diff --git a/README.md b/README.md index 5842741437347c7aaf48f81ad7ee270c43ecb1dc..06159f8759948c7f77fa526e5a05ae36970210ee 100644 --- a/README.md +++ b/README.md @@ -9,7 +9,7 @@ 飞桨图像分类套件PaddleClas是飞桨为工业界和学术界所准备的一个图像分类任务的工具集，助力使用者训练出更好的视觉模型和应用落地。
- +
## 丰富的模型库 @@ -17,21 +17,20 @@ 基于ImageNet1k分类数据集，PaddleClas提供ResNet、ResNet_vd、Res2Net、HRNet、MobileNetV3等23种系列的分类网络结构的简单介绍、论文指标复现配置，以及在复现过程中的训练技巧。与此同时，也提供了对应的117个图像分类预训练模型，并且基于TensorRT评估了服务器端模型的GPU预测时间，以及在骁龙855（SD855）上评估了移动端模型的CPU预测时间和存储大小。支持的***预训练模型列表、下载地址以及更多信息***请见文档教程中的[**模型库章节**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)。
- +
-上图对比了一些最新的面向服务器端应用场景的模型，在使用V100，FP32和TensorRT预测一张图像的时间和其准确率，图中准确率82.4%的ResNet50_vd_ssld和83.7%的ResNet101_vd_ssld，是采用PaddleClas提供的SSLD知识蒸馏方案训练的模型。图中相同颜色和符号的点代表同一系列不同规模的模型。不同模型的简介、FLOPS、Parameters以及详细的GPU预测时间请参考文档教程中的[**模型库章节**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)。 +上图对比了一些最新的面向服务器端应用场景的模型，在使用V100，FP32和TensorRT，batch size为1时的预测时间及其准确率，图中准确率82.4%的ResNet50_vd_ssld和83.7%的ResNet101_vd_ssld，是采用PaddleClas提供的SSLD知识蒸馏方案训练的模型。图中相同颜色和符号的点代表同一系列不同规模的模型。不同模型的简介、FLOPS、Parameters以及详细的GPU预测时间(包括不同batchsize的T4卡预测速度)请参考文档教程中的[**模型库章节**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)。
+src="./docs/images/models/mobile_arm_top1.png" width="700">
上图对比了一些最新的面向移动端应用场景的模型，在骁龙855（SD855）上预测一张图像的时间和其准确率，包括MobileNetV1系列、MobileNetV2系列、MobileNetV3系列和ShuffleNetV2系列。图中准确率79%的MV3_large_x1_0_ssld（M是MobileNet的简称），71.3%的MV3_small_x1_0_ssld、76.74%的MV2_ssld和77.89%的MV1_ssld，是采用PaddleClas提供的SSLD蒸馏方法训练的模型。MV3_large_x1_0_ssld_int8是进一步进行INT8量化的模型。不同模型的简介、FLOPS、Parameters和模型存储大小请参考文档教程中的[**模型库章节**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)。 - TODO - [ ] EfficientLite、GhostNet、RegNet论文指标复现和性能评估 -- [ ] The speed benchmark on P4/T4 ## 高阶优化支持 除了提供丰富的分类网络结构和预训练模型，PaddleClas也支持了一系列有助于图像分类任务效果和效率提升的算法或工具。 @@ -41,14 +40,14 @@ src="docs/images/models/mobile_arm_top1.png" width="700">
+src="./docs/images/distillation/distillation_perform_s.jpg" width="700">
以在ImageNet1K蒸馏模型为例，SSLD知识蒸馏方案框架图如下，该方案的核心关键点包括教师模型的选择、loss计算方式、迭代轮数、无标签数据的使用、以及ImageNet1k蒸馏finetune，每部分的详细介绍以及实验介绍请参考文档教程中的[**知识蒸馏章节**](https://paddleclas.readthedocs.io/zh_CN/latest/advanced_tutorials/distillation/index.html)。
+src="./docs/images/distillation/ppcls_distillation_s.jpg" width="700">
### 数据增广 @@ -57,14 +56,14 @@ src="docs/images/distillation/ppcls_distillation.png" width="700">
+src="./docs/images/image_aug/image_aug_samples_s.jpg" width="800">
PaddleClas提供了上述8种数据增广算法的复现和在统一实验环境下的效果评估。下图展示了不同数据增广方式在ResNet50上的表现, 与标准变换相比，采用数据增广，识别准确率最高可以提升1%。每种数据增广方法的详细介绍、对比的实验环境请参考文档教程中的[**数据增广章节**](https://paddleclas.readthedocs.io/zh_CN/latest/advanced_tutorials/image_augmentation/index.html)。
+src="./docs/images/image_aug/main_image_aug_s.jpg" width="600">
## 30分钟玩转PaddleClas @@ -80,7 +79,7 @@ PaddleClas的安装说明、模型训练、预测、评估以及模型微调（f ### 10万类图像分类预训练模型 在实际应用中，由于训练数据匮乏，往往将ImageNet1K数据集训练的分类模型作为预训练模型，进行图像分类的迁移学习。然而ImageNet1K数据集的类别只有1000种，预训练模型的特征迁移能力有限。因此百度自研了一个有语义体系的、粒度有粗有细的10w级别的Tag体系，通过人工或半监督方式，至今收集到 5500w+图片训练数据；该系统是国内甚至世界范围内最大规模的图片分类体系和训练集合。PaddleClas提供了在该数据集上训练的ResNet50_vd的模型。下表显示了一些实际应用场景中，使用ImageNet预训练模型和上述10万类图像分类预训练模型的效果比对，使用10万类图像分类预训练模型，识别准确率最高可以提升30%。 - + | 数据集 | 数据统计 | ImageNet预训练模型 | 10万类图像分类预训练模型 | |:--:|:--:|:--:|:--:| | 花卉 | class_num:102
train/val:5789/2396 | 0.7779 | 0.9892 | @@ -100,7 +99,7 @@ PaddleClas的安装说明、模型训练、预测、评估以及模型微调（f
+src="./docs/images/det/pssdet.png" width="500">
