ out = videomodel.net(input=self.feature_input[0],
+ class_dim=cfg['class_dim'])
+ self.network_outputs = [out]
+
+ def optimizer(self):
+ assert self.mode == 'train', "optimizer only can be get in train mode"
+ epoch_points = [self.num_epochs / 3, self.num_epochs * 2 / 3]
+ total_videos = self.total_videos
+ step = int(total_videos / self.batch_size + 1)
+ bd = [e * step for e in epoch_points]
+ base_lr = self.base_learning_rate
+ lr_decay = self.learning_rate_decay
+ lr = [base_lr, base_lr * lr_decay, base_lr * lr_decay * lr_decay]
+ l2_weight_decay = self.l2_weight_decay
+ momentum = self.momentum
+ optimizer = fluid.optimizer.Momentum(
+ learning_rate=fluid.layers.piecewise_decay(
+ boundaries=bd, values=lr),
+ momentum=momentum,
+ regularization=fluid.regularizer.L2Decay(l2_weight_decay))
+
+ return optimizer
+
+ def loss(self):
+ assert self.mode != 'infer', "invalid loss calculationg in infer mode"
+ cost = fluid.layers.cross_entropy(input=self.network_outputs[0], \
+ label=self.label_input, ignore_index=-1)
+ self.loss_ = fluid.layers.mean(x=cost)
+ return self.loss_
+
+ def outputs(self):
+ return self.network_outputs
+
+ def feeds(self):
+ return self.feature_input if self.mode == 'infer' else self.feature_input + [
+ self.label_input
+ ]
+
+ def pretrain_info(self):
+ return ('ResNet50_pretrained', 'https://paddlemodels.bj.bcebos.com/video_classification/ResNet50_pretrained.tar.gz')
+
+ def weights_info(self):
+ return ('tsn_kinetics',
+ 'https://paddlemodels.bj.bcebos.com/video_classification/tsn_kinetics.tar.gz')
+
+ def load_pretrain_params(self, exe, pretrain, prog, place):
+ def is_parameter(var):
+ return isinstance(var, fluid.framework.Parameter) and (not ("fc_0" in var.name))
+
+ logger.info("Load pretrain weights from {}, exclude fc layer.".format(pretrain))
+ vars = filter(is_parameter, prog.list_vars())
+ fluid.io.load_vars(exe, pretrain, vars=vars, main_program=prog)
+
diff --git a/fluid/PaddleCV/video/models/tsn/tsn_res_model.py b/fluid/PaddleCV/video/models/tsn/tsn_res_model.py
new file mode 100644
index 0000000000000000000000000000000000000000..09dc54893f3305a0a1a94fe6e73aff32680915d9
--- /dev/null
+++ b/fluid/PaddleCV/video/models/tsn/tsn_res_model.py
@@ -0,0 +1,158 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import time
+import sys
+import paddle.fluid as fluid
+import math
+
+
+class TSN_ResNet():
+ def __init__(self, layers=50, seg_num=7, is_training=True):
+ self.layers = layers
+ self.seg_num = seg_num
+ self.is_training = is_training
+
+ def conv_bn_layer(self,
+ input,
+ num_filters,
+ filter_size,
+ stride=1,
+ groups=1,
+ act=None,
+ name=None):
+ conv = fluid.layers.conv2d(
+ input=input,
+ num_filters=num_filters,
+ filter_size=filter_size,
+ stride=stride,
+ padding=(filter_size - 1) // 2,
+ groups=groups,
+ act=None,
+ param_attr=fluid.param_attr.ParamAttr(name=name + "_weights"),
+ bias_attr=False)
+ if name == "conv1":
+ bn_name = "bn_" + name
+ else:
+ bn_name = "bn" + name[3:]
+
+ return fluid.layers.batch_norm(
+ input=conv,
+ act=act,
+ is_test=(not self.is_training),
+ param_attr=fluid.param_attr.ParamAttr(name=bn_name + "_scale"),
+ bias_attr=fluid.param_attr.ParamAttr(bn_name + '_offset'),
+ moving_mean_name=bn_name + "_mean",
+ moving_variance_name=bn_name + '_variance')
+
+ def shortcut(self, input, ch_out, stride, name):
+ ch_in = input.shape[1]
+ if ch_in != ch_out or stride != 1:
+ return self.conv_bn_layer(input, ch_out, 1, stride, name=name)
+ else:
+ return input
+
+ def bottleneck_block(self, input, num_filters, stride, name):
+ conv0 = self.conv_bn_layer(
+ input=input,
+ num_filters=num_filters,
+ filter_size=1,
+ act='relu',
+ name=name + "_branch2a")
+ conv1 = self.conv_bn_layer(
+ input=conv0,
+ num_filters=num_filters,
+ filter_size=3,
+ stride=stride,
+ act='relu',
+ name=name + "_branch2b")
+ conv2 = self.conv_bn_layer(
+ input=conv1,
+ num_filters=num_filters * 4,
+ filter_size=1,
+ act=None,
+ name=name + "_branch2c")
+
+ short = self.shortcut(
+ input, num_filters * 4, stride, name=name + "_branch1")
+
+ return fluid.layers.elementwise_add(x=short, y=conv2, act='relu')
+
+ def net(self, input, class_dim=101):
+ layers = self.layers
+ seg_num = self.seg_num
+ supported_layers = [50, 101, 152]
+ assert layers in supported_layers, \
+ "supported layers are {} but input layer is {}".format(supported_layers, layers)
+
+ # reshape input
+ channels = input.shape[2]
+ short_size = input.shape[3]
+ input = fluid.layers.reshape(
+ x=input, shape=[-1, channels, short_size, short_size])
+
+ if layers == 50:
+ depth = [3, 4, 6, 3]
+ elif layers == 101:
+ depth = [3, 4, 23, 3]
+ elif layers == 152:
+ depth = [3, 8, 36, 3]
+ num_filters = [64, 128, 256, 512]
+
+ conv = self.conv_bn_layer(
+ input=input,
+ num_filters=64,
+ filter_size=7,
+ stride=2,
+ act='relu',
+ name='conv1')
+ conv = fluid.layers.pool2d(
+ input=conv,
+ pool_size=3,
+ pool_stride=2,
+ pool_padding=1,
+ pool_type='max')
+
+ for block in range(len(depth)):
+ for i in range(depth[block]):
+ if layers in [101, 152] and block == 2:
+ if i == 0:
+ conv_name = "res" + str(block + 2) + "a"
+ else:
+ conv_name = "res" + str(block + 2) + "b" + str(i)
+ else:
+ conv_name = "res" + str(block + 2) + chr(97 + i)
+
+ conv = self.bottleneck_block(
+ input=conv,
+ num_filters=num_filters[block],
+ stride=2 if i == 0 and block != 0 else 1,
+ name=conv_name)
+
+ pool = fluid.layers.pool2d(
+ input=conv, pool_size=7, pool_type='avg', global_pooling=True)
+
+ feature = fluid.layers.reshape(
+ x=pool, shape=[-1, seg_num, pool.shape[1]])
+ out = fluid.layers.reduce_mean(feature, dim=1)
+
+ stdv = 1.0 / math.sqrt(pool.shape[1] * 1.0)
+ out = fluid.layers.fc(input=out,
+ size=class_dim,
+ act='softmax',
+ param_attr=fluid.param_attr.ParamAttr(
+ initializer=fluid.initializer.Uniform(-stdv,
+ stdv)))
+ return out
diff --git a/fluid/PaddleCV/video/models/utils.py b/fluid/PaddleCV/video/models/utils.py
new file mode 100755
index 0000000000000000000000000000000000000000..b02abfdf134c869fe4805f4a746d7357efd0b7b1
--- /dev/null
+++ b/fluid/PaddleCV/video/models/utils.py
@@ -0,0 +1,47 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import wget
+import tarfile
+
+__all__ = ['decompress', 'download', 'AttrDict']
+
+
+def decompress(path):
+ t = tarfile.open(path)
+ t.extractall(path='/'.join(path.split('/')[:-1]))
+ t.close()
+ os.remove(path)
+
+
+def download(url, path):
+ weight_dir = '/'.join(path.split('/')[:-1])
+ if not os.path.exists(weight_dir):
+ os.makedirs(weight_dir)
+
+ path = path + ".tar.gz"
+ wget.download(url, path)
+ decompress(path)
+
+
+class AttrDict(dict):
+ def __getattr__(self, key):
+ return self[key]
+
+ def __setattr__(self, key, value):
+ if key in self.__dict__:
+ self.__dict__[key] = value
+ else:
+ self[key] = value
diff --git a/fluid/PaddleCV/video/scripts/infer/infer_attention_cluster.sh b/fluid/PaddleCV/video/scripts/infer/infer_attention_cluster.sh
new file mode 100644
index 0000000000000000000000000000000000000000..be6045db83d29426d363bacc524434f60d45ea57
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/infer/infer_attention_cluster.sh
@@ -0,0 +1,4 @@
+python infer.py --model-name="AttentionCluster" --config=./configs/attention_cluster.txt \
+ --filelist=./data/youtube8m/infer.list \
+ --weights=./checkpoints/AttentionCluster_epoch0 \
+ --save-dir="./save"
diff --git a/fluid/PaddleCV/video/scripts/infer/infer_attention_lstm.sh b/fluid/PaddleCV/video/scripts/infer/infer_attention_lstm.sh
new file mode 100644
index 0000000000000000000000000000000000000000..019bb346617d844e33d4dc883d1a6f96a7a91f3a
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/infer/infer_attention_lstm.sh
@@ -0,0 +1,4 @@
+python infer.py --model-name="AttentionLSTM" --config=./configs/attention_lstm.txt \
+ --filelist=./data/youtube8m/infer.list \
+ --weights=./checkpoints/AttentionLSTM_epoch0 \
+ --save-dir="./save"
diff --git a/fluid/PaddleCV/video/scripts/infer/infer_nextvlad.sh b/fluid/PaddleCV/video/scripts/infer/infer_nextvlad.sh
new file mode 100644
index 0000000000000000000000000000000000000000..1a96980106bbda1fc3678785323904a4ccecaa65
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/infer/infer_nextvlad.sh
@@ -0,0 +1,3 @@
+python infer.py --model-name="NEXTVLAD" --config=./configs/nextvlad.txt --filelist=./data/youtube8m/infer.list \
+ --weights=./checkpoints/NEXTVLAD_epoch0 \
+ --save-dir="./save"
diff --git a/fluid/PaddleCV/video/scripts/infer/infer_stnet.sh b/fluid/PaddleCV/video/scripts/infer/infer_stnet.sh
new file mode 100644
index 0000000000000000000000000000000000000000..8b27a234d9a650b3e4acf8d9dae5ba1bb68fc71b
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/infer/infer_stnet.sh
@@ -0,0 +1,2 @@
+python infer.py --model-name="STNET" --config=./configs/stnet.txt --filelist=./data/kinetics/infer.list \
+ --log-interval=10 --weights=./checkpoints/STNET_epoch0 --save-dir=./save
diff --git a/fluid/PaddleCV/video/scripts/infer/infer_tsn.sh b/fluid/PaddleCV/video/scripts/infer/infer_tsn.sh
new file mode 100644
index 0000000000000000000000000000000000000000..515feaf4a502bb35691d357a038f702345e9b9a2
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/infer/infer_tsn.sh
@@ -0,0 +1,2 @@
+python infer.py --model-name="TSN" --config=./configs/tsn.txt --filelist=./data/kinetics/infer.list \
+ --log-interval=10 --weights=./checkpoints/TSN_epoch0 --save-dir=./save
diff --git a/fluid/PaddleCV/video/scripts/test/test_attention_cluster.sh b/fluid/PaddleCV/video/scripts/test/test_attention_cluster.sh
new file mode 100644
index 0000000000000000000000000000000000000000..21df131934cb4306d185fd76374b4314767add68
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/test/test_attention_cluster.sh
@@ -0,0 +1,2 @@
+python test.py --model-name="AttentionCluster" --config=./configs/attention_cluster.txt \
+ --log-interval=5 --weights=./checkpoints/AttentionCluster_epoch0
diff --git a/fluid/PaddleCV/video/scripts/test/test_attention_lstm.sh b/fluid/PaddleCV/video/scripts/test/test_attention_lstm.sh
new file mode 100644
index 0000000000000000000000000000000000000000..d728dbd1c172a5d3c9e19dd7dce457136a90f3d5
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/test/test_attention_lstm.sh
@@ -0,0 +1,2 @@
+python test.py --model-name="AttentionLSTM" --config=./configs/attention_lstm.txt \
+ --log-interval=5 --weights=./checkpoints/AttentionLSTM_epoch0
diff --git a/fluid/PaddleCV/video/scripts/test/test_nextvlad.sh b/fluid/PaddleCV/video/scripts/test/test_nextvlad.sh
new file mode 100644
index 0000000000000000000000000000000000000000..239e9980153303a161511a217a09a4d63b216e3b
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/test/test_nextvlad.sh
@@ -0,0 +1,2 @@
+python test.py --model-name="NEXTVLAD" --config=./configs/nextvlad.txt \
+ --log-interval=10 --weights=./checkpoints/NEXTVLAD_epoch0
diff --git a/fluid/PaddleCV/video/scripts/test/test_stnet.sh b/fluid/PaddleCV/video/scripts/test/test_stnet.sh
new file mode 100644
index 0000000000000000000000000000000000000000..6913ea6970f5448d83b72c5fe8f3b9c05925d9a8
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/test/test_stnet.sh
@@ -0,0 +1,2 @@
+python test.py --model-name="STNET" --config=./configs/stnet.txt \
+ --log-interval=10 --weights=./checkpoints/STNET_epoch0
diff --git a/fluid/PaddleCV/video/scripts/test/test_tsn.sh b/fluid/PaddleCV/video/scripts/test/test_tsn.sh
new file mode 100644
index 0000000000000000000000000000000000000000..b66bcb2cf08fbcccef1954369dd53d6e61b0894a
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/test/test_tsn.sh
@@ -0,0 +1,2 @@
+python test.py --model-name="TSN" --config=./configs/tsn.txt \
+ --log-interval=10 --weights=./checkpoints/TSN_epoch0
diff --git a/fluid/PaddleCV/video/scripts/train/train_attention_cluster.sh b/fluid/PaddleCV/video/scripts/train/train_attention_cluster.sh
new file mode 100644
index 0000000000000000000000000000000000000000..0a0b0bbb33ede34f56e7bda9f0dbce007e197aed
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/train/train_attention_cluster.sh
@@ -0,0 +1,2 @@
+python train.py --model-name="AttentionCluster" --config=./configs/attention_cluster.txt --epoch-num=5 \
+ --valid-interval=1 --log-interval=10
diff --git a/fluid/PaddleCV/video/scripts/train/train_attention_lstm.sh b/fluid/PaddleCV/video/scripts/train/train_attention_lstm.sh
new file mode 100644
index 0000000000000000000000000000000000000000..bb855b19cf1122fea3cbee0171531e6003fb64a9
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/train/train_attention_lstm.sh
@@ -0,0 +1,2 @@
+python train.py --model-name="AttentionLSTM" --config=./configs/attention_lstm.txt --epoch-num=10 \
+ --valid-interval=1 --log-interval=10
diff --git a/fluid/PaddleCV/video/scripts/train/train_nextvlad.sh b/fluid/PaddleCV/video/scripts/train/train_nextvlad.sh
new file mode 100644
index 0000000000000000000000000000000000000000..b5857e9f35a47d89ce5185c6d42b2ed51207e390
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/train/train_nextvlad.sh
@@ -0,0 +1,3 @@
+export CUDA_VISIBLE_DEVICES=0,1,2,3
+python train.py --model-name="NEXTVLAD" --config=./configs/nextvlad.txt --epoch-num=6 \
+ --valid-interval=1 --log-interval=10
diff --git a/fluid/PaddleCV/video/scripts/train/train_stnet.sh b/fluid/PaddleCV/video/scripts/train/train_stnet.sh
new file mode 100644
index 0000000000000000000000000000000000000000..c595c10c025c517e9cdc4d70a9d316b853768fa9
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/train/train_stnet.sh
@@ -0,0 +1,2 @@
+python train.py --model-name="STNET" --config=./configs/stnet.txt --epoch-num=60 \
+ --valid-interval=1 --log-interval=10
diff --git a/fluid/PaddleCV/video/scripts/train/train_tsn.sh b/fluid/PaddleCV/video/scripts/train/train_tsn.sh
new file mode 100644
index 0000000000000000000000000000000000000000..e476744d4626b0354cbf0ebb4f7c4b4ffa55a7f1
--- /dev/null
+++ b/fluid/PaddleCV/video/scripts/train/train_tsn.sh
@@ -0,0 +1,2 @@
+python train.py --model-name="TSN" --config=./configs/tsn.txt --epoch-num=45 \
+ --valid-interval=1 --log-interval=10
diff --git a/fluid/PaddleCV/video/test.py b/fluid/PaddleCV/video/test.py
new file mode 100755
index 0000000000000000000000000000000000000000..9698caecc21a26dc38256b145dd54d04a2e13c88
--- /dev/null
+++ b/fluid/PaddleCV/video/test.py
@@ -0,0 +1,124 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import sys
+import time
+import logging
+import argparse
+import numpy as np
+import paddle.fluid as fluid
+
+from config import *
+import models
+from datareader import get_reader
+from metrics import get_metrics
+
+logging.root.handlers = []
+FORMAT = '[%(levelname)s: %(filename)s: %(lineno)4d]: %(message)s'
+logging.basicConfig(level=logging.INFO, format=FORMAT, stream=sys.stdout)
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+ parser = argparse.ArgumentParser()
+ parser.add_argument(
+ '--model-name',
+ type=str,
+ default='AttentionCluster',
+ help='name of model to train.')
+ parser.add_argument(
+ '--config',
+ type=str,
+ default='configs/attention_cluster.txt',
+ help='path to config file of model')
+ parser.add_argument(
+ '--batch-size',
+ type=int,
+ default=None,
+ help='traing batch size per GPU. None to use config file setting.')
+ parser.add_argument(
+ '--use-gpu', type=bool, default=True, help='default use gpu.')
+ parser.add_argument(
+ '--weights',
+ type=str,
+ default=None,
+ help='weight path, None to use weights from Paddle.')
+ parser.add_argument(
+ '--log-interval',
+ type=int,
+ default=1,
+ help='mini-batch interval to log.')
+ args = parser.parse_args()
+ return args
+
+
+def test(args):
+ # parse config
+ config = parse_config(args.config)
+ test_config = merge_configs(config, 'test', vars(args))
+
+ # build model
+ test_model = models.get_model(args.model_name, test_config, mode='test')
+ test_model.build_input(use_pyreader=False)
+ test_model.build_model()
+ test_feeds = test_model.feeds()
+ test_outputs = test_model.outputs()
+ loss = test_model.loss()
+
+ place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
+ exe = fluid.Executor(place)
+
+ if args.weights:
+ assert os.path.exists(
+ args.weights), "Given weight dir {} not exist.".format(args.weights)
+ weights = args.weights or test_model.get_weights()
+
+ def if_exist(var):
+ return os.path.exists(os.path.join(weights, var.name))
+
+ fluid.io.load_vars(exe, weights, predicate=if_exist)
+
+ # get reader and metrics
+ test_reader = get_reader(args.model_name.upper(), 'test', test_config)
+ test_metrics = get_metrics(args.model_name.upper(), 'test', test_config)
+
+ test_feeder = fluid.DataFeeder(place=place, feed_list=test_feeds)
+ fetch_list = [loss.name] + [x.name
+ for x in test_outputs] + [test_feeds[-1].name]
+
+ epoch_period = []
+ for test_iter, data in enumerate(test_reader()):
+ cur_time = time.time()
+ test_outs = exe.run(fetch_list=fetch_list,
+ feed=test_feeder.feed(data))
+ period = time.time() - cur_time
+ epoch_period.append(period)
+ loss = np.array(test_outs[0])
+ pred = np.array(test_outs[1])
+ label = np.array(test_outs[-1])
+ test_metrics.accumulate(loss, pred, label)
+
+ # metric here
+ if args.log_interval > 0 and test_iter % args.log_interval == 0:
+ info_str = '[EVAL] Batch {}'.format(test_iter)
+ test_metrics.calculate_and_log_out(loss, pred, label, info_str)
+ test_metrics.finalize_and_log_out("[EVAL] eval finished. ")
+
+
+if __name__ == "__main__":
+ args = parse_args()
+ logger.info(args)
+
+ test(args)
diff --git a/fluid/PaddleCV/video/tools/__init__.py b/fluid/PaddleCV/video/tools/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391
diff --git a/fluid/PaddleCV/video/tools/train_utils.py b/fluid/PaddleCV/video/tools/train_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..ad2b10f94827dacf1aad4822c5ce4dcb13fa94e9
--- /dev/null
+++ b/fluid/PaddleCV/video/tools/train_utils.py
@@ -0,0 +1,134 @@
+import os
+import time
+import numpy as np
+import paddle
+import paddle.fluid as fluid
+import logging
+import shutil
+
+logger = logging.getLogger(__name__)
+
+
+def test_without_pyreader(test_exe,
+ test_reader,
+ test_feeder,
+ test_fetch_list,
+ test_metrics,
+ log_interval=0):
+ test_metrics.reset()
+ for test_iter, data in enumerate(test_reader()):
+ test_outs = test_exe.run(test_fetch_list, feed=test_feeder.feed(data))
+ loss = np.array(test_outs[0])
+ pred = np.array(test_outs[1])
+ label = np.array(test_outs[-1])
+ test_metrics.accumulate(loss, pred, label)
+ if log_interval > 0 and test_iter % log_interval == 0:
+ test_metrics.calculate_and_log_out(loss, pred, label, \
+ info = '[TEST] test_iter {} '.format(test_iter))
+ test_metrics.finalize_and_log_out("[TEST] Finish")
+
+
+def test_with_pyreader(test_exe,
+ test_pyreader,
+ test_fetch_list,
+ test_metrics,
+ log_interval=0):
+ if not test_pyreader:
+ logger.error("[TEST] get pyreader failed.")
+ test_pyreader.start()
+ test_metrics.reset()
+ test_iter = 0
+ try:
+ while True:
+ test_outs = test_exe.run(fetch_list=test_fetch_list)
+ loss = np.array(test_outs[0])
+ pred = np.array(test_outs[1])
+ label = np.array(test_outs[-1])
+ test_metrics.accumulate(loss, pred, label)
+ if log_interval > 0 and test_iter % log_interval == 0:
+ test_metrics.calculate_and_log_out(loss, pred, label, \
+ info = '[TEST] test_iter {} '.format(test_iter))
+ test_iter += 1
+ except fluid.core.EOFException:
+ test_metrics.finalize_and_log_out("[TEST] Finish")
+ finally:
+ test_pyreader.reset()
+
+
+def train_without_pyreader(exe, train_prog, train_exe, train_reader, train_feeder, \
+ train_fetch_list, train_metrics, epochs = 10, \
+ log_interval = 0, valid_interval = 0, save_dir = './', \
+ save_model_name = 'model', test_exe = None, test_reader = None, \
+ test_feeder = None, test_fetch_list = None, test_metrics = None):
+ for epoch in range(epochs):
+ epoch_periods = []
+ for train_iter, data in enumerate(train_reader()):
+ cur_time = time.time()
+ train_outs = train_exe.run(train_fetch_list,
+ feed=train_feeder.feed(data))
+ period = time.time() - cur_time
+ epoch_periods.append(period)
+ loss = np.array(train_outs[0])
+ pred = np.array(train_outs[1])
+ label = np.array(train_outs[-1])
+ if log_interval > 0 and (train_iter % log_interval == 0):
+ # eval here
+ train_metrics.calculate_and_log_out(loss, pred, label, \
+ info = '[TRAIN] Epoch {}, iter {} '.format(epoch, train_iter))
+ train_iter += 1
+ logger.info('[TRAIN] Epoch {} training finished, average time: {}'.
+ format(epoch, np.mean(epoch_periods)))
+ save_model(exe, train_prog, save_dir, save_model_name,
+ "_epoch{}".format(epoch))
+ if test_exe and valid_interval > 0 and (epoch + 1) % valid_interval == 0:
+ test_without_pyreader(test_exe, test_reader, test_feeder,
+ test_fetch_list, test_metrics, log_interval)
+
+
+
+def train_with_pyreader(exe, train_prog, train_exe, train_pyreader, \
+ train_fetch_list, train_metrics, epochs = 10, \
+ log_interval = 0, valid_interval = 0, \
+ save_dir = './', save_model_name = 'model', \
+ test_exe = None, test_pyreader = None, \
+ test_fetch_list = None, test_metrics = None):
+ if not train_pyreader:
+ logger.error("[TRAIN] get pyreader failed.")
+ for epoch in range(epochs):
+ train_pyreader.start()
+ train_metrics.reset()
+ try:
+ train_iter = 0
+ epoch_periods = []
+ while True:
+ cur_time = time.time()
+ train_outs = train_exe.run(fetch_list=train_fetch_list)
+ period = time.time() - cur_time
+ epoch_periods.append(period)
+ loss = np.array(train_outs[0])
+ pred = np.array(train_outs[1])
+ label = np.array(train_outs[-1])
+ if log_interval > 0 and (train_iter % log_interval == 0):
+ # eval here
+ train_metrics.calculate_and_log_out(loss, pred, label, \
+ info = '[TRAIN] Epoch {}, iter {} '.format(epoch, train_iter))
+ train_iter += 1
+ except fluid.core.EOFException:
+ # eval here
+ logger.info('[TRAIN] Epoch {} training finished, average time: {}'.
+ format(epoch, np.mean(epoch_periods)))
+ save_model(exe, train_prog, save_dir, save_model_name,
+ "_epoch{}".format(epoch))
+ if test_exe and valid_interval > 0 and (epoch + 1) % valid_interval == 0:
+ test_with_pyreader(test_exe, test_pyreader, test_fetch_list,
+ test_metrics, log_interval)
+ finally:
+ epoch_period = []
+ train_pyreader.reset()
+
+
+def save_model(exe, program, save_dir, model_name, postfix=None):
+ model_path = os.path.join(save_dir, model_name + postfix)
+ if os.path.isdir(model_path):
+ shutil.rmtree(model_path)
+ fluid.io.save_persistables(exe, model_path, main_program=program)
diff --git a/fluid/PaddleCV/video/train.py b/fluid/PaddleCV/video/train.py
new file mode 100755
index 0000000000000000000000000000000000000000..154c51edd431286555b0e11d42a2c7a50ff4ee42
--- /dev/null
+++ b/fluid/PaddleCV/video/train.py
@@ -0,0 +1,226 @@
+# Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
+#
+#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
+import sys
+import time
+import argparse
+import logging
+import numpy as np
+import paddle.fluid as fluid
+
+from tools.train_utils import train_with_pyreader, train_without_pyreader
+import models
+from config import *
+from datareader import get_reader
+from metrics import get_metrics
+
+logging.root.handlers = []
+FORMAT = '[%(levelname)s: %(filename)s: %(lineno)4d]: %(message)s'
+logging.basicConfig(level=logging.INFO, format=FORMAT, stream=sys.stdout)
+logger = logging.getLogger(__name__)
+
+
+def parse_args():
+ parser = argparse.ArgumentParser("Paddle Video train script")
+ parser.add_argument(
+ '--model-name',
+ type=str,
+ default='AttentionCluster',
+ help='name of model to train.')
+ parser.add_argument(
+ '--config',
+ type=str,
+ default='configs/attention_cluster.txt',
+ help='path to config file of model')
+ parser.add_argument(
+ '--batch-size',
+ type=int,
+ default=None,
+ help='training batch size. None to use config file setting.')
+ parser.add_argument(
+ '--learning-rate',
+ type=float,
+ default=None,
+ help='learning rate use for training. None to use config file setting.')
+ parser.add_argument(
+ '--pretrain',
+ type=str,
+ default=None,
+ help='path to pretrain weights. None to use default weights path in ~/.paddle/weights.'
+ )
+ parser.add_argument(
+ '--resume',
+ type=str,
+ default=None,
+ help='path to resume training based on previous checkpoints. '
+ 'None for not resuming any checkpoints.'
+ )
+ parser.add_argument(
+ '--use-gpu', type=bool, default=True, help='default use gpu.')
+ parser.add_argument(
+ '--no-use-pyreader',
+ action='store_true',
+ default=False,
+ help='whether to use pyreader')
+ parser.add_argument(
+ '--no-memory-optimize',
+ action='store_true',
+ default=False,
+ help='whether to use memory optimize in train')
+ parser.add_argument(
+ '--epoch-num',
+ type=int,
+ default=0,
+ help='epoch number, 0 for read from config file')
+ parser.add_argument(
+ '--valid-interval',
+ type=int,
+ default=1,
+ help='validation epoch interval, 0 for no validation.')
+ parser.add_argument(
+ '--save-dir',
+ type=str,
+ default='checkpoints',
+ help='directory name to save train snapshoot')
+ parser.add_argument(
+ '--log-interval',
+ type=int,
+ default=10,
+ help='mini-batch interval to log.')