- TODO diff --git a/configs/ResNet/ResNet50_fp16.yml b/configs/ResNet/ResNet50_fp16.yml new file mode 100644 index 0000000000000000000000000000000000000000..a952833221a193cefe32c004f40181eaa409188d --- /dev/null +++ b/configs/ResNet/ResNet50_fp16.yml @@ -0,0 +1,81 @@ +mode: 'train' +ARCHITECTURE: + name: 'ResNet50' + +pretrained_model: "" +model_save_dir: "./output/" +classes_num: 1000 +total_images: 1281167 +save_interval: 1 +validate: True +valid_interval: 1 +epochs: 120 +topk: 5 +image_shape: [3, 224, 224] + +# mixed precision training +use_fp16: True +amp_scale_loss: 128.0 +use_dynamic_loss_scaling: True + +use_mix: False +ls_epsilon: -1 + +LEARNING_RATE: + function: 'Piecewise' + params: + lr: 0.1 + decay_epochs: [30, 60, 90] + gamma: 0.1 + +OPTIMIZER: + function: 'Momentum' + params: + momentum: 0.9 + regularizer: + function: 'L2' + factor: 0.000100 + +TRAIN: + batch_size: 256 + num_workers: 4 + file_list: "./dataset/ILSVRC2012/train_list.txt" + data_dir: "./dataset/ILSVRC2012/" + shuffle_seed: 0 + transforms: + - DecodeImage: + to_rgb: True + to_np: False + channel_first: False + - RandCropImage: + size: 224 + - RandFlipImage: + flip_code: 1 + - NormalizeImage: + scale: 1./255. + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: '' + - ToCHWImage: + +VALID: + batch_size: 64 + num_workers: 4 + file_list: "./dataset/ILSVRC2012/val_list.txt" + data_dir: "./dataset/ILSVRC2012/" + shuffle_seed: 0 + transforms: + - DecodeImage: + to_rgb: True + to_np: False + channel_first: False + - ResizeImage: + resize_short: 256 + - CropImage: + size: 224 + - NormalizeImage: + scale: 1.0/255.0 + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: '' + - ToCHWImage: diff --git a/docs/images/distillation/distillation_perform_s.jpg b/docs/images/distillation/distillation_perform_s.jpg new file mode 100644 index 0000000000000000000000000000000000000000..07b1bf6790d4748b625529f5669c5ff581dab58d Binary files /dev/null and b/docs/images/distillation/distillation_perform_s.jpg differ diff --git a/docs/images/distillation/ppcls_distillation_s.jpg b/docs/images/distillation/ppcls_distillation_s.jpg new file mode 100644 index 0000000000000000000000000000000000000000..95402d401d265a008c5710dfcc711f9a16e6d962 Binary files /dev/null and b/docs/images/distillation/ppcls_distillation_s.jpg differ diff --git a/docs/images/image_aug/image_aug_samples_s.jpg b/docs/images/image_aug/image_aug_samples_s.jpg new file mode 100644 index 0000000000000000000000000000000000000000..aad144ae635cc2c2e01c43e73496c279a96d7f09 Binary files /dev/null and b/docs/images/image_aug/image_aug_samples_s.jpg differ diff --git a/docs/images/image_aug/main_image_aug_s.jpg b/docs/images/image_aug/main_image_aug_s.jpg new file mode 100644 index 0000000000000000000000000000000000000000..efaa27c640be17465acb77b3311752ea944c0acd Binary files /dev/null and b/docs/images/image_aug/main_image_aug_s.jpg differ diff --git a/docs/images/main_features_s.png b/docs/images/main_features_s.png new file mode 100644 index 0000000000000000000000000000000000000000..d1e2acc515c6753193ce1b5f1a89a4435f905d74 Binary files /dev/null and b/docs/images/main_features_s.png differ diff --git a/docs/images/models/DPN.png.flops.png b/docs/images/models/DPN.png.flops.png deleted file mode 100644 index 72bb96f49812711035ec09ce0d8d44202d17cfcb..0000000000000000000000000000000000000000 Binary files a/docs/images/models/DPN.png.flops.png and /dev/null differ diff --git a/docs/images/models/DPN.png.fp32.png b/docs/images/models/DPN.png.fp32.png