+ args = parser.parse_args()
+ return args
+
+
+def train(args):
+ # parse config
+ config = parse_config(args.config)
+ train_config = merge_configs(config, 'train', vars(args))
+ valid_config = merge_configs(config, 'valid', vars(args))
+ train_model = models.get_model(args.model_name, train_config, mode='train')
+ valid_model = models.get_model(args.model_name, valid_config, mode='valid')
+
+ # build model
+ startup = fluid.Program()
+ train_prog = fluid.Program()
+ with fluid.program_guard(train_prog, startup):
+ with fluid.unique_name.guard():
+ train_model.build_input(not args.no_use_pyreader)
+ train_model.build_model()
+ # for the input, has the form [data1, data2,..., label], so train_feeds[-1] is label
+ train_feeds = train_model.feeds()
+ train_feeds[-1].persistable = True
+ # for the output of classification model, has the form [pred]
+ train_outputs = train_model.outputs()
+ for output in train_outputs:
+ output.persistable = True
+ train_loss = train_model.loss()
+ train_loss.persistable = True
+ # outputs, loss, label should be fetched, so set persistable to be true
+ optimizer = train_model.optimizer()
+ optimizer.minimize(train_loss)
+ train_pyreader = train_model.pyreader()
+
+ if not args.no_memory_optimize:
+ fluid.memory_optimize(train_prog)
+
+ valid_prog = fluid.Program()
+ with fluid.program_guard(valid_prog, startup):
+ with fluid.unique_name.guard():
+ valid_model.build_input(not args.no_use_pyreader)
+ valid_model.build_model()
+ valid_feeds = valid_model.feeds()
+ valid_outputs = valid_model.outputs()
+ valid_loss = valid_model.loss()
+ valid_pyreader = valid_model.pyreader()
+
+ place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace()
+ exe = fluid.Executor(place)
+ exe.run(startup)
+
+ if args.resume:
+ # if resume weights is given, load resume weights directly
+ assert os.path.exists(args.resume), \
+ "Given resume weight dir {} not exist.".format(args.resume)
+ def if_exist(var):
+ return os.path.exists(os.path.join(args.resume, var.name))
+ fluid.io.load_vars(exe, args.resume, predicate=if_exist, main_program=train_prog)
+ else:
+ # if not in resume mode, load pretrain weights
+ if args.pretrain:
+ assert os.path.exists(args.pretrain), \
+ "Given pretrain weight dir {} not exist.".format(args.pretrain)
+ pretrain = args.pretrain or train_model.get_pretrain_weights()
+ if pretrain:
+ train_model.load_pretrain_params(exe, pretrain, train_prog, place)
+
+ train_exe = fluid.ParallelExecutor(
+ use_cuda=args.use_gpu,
+ loss_name=train_loss.name,
+ main_program=train_prog)
+ valid_exe = fluid.ParallelExecutor(
+ use_cuda=args.use_gpu,
+ share_vars_from=train_exe,
+ main_program=valid_prog)
+
+ # get reader
+ bs_denominator = 1
+ if (not args.no_use_pyreader) and args.use_gpu:
+ bs_denominator = train_config.TRAIN.num_gpus
+ train_config.TRAIN.batch_size = int(train_config.TRAIN.batch_size /
+ bs_denominator)
+ valid_config.VALID.batch_size = int(valid_config.VALID.batch_size /
+ bs_denominator)
+ train_reader = get_reader(args.model_name.upper(), 'train', train_config)
+ valid_reader = get_reader(args.model_name.upper(), 'valid', valid_config)
+
+ # get metrics
+ train_metrics = get_metrics(args.model_name.upper(), 'train', train_config)
+ valid_metrics = get_metrics(args.model_name.upper(), 'valid', valid_config)
+
+ train_fetch_list = [train_loss.name] + [x.name for x in train_outputs
+ ] + [train_feeds[-1].name]
+ valid_fetch_list = [valid_loss.name] + [x.name for x in valid_outputs
+ ] + [valid_feeds[-1].name]
+
+ epochs = args.epoch_num or train_model.epoch_num()
+
+ if args.no_use_pyreader:
+ train_feeder = fluid.DataFeeder(place=place, feed_list=train_feeds)
+ valid_feeder = fluid.DataFeeder(place=place, feed_list=valid_feeds)
+ train_without_pyreader(exe, train_prog, train_exe, train_reader, train_feeder,
+ train_fetch_list, train_metrics, epochs = epochs,
+ log_interval = args.log_interval, valid_interval = args.valid_interval,
+ save_dir = args.save_dir, save_model_name = args.model_name,
+ test_exe = valid_exe, test_reader = valid_reader, test_feeder = valid_feeder,
+ test_fetch_list = valid_fetch_list, test_metrics = valid_metrics)
+ else:
+ train_pyreader.decorate_paddle_reader(train_reader)
+ valid_pyreader.decorate_paddle_reader(valid_reader)
+ train_with_pyreader(exe, train_prog, train_exe, train_pyreader, train_fetch_list, train_metrics,
+ epochs = epochs, log_interval = args.log_interval,
+ valid_interval = args.valid_interval,
+ save_dir = args.save_dir, save_model_name = args.model_name,
+ test_exe = valid_exe, test_pyreader = valid_pyreader,
+ test_fetch_list = valid_fetch_list, test_metrics = valid_metrics)
+
+
+if __name__ == "__main__":
+ args = parse_args()
+ logger.info(args)
+
+ if not os.path.exists(args.save_dir):
+ os.makedirs(args.save_dir)
+
+ train(args)
diff --git a/fluid/PaddleCV/video/utils.py b/fluid/PaddleCV/video/utils.py
new file mode 100755
index 0000000000000000000000000000000000000000..3b07d606c60b9834429fef94d43c0a5619cd1db5
--- /dev/null
+++ b/fluid/PaddleCV/video/utils.py
@@ -0,0 +1,25 @@
+# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
+#
+#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.
+
+__all__ = ['AttrDict']
+
+class AttrDict(dict):
+ def __getattr__(self, key):
+ return self[key]
+
+ def __setattr__(self, key, value):
+ if key in self.__dict__:
+ self.__dict__[key] = value
+ else:
+ self[key] = value
diff --git a/fluid/PaddleNLP/LAC b/fluid/PaddleNLP/LAC
index d2fc9e0b45b4e6cfc93e73054026fc5a8abfbfb9..a4eb73b2fb64d8aab8499a1184edf4fc386f8268 160000
--- a/fluid/PaddleNLP/LAC
+++ b/fluid/PaddleNLP/LAC
@@ -1 +1 @@
-Subproject commit d2fc9e0b45b4e6cfc93e73054026fc5a8abfbfb9
+Subproject commit a4eb73b2fb64d8aab8499a1184edf4fc386f8268
diff --git a/fluid/PaddleNLP/Senta b/fluid/PaddleNLP/Senta
index 733c1d02085a3092dd262c4f396563962a514c3e..dc1af6a83dd1372055158ac6d17f6d14b3a0f0f8 160000
--- a/fluid/PaddleNLP/Senta
+++ b/fluid/PaddleNLP/Senta
@@ -1 +1 @@
-Subproject commit 733c1d02085a3092dd262c4f396563962a514c3e
+Subproject commit dc1af6a83dd1372055158ac6d17f6d14b3a0f0f8
diff --git a/fluid/PaddleNLP/SimNet b/fluid/PaddleNLP/SimNet
index 60b698a294c34420a7f0aab3112f27649aed1445..57b93859aa070ae6d96f10a470b1bdf2cfaea052 160000
--- a/fluid/PaddleNLP/SimNet
+++ b/fluid/PaddleNLP/SimNet
@@ -1 +1 @@
-Subproject commit 60b698a294c34420a7f0aab3112f27649aed1445
+Subproject commit 57b93859aa070ae6d96f10a470b1bdf2cfaea052
diff --git a/fluid/PaddleNLP/chinese_ner/infer.py b/fluid/PaddleNLP/chinese_ner/infer.py
index a15fdb53d89f2f7845e6bb54aa32fe922bb64682..dd0d156b015423cb6805b3eb946383905328c835 100644
--- a/fluid/PaddleNLP/chinese_ner/infer.py
+++ b/fluid/PaddleNLP/chinese_ner/infer.py
@@ -25,7 +25,7 @@ def parse_args():
parser.add_argument(
'--model_path',
type=str,
- default='model/params_pass_0',
+ default='output/params_pass_0',
help='A path to the model. (default: %(default)s)')
parser.add_argument(
'--test_data_dir',
diff --git a/fluid/PaddleNLP/deep_attention_matching_net/test_and_evaluate.py b/fluid/PaddleNLP/deep_attention_matching_net/test_and_evaluate.py
index 998914d1fdcdb4bf3c442fb5276274ebf0aae038..98239220054af17b136deb32da71b7ec81366d86 100644
--- a/fluid/PaddleNLP/deep_attention_matching_net/test_and_evaluate.py
+++ b/fluid/PaddleNLP/deep_attention_matching_net/test_and_evaluate.py
@@ -130,13 +130,13 @@ def test(args):
loss, logits = dam.create_network()
loss.persistable = True
+ logits.persistable = True
# gradient clipping
fluid.clip.set_gradient_clip(clip=fluid.clip.GradientClipByValue(
max=1.0, min=-1.0))
test_program = fluid.default_main_program().clone(for_test=True)
-
optimizer = fluid.optimizer.Adam(
learning_rate=fluid.layers.exponential_decay(
learning_rate=args.learning_rate,
@@ -145,7 +145,6 @@ def test(args):
staircase=True))
optimizer.minimize(loss)
- # The fethced loss is wrong when mem opt is enabled
fluid.memory_optimize(fluid.default_main_program())
if args.use_cuda:
@@ -173,8 +172,10 @@ def test(args):
if args.ext_eval:
import utils.douban_evaluation as eva
+ eval_metrics = ["MAP", "MRR", "P@1", "R_{10}@1", "R_{10}@2", "R_{10}@5"]
else:
import utils.evaluation as eva
+ eval_metrics = ["R_2@1", "R_{10}@1", "R_{10}@2", "R_{10}@5"]
test_batches = reader.build_batches(test_data, data_conf)
@@ -214,8 +215,8 @@ def test(args):
result = eva.evaluate(score_path)
result_file_path = os.path.join(args.save_path, 'result.txt')
with open(result_file_path, 'w') as out_file:
- for p_at in result:
- out_file.write(str(p_at) + '\n')
+ for metric, p_at in zip(eval_metrics, result):
+ out_file.write(metric + ": " + str(p_at) + '\n')
print('finish test')
print(time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time())))
diff --git a/fluid/PaddleNLP/machine_reading_comprehension/run.py b/fluid/PaddleNLP/machine_reading_comprehension/run.py
index 884549d106af7f44789728fb488b5e60e149e118..e9ba1d0b14023f75d7551728dd22571cf8b72fa4 100644
--- a/fluid/PaddleNLP/machine_reading_comprehension/run.py
+++ b/fluid/PaddleNLP/machine_reading_comprehension/run.py
@@ -523,8 +523,8 @@ def evaluate(logger, args):
inference_program = main_program.clone(for_test=True)
eval_loss, bleu_rouge = validation(
- inference_program, avg_cost, s_probs, e_probs, feed_order,
- place, dev_count, vocab, brc_data, logger, args)
+ inference_program, avg_cost, s_probs, e_probs, match,
+ feed_order, place, dev_count, vocab, brc_data, logger, args)
logger.info('Dev eval loss {}'.format(eval_loss))
logger.info('Dev eval result: {}'.format(bleu_rouge))
logger.info('Predicted answers are saved to {}'.format(
diff --git a/fluid/PaddleNLP/neural_machine_translation/rnn_search/README.md b/fluid/PaddleNLP/neural_machine_translation/rnn_search/README.md
index 59860114a101b54d8c5f148bd8d725d9bfe778bc..86d4a021baf11e04a9fd07c05dbf50425451efab 100644
--- a/fluid/PaddleNLP/neural_machine_translation/rnn_search/README.md
+++ b/fluid/PaddleNLP/neural_machine_translation/rnn_search/README.md
@@ -1,4 +1,4 @@
-运行本目录下的范例模型需要安装PaddlePaddle Fluid 1.0版。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](http://www.paddlepaddle.org/docs/develop/documentation/zh/build_and_install/pip_install_cn.html)中的说明更新 PaddlePaddle 安装版本。
+运行本目录下的范例模型需要安装PaddlePaddle Fluid 1.0版。如果您的 PaddlePaddle 安装版本低于此要求，请按照[安装文档](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/install/index_cn.html)中的说明更新 PaddlePaddle 安装版本。
# 机器翻译：RNN Search
@@ -24,7 +24,7 @@
本目录下此范例模型的实现，旨在展示如何用Paddle Fluid实现一个带有注意力机制（Attention）的RNN模型来解决Seq2Seq类问题，以及如何使用带有Beam Search算法的解码器。如果您仅仅只是需要在机器翻译方面有着较好翻译效果的模型，则建议您参考[Transformer的Paddle Fluid实现](https://github.com/PaddlePaddle/models/tree/develop/fluid/neural_machine_translation/transformer)。
## 模型概览
-RNN Search模型使用了经典的编码器-解码器（Encoder-Decoder）的框架结构来解决Seq2Seq类问题。这种方法先用编码器将源序列编码成vector，再用解码器将该vector解码为目标序列。这其实模拟了人类在进行翻译类任务时的行为：先解析源语言，理解其含义，再根据该含义来写出目标语言的语句。编码器和解码器往往都使用RNN来实现。关于此方法的具体原理和数学表达式，可以参考[深度学习101](http://www.paddlepaddle.org/documentation/docs/zh/0.15.0/beginners_guide/basics/machine_translation/index.html).
+RNN Search模型使用了经典的编码器-解码器（Encoder-Decoder）的框架结构来解决Seq2Seq类问题。这种方法先用编码器将源序列编码成vector，再用解码器将该vector解码为目标序列。这其实模拟了人类在进行翻译类任务时的行为：先解析源语言，理解其含义，再根据该含义来写出目标语言的语句。编码器和解码器往往都使用RNN来实现。关于此方法的具体原理和数学表达式，可以参考[深度学习101](http://paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/basics/machine_translation/index.html).
本模型中，在编码器方面，我们的实现使用了双向循环神经网络（Bi-directional Recurrent Neural Network）；在解码器方面，我们使用了带注意力（Attention）机制的RNN解码器，并同时提供了一个不带注意力机制的解码器实现作为对比；而在预测方面我们使用柱搜索（beam search）算法来生成翻译的目标语句。以下将分别介绍用到的这些方法。
@@ -45,7 +45,7 @@ RNN Search模型使用了经典的编码器-解码器（Encoder-Decoder）的框
### 注意力机制
如果编码阶段的输出是一个固定维度的向量，会带来以下两个问题：1）不论源语言序列的长度是5个词还是50个词，如果都用固定维度的向量去编码其中的语义和句法结构信息，对模型来说是一个非常高的要求，特别是对长句子序列而言；2）直觉上，当人类翻译一句话时，会对与当前译文更相关的源语言片段上给予更多关注，且关注点会随着翻译的进行而改变。而固定维度的向量则相当于，任何时刻都对源语言所有信息给予了同等程度的关注，这是不合理的。因此，Bahdanau等人\[[4](#参考文献)\]引入注意力（attention）机制，可以对编码后的上下文片段进行解码，以此来解决长句子的特征学习问题。下面介绍在注意力机制下的解码器结构。
-与简单的解码器不同，这里$z_i$的计算公式为：
+与简单的解码器不同，这里$z_i$的计算公式为 （由于Github原生不支持LaTeX公式，请您移步[这里](http://www.paddlepaddle.org/documentation/docs/zh/1.2/beginners_guide/basics/machine_translation/index.html)查看）：
$$z_{i+1}=\phi _{\theta '}\left ( c_i,u_i,z_i \right )$$
@@ -131,4 +131,4 @@ python infer.py
5. Papineni K, Roukos S, Ward T, et al. [BLEU: a method for automatic evaluation of machine translation](http://dl.acm.org/citation.cfm?id=1073135)[C]//Proceedings of the 40th annual meeting on association for computational linguistics. Association for Computational Linguistics, 2002: 311-318.
-
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。
\ No newline at end of file
+
本教程 由 PaddlePaddle 创作，采用 知识共享 署名-相同方式共享 4.0 国际 许可协议进行许可。
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md b/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md
index 7e7a09e7a1e4e8dcfddc5dbbc27c94a757d80d9e..bdac7cb0b7c4f9d51bbc281b351232c6edc75a36 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/README_cn.md
@@ -69,9 +69,9 @@ WMT 数据集是机器翻译领域公认的主流数据集，[WMT'16 EN-DE 数
└── subword-nmt # BPE 编码的代码
```
-`gen_data/wmt16_ende_data_bpe` 中是我们最终使用的英德翻译数据，其中 `train.tok.clean.bpe.32000.en-de` 为训练数据，`newstest2016.tok.bpe.32000.en-de` 等为验证和测试数据，。`vocab_all.bpe.32000` 为相应的词典文件（已加入 `` 、`` 和 `` 这三个特殊符号，源语言和目标语言共享该词典文件）。
+`gen_data/wmt16_ende_data_bpe` 中是我们最终使用的英德翻译数据，其中 `train.tok.clean.bpe.32000.en-de` 为训练数据，`newstest2016.tok.bpe.32000.en-de` 等为验证和测试数据，`vocab_all.bpe.32000` 为相应的词典文件（已加入 `` 、`` 和 `` 这三个特殊符号，源语言和目标语言共享该词典文件）。另外我们也整理提供了一份处理好的 WMT'16 EN-DE 数据以供[下载](https://transformer-res.bj.bcebos.com/wmt16_ende_data_bpe_clean.tar.gz)使用（包含训练所需 BPE 数据和词典以及预测和评估所需的 BPE 数据和 tokenize 的数据）。
-对于其他自定义数据，转换为类似 `train.tok.clean.bpe.32000.en-de` 的数据格式（`\t` 分隔的源语言和目标语言句子对，句子中的 token 之间使用空格分隔）即可；如需使用 BPE 编码，可参考，亦可以使用类似 WMT，使用 `gen_data.sh` 进行处理。
+对于其他自定义数据，转换为类似 `train.tok.clean.bpe.32000.en-de` 的数据格式（`\t` 分隔的源语言和目标语言句子对，句子中的 token 之间使用空格分隔）即可；如需使用 BPE 编码，亦可以使用类似 WMT'16 EN-DE 原始数据的格式，参照 `gen_data.sh` 进行处理。
### 模型训练
@@ -110,11 +110,9 @@ python -u train.py \
--batch_size 3200 \
--sort_type pool \
--pool_size 200000 \
- n_layer 6 \
n_head 16 \
d_model 1024 \
d_inner_hid 4096 \
- n_head 16 \
prepostprocess_dropout 0.3
```
有关这些参数更详细信息的请参考 `config.py` 中的注释说明。
@@ -144,30 +142,53 @@ python -u infer.py \
--token_delimiter ' ' \
--batch_size 32 \
model_path trained_models/iter_100000.infer.model \
- beam_size 4 \
+ beam_size 5 \
max_out_len 255
```
-和模型训练时类似，预测时也需要设置数据和 reader 相关的参数，并可以执行 `python infer.py --help` 查看这些参数的说明（部分参数意义和训练时略有不同）；同样可以在预测命令中设置模型超参数，但应与模型训练时的设置一致；此外相比于模型训练，预测时还有一些额外的参数，如需要设置 `model_path` 来给出模型所在目录，可以设置 `beam_size` 和 `max_out_len` 来指定 Beam Search 算法的搜索宽度和最大深度（翻译长度），这些参数也可以在 `config.py` 中的 `InferTaskConfig` 内查阅注释说明并进行更改设置。
+和模型训练时类似，预测时也需要设置数据和 reader 相关的参数，并可以执行 `python infer.py --help` 查看这些参数的说明（部分参数意义和训练时略有不同）；同样可以在预测命令中设置模型超参数，但应与模型训练时的设置一致，如训练时使用 big model 的参数设置，则预测时对应类似如下命令：
+```sh
+python -u infer.py \
+ --src_vocab_fpath gen_data/wmt16_ende_data_bpe/vocab_all.bpe.32000 \
+ --trg_vocab_fpath gen_data/wmt16_ende_data_bpe/vocab_all.bpe.32000 \
+ --special_token '' '' '' \
+ --test_file_pattern gen_data/wmt16_ende_data_bpe/newstest2016.tok.bpe.32000.en-de \
+ --token_delimiter ' ' \
+ --batch_size 32 \
+ model_path trained_models/iter_100000.infer.model \
+ n_head 16 \
+ d_model 1024 \
+ d_inner_hid 4096 \
+ prepostprocess_dropout 0.3 \
+ beam_size 5 \
+ max_out_len 255
+```
+此外相比于模型训练，预测时还有一些额外的参数，如需要设置 `model_path` 来给出模型所在目录，可以设置 `beam_size` 和 `max_out_len` 来指定 Beam Search 算法的搜索宽度和最大深度（翻译长度），这些参数也可以在 `config.py` 中的 `InferTaskConfig` 内查阅注释说明并进行更改设置。
执行以上预测命令会打印翻译结果到标准输出，每行输出是对应行输入的得分最高的翻译。对于使用 BPE 的英德数据，预测出的翻译结果也将是 BPE 表示的数据，要还原成原始的数据（这里指 tokenize 后的数据）才能进行正确的评估，可以使用以下命令来恢复 `predict.txt` 内的翻译结果到 `predict.tok.txt` 中（无需再次 tokenize 处理）：
```sh
sed -r 's/(@@ )|(@@ ?$)//g' predict.txt > predict.tok.txt
```
-接下来就可以使用参考翻译对翻译结果进行 BLEU 指标的评估了。以英德翻译 `newstest2016.tok.de` 数据为例，执行如下命令：
+接下来就可以使用参考翻译对翻译结果进行 BLEU 指标的评估了，评估需要用到 mosesdecoder 中的脚本，可以通过以下命令获取：
+```sh
+git clone https://github.com/moses-smt/mosesdecoder.git
+```
+以英德翻译 `newstest2014.tok.de` 数据为例，获取 mosesdecoder 后使用 `multi-bleu.perl` 执行如下命令进行翻译结果评估：
```sh
-perl gen_data/mosesdecoder/scripts/generic/multi-bleu.perl gen_data/wmt16_ende_data/newstest2016.tok.de < predict.tok.txt
+perl gen_data/mosesdecoder/scripts/generic/multi-bleu.perl gen_data/wmt16_ende_data/newstest2014.tok.de < predict.tok.txt
```
-可以看到类似如下的结果（为单机两卡训练 200K 个 iteration 后模型的预测结果）。
+可以看到类似如下的结果：
```
-BLEU = 33.08, 64.2/39.2/26.4/18.5 (BP=0.994, ratio=0.994, hyp_len=61971, ref_len=62362)
+BLEU = 26.35, 57.7/32.1/20.0/13.0 (BP=1.000, ratio=1.013, hyp_len=63903, ref_len=63078)
```
-目前在未使用 model average 的情况下，英德翻译 base model 八卡训练 100K 个 iteration 后测试 BLEU 值如下：
+目前在未使用 model average 的情况下，英德翻译 base model 和 big model 八卡训练 100K 个 iteration 后测试 BLEU 值如下：
| 测试集 | newstest2014 | newstest2015 | newstest2016 |
|-|-|-|-|
-| BLEU | 26.25 | 29.15 | 33.64 |
+| Base | 26.35 | 29.07 | 33.30 |
+| Big | 27.07 | 30.09 | 34.38 |
+我们这里也提供了以上 [base model](https://transformer-res.bj.bcebos.com/base_model.tar.gz) 和 [big model](https://transformer-res.bj.bcebos.com/big_model.tar.gz) 模型的下载以供使用。
### 分布式训练
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/config.py b/fluid/PaddleNLP/neural_machine_translation/transformer/config.py
index ca119aa6fd0878b1e2cea5c0eaba050b54348f79..823341ed9084e80b5fe74655bf8db897d72175f0 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/config.py
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/config.py
@@ -164,7 +164,10 @@ input_descs = {
# [batch_size * max_trg_len_in_batch, 1]
"lbl_weight": [(batch_size * seq_len, 1), "float32"],
# This input is used in beam-search decoder.
- "init_score": [(batch_size, 1), "float32"],
+ "init_score": [(batch_size, 1), "float32", 2],
+ # This input is used in beam-search decoder for the first gather
+ # (cell states updation)
+ "init_idx": [(batch_size, ), "int32"],
}
# Names of word embedding table which might be reused for weight sharing.
@@ -194,4 +197,5 @@ label_data_input_fields = (
fast_decoder_data_input_fields = (
"trg_word",
"init_score",
+ "init_idx",
"trg_src_attn_bias", )
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/images/attention_formula.png b/fluid/PaddleNLP/neural_machine_translation/transformer/images/attention_formula.png
new file mode 100644
index 0000000000000000000000000000000000000000..249857f524b4137bafc2d4d1b779ed62d1437b6d
Binary files /dev/null and b/fluid/PaddleNLP/neural_machine_translation/transformer/images/attention_formula.png differ
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/infer.py b/fluid/PaddleNLP/neural_machine_translation/transformer/infer.py
index 6fc04a9422c136d941559d1b45af8bd88c2d2460..57ea546a5c6eb6ef484bbd975312cf99fd0ba18b 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/infer.py
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/infer.py
@@ -1,18 +1,20 @@
import argparse
import ast
+import multiprocessing
import numpy as np
+import os
from functools import partial
import paddle
import paddle.fluid as fluid
import model
+import reader
+from config import *
from model import wrap_encoder as encoder
from model import wrap_decoder as decoder
from model import fast_decode as fast_decoder
-from config import *
-from train import pad_batch_data
-import reader
+from train import pad_batch_data, prepare_data_generator
def parse_args():
@@ -54,6 +56,21 @@ def parse_args():
default=" ",
help="The delimiter used to split tokens in source or target sentences. "
"For EN-DE BPE data we provided, use spaces as token delimiter. ")
+ parser.add_argument(
+ "--use_mem_opt",
+ type=ast.literal_eval,
+ default=True,
+ help="The flag indicating whether to use memory optimization.")
+ parser.add_argument(
+ "--use_py_reader",
+ type=ast.literal_eval,
+ default=True,
+ help="The flag indicating whether to use py_reader.")
+ parser.add_argument(
+ "--use_parallel_exe",
+ type=ast.literal_eval,
+ default=False,
+ help="The flag indicating whether to use ParallelExecutor.")
parser.add_argument(
'opts',
help='See config.py for all options',
@@ -123,106 +140,185 @@ def prepare_batch_input(insts, data_input_names, src_pad_idx, bos_idx, n_head,
trg_word, dtype="float32").reshape(-1, 1),
place, [range(trg_word.shape[0] + 1)] * 2)
trg_word = to_lodtensor(trg_word, place, [range(trg_word.shape[0] + 1)] * 2)
+ init_idx = np.asarray(range(len(insts)), dtype="int32")
data_input_dict = dict(
zip(data_input_names, [
src_word, src_pos, src_slf_attn_bias, trg_word, init_score,
- trg_src_attn_bias
+ init_idx, trg_src_attn_bias
]))
+ return data_input_dict
+
+
+def prepare_feed_dict_list(data_generator, count, place):
+ """
+ Prepare the list of feed dict for multi-devices.
+ """
+ feed_dict_list = []
+ if data_generator is not None: # use_py_reader == False
+ data_input_names = encoder_data_input_fields + fast_decoder_data_input_fields
+ data = next(data_generator)
+ for idx, data_buffer in enumerate(data):
+ data_input_dict = prepare_batch_input(
+ data_buffer, data_input_names, ModelHyperParams.eos_idx,
+ ModelHyperParams.bos_idx, ModelHyperParams.n_head,
+ ModelHyperParams.d_model, place)
+ feed_dict_list.append(data_input_dict)
+ return feed_dict_list if len(feed_dict_list) == count else None
+
+
+def py_reader_provider_wrapper(data_reader, place):
+ """
+ Data provider needed by fluid.layers.py_reader.
+ """
- input_dict = dict(data_input_dict.items())
- return input_dict
+ def py_reader_provider():
+ data_input_names = encoder_data_input_fields + fast_decoder_data_input_fields
+ for batch_id, data in enumerate(data_reader()):
+ data_input_dict = prepare_batch_input(
+ data, data_input_names, ModelHyperParams.eos_idx,
+ ModelHyperParams.bos_idx, ModelHyperParams.n_head,
+ ModelHyperParams.d_model, place)
+ yield [data_input_dict[item] for item in data_input_names]
+ return py_reader_provider
-def fast_infer(test_data, trg_idx2word):
+
+def fast_infer(args):
"""
Inference by beam search decoder based solely on Fluid operators.
"""
- place = fluid.CUDAPlace(0) if InferTaskConfig.use_gpu else fluid.CPUPlace()
- exe = fluid.Executor(place)
+ out_ids, out_scores, pyreader = fast_decoder(
+ ModelHyperParams.src_vocab_size,
+ ModelHyperParams.trg_vocab_size,
+ ModelHyperParams.max_length + 1,
+ ModelHyperParams.n_layer,
+ ModelHyperParams.n_head,
+ ModelHyperParams.d_key,
+ ModelHyperParams.d_value,
+ ModelHyperParams.d_model,
+ ModelHyperParams.d_inner_hid,
+ ModelHyperParams.prepostprocess_dropout,
+ ModelHyperParams.attention_dropout,
+ ModelHyperParams.relu_dropout,
+ ModelHyperParams.preprocess_cmd,
+ ModelHyperParams.postprocess_cmd,
+ ModelHyperParams.weight_sharing,
+ InferTaskConfig.beam_size,
+ InferTaskConfig.max_out_len,
+ ModelHyperParams.eos_idx,
+ use_py_reader=args.use_py_reader)
+
+ # This is used here to set dropout to the test mode.
+ infer_program = fluid.default_main_program().clone(for_test=True)
- out_ids, out_scores = fast_decoder(
- ModelHyperParams.src_vocab_size, ModelHyperParams.trg_vocab_size,
- ModelHyperParams.max_length + 1, ModelHyperParams.n_layer,
- ModelHyperParams.n_head, ModelHyperParams.d_key,
- ModelHyperParams.d_value, ModelHyperParams.d_model,
- ModelHyperParams.d_inner_hid, ModelHyperParams.prepostprocess_dropout,
- ModelHyperParams.attention_dropout, ModelHyperParams.relu_dropout,
- ModelHyperParams.preprocess_cmd, ModelHyperParams.postprocess_cmd,
- ModelHyperParams.weight_sharing, InferTaskConfig.beam_size,
- InferTaskConfig.max_out_len, ModelHyperParams.eos_idx)
+ if args.use_mem_opt:
+ fluid.memory_optimize(infer_program)
+
+ if InferTaskConfig.use_gpu:
+ place = fluid.CUDAPlace(0)
+ dev_count = fluid.core.get_cuda_device_count()
+ else:
+ place = fluid.CPUPlace()
+ dev_count = int(os.environ.get('CPU_NUM', multiprocessing.cpu_count()))
+ exe = fluid.Executor(place)
+ exe.run(fluid.default_startup_program())
fluid.io.load_vars(
exe,
InferTaskConfig.model_path,
vars=[
- var for var in fluid.default_main_program().list_vars()
+ var for var in infer_program.list_vars()
if isinstance(var, fluid.framework.Parameter)
])
- # This is used here to set dropout to the test mode.
- infer_program = fluid.default_main_program().clone(for_test=True)
+ exec_strategy = fluid.ExecutionStrategy()
+ # For faster executor
+ exec_strategy.use_experimental_executor = True
+ exec_strategy.num_threads = 1
+ build_strategy = fluid.BuildStrategy()
+ infer_exe = fluid.ParallelExecutor(
+ use_cuda=TrainTaskConfig.use_gpu,
+ main_program=infer_program,
+ build_strategy=build_strategy,
+ exec_strategy=exec_strategy)
- for batch_id, data in enumerate(test_data.batch_generator()):
- data_input = prepare_batch_input(
- data, encoder_data_input_fields + fast_decoder_data_input_fields,
- ModelHyperParams.eos_idx, ModelHyperParams.bos_idx,
- ModelHyperParams.n_head, ModelHyperParams.d_model, place)
- seq_ids, seq_scores = exe.run(infer_program,
- feed=data_input,
- fetch_list=[out_ids, out_scores],
- return_numpy=False)
- # How to parse the results:
- # Suppose the lod of seq_ids is:
- # [[0, 3, 6], [0, 12, 24, 40, 54, 67, 82]]
- # then from lod[0]:
- # there are 2 source sentences, beam width is 3.