deleted file mode 100644 index bcd524c781d305fdacfd5a07886851fadbeeeddc..0000000000000000000000000000000000000000 Binary files a/docs/images/models/DPN.png.fp32.png and /dev/null differ diff --git a/docs/images/models/DPN.png.params.png b/docs/images/models/DPN.png.params.png deleted file mode 100644 index 818f1961bbaf98cd0aff0c5405113f107fa02e54..0000000000000000000000000000000000000000 Binary files a/docs/images/models/DPN.png.params.png and /dev/null differ diff --git a/docs/images/models/EfficientNet.png b/docs/images/models/EfficientNet.png deleted file mode 100644 index 5556481c960432cab6080c243644ab43783ceabb..0000000000000000000000000000000000000000 Binary files a/docs/images/models/EfficientNet.png and /dev/null differ diff --git a/docs/images/models/EfficientNet.png.flops.png b/docs/images/models/EfficientNet.png.flops.png deleted file mode 100644 index dd3c36ced2973133bdd0bb6b300125b25fdcefe4..0000000000000000000000000000000000000000 Binary files a/docs/images/models/EfficientNet.png.flops.png and /dev/null differ diff --git a/docs/images/models/EfficientNet.png.fp32.png b/docs/images/models/EfficientNet.png.fp32.png deleted file mode 100644 index eca753f7d84699cebee13d15484291d96f0a9b6f..0000000000000000000000000000000000000000 Binary files a/docs/images/models/EfficientNet.png.fp32.png and /dev/null differ diff --git a/docs/images/models/EfficientNet.png.params.png b/docs/images/models/EfficientNet.png.params.png deleted file mode 100644 index 2348c55013998bcb2ed2c3edff282493243ee37a..0000000000000000000000000000000000000000 Binary files a/docs/images/models/EfficientNet.png.params.png and /dev/null differ diff --git a/docs/images/models/HRNet.png.flops.png b/docs/images/models/HRNet.png.flops.png deleted file mode 100644 index 5f8ce9cd2c1c8ed9e8fb775bd89a8acd3a8e9402..0000000000000000000000000000000000000000 Binary files a/docs/images/models/HRNet.png.flops.png and /dev/null differ diff --git a/docs/images/models/HRNet.png.fp32.png b/docs/images/models/HRNet.png.fp32.png deleted file mode 100644 index 0e73fb4b57dc374560efa92429a2dd457c73369e..0000000000000000000000000000000000000000 Binary files a/docs/images/models/HRNet.png.fp32.png and /dev/null differ diff --git a/docs/images/models/HRNet.png.params.png b/docs/images/models/HRNet.png.params.png deleted file mode 100644 index e4443a770ac0ca910d1158fe9adaf6dd92e680aa..0000000000000000000000000000000000000000 Binary files a/docs/images/models/HRNet.png.params.png and /dev/null differ diff --git a/docs/images/models/Inception.png.flops.png b/docs/images/models/Inception.png.flops.png deleted file mode 100644 index 589f3931c1feef3e1c0245566cd0c7e0a22782d8..0000000000000000000000000000000000000000 Binary files a/docs/images/models/Inception.png.flops.png and /dev/null differ diff --git a/docs/images/models/Inception.png.fp32.png b/docs/images/models/Inception.png.fp32.png deleted file mode 100644 index b9245800a2d7ca6fad5ed7457e55356d579a81d0..0000000000000000000000000000000000000000 Binary files a/docs/images/models/Inception.png.fp32.png and /dev/null differ diff --git a/docs/images/models/Inception.png.params.png b/docs/images/models/Inception.png.params.png deleted file mode 100644 index 657c4360451b24905b9a1ce170e4da719d35d917..0000000000000000000000000000000000000000 Binary files a/docs/images/models/Inception.png.params.png and /dev/null differ diff --git a/docs/images/models/ResNet.png.flops.png