- # from lod[1]:
- # the first source sentence has 3 hyps; the lengths are 12, 12, 16
- # the second source sentence has 3 hyps; the lengths are 14, 13, 15
- hyps = [[] for i in range(len(data))]
- scores = [[] for i in range(len(data))]
- for i in range(len(seq_ids.lod()[0]) - 1): # for each source sentence
- start = seq_ids.lod()[0][i]
- end = seq_ids.lod()[0][i + 1]
- for j in range(end - start): # for each candidate
- sub_start = seq_ids.lod()[1][start + j]
- sub_end = seq_ids.lod()[1][start + j + 1]
- hyps[i].append(" ".join([
- trg_idx2word[idx]
- for idx in post_process_seq(
- np.array(seq_ids)[sub_start:sub_end])
- ]))
- scores[i].append(np.array(seq_scores)[sub_end - 1])
- print(hyps[i][-1])
- if len(hyps[i]) >= InferTaskConfig.n_best:
- break
-
-
-def infer(args, inferencer=fast_infer):
- place = fluid.CUDAPlace(0) if InferTaskConfig.use_gpu else fluid.CPUPlace()
- test_data = reader.DataReader(
- src_vocab_fpath=args.src_vocab_fpath,
- trg_vocab_fpath=args.trg_vocab_fpath,
- fpattern=args.test_file_pattern,
- token_delimiter=args.token_delimiter,
- use_token_batch=False,
- batch_size=args.batch_size,
- pool_size=args.pool_size,
- sort_type=reader.SortType.NONE,
- shuffle=False,
- shuffle_batch=False,
- start_mark=args.special_token[0],
- end_mark=args.special_token[1],
- unk_mark=args.special_token[2],
- # count start and end tokens out
- max_length=ModelHyperParams.max_length - 2,
- clip_last_batch=False)
- trg_idx2word = test_data.load_dict(
+ # data reader settings for inference
+ args.train_file_pattern = args.test_file_pattern
+ args.use_token_batch = False
+ args.sort_type = reader.SortType.NONE
+ args.shuffle = False
+ args.shuffle_batch = False
+ test_data = prepare_data_generator(
+ args,
+ is_test=False,
+ count=dev_count,
+ pyreader=pyreader,
+ py_reader_provider_wrapper=py_reader_provider_wrapper,
+ place=place)
+ if args.use_py_reader:
+ pyreader.start()
+ data_generator = None
+ else:
+ data_generator = test_data()
+ trg_idx2word = reader.DataReader.load_dict(
dict_path=args.trg_vocab_fpath, reverse=True)
- inferencer(test_data, trg_idx2word)
+
+ while True:
+ try:
+ feed_dict_list = prepare_feed_dict_list(data_generator, dev_count,
+ place)
+ if args.use_parallel_exe:
+ seq_ids, seq_scores = infer_exe.run(
+ fetch_list=[out_ids.name, out_scores.name],
+ feed=feed_dict_list,
+ return_numpy=False)
+ else:
+ seq_ids, seq_scores = exe.run(
+ program=infer_program,
+ fetch_list=[out_ids.name, out_scores.name],
+ feed=feed_dict_list[0]
+ if feed_dict_list is not None else None,
+ return_numpy=False,
+ use_program_cache=True)
+ seq_ids_list, seq_scores_list = [seq_ids], [
+ seq_scores
+ ] if isinstance(
+ seq_ids, paddle.fluid.core.LoDTensor) else (seq_ids, seq_scores)
+ for seq_ids, seq_scores in zip(seq_ids_list, seq_scores_list):
+ # How to parse the results:
+ # Suppose the lod of seq_ids is:
+ # [[0, 3, 6], [0, 12, 24, 40, 54, 67, 82]]
+ # then from lod[0]:
+ # there are 2 source sentences, beam width is 3.
+ # from lod[1]:
+ # the first source sentence has 3 hyps; the lengths are 12, 12, 16
+ # the second source sentence has 3 hyps; the lengths are 14, 13, 15
+ hyps = [[] for i in range(len(seq_ids.lod()[0]) - 1)]
+ scores = [[] for i in range(len(seq_scores.lod()[0]) - 1)]
+ for i in range(len(seq_ids.lod()[0]) -
+ 1): # for each source sentence
+ start = seq_ids.lod()[0][i]
+ end = seq_ids.lod()[0][i + 1]
+ for j in range(end - start): # for each candidate
+ sub_start = seq_ids.lod()[1][start + j]
+ sub_end = seq_ids.lod()[1][start + j + 1]
+ hyps[i].append(" ".join([
+ trg_idx2word[idx]
+ for idx in post_process_seq(
+ np.array(seq_ids)[sub_start:sub_end])
+ ]))
+ scores[i].append(np.array(seq_scores)[sub_end - 1])
+ print(hyps[i][-1])
+ if len(hyps[i]) >= InferTaskConfig.n_best:
+ break
+ except (StopIteration, fluid.core.EOFException):
+ # The data pass is over.
+ if args.use_py_reader:
+ pyreader.reset()
+ break
if __name__ == "__main__":
args = parse_args()
- infer(args)
+ fast_infer(args)
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/model.py b/fluid/PaddleNLP/neural_machine_translation/transformer/model.py
index 1e510bc620dc56f82e8e7303a56ca44a44b74650..bf68e089004cf53bfb8d2910b48664e25e44c10d 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/model.py
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/model.py
@@ -7,6 +7,43 @@ import paddle.fluid.layers as layers
from config import *
+def wrap_layer_with_block(layer, block_idx):
+ """
+ Make layer define support indicating block, by which we can add layers
+ to other blocks within current block. This will make it easy to define
+ cache among while loop.
+ """
+
+ class BlockGuard(object):
+ """
+ BlockGuard class.
+
+ BlockGuard class is used to switch to the given block in a program by
+ using the Python `with` keyword.
+ """
+
+ def __init__(self, block_idx=None, main_program=None):
+ self.main_program = fluid.default_main_program(
+ ) if main_program is None else main_program
+ self.old_block_idx = self.main_program.current_block().idx
+ self.new_block_idx = block_idx
+
+ def __enter__(self):
+ self.main_program.current_block_idx = self.new_block_idx
+
+ def __exit__(self, exc_type, exc_val, exc_tb):
+ self.main_program.current_block_idx = self.old_block_idx
+ if exc_type is not None:
+ return False # re-raise exception
+ return True
+
+ def layer_wrapper(*args, **kwargs):
+ with BlockGuard(block_idx):
+ return layer(*args, **kwargs)
+
+ return layer_wrapper
+
+
def position_encoding_init(n_position, d_pos_vec):
"""
Generate the initial values for the sinusoid position encoding table.
@@ -35,7 +72,9 @@ def multi_head_attention(queries,
d_model,
n_head=1,
dropout_rate=0.,
- cache=None):
+ cache=None,
+ gather_idx=None,
+ static_kv=False):
"""
Multi-Head Attention. Note that attn_bias is added to the logit before
computing softmax activiation to mask certain selected positions so that
@@ -56,42 +95,86 @@ def multi_head_attention(queries,
size=d_key * n_head,
bias_attr=False,
num_flatten_dims=2)
- k = layers.fc(input=keys,
- size=d_key * n_head,
- bias_attr=False,
- num_flatten_dims=2)
- v = layers.fc(input=values,
- size=d_value * n_head,
- bias_attr=False,
- num_flatten_dims=2)
+ # For encoder-decoder attention in inference, insert the ops and vars
+ # into global block to use as cache among beam search.
+ fc_layer = wrap_layer_with_block(
+ layers.fc, fluid.default_main_program().current_block()
+ .parent_idx) if cache is not None and static_kv else layers.fc
+ k = fc_layer(
+ input=keys,
+ size=d_key * n_head,
+ bias_attr=False,
+ num_flatten_dims=2)
+ v = fc_layer(
+ input=values,
+ size=d_value * n_head,
+ bias_attr=False,
+ num_flatten_dims=2)
return q, k, v
- def __split_heads(x, n_head):
+ def __split_heads_qkv(queries, keys, values, n_head, d_key, d_value):
"""
- Reshape the last dimension of inpunt tensor x so that it becomes two
- dimensions and then transpose. Specifically, input a tensor with shape
- [bs, max_sequence_length, n_head * hidden_dim] then output a tensor
+ Reshape input tensors at the last dimension to split multi-heads
+ and then transpose. Specifically, transform the input tensor with shape
+ [bs, max_sequence_length, n_head * hidden_dim] to the output tensor
with shape [bs, n_head, max_sequence_length, hidden_dim].
"""
- if n_head == 1:
- return x
-
- hidden_size = x.shape[-1]
# The value 0 in shape attr means copying the corresponding dimension
# size of the input as the output dimension size.
- reshaped = layers.reshape(
- x=x, shape=[0, 0, n_head, hidden_size // n_head], inplace=True)
-
+ reshaped_q = layers.reshape(
+ x=queries, shape=[0, 0, n_head, d_key], inplace=True)
# permuate the dimensions into:
# [batch_size, n_head, max_sequence_len, hidden_size_per_head]
- return layers.transpose(x=reshaped, perm=[0, 2, 1, 3])
+ q = layers.transpose(x=reshaped_q, perm=[0, 2, 1, 3])
+ # For encoder-decoder attention in inference, insert the ops and vars
+ # into global block to use as cache among beam search.
+ reshape_layer = wrap_layer_with_block(
+ layers.reshape,
+ fluid.default_main_program().current_block()
+ .parent_idx) if cache is not None and static_kv else layers.reshape
+ transpose_layer = wrap_layer_with_block(
+ layers.transpose,
+ fluid.default_main_program().current_block().
+ parent_idx) if cache is not None and static_kv else layers.transpose
+ reshaped_k = reshape_layer(
+ x=keys, shape=[0, 0, n_head, d_key], inplace=True)
+ k = transpose_layer(x=reshaped_k, perm=[0, 2, 1, 3])
+ reshaped_v = reshape_layer(
+ x=values, shape=[0, 0, n_head, d_value], inplace=True)
+ v = transpose_layer(x=reshaped_v, perm=[0, 2, 1, 3])
+
+ if cache is not None: # only for faster inference
+ if static_kv: # For encoder-decoder attention in inference
+ cache_k, cache_v = cache["static_k"], cache["static_v"]
+ # To init the static_k and static_v in cache.
+ # Maybe we can use condition_op(if_else) to do these at the first
+ # step in while loop to replace these, however it might be less
+ # efficient.
+ static_cache_init = wrap_layer_with_block(
+ layers.assign,
+ fluid.default_main_program().current_block().parent_idx)
+ static_cache_init(k, cache_k)
+ static_cache_init(v, cache_v)
+ else: # For decoder self-attention in inference
+ cache_k, cache_v = cache["k"], cache["v"]
+ # gather cell states corresponding to selected parent
+ select_k = layers.gather(cache_k, index=gather_idx)
+ select_v = layers.gather(cache_v, index=gather_idx)
+ if not static_kv:
+ # For self attention in inference, use cache and concat time steps.
+ select_k = layers.concat([select_k, k], axis=2)
+ select_v = layers.concat([select_v, v], axis=2)
+ # update cell states(caches) cached in global block
+ layers.assign(select_k, cache_k)
+ layers.assign(select_v, cache_v)
+ return q, select_k, select_v
+ return q, k, v
def __combine_heads(x):
"""
Transpose and then reshape the last two dimensions of inpunt tensor x
so that it becomes one dimension, which is reverse to __split_heads.
"""
- if len(x.shape) == 3: return x
if len(x.shape) != 4:
raise ValueError("Input(x) should be a 4-D Tensor.")
@@ -107,8 +190,7 @@ def multi_head_attention(queries,
"""
Scaled Dot-Product Attention
"""
- scaled_q = layers.scale(x=q, scale=d_key**-0.5)
- product = layers.matmul(x=scaled_q, y=k, transpose_y=True)
+ product = layers.matmul(x=q, y=k, transpose_y=True, alpha=d_key**-0.5)
if attn_bias:
product += attn_bias
weights = layers.softmax(product)
@@ -122,23 +204,7 @@ def multi_head_attention(queries,
return out
q, k, v = __compute_qkv(queries, keys, values, n_head, d_key, d_value)
-
- if cache is not None: # use cache and concat time steps
- # Since the inplace reshape in __split_heads changes the shape of k and
- # v, which is the cache input for next time step, reshape the cache
- # input from the previous time step first.
- k = cache["k"] = layers.concat(
- [layers.reshape(
- cache["k"], shape=[0, 0, d_key * n_head]), k],
- axis=1)
- v = cache["v"] = layers.concat(
- [layers.reshape(
- cache["v"], shape=[0, 0, d_value * n_head]), v],
- axis=1)
-
- q = __split_heads(q, n_head)
- k = __split_heads(k, n_head)
- v = __split_heads(v, n_head)
+ q, k, v = __split_heads_qkv(q, k, v, n_head, d_key, d_value)
ctx_multiheads = scaled_dot_product_attention(q, k, v, attn_bias, d_model,
dropout_rate)
@@ -327,7 +393,8 @@ def decoder_layer(dec_input,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
- cache=None):
+ cache=None,
+ gather_idx=None):
""" The layer to be stacked in decoder part.
The structure of this module is similar to that in the encoder part except
a multi-head attention is added to implement encoder-decoder attention.
@@ -342,7 +409,8 @@ def decoder_layer(dec_input,
d_model,
n_head,
attention_dropout,
- cache, )
+ cache=cache,
+ gather_idx=gather_idx)
slf_attn_output = post_process_layer(
dec_input,
slf_attn_output,
@@ -358,7 +426,10 @@ def decoder_layer(dec_input,
d_value,
d_model,
n_head,
- attention_dropout, )
+ attention_dropout,
+ cache=cache,
+ gather_idx=gather_idx,
+ static_kv=True)
enc_attn_output = post_process_layer(
slf_attn_output,
enc_attn_output,
@@ -393,7 +464,8 @@ def decoder(dec_input,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
- caches=None):
+ caches=None,
+ gather_idx=None):
"""
The decoder is composed of a stack of identical decoder_layer layers.
"""
@@ -413,7 +485,8 @@ def decoder(dec_input,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
- cache=None if caches is None else caches[i])
+ cache=None if caches is None else caches[i],
+ gather_idx=gather_idx)
dec_input = dec_output
dec_output = pre_process_layer(dec_output, preprocess_cmd,
prepostprocess_dropout)
@@ -610,7 +683,8 @@ def wrap_decoder(trg_vocab_size,
weight_sharing,
dec_inputs=None,
enc_output=None,
- caches=None):
+ caches=None,
+ gather_idx=None):
"""
The wrapper assembles together all needed layers for the decoder.
"""
@@ -646,7 +720,8 @@ def wrap_decoder(trg_vocab_size,
relu_dropout,
preprocess_cmd,
postprocess_cmd,
- caches=caches)
+ caches=caches,
+ gather_idx=gather_idx)
# Reshape to 2D tensor to use GEMM instead of BatchedGEMM
dec_output = layers.reshape(
dec_output, shape=[-1, dec_output.shape[-1]], inplace=True)
@@ -666,9 +741,43 @@ def wrap_decoder(trg_vocab_size,
return predict
-def fast_decode(
+def fast_decode(src_vocab_size,
+ trg_vocab_size,
+ max_in_len,
+ n_layer,
+ n_head,
+ d_key,
+ d_value,
+ d_model,
+ d_inner_hid,
+ prepostprocess_dropout,
+ attention_dropout,
+ relu_dropout,
+ preprocess_cmd,
+ postprocess_cmd,
+ weight_sharing,
+ beam_size,
+ max_out_len,
+ eos_idx,
+ use_py_reader=False):
+ """
+ Use beam search to decode. Caches will be used to store states of history
+ steps which can make the decoding faster.
+ """
+ data_input_names = encoder_data_input_fields + fast_decoder_data_input_fields
+
+ if use_py_reader:
+ all_inputs, reader = make_all_py_reader_inputs(data_input_names)
+ else:
+ all_inputs = make_all_inputs(data_input_names)
+
+ enc_inputs_len = len(encoder_data_input_fields)
+ dec_inputs_len = len(fast_decoder_data_input_fields)
+ enc_inputs = all_inputs[0:enc_inputs_len]
+ dec_inputs = all_inputs[enc_inputs_len:enc_inputs_len + dec_inputs_len]
+
+ enc_output = wrap_encoder(
src_vocab_size,
- trg_vocab_size,
max_in_len,
n_layer,
n_head,
@@ -682,64 +791,60 @@ def fast_decode(
preprocess_cmd,
postprocess_cmd,
weight_sharing,
- beam_size,
- max_out_len,
- eos_idx, ):
- """
- Use beam search to decode. Caches will be used to store states of history
- steps which can make the decoding faster.
- """
- enc_output = wrap_encoder(
- src_vocab_size, max_in_len, n_layer, n_head, d_key, d_value, d_model,
- d_inner_hid, prepostprocess_dropout, attention_dropout, relu_dropout,
- preprocess_cmd, postprocess_cmd, weight_sharing)
- start_tokens, init_scores, trg_src_attn_bias = make_all_inputs(
- fast_decoder_data_input_fields)
+ enc_inputs, )
+ start_tokens, init_scores, parent_idx, trg_src_attn_bias = dec_inputs
def beam_search():
max_len = layers.fill_constant(
- shape=[1], dtype=start_tokens.dtype, value=max_out_len)
+ shape=[1],
+ dtype=start_tokens.dtype,
+ value=max_out_len,
+ force_cpu=True)
step_idx = layers.fill_constant(
- shape=[1], dtype=start_tokens.dtype, value=0)
- cond = layers.less_than(x=step_idx, y=max_len)
+ shape=[1], dtype=start_tokens.dtype, value=0, force_cpu=True)
+ cond = layers.less_than(x=step_idx, y=max_len) # default force_cpu=True
while_op = layers.While(cond)
# array states will be stored for each step.
ids = layers.array_write(
layers.reshape(start_tokens, (-1, 1)), step_idx)
scores = layers.array_write(init_scores, step_idx)
# cell states will be overwrited at each step.
- # caches contains states of history steps to reduce redundant
- # computation in decoder.