b/docs/images/models/ResNet.png.flops.png deleted file mode 100644 index da1fd2eb359f57fbd545a5436b737fe23df6e891..0000000000000000000000000000000000000000 Binary files a/docs/images/models/ResNet.png.flops.png and /dev/null differ diff --git a/docs/images/models/ResNet.png.fp32.png b/docs/images/models/ResNet.png.fp32.png deleted file mode 100644 index 05020997f2b6eb2c926d2a8948ed69b393b6cd3b..0000000000000000000000000000000000000000 Binary files a/docs/images/models/ResNet.png.fp32.png and /dev/null differ diff --git a/docs/images/models/ResNet.png.params.png b/docs/images/models/ResNet.png.params.png deleted file mode 100644 index 6fcbb69cc1e1e9a3402f2849fe016a73312d9a79..0000000000000000000000000000000000000000 Binary files a/docs/images/models/ResNet.png.params.png and /dev/null differ diff --git a/docs/images/models/SeResNeXt.png.flops.png b/docs/images/models/SeResNeXt.png.flops.png deleted file mode 100644 index 51d6d6497e9cd582ad671a79b9e24bbdc1a9bdae..0000000000000000000000000000000000000000 Binary files a/docs/images/models/SeResNeXt.png.flops.png and /dev/null differ diff --git a/docs/images/models/SeResNeXt.png.fp32.png b/docs/images/models/SeResNeXt.png.fp32.png deleted file mode 100644 index 452488955096f896ecb9dafe07885f666c92d8ad..0000000000000000000000000000000000000000 Binary files a/docs/images/models/SeResNeXt.png.fp32.png and /dev/null differ diff --git a/docs/images/models/SeResNeXt.png.params.png b/docs/images/models/SeResNeXt.png.params.png deleted file mode 100644 index 9898f52fb0be6bdfb22d39a1f4a16c98e20ab510..0000000000000000000000000000000000000000 Binary files a/docs/images/models/SeResNeXt.png.params.png and /dev/null differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs1.EfficientNet.png b/docs/images/models/T4_benchmark/t4.fp16.bs1.EfficientNet.png new file mode 100644 index 0000000000000000000000000000000000000000..1c87d711ea11f3d662cdf78e34959ddd1f355f76 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs1.EfficientNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs4.DPN.png b/docs/images/models/T4_benchmark/t4.fp16.bs4.DPN.png new file mode 100644 index 0000000000000000000000000000000000000000..1eb393903bbf3b4e02b83962b33c0e0a3b4e341a Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs4.DPN.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs4.HRNet.png b/docs/images/models/T4_benchmark/t4.fp16.bs4.HRNet.png new file mode 100644 index 0000000000000000000000000000000000000000..f21d63cd1a1e24481947875db23f8447af1a65ca Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs4.HRNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs4.Inception.png b/docs/images/models/T4_benchmark/t4.fp16.bs4.Inception.png new file mode 100644 index 0000000000000000000000000000000000000000..8095a3c0253170c00d8ae74af4dec25c4f9544eb Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs4.Inception.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs4.ResNet.png b/docs/images/models/T4_benchmark/t4.fp16.bs4.ResNet.png new file mode 100644 index 0000000000000000000000000000000000000000..53a603ecad2e36df580167e134ef036df14d5596 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs4.ResNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp16.bs4.SeResNeXt.png b/docs/images/models/T4_benchmark/t4.fp16.bs4.SeResNeXt.png new file mode 100644 index 0000000000000000000000000000000000000000..99b8a039e0fda22053e7d7cb971d5e83b208ec6b Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp16.bs4.SeResNeXt.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs1.EfficientNet.png b/docs/images/models/T4_benchmark/t4.fp32.bs1.EfficientNet.png new file mode 100644 index 0000000000000000000000000000000000000000..395a32f5c7e28ed09ee2b7e12e4a3ea2e9094154 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs1.EfficientNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..24aabf8c3fb6607e4bb17f4d4dcc72d733476c48 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.params.png new file mode 100644 index 0000000000000000000000000000000000000000..689e73d31d70cf8566c356142080d79a0f64d6e3 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.png new file mode 100644 index 0000000000000000000000000000000000000000..dc3922d2e2347f11b193057de9bbf730489b9cc1 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.DPN.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..deacfaca6a279fed852dfcbf0006dc497191a7a8 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.params.png new file mode 100644 index 0000000000000000000000000000000000000000..7177bbc56b374dd71c66231132ead01c8b141732 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..062ecd79d2fc3ab788212a238b8ca4627b0dd14f Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.params.png new file mode 100644 index 0000000000000000000000000000000000000000..4bb3f76caecbdbd43f04ed62507985476f5cac40 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.png new file mode 100644 index 0000000000000000000000000000000000000000..3905f0b38b7cffa856d944d99ef0035cf6d4b489 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.HRNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..6fdc94b27130924d0e71a7a6954f239e632a4215 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.params.png new file mode 100644 index 0000000000000000000000000000000000000000..25a5d1648e1cf220a383377a8d9fd5ae3c9eceea Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.png new file mode 100644 index 0000000000000000000000000000000000000000..7ef4f339ae2b9414b4cf71a8772cc4c9a92a0ad1 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.Inception.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..755adb5684562c341816ca98856490d270735ac5 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.params.png new file mode 100644 index 0000000000000000000000000000000000000000..44e03fdd20df450573889e5c4eca85cf5b686d9b Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.png new file mode 100644 index 0000000000000000000000000000000000000000..a461b9ec281c2340780da6d69e625a0947ffa42e Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.ResNet.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..197522d16401ef6430313d36ac235a7b37e1d7f9 Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.params.png new file mode 100644 index 0000000000000000000000000000000000000000..6943fc056c1cd22b4b4d1767acf72fd94a209d0d Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.params.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.png new file mode 100644 index 0000000000000000000000000000000000000000..8476efb33b73acd836993a5fec3967122da319ab Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png new file mode 100644 index 0000000000000000000000000000000000000000..965efff5c1a3bd32c9bc2da9c1b3034dc2cd55ef Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.flops.