- caches = [{
- "k": layers.fill_constant_batch_size_like(
- input=start_tokens,
- shape=[-1, 0, d_model],
- dtype=enc_output.dtype,
- value=0),
- "v": layers.fill_constant_batch_size_like(
- input=start_tokens,
- shape=[-1, 0, d_model],
- dtype=enc_output.dtype,
- value=0)
- } for i in range(n_layer)]
+ # caches contains states of history steps in decoder self-attention
+ # and static encoder output projections in encoder-decoder attention
+ # to reduce redundant computation.
+ caches = [
+ {
+ "k": # for self attention
+ layers.fill_constant_batch_size_like(
+ input=start_tokens,
+ shape=[-1, n_head, 0, d_key],
+ dtype=enc_output.dtype,
+ value=0),
+ "v": # for self attention
+ layers.fill_constant_batch_size_like(
+ input=start_tokens,
+ shape=[-1, n_head, 0, d_value],
+ dtype=enc_output.dtype,
+ value=0),
+ "static_k": # for encoder-decoder attention
+ layers.create_tensor(dtype=enc_output.dtype),
+ "static_v": # for encoder-decoder attention
+ layers.create_tensor(dtype=enc_output.dtype)
+ } for i in range(n_layer)
+ ]
+
with while_op.block():
pre_ids = layers.array_read(array=ids, i=step_idx)
- pre_ids = layers.reshape(pre_ids, (-1, 1, 1))
+ # Since beam_search_op dosen't enforce pre_ids' shape, we can do
+ # inplace reshape here which actually change the shape of pre_ids.
+ pre_ids = layers.reshape(pre_ids, (-1, 1, 1), inplace=True)
pre_scores = layers.array_read(array=scores, i=step_idx)
- # sequence_expand can gather sequences according to lod thus can be
- # used in beam search to sift states corresponding to selected ids.
- pre_src_attn_bias = layers.sequence_expand(
- x=trg_src_attn_bias, y=pre_scores)
- pre_enc_output = layers.sequence_expand(x=enc_output, y=pre_scores)
- pre_caches = [{
- "k": layers.sequence_expand(
- x=cache["k"], y=pre_scores),
- "v": layers.sequence_expand(
- x=cache["v"], y=pre_scores),
- } for cache in caches]
+ # gather cell states corresponding to selected parent
+ pre_src_attn_bias = layers.gather(
+ trg_src_attn_bias, index=parent_idx)
pre_pos = layers.elementwise_mul(
x=layers.fill_constant_batch_size_like(
- input=pre_enc_output, # cann't use pre_ids here since it has lod
+ input=pre_src_attn_bias, # cann't use lod tensor here
value=1,
shape=[-1, 1, 1],
dtype=pre_ids.dtype),
@@ -761,35 +866,33 @@ def fast_decode(
postprocess_cmd,
weight_sharing,
dec_inputs=(pre_ids, pre_pos, None, pre_src_attn_bias),
- enc_output=pre_enc_output,
- caches=pre_caches)
-
+ enc_output=enc_output,
+ caches=caches,
+ gather_idx=parent_idx)
+ # intra-beam topK
topk_scores, topk_indices = layers.topk(
input=layers.softmax(logits), k=beam_size)
accu_scores = layers.elementwise_add(
- x=layers.log(topk_scores),
- y=layers.reshape(
- pre_scores, shape=[-1]),
- axis=0)
- # beam_search op uses lod to distinguish branches.
+ x=layers.log(topk_scores), y=pre_scores, axis=0)
+ # beam_search op uses lod to differentiate branches.
topk_indices = layers.lod_reset(topk_indices, pre_ids)
- selected_ids, selected_scores = layers.beam_search(
+ # topK reduction across beams, also contain special handle of
+ # end beams and end sentences(batch reduction)
+ selected_ids, selected_scores, gather_idx = layers.beam_search(
pre_ids=pre_ids,
pre_scores=pre_scores,
ids=topk_indices,
scores=accu_scores,
beam_size=beam_size,
- end_id=eos_idx)
-
+ end_id=eos_idx,
+ return_parent_idx=True)
layers.increment(x=step_idx, value=1.0, in_place=True)
- # update states
+ # cell states(caches) have been updated in wrap_decoder,
+ # only need to update beam search states here.
layers.array_write(selected_ids, i=step_idx, array=ids)
layers.array_write(selected_scores, i=step_idx, array=scores)
+ layers.assign(gather_idx, parent_idx)
layers.assign(pre_src_attn_bias, trg_src_attn_bias)
- layers.assign(pre_enc_output, enc_output)
- for i in range(n_layer):
- layers.assign(pre_caches[i]["k"], caches[i]["k"])
- layers.assign(pre_caches[i]["v"], caches[i]["v"])
length_cond = layers.less_than(x=step_idx, y=max_len)
finish_cond = layers.logical_not(layers.is_empty(x=selected_ids))
layers.logical_and(x=length_cond, y=finish_cond, out=cond)
@@ -799,4 +902,4 @@ def fast_decode(
return finished_ids, finished_scores
finished_ids, finished_scores = beam_search()
- return finished_ids, finished_scores
+ return finished_ids, finished_scores, reader if use_py_reader else None
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/profile.py b/fluid/PaddleNLP/neural_machine_translation/transformer/profile.py
index 9a437725cb27c29b0233d6297e84781f5343aff1..76711ece132113863f1e42d4ac1529f63ed90ff3 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/profile.py
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/profile.py
@@ -186,7 +186,7 @@ def main(args):
# Since the token number differs among devices, customize gradient scale to
# use token average cost among multi-devices. and the gradient scale is
# `1 / token_number` for average cost.
- build_strategy.gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.Customized
+ # build_strategy.gradient_scale_strategy = fluid.BuildStrategy.GradientScaleStrategy.Customized
train_exe = fluid.ParallelExecutor(
use_cuda=TrainTaskConfig.use_gpu,
loss_name=avg_cost.name,
diff --git a/fluid/PaddleNLP/neural_machine_translation/transformer/train.py b/fluid/PaddleNLP/neural_machine_translation/transformer/train.py
index 16d48238941a03309cc9ba269cd619bd21e0f561..4313f8b441ee194935c7c47abc52271589c7765d 100644
--- a/fluid/PaddleNLP/neural_machine_translation/transformer/train.py
+++ b/fluid/PaddleNLP/neural_machine_translation/transformer/train.py
@@ -10,7 +10,6 @@ import time
import numpy as np
import paddle.fluid as fluid
-from paddle.fluid.transpiler.details import program_to_code
import reader
from config import *
@@ -258,7 +257,12 @@ def prepare_batch_input(insts, data_input_names, src_pad_idx, trg_pad_idx,
return data_input_dict, np.asarray([num_token], dtype="float32")
-def prepare_data_generator(args, is_test, count, pyreader):
+def prepare_data_generator(args,
+ is_test,
+ count,
+ pyreader,
+ py_reader_provider_wrapper,
+ place=None):
"""
Data generator wrapper for DataReader. If use py_reader, set the data
provider for py_reader
@@ -319,7 +323,7 @@ def prepare_data_generator(args, is_test, count, pyreader):
data_reader = split(data_reader, count)
if args.use_py_reader:
pyreader.decorate_tensor_provider(
- py_reader_provider_wrapper(data_reader))
+ py_reader_provider_wrapper(data_reader, place))
data_reader = None
else: # Data generator for multi-devices
data_reader = stack(data_reader, count)
@@ -357,7 +361,7 @@ def prepare_feed_dict_list(data_generator, init_flag, count):
return feed_dict_list if len(feed_dict_list) == count else None
-def py_reader_provider_wrapper(data_reader):
+def py_reader_provider_wrapper(data_reader, place):
"""
Data provider needed by fluid.layers.py_reader.
"""
@@ -370,8 +374,7 @@ def py_reader_provider_wrapper(data_reader):
data, data_input_names, ModelHyperParams.eos_idx,
ModelHyperParams.eos_idx, ModelHyperParams.n_head,
ModelHyperParams.d_model)
- total_dict = dict(data_input_dict.items())
- yield [total_dict[item] for item in data_input_names]
+ yield [data_input_dict[item] for item in data_input_names]
return py_reader_provider
@@ -406,7 +409,11 @@ def test_context(exe, train_exe, dev_count):
is_test=True)
test_prog = test_prog.clone(for_test=True)
test_data = prepare_data_generator(
- args, is_test=True, count=dev_count, pyreader=pyreader)
+ args,
+ is_test=True,
+ count=dev_count,
+ pyreader=pyreader,
+ py_reader_provider_wrapper=py_reader_provider_wrapper)
exe.run(startup_prog) # to init pyreader for testing
if TrainTaskConfig.ckpt_path:
@@ -477,7 +484,11 @@ def train_loop(exe,
logging.info("begin reader")
train_data = prepare_data_generator(
- args, is_test=False, count=dev_count, pyreader=pyreader)
+ args,
+ is_test=False,
+ count=dev_count,
+ pyreader=pyreader,
+ py_reader_provider_wrapper=py_reader_provider_wrapper)
# For faster executor
exec_strategy = fluid.ExecutionStrategy()
diff --git a/fluid/PaddleNLP/sequence_tagging_for_ner/train.py b/fluid/PaddleNLP/sequence_tagging_for_ner/train.py
index b77c081ba38015e1829fcc6c633e7fbaa4376bb1..68e621371e09b654007134c8ce449e3491b9516f 100644
--- a/fluid/PaddleNLP/sequence_tagging_for_ner/train.py
+++ b/fluid/PaddleNLP/sequence_tagging_for_ner/train.py
@@ -136,8 +136,9 @@ def main(train_data_file,
" pass_f1_score:" + str(test_pass_f1_score))
save_dirname = os.path.join(model_save_dir, "params_pass_%d" % pass_id)
- fluid.io.save_inference_model(save_dirname, ['word', 'mark'],
- crf_decode, exe)
+ if "CE_MODE_X" not in os.environ:
+ fluid.io.save_inference_model(save_dirname, ['word', 'mark'],
+ crf_decode, exe)
if "CE_MODE_X" in os.environ:
print("kpis train_precision %f" % pass_precision)
diff --git a/fluid/PaddleNLP/text_classification/async_executor/data_generator/build_raw_data.py b/fluid/PaddleNLP/text_classification/async_executor/data_generator/build_raw_data.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c0c0981c93b3b1e9231c7efe1f0b49e178c060f
--- /dev/null
+++ b/fluid/PaddleNLP/text_classification/async_executor/data_generator/build_raw_data.py
@@ -0,0 +1,62 @@
+# Copyright (c) 2018 PaddlePaddle 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.
+"""
+Build raw data
+"""
+from __future__ import print_function
+import sys
+import os
+import random
+import re
+data_type = sys.argv[1]
+
+if not (data_type == "train" or data_type == "test"):
+ print("python %s [test/train]" % sys.argv[0], file=sys.stderr)
+ sys.exit(-1)
+
+pos_folder = "aclImdb/" + data_type + "/pos/"
+neg_folder = "aclImdb/" + data_type + "/neg/"
+
+pos_train_list = [(pos_folder + x, "1") for x in os.listdir(pos_folder)]
+neg_train_list = [(neg_folder + x, "0") for x in os.listdir(neg_folder)]
+
+all_train_list = pos_train_list + neg_train_list
+random.shuffle(all_train_list)
+
+
+def load_dict(dictfile):
+ """
+ Load word id dict
+ """
+ vocab = {}
+ wid = 0
+ with open(dictfile) as f:
+ for line in f:
+ vocab[line.strip()] = str(wid)
+ wid += 1
+ return vocab
+
+
+vocab = load_dict("aclImdb/imdb.vocab")
+unk_id = str(len(vocab))
+print("vocab size: ", len(vocab), file=sys.stderr)
+pattern = re.compile(r'(;|,|\.|\?|!|\s|\(|\))')
+
+for fitem in all_train_list:
+ label = str(fitem[1])
+ fname = fitem[0]
+ with open(fname) as f:
+ sent = f.readline().lower().replace("
", " ").strip()
+ out_s = "%s | %s" % (sent, label)
+ print(out_s, file=sys.stdout)
diff --git a/fluid/PaddleRec/ctr/.run_ce.sh b/fluid/PaddleRec/ctr/.run_ce.sh
new file mode 100755
index 0000000000000000000000000000000000000000..fc1e1303a06f96086eea679142770e3aefac3ff3
--- /dev/null
+++ b/fluid/PaddleRec/ctr/.run_ce.sh
@@ -0,0 +1,22 @@
+#!/bin/bash
+
+export MKL_NUM_THREADS=1
+export OMP_NUM_THREADS=1
+
+
+#cudaid=${face_detection:=0} # use 0-th card as default
+#export CUDA_VISIBLE_DEVICES=$cudaid
+export CPU_NUM=1
+export NUM_THREADS=1
+
+FLAGS_benchmark=true python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
+
+export CPU_NUM=1
+export NUM_THREADS=8
+
+FLAGS_benchmark=true python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
+
+export CPU_NUM=8
+export NUM_THREADS=8
+
+FLAGS_benchmark=true python train.py --is_local 1 --cloud_train 0 --train_data_path data/raw/train.txt --enable_ce | python _ce.py
diff --git a/fluid/PaddleRec/ctr/_ce.py b/fluid/PaddleRec/ctr/_ce.py
new file mode 100644
index 0000000000000000000000000000000000000000..91867d036a050f03b7c685ba73c3051eca97c9aa
--- /dev/null
+++ b/fluid/PaddleRec/ctr/_ce.py
@@ -0,0 +1,78 @@
+# this file is only used for continuous evaluation test!
+
+import os
+import sys
+sys.path.append(os.environ['ceroot'])
+from kpi import CostKpi
+from kpi import DurationKpi
+from kpi import AccKpi
+
+
+each_pass_duration_cpu1_thread1_kpi = DurationKpi('each_pass_duration_cpu1_thread1', 0.08, 0, actived=True)
+train_loss_cpu1_thread1_kpi = CostKpi('train_loss_cpu1_thread1', 0.08, 0)
+train_auc_val_cpu1_thread1_kpi = AccKpi('train_auc_val_cpu1_thread1', 0.08, 0)
+train_batch_auc_val_cpu1_thread1_kpi = AccKpi('train_batch_auc_val_cpu1_thread1', 0.08, 0)
+each_pass_duration_cpu1_thread8_kpi = DurationKpi('each_pass_duration_cpu1_thread8', 0.08, 0, actived=True)
+train_loss_cpu1_thread8_kpi = CostKpi('train_loss_cpu1_thread8', 0.08, 0)
+train_auc_val_cpu1_thread8_kpi = AccKpi('train_auc_val_cpu1_thread8', 0.08, 0)
+train_batch_auc_val_cpu1_thread8_kpi = AccKpi('train_batch_auc_val_cpu1_thread8', 0.08, 0)
+each_pass_duration_cpu8_thread8_kpi = DurationKpi('each_pass_duration_cpu8_thread8', 0.08, 0, actived=True)
+train_loss_cpu8_thread8_kpi = CostKpi('train_loss_cpu8_thread8', 0.08, 0)
+train_auc_val_cpu8_thread8_kpi = AccKpi('train_auc_val_cpu8_thread8', 0.08, 0)
+train_batch_auc_val_cpu8_thread8_kpi = AccKpi('train_batch_auc_val_cpu8_thread8', 0.08, 0)
+
+tracking_kpis = [
+ each_pass_duration_cpu1_thread1_kpi,
+ train_loss_cpu1_thread1_kpi,
+ train_auc_val_cpu1_thread1_kpi,
+ train_batch_auc_val_cpu1_thread1_kpi,
+ each_pass_duration_cpu1_thread8_kpi,
+ train_loss_cpu1_thread8_kpi,
+ train_auc_val_cpu1_thread8_kpi,
+ train_batch_auc_val_cpu1_thread8_kpi,
+ each_pass_duration_cpu8_thread8_kpi,
+ train_loss_cpu8_thread8_kpi,
+ train_auc_val_cpu8_thread8_kpi,
+ train_batch_auc_val_cpu8_thread8_kpi,
+ ]
+
+
+def parse_log(log):
+ '''
+ This method should be implemented by model developers.