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.flops.png new file mode 100644 index 0000000000000000000000000000000000000000..8c1be3ae9b32773ca56adac9dc1c15d9532e8f1a Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.flops.png differ diff --git a/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.params.png b/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.params.png new file mode 100644 index 0000000000000000000000000000000000000000..a41a325b5f5e9e85b86abfae2e62c9e2132fc7ec Binary files /dev/null and b/docs/images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.params.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.DPN.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.DPN.png new file mode 100644 index 0000000000000000000000000000000000000000..10f542b7fc989da80af12244cd45662bccfe677b Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.DPN.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.EfficientNet.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.EfficientNet.png new file mode 100644 index 0000000000000000000000000000000000000000..0491ff3e6f4795fec2e131a4df31b4a3b102900e Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.EfficientNet.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.HRNet.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.HRNet.png new file mode 100644 index 0000000000000000000000000000000000000000..284088ec5555e2128eab564cf8331532cfe08370 Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.HRNet.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.Inception.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.Inception.png new file mode 100644 index 0000000000000000000000000000000000000000..0e0e42e7ccc319bd2c9842c3d99c849b547ac603 Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.Inception.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.ResNet.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.ResNet.png new file mode 100644 index 0000000000000000000000000000000000000000..2332791866c06b9cdcfc576947280de27ba3667c Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.ResNet.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.SeResNeXt.png b/docs/images/models/V100_benchmark/v100.fp32.bs1.SeResNeXt.png new file mode 100644 index 0000000000000000000000000000000000000000..610c846347f008216589ed080d834588d7fedfe0 Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.SeResNeXt.png differ diff --git a/docs/images/models/V100_benchmark/v100.fp32.bs1.main_fps_top1_s.jpg b/docs/images/models/V100_benchmark/v100.fp32.bs1.main_fps_top1_s.jpg new file mode 100644 index 0000000000000000000000000000000000000000..407d4cff7d39b141dc210a523cd87a6d2f288bf2 Binary files /dev/null and b/docs/images/models/V100_benchmark/v100.fp32.bs1.main_fps_top1_s.jpg differ diff --git a/docs/images/models/main_fps_top1.png b/docs/images/models/main_fps_top1.png deleted file mode 100755 index a7149d7de16c71dbd3359efdbcee519a486a2278..0000000000000000000000000000000000000000 Binary files a/docs/images/models/main_fps_top1.png and /dev/null differ diff --git a/docs/images/models/mobile_arm_storage.png b/docs/images/models/mobile_arm_storage.png old mode 100755 new mode 100644 index 350fd3ced05e802250a461a747f60c2e5cf1be0c..07e1f4f3fe95ed7e9a358b212f4d5d939fb94a02 Binary files a/docs/images/models/mobile_arm_storage.png and b/docs/images/models/mobile_arm_storage.png differ diff --git a/docs/images/models/mobile_arm_top1.png b/docs/images/models/mobile_arm_top1.png old mode 100755 new mode 100644 index 37091dd2c91a334c2888d5facec7598ad3219e84..06add75fe630510e6ab9af62b3fa93dc166b7944 