+
+ The suggestion:
+
+ each line in the log should be key, value, for example:
+
+ "
+ train_cost\t1.0
+ test_cost\t1.0
+ train_cost\t1.0
+ train_cost\t1.0
+ train_acc\t1.2
+ "
+ '''
+ for line in log.split('\n'):
+ fs = line.strip().split('\t')
+ print(fs)
+ if len(fs) == 3 and fs[0] == 'kpis':
+ kpi_name = fs[1]
+ kpi_value = float(fs[2])
+ yield kpi_name, kpi_value
+
+
+def log_to_ce(log):
+ kpi_tracker = {}
+ for kpi in tracking_kpis:
+ kpi_tracker[kpi.name] = kpi
+
+ for (kpi_name, kpi_value) in parse_log(log):
+ print(kpi_name, kpi_value)
+ kpi_tracker[kpi_name].add_record(kpi_value)
+ kpi_tracker[kpi_name].persist()
+
+
+if __name__ == '__main__':
+ log = sys.stdin.read()
+ log_to_ce(log)
diff --git a/fluid/PaddleRec/ctr/infer.py b/fluid/PaddleRec/ctr/infer.py
index 24f4f7bf9be694e7cb3442632c4945dd63a17c4e..19f7013b305e3900fe05b7575cfc20c331ca5daf 100644
--- a/fluid/PaddleRec/ctr/infer.py
+++ b/fluid/PaddleRec/ctr/infer.py
@@ -61,14 +61,14 @@ def infer():
startup_program = fluid.framework.Program()
test_program = fluid.framework.Program()
with fluid.framework.program_guard(test_program, startup_program):
- loss, data_list, auc_var, batch_auc_var = ctr_dnn_model(args.embedding_size, args.sparse_feature_dim)
+ loss, auc_var, batch_auc_var, _, data_list = ctr_dnn_model(args.embedding_size, args.sparse_feature_dim, False)
- exe = fluid.Executor(place)
-
- feeder = fluid.DataFeeder(feed_list=data_list, place=place)
-
- with fluid.scope_guard(inference_scope):
- [inference_program, _, fetch_targets] = fluid.io.load_inference_model(args.model_path, exe)
+ exe = fluid.Executor(place)
+
+ feeder = fluid.DataFeeder(feed_list=data_list, place=place)
+
+ fluid.io.load_persistables(executor=exe, dirname=args.model_path,
+ main_program=fluid.default_main_program())
def set_zero(var_name):
param = inference_scope.var(var_name).get_tensor()
@@ -80,9 +80,9 @@ def infer():
set_zero(name)
for batch_id, data in enumerate(test_reader()):
- loss_val, auc_val = exe.run(inference_program,
+ loss_val, auc_val = exe.run(test_program,
feed=feeder.feed(data),
- fetch_list=fetch_targets)
+ fetch_list=[loss, auc_var])
if batch_id % 100 == 0:
logger.info("TEST --> batch: {} loss: {} auc: {}".format(batch_id, loss_val/args.batch_size, auc_val))
diff --git a/fluid/PaddleRec/ctr/network_conf.py b/fluid/PaddleRec/ctr/network_conf.py
index 54dd855928d0a25dff99c0b5165c1b1343732138..569e84f46b327e84f70c622f06e99ddba031b4f6 100644
--- a/fluid/PaddleRec/ctr/network_conf.py
+++ b/fluid/PaddleRec/ctr/network_conf.py
@@ -104,7 +104,7 @@ def ctr_deepfm_model(factor_size, sparse_feature_dim, dense_feature_dim, sparse_
return avg_cost, auc_var, batch_auc_var, py_reader
-def ctr_dnn_model(embedding_size, sparse_feature_dim):
+def ctr_dnn_model(embedding_size, sparse_feature_dim, use_py_reader=True):
def embedding_layer(input):
return fluid.layers.embedding(
@@ -126,13 +126,15 @@ def ctr_dnn_model(embedding_size, sparse_feature_dim):
label = fluid.layers.data(name='label', shape=[1], dtype='int64')
- datas = [dense_input] + sparse_input_ids + [label]
+ words = [dense_input] + sparse_input_ids + [label]
- py_reader = fluid.layers.create_py_reader_by_data(capacity=64,
- feed_list=datas,
- name='py_reader',
- use_double_buffer=True)
- words = fluid.layers.read_file(py_reader)
+ py_reader = None
+ if use_py_reader:
+ py_reader = fluid.layers.create_py_reader_by_data(capacity=64,
+ feed_list=words,
+ name='py_reader',
+ use_double_buffer=True)
+ words = fluid.layers.read_file(py_reader)
sparse_embed_seq = list(map(embedding_layer, words[1:-1]))
concated = fluid.layers.concat(sparse_embed_seq + words[0:1], axis=1)
@@ -156,4 +158,4 @@ def ctr_dnn_model(embedding_size, sparse_feature_dim):
auc_var, batch_auc_var, auc_states = \
fluid.layers.auc(input=predict, label=words[-1], num_thresholds=2 ** 12, slide_steps=20)
- return avg_cost, auc_var, batch_auc_var, py_reader
+ return avg_cost, auc_var, batch_auc_var, py_reader, words
diff --git a/fluid/PaddleRec/ctr/reader.py b/fluid/PaddleRec/ctr/reader.py
index affd1d0e722ab957bd990b01702d0787c84470bd..e6bcc11d4f465d48f99ef2aa8059a952a3f8dfd4 100644
--- a/fluid/PaddleRec/ctr/reader.py
+++ b/fluid/PaddleRec/ctr/reader.py
@@ -46,7 +46,7 @@ class CriteoDataset(Dataset):
return self._reader_creator(file_list, True, trainer_num, trainer_id)
def test(self, file_list):
- return self._reader_creator(file_list, False, -1)
+ return self._reader_creator(file_list, False, 1, 0)
def infer(self, file_list):
- return self._reader_creator(file_list, False, -1)
+ return self._reader_creator(file_list, False, 1, 0)
diff --git a/fluid/PaddleRec/ctr/train.py b/fluid/PaddleRec/ctr/train.py
index 29581528be940f2e4ca8ba01f74da32a32a00024..93ce357b45e9bab4a0d46d43cfaf7573ddb31adc 100644
--- a/fluid/PaddleRec/ctr/train.py
+++ b/fluid/PaddleRec/ctr/train.py
@@ -107,12 +107,22 @@ def parse_args():
type=int,
default=1,
help='The num of trianers, (default: 1)')
+ parser.add_argument(
+ '--enable_ce',
+ action='store_true',
+ help='If set, run the task with continuous evaluation logs.')
return parser.parse_args()
def train_loop(args, train_program, py_reader, loss, auc_var, batch_auc_var,
trainer_num, trainer_id):
+
+ if args.enable_ce:
+ SEED = 102
+ train_program.random_seed = SEED
+ fluid.default_startup_program().random_seed = SEED
+
dataset = reader.CriteoDataset(args.sparse_feature_dim)
train_reader = paddle.batch(
paddle.reader.shuffle(
@@ -146,6 +156,7 @@ def train_loop(args, train_program, py_reader, loss, auc_var, batch_auc_var,
exe.run(fluid.default_startup_program())
+ total_time = 0
for pass_id in range(args.num_passes):
pass_start = time.time()
batch_id = 0
@@ -163,16 +174,33 @@ def train_loop(args, train_program, py_reader, loss, auc_var, batch_auc_var,
if batch_id % 1000 == 0 and batch_id != 0:
model_dir = args.model_output_dir + '/batch-' + str(batch_id)
if args.trainer_id == 0:
- fluid.io.save_inference_model(model_dir, data_name_list, [loss, auc_var], exe)
+ fluid.io.save_persistables(executor=exe, dirname=model_dir,
+ main_program=fluid.default_main_program())
batch_id += 1
except fluid.core.EOFException:
py_reader.reset()
print("pass_id: %d, pass_time_cost: %f" % (pass_id, time.time() - pass_start))
+ total_time += time.time() - pass_start
+
model_dir = args.model_output_dir + '/pass-' + str(pass_id)
if args.trainer_id == 0:
- fluid.io.save_inference_model(model_dir, data_name_list, [loss, auc_var], exe)
+ fluid.io.save_persistables(executor=exe, dirname=model_dir,
+ main_program=fluid.default_main_program())
+ # only for ce
+ if args.enable_ce:
+ threads_num, cpu_num = get_cards(args)
+ epoch_idx = args.num_passes
+ print("kpis\teach_pass_duration_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, total_time / epoch_idx))
+ print("kpis\ttrain_loss_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, loss_val/args.batch_size))
+ print("kpis\ttrain_auc_val_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, auc_val))
+ print("kpis\ttrain_batch_auc_val_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, batch_auc_val))
+
def train():
args = parse_args()
@@ -180,7 +208,7 @@ def train():
if not os.path.isdir(args.model_output_dir):
os.mkdir(args.model_output_dir)
- loss, auc_var, batch_auc_var, py_reader = ctr_dnn_model(args.embedding_size, args.sparse_feature_dim)
+ loss, auc_var, batch_auc_var, py_reader, _ = ctr_dnn_model(args.embedding_size, args.sparse_feature_dim)
optimizer = fluid.optimizer.Adam(learning_rate=1e-4)
optimizer.minimize(loss)
if args.cloud_train:
@@ -224,5 +252,11 @@ def train():
)
+def get_cards(args):
+ threads_num = os.environ.get('NUM_THREADS', 1)
+ cpu_num = os.environ.get('CPU_NUM', 1)
+ return int(threads_num), int(cpu_num)
+
+
if __name__ == '__main__':
train()
diff --git a/fluid/PaddleRec/gru4rec/README.md b/fluid/PaddleRec/gru4rec/README.md
index 0ea3f838eaf9e2f46b7d1551a36aa1f6b462ce44..9c4a5247fe2ecb64e79ba96e0922f0fd1750aa8e 100644
--- a/fluid/PaddleRec/gru4rec/README.md
+++ b/fluid/PaddleRec/gru4rec/README.md
@@ -79,7 +79,7 @@ SessionId ItemId Time
2 214757407 1396850438.247
```
-数据格式需要转换 运行脚本
+数据格式需要转换, 运行脚本如下
```
python convert_format.py
```
@@ -101,7 +101,7 @@ python convert_format.py
根据训练和测试文件生成字典和对应的paddle输入文件
-注意需要将训练文件放到一个目录下面，测试文件放到一个目录下面,同时支持多训练文件
+需要将训练文件放到目录raw_train_data下，测试文件放到目录raw_test_data下，并生成对应的train_data，test_data和vocab.txt文件
```
python text2paddle.py raw_train_data/ raw_test_data/ train_data test_data vocab.txt
```
diff --git a/fluid/PaddleRec/gru4rec/cluster_train.py b/fluid/PaddleRec/gru4rec/cluster_train.py
index b9b0820d0293700f379abf135e8d149311a7e1a1..f50542bf011d0caacddb3831368493df106463f5 100644
--- a/fluid/PaddleRec/gru4rec/cluster_train.py
+++ b/fluid/PaddleRec/gru4rec/cluster_train.py
@@ -13,22 +13,26 @@ import net
SEED = 102
+
def parse_args():
parser = argparse.ArgumentParser("gru4rec benchmark.")
parser.add_argument(
- '--train_dir', type=str, default='train_data', help='train file address')
- parser.add_argument(
- '--vocab_path', type=str, default='vocab.txt', help='vocab file address')
- parser.add_argument(
- '--is_local', type=int, default=1, help='whether local')
+ '--train_dir',
+ type=str,
+ default='train_data',
+ help='train file address')
parser.add_argument(
- '--hid_size', type=int, default=100, help='hid size')
+ '--vocab_path',
+ type=str,
+ default='vocab.txt',
+ help='vocab file address')
+ parser.add_argument('--is_local', type=int, default=1, help='whether local')
+ parser.add_argument('--hid_size', type=int, default=100, help='hid size')
parser.add_argument(
'--model_dir', type=str, default='model_recall20', help='model dir')
parser.add_argument(
'--batch_size', type=int, default=5, help='num of batch size')
- parser.add_argument(
- '--pass_num', type=int, default=10, help='num of epoch')
+ parser.add_argument('--pass_num', type=int, default=10, help='num of epoch')
parser.add_argument(
'--print_batch', type=int, default=10, help='num of print batch')
parser.add_argument(
@@ -40,19 +44,33 @@ def parse_args():
parser.add_argument(
'--role', type=str, default='pserver', help='trainer or pserver')
parser.add_argument(
- '--endpoints', type=str, default='127.0.0.1:6000', help='The pserver endpoints, like: 127.0.0.1:6000, 127.0.0.1:6001')
- parser.add_argument(
- '--current_endpoint', type=str, default='127.0.0.1:6000', help='The current_endpoint')
- parser.add_argument(
- '--trainer_id', type=int, default=0, help='trainer id ,only trainer_id=0 save model')
- parser.add_argument(
- '--trainers', type=int, default=1, help='The num of trianers, (default: 1)')
+ '--endpoints',
+ type=str,
+ default='127.0.0.1:6000',
+ help='The pserver endpoints, like: 127.0.0.1:6000, 127.0.0.1:6001')
+ parser.add_argument(
+ '--current_endpoint',
+ type=str,
+ default='127.0.0.1:6000',
+ help='The current_endpoint')
+ parser.add_argument(
+ '--trainer_id',
+ type=int,
+ default=0,
+ help='trainer id ,only trainer_id=0 save model')
+ parser.add_argument(
+ '--trainers',
+ type=int,
+ default=1,
+ help='The num of trianers, (default: 1)')
args = parser.parse_args()
return args
+
def get_cards(args):
return args.num_devices
+
def train():
""" do training """
args = parse_args()
@@ -67,12 +85,13 @@ def train():
buffer_size=1000, word_freq_threshold=0, is_train=True)
# Train program
- src_wordseq, dst_wordseq, avg_cost, acc = net.network(vocab_size=vocab_size, hid_size=hid_size)
+ src_wordseq, dst_wordseq, avg_cost, acc = net.all_vocab_network(
+ vocab_size=vocab_size, hid_size=hid_size)
# Optimization to minimize lost
sgd_optimizer = fluid.optimizer.SGD(learning_rate=args.base_lr)
sgd_optimizer.minimize(avg_cost)
-
+
def train_loop(main_program):
""" train network """
pass_num = args.pass_num
@@ -97,9 +116,11 @@ def train():
lod_dst_wordseq = utils.to_lodtensor([dat[1] for dat in data],
place)
ret_avg_cost = exe.run(main_program,
- feed={ "src_wordseq": lod_src_wordseq,
- "dst_wordseq": lod_dst_wordseq},
- fetch_list=fetch_list)
+ feed={
+ "src_wordseq": lod_src_wordseq,
+ "dst_wordseq": lod_dst_wordseq
+ },
+ fetch_list=fetch_list)
avg_ppl = np.exp(ret_avg_cost[0])
newest_ppl = np.mean(avg_ppl)
if i % args.print_batch == 0:
@@ -113,7 +134,8 @@ def train():
feed_var_names = ["src_wordseq", "dst_wordseq"]
fetch_vars = [avg_cost, acc]
if args.trainer_id == 0:
- fluid.io.save_inference_model(save_dir, feed_var_names, fetch_vars, exe)
+ fluid.io.save_inference_model(save_dir, feed_var_names,
+ fetch_vars, exe)
print("model saved in %s" % save_dir)
print("finish training")
@@ -123,7 +145,8 @@ def train():
else:
print("run distribute training")
t = fluid.DistributeTranspiler()
- t.transpile(args.trainer_id, pservers=args.endpoints, trainers=args.trainers)
+ t.transpile(
+ args.trainer_id, pservers=args.endpoints, trainers=args.trainers)
if args.role == "pserver":
print("run psever")
pserver_prog = t.get_pserver_program(args.current_endpoint)
@@ -136,5 +159,6 @@ def train():
print("run trainer")
train_loop(t.get_trainer_program())
+
if __name__ == "__main__":
train()
diff --git a/fluid/PaddleRec/gru4rec/infer.py b/fluid/PaddleRec/gru4rec/infer.py
index a39eb5db1607e661923ed08fba93e0606e9a1750..a5da585ac23e6f65213b358c9d16ed10fb091b05 100644
--- a/fluid/PaddleRec/gru4rec/infer.py
+++ b/fluid/PaddleRec/gru4rec/infer.py
@@ -11,23 +11,27 @@ import paddle
import utils
+
def parse_args():
parser = argparse.ArgumentParser("gru4rec benchmark.")
parser.add_argument(
'--test_dir', type=str, default='test_data', help='test file address')
parser.add_argument(
- '--start_index', type=int, default='1', help='start index')
+ '--start_index', type=int, default='1', help='start index')
+ parser.add_argument(
+ '--last_index', type=int, default='10', help='end index')
parser.add_argument(
- '--last_index', type=int, default='10', help='end index')
+ '--model_dir', type=str, default='model_recall20', help='model dir')
parser.add_argument(
- '--model_dir', type=str, default='model_recall20', help='model dir')
+ '--use_cuda', type=int, default='0', help='whether use cuda')
parser.add_argument(
- '--use_cuda', type=int, default='1', help='whether use cuda')
+ '--batch_size', type=int, default='5', help='batch_size')
parser.add_argument(
- '--batch_size', type=int, default='5', help='batch_size')
+ '--vocab_path', type=str, default='vocab.txt', help='vocab file')
args = parser.parse_args()
return args
+
def infer(test_reader, use_cuda, model_path):
""" inference function """
place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
@@ -72,11 +76,16 @@ if __name__ == "__main__":
test_dir = args.test_dir
model_dir = args.model_dir
batch_size = args.batch_size
+ vocab_path = args.vocab_path
use_cuda = True if args.use_cuda else False
- print("start index: ", start_index, " last_index:" ,last_index)
+ print("start index: ", start_index, " last_index:", last_index)
vocab_size, test_reader = utils.prepare_data(
- test_dir, "", batch_size=batch_size,
- buffer_size=1000, word_freq_threshold=0, is_train=False)
+ test_dir,
+ vocab_path,
+ batch_size=batch_size,
+ buffer_size=1000,
+ word_freq_threshold=0,
+ is_train=False)
for epoch in range(start_index, last_index + 1):
epoch_path = model_dir + "/epoch_" + str(epoch)
diff --git a/fluid/PaddleRec/gru4rec/net.py b/fluid/PaddleRec/gru4rec/net.py
index ebb512377eae865b90f3d0360931a744b1a0ad07..6a715443ff1e72ae77aba51d5eaffe4eefee9687 100644
--- a/fluid/PaddleRec/gru4rec/net.py
+++ b/fluid/PaddleRec/gru4rec/net.py
@@ -171,7 +171,8 @@ def train_cross_entropy_network(vocab_size, neg_size, hid_size, drop_out=0.2):
ele_mul = fluid.layers.elementwise_mul(emb_label_drop, gru)
red_sum = fluid.layers.reduce_sum(input=ele_mul, dim=1, keep_dim=True)
- pre = fluid.layers.sequence_reshape(input=red_sum, new_dim=(neg_size + 1))
+ pre_ = fluid.layers.sequence_reshape(input=red_sum, new_dim=(neg_size + 1))
+ pre = fluid.layers.softmax(input=pre_)
cost = fluid.layers.cross_entropy(input=pre, label=pos_label)
cost_sum = fluid.layers.reduce_sum(input=cost)
diff --git a/fluid/PaddleRec/gru4rec/train_sample_neg.py b/fluid/PaddleRec/gru4rec/train_sample_neg.py
index eb7ec3d4901d9ab7916546d83570534c81a8b0ff..1b1736cf937723bc86693c0d8cd39e579735f129 100644
--- a/fluid/PaddleRec/gru4rec/train_sample_neg.py
+++ b/fluid/PaddleRec/gru4rec/train_sample_neg.py
@@ -68,9 +68,11 @@ def train():
# Train program
if args.loss == 'bpr':
+ print('bpr loss')
src, pos_label, label, avg_cost = net.train_bpr_network(
neg_size=args.neg_size, vocab_size=vocab_size, hid_size=hid_size)
else:
+ print('cross-entory loss')
src, pos_label, label, avg_cost = net.train_cross_entropy_network(
neg_size=args.neg_size, vocab_size=vocab_size, hid_size=hid_size)
diff --git a/fluid/PaddleRec/gru4rec/utils.py b/fluid/PaddleRec/gru4rec/utils.py
index 429026b831454e44869238744c890e6139a6074d..1cd6a313b2a5097b16c473722737e0e6936f4e31 100644
--- a/fluid/PaddleRec/gru4rec/utils.py
+++ b/fluid/PaddleRec/gru4rec/utils.py
@@ -45,8 +45,8 @@ def to_lodtensor_bpr(raw_data, neg_size, vocab_size, place):
neg_data = np.tile(pos_data, neg_size)
np.random.shuffle(neg_data)
for ii in range(length * neg_size):
- if neg_data[ii] == pos_data[ii / neg_size]:
- neg_data[ii] = pos_data[length - 1 - ii / neg_size]
+ if neg_data[ii] == pos_data[ii // neg_size]:
+ neg_data[ii] = pos_data[length - 1 - ii // neg_size]
label_data = np.column_stack(
(pos_data.reshape(length, 1), neg_data.reshape(length, neg_size)))
diff --git a/fluid/PaddleRec/multiview_simnet/.run_ce.sh b/fluid/PaddleRec/multiview_simnet/.run_ce.sh
new file mode 100755
index 0000000000000000000000000000000000000000..4aaed27e9eab72c7043925cd051a7e528e2205bc
--- /dev/null
+++ b/fluid/PaddleRec/multiview_simnet/.run_ce.sh
@@ -0,0 +1,11 @@
+#!/bin/bash
+
+export MKL_NUM_THREADS=1
+export OMP_NUM_THREADS=1
+
+
+export CPU_NUM=1
+export NUM_THREADS=1
+
+FLAGS_benchmark=true python train.py --enable_ce | python _ce.py
+
diff --git a/fluid/PaddleRec/multiview_simnet/_ce.py b/fluid/PaddleRec/multiview_simnet/_ce.py
new file mode 100644
index 0000000000000000000000000000000000000000..ce6cf7cfa41748ab69ca9aa0955461371c42fe36
--- /dev/null
+++ b/fluid/PaddleRec/multiview_simnet/_ce.py
@@ -0,0 +1,58 @@
+# this file is only used for continuous evaluation test!
+
+import os
+import sys
+sys.path.append(os.environ['ceroot'])
+from kpi import CostKpi
+from kpi import DurationKpi
+from kpi import AccKpi
+
+
+each_pass_duration_cpu1_thread1_kpi = DurationKpi('each_pass_duration_cpu1_thread1', 0.08, 0, actived=True)
+train_loss_cpu1_thread1_kpi = CostKpi('train_loss_cpu1_thread1', 0.08, 0)
+
+tracking_kpis = [
+ each_pass_duration_cpu1_thread1_kpi,
+ train_loss_cpu1_thread1_kpi,
+ ]
+
+
+def parse_log(log):
+ '''
+ This method should be implemented by model developers.
+
+ The suggestion:
+
+ each line in the log should be key, value, for example:
+
+ "
+ train_cost\t1.0
+ test_cost\t1.0
+ train_cost\t1.0
+ train_cost\t1.0
+ train_acc\t1.2
+ "
+ '''
+ for line in log.split('\n'):
+ fs = line.strip().split('\t')
+ print(fs)
+ if len(fs) == 3 and fs[0] == 'kpis':
+ kpi_name = fs[1]
+ kpi_value = float(fs[2])
+ yield kpi_name, kpi_value
+
+
+def log_to_ce(log):
+ kpi_tracker = {}
+ for kpi in tracking_kpis:
+ kpi_tracker[kpi.name] = kpi
+
+ for (kpi_name, kpi_value) in parse_log(log):
+ print(kpi_name, kpi_value)
+ kpi_tracker[kpi_name].add_record(kpi_value)
+ kpi_tracker[kpi_name].persist()
+
+
+if __name__ == '__main__':
+ log = sys.stdin.read()
+ log_to_ce(log)
diff --git a/fluid/PaddleRec/multiview_simnet/train.py b/fluid/PaddleRec/multiview_simnet/train.py
index b4a566d39333d871d30e5996c45d7ea9ef7b1531..f098fd109e8813ffbfb40753122acbef3cd896a6 100644
--- a/fluid/PaddleRec/multiview_simnet/train.py
+++ b/fluid/PaddleRec/multiview_simnet/train.py
@@ -81,10 +81,19 @@ def parse_args():
"for index processing")
parser.add_argument(
"--hidden_size", type=int, default=128, help="Hidden dim")
+ parser.add_argument(
+ '--enable_ce',
+ action='store_true',
+ help='If set, run the task with continuous evaluation logs.')
return parser.parse_args()
def start_train(args):
+ if args.enable_ce:
+ SEED = 102
+ fluid.default_startup_program().random_seed = SEED
+ fluid.default_startup_program().random_seed = SEED
+
dataset = reader.SyntheticDataset(args.sparse_feature_dim, args.query_slots,
args.title_slots)
train_reader = paddle.batch(
@@ -115,7 +124,10 @@ def start_train(args):
exe = fluid.Executor(place)
exe.run(startup_program)
+ total_time = 0
+ ce_info = []
for pass_id in range(args.epochs):
+ start_time = time.time()
for batch_id, data in enumerate(train_reader()):
loss_val, correct_val = exe.run(loop_program,
feed=feeder.feed(data),
@@ -123,10 +135,34 @@ def start_train(args):
logger.info("TRAIN --> pass: {} batch_id: {} avg_cost: {}, acc: {}"
.format(pass_id, batch_id, loss_val,
float(correct_val) / args.batch_size))
+ ce_info.append(loss_val[0])
+ end_time = time.time()
+ total_time += end_time - start_time
fluid.io.save_inference_model(args.model_output_dir,
[val.name for val in all_slots],
[avg_cost, correct], exe)
+ # only for ce
+ if args.enable_ce:
+ threads_num, cpu_num = get_cards(args)
+ epoch_idx = args.epochs
+ ce_loss = 0
+ try:
+ ce_loss = ce_info[-2]
+ except:
+ logger.error("ce info error")
+
+ print("kpis\teach_pass_duration_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, total_time / epoch_idx))
+ print("kpis\ttrain_loss_cpu%s_thread%s\t%s" %
+ (cpu_num, threads_num, ce_loss))
+
+
+def get_cards(args):
+ threads_num = os.environ.get('NUM_THREADS', 1)
+ cpu_num = os.environ.get('CPU_NUM', 1)
+ return int(threads_num), int(cpu_num)
+
def main():
args = parse_args()
diff --git a/fluid/PaddleRec/ssr/infer.py b/fluid/PaddleRec/ssr/infer.py
index d5c9ee1b5dc95eb403932e0ff7534bfadc7568d7..38fb5cd762e117409b12ce8bd202f110a1cdfcb4 100644
--- a/fluid/PaddleRec/ssr/infer.py
+++ b/fluid/PaddleRec/ssr/infer.py
@@ -81,7 +81,7 @@ def infer(args, vocab_size, test_reader):
start_up_program = fluid.Program()
with fluid.program_guard(main_program, start_up_program):
acc = model(vocab_size, emb_size, hid_size)
- for epoch in xrange(start_index, last_index + 1):
+ for epoch in range(start_index, last_index + 1):
copy_program = main_program.clone()
model_path = model_dir + "/epoch_" + str(epoch)
fluid.io.load_params(
diff --git a/fluid/README.cn.rst b/fluid/README.cn.rst
index 811038c8aaedcce2b55fca54e647dcda98924db9..115b3e157e26c764ab44dcab71fffc42b2fb8dca 100644
--- a/fluid/README.cn.rst
+++ b/fluid/README.cn.rst
@@ -33,11 +33,14 @@ Fluid模型配置和参数文件的工具。
VOC `__\ 、\ `MS
COCO `__\ 数据训练通用物体检测模型，当前介绍了SSD算法，SSD全称Single Shot MultiBox Detector，是目标检测领域较新且效果较好的检测算法之一，具有检测速度快且检测精度高的特点。
-开放环境中的检测人脸，尤其是小的、模糊的和部分遮挡的人脸也是一个具有挑战的任务。我们也介绍了如何基于 `WIDER FACE `_ 数据训练百度自研的人脸检测PyramidBox模型，该算法于2018年3月份在WIDER FACE的多项评测中均获得 `第一名 `_。
+开放环境中的检测人脸，尤其是小的、模糊的和部分遮挡的人脸也是一个具有挑战的任务。我们也介绍了如何基于 `WIDER FACE `_ 数据训练百度自研的人脸检测PyramidBox模型，该算法于2018年3月份在WIDER FACE的多项评测中均获得 `第一名 `_ 。
+
+RCNN系列模型是典型的两阶段目标检测器，相较于传统提取区域的方法，RCNN中RPN网络通过共享卷积层参数大幅提高提取区域的效率，并提出高质量的候选区域。其中典型模型包括Faster RCNN和Mask RCNN。
- `Single Shot MultiBox
Detector `__
- `Face Detector: PyramidBox `_
+- `RCNN `_
图像语义分割
------------
diff --git a/fluid/README.md b/fluid/README.md
index 9bbcb9623695319d34d0b986ad85d9029ef0b0a5..bb76fea7cf2a21064ba669c50a8cbda1ce1a109b 100644
--- a/fluid/README.md
+++ b/fluid/README.md
@@ -28,11 +28,14 @@ Fluid模型配置和参数文件的工具。
开放环境中的检测人脸，尤其是小的、模糊的和部分遮挡的人脸也是一个具有挑战的任务。我们也介绍了如何基于 [WIDER FACE](http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace) 数据训练百度自研的人脸检测PyramidBox模型，该算法于2018年3月份在WIDER FACE的多项评测中均获得 [第一名](http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/WiderFace_Results.html)。
-Faster RCNN 是典型的两阶段目标检测器，相较于传统提取区域的方法，Faster RCNN中RPN网络通过共享卷积层参数大幅提高提取区域的效率，并提出高质量的候选区域。
+Faster RCNN模型是典型的两阶段目标检测器，相较于传统提取区域的方法，通过RPN网络共享卷积层参数大幅提高提取区域的效率，并提出高质量的候选区域。
+
+Mask RCNN模型是基于Faster RCNN模型的经典实例分割模型，在原有Faster RCNN模型基础上添加分割分支，得到掩码结果，实现了掩码和类别预测关系的解藕。
- [Single Shot MultiBox Detector](https://github.com/PaddlePaddle/models/blob/develop/fluid/PaddleCV/object_detection/README_cn.md)
- [Face Detector: PyramidBox](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/face_detection/README_cn.md)
-- [Faster RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/faster_rcnn/README_cn.md)
+- [Faster RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/rcnn/README_cn.md)
+- [Mask RCNN](https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/rcnn/README_cn.md)
图像语义分割
------------