Binary files a/docs/images/models/mobile_arm_top1.png and b/docs/images/models/mobile_arm_top1.png differ diff --git a/docs/images/models/mobile_trt.png b/docs/images/models/mobile_trt.png deleted file mode 100644 index d722548bae56b9aeca081cffe3e7d34494864033..0000000000000000000000000000000000000000 Binary files a/docs/images/models/mobile_trt.png and /dev/null differ diff --git a/docs/images/models/mobile_trt.png.flops.png b/docs/images/models/mobile_trt.png.flops.png deleted file mode 100644 index 6e7010906c58ef9e1f4f286725442ca85e638c51..0000000000000000000000000000000000000000 Binary files a/docs/images/models/mobile_trt.png.flops.png and /dev/null differ diff --git a/docs/images/models/mobile_trt.png.params.png b/docs/images/models/mobile_trt.png.params.png deleted file mode 100644 index 35b78a65b390f9843def8870bbd63f656070cdcc..0000000000000000000000000000000000000000 Binary files a/docs/images/models/mobile_trt.png.params.png and /dev/null differ diff --git a/docs/zh_CN/extension/paddle_mobile_inference.md b/docs/zh_CN/extension/paddle_mobile_inference.md index e076fc7d471e8fd3ada051e3df171b23953884fe..833231f736ffc19cb3b705025d7511dc8eba6bc4 100644 --- a/docs/zh_CN/extension/paddle_mobile_inference.md +++ b/docs/zh_CN/extension/paddle_mobile_inference.md @@ -1,5 +1,119 @@ # Paddle-Lite +## 一、简介 + [Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite) 是飞桨推出的一套功能完善、易用性强且性能卓越的轻量化推理引擎。 轻量化体现在使用较少比特数用于表示神经网络的权重和激活，能够大大降低模型的体积，解决终端设备存储空间有限的问题，推理性能也整体优于其他框架。 -[PaddleClas](https://github.com/PaddlePaddle/PaddleClas) 使用 Paddle-Lite 进行了[移动端模型的性能评估](../models/Mobile.md)，具体流程参考 [Paddle-Lite 文档](https://paddle-lite.readthedocs.io/zh/latest/)。 +[PaddleClas](https://github.com/PaddlePaddle/PaddleClas) 使用 Paddle-Lite 进行了[移动端模型的性能评估](../models/Mobile.md)，本部分以`ImageNet1k`数据集的`MobileNetV1`模型为例，介绍怎样使用`Paddle-Lite`，在移动端(基于骁龙855的安卓开发平台)对进行模型速度评估。 + + +## 二、评估步骤 + +### 2.1 导出inference模型 + +* 首先需要将训练过程中保存的模型存储为用于预测部署的固化模型，可以使用`tools/export_model.py`导出inference模型，具体使用方法如下。 + +```shell +python tools/export_model.py -m MobileNetV1 -p pretrained/MobileNetV1_pretrained/ -o inference/MobileNetV1 +``` + +最终在`inference/MobileNetV1`文件夹下会保存得到`model`与`parmas`文件。 + + +### 2.2 benchmark二进制文件下载 + +* 使用adb(Android Debug Bridge)工具可以连接Android手机与PC端，并进行开发调试等。安装好adb，并确保PC端和手机连接成功后，使用以下命令可以查看手机的ARM版本，并基于此选择合适的预编译库。 + +```shell +adb shell getprop ro.product.cpu.abi +``` + +* 下载benchmark_bin文件 + +```shell +wget -c https://paddle-inference-dist.bj.bcebos.com/PaddleLite/benchmark_0/benchmark_bin_v8 +``` + +如果查看的ARM版本为v7，则需要下载v7版本的benchmark_bin文件，下载命令如下。 + +```shell +wget -c https://paddle-inference-dist.bj.bcebos.com/PaddleLite/benchmark_0/benchmark_bin_v7 +``` + +### 2.3 模型速度benchmark + +PC端和手机连接成功后，使用下面的命令开始模型评估。 + +``` +sh tools/lite/benchmark.sh ./benchmark_bin_v8 ./inference result_armv8.txt true +``` + +其中`./benchmark_bin_v8`为benchmark二进制文件路径，`./inference`为所有需要评测的模型的路径，`result_armv8.txt`为保存的结果文件，最后的参数`true`表示在评估之后会首先进行模型优化。最终在当前文件夹下会输出`result_armv8.txt`的评估结果文件，具体信息如下。 + +``` +PaddleLite Benchmark +Threads=1 Warmup=10 Repeats=30 +MobileNetV1 min = 30.89100 max = 30.73600 average = 30.79750 + +Threads=2 Warmup=10 Repeats=30 +MobileNetV1 min = 18.26600 max = 18.14000 average = 18.21637 + +Threads=4 Warmup=10 Repeats=30 +MobileNetV1 min = 10.03200 max = 9.94300 average = 9.97627 +``` + +这里给出了不同线程数下的模型预测速度，单位为FPS，以线程数为1为例，MobileNetV1在骁龙855上的平均速度为`30.79750FPS`。 + + +### 2.4 模型优化与速度评估 + + +* 在2.3节中提到了在模型评估之前对其进行优化，在这里也可以首先对模型进行优化，再直接加载优化后的模型进行速度评估。 + +* Paddle-Lite 提供了多种策略来自动优化原始的训练模型，其中包括量化、子图融合、混合调度、Kernel优选等等方法。为了使优化过程更加方便易用，Paddle-Lite提供了opt 工具来自动完成优化步骤，输出一个轻量的、最优的可执行模型。可以在[Paddle-Lite模型优化工具页面](https://paddle-lite.readthedocs.io/zh/latest/user_guides/model_optimize_tool.html)下载。在这里以`MacOS`开发环境为例，下载[opt_mac](https://paddlelite-data.bj.bcebos.com/model_optimize_tool/opt_mac)模型优化工具，并使用下面的命令对模型进行优化。 + + + +```shell +model_file="../MobileNetV1/model" +param_file="../MobileNetV1/params" +opt_models_dir="./opt_models" +mkdir ${opt_models_dir} +./opt_mac --model_file=${model_file} \ + --param_file=${param_file} \ + --valid_targets=arm \ + --optimize_out_type=naive_buffer \ + --prefer_int8_kernel=false \ + --optimize_out=${opt_models_dir}/MobileNetV1 +``` + +其中`model_file`与`param_file`分别是导出的inference模型结构文件与参数文件地址，转换成功后，会在`opt_models`文件夹下生成`MobileNetV1.nb`文件。 + + + + +使用benchmark_bin文件加载优化后的模型进行评估，具体的命令如下。 + +```shell +bash benchmark.sh ./benchmark_bin_v8 ./opt_models result_armv8.txt +``` + +最终`result_armv8.txt`中结果如下。 + +``` +PaddleLite Benchmark +Threads=1 Warmup=10 Repeats=30 +MobileNetV1_lite min = 30.89500 max = 30.78500 average = 30.84173 + +Threads=2 Warmup=10 Repeats=30 +MobileNetV1_lite min = 18.25300 max = 18.11000 average = 18.18017 + +Threads=4 Warmup=10 Repeats=30 +MobileNetV1_lite min = 10.00600 max = 9.90000 average = 9.96177 +``` + + +以线程数为1为例，MobileNetV1在骁龙855上的平均速度为`30.84173FPS`。 + + +更加具体的参数解释与Paddle-Lite使用方法可以参考 [Paddle-Lite 文档](https://paddle-lite.readthedocs.io/zh/latest/)。 diff --git a/docs/zh_CN/extension/paddle_serving.md b/docs/zh_CN/extension/paddle_serving.md index 62b8cda8b74283fea701a86f6f05e5194b0a6da8..7b9102042b5018ab972bbd2542526186b5b402df 100644 --- a/docs/zh_CN/extension/paddle_serving.md +++ b/docs/zh_CN/extension/paddle_serving.md @@ -1,4 +1,65 @@ # 模型服务化部署 -[Paddle Serving](https://github.com/PaddlePaddle/Serving) 旨在帮助深度学习开发者轻易部署在线预测服务，支持一键部署工业级的服务能力、客户端和服务端之间高并发和高效通信、并支持多种编程语言开发客户端等特点，详细使用请参考 [Paddle Serving 相关文档](https://github.com/PaddlePaddle/Serving)。 +## 一、简介 +[Paddle Serving](https://github.com/PaddlePaddle/Serving) 旨在帮助深度学习开发者轻易部署在线预测服务，支持一键部署工业级的服务能力、客户端和服务端之间高并发和高效通信、并支持多种编程语言开发客户端。 +该部分以HTTP预测服务部署为例，介绍怎样在PaddleClas中使用PaddleServing部署模型服务。 + + +## 二、Serving安装 + +Serving官网推荐使用docker安装并部署Serving环境。首先需要拉取docker环境并创建基于Serving的docker。 + +```shell +nvidia-docker pull hub.baidubce.com/paddlepaddle/serving:0.2.0-gpu +nvidia-docker run -p 9292:9292 --name test -dit hub.baidubce.com/paddlepaddle/serving:0.2.0-gpu +nvidia-docker exec -it test bash +``` + +进入docker后，需要安装Serving相关的python包。 + +```shell +pip install paddlepaddle-gpu +pip install paddle-serving-client +pip install paddle-serving-server-gpu +``` + +* 如果安装速度太慢，可以通过`-i https://pypi.tuna.tsinghua.edu.cn/simple`更换源，加速安装过程。 + +* 如果希望部署CPU服务，可以安装serving-server的cpu版本，安装命令如下。 + +```shell +pip install paddle-serving-server +``` + +### 三、导出模型 + +使用`tools/export_serving_model.py`脚本导出Serving模型，以`ResNet50_vd`为例，使用方法如下。 + +```shell +python tools/export_serving_model.py -m ResNet50_vd -p ./pretrained/ResNet50_vd_pretrained/ -o serving +``` + +最终在serving文件夹下会生成`ppcls_client_conf`与`ppcls_model`两个文件夹，分别存储了client配置、模型参数与结构文件。 + + +### 四、服务部署与请求 + +* 使用下面的方式启动Serving服务。 + +```shell +python tools/serving/image_service_gpu.py serving/ppcls_model workdir 9292 +``` + +其中`serving/ppcls_model`为刚才保存的Serving模型地址，`workdir`为为工作目录，`9292`为服务的端口号。 + + +* 使用下面的脚本向Serving服务发送识别请求，并返回结果。 + +``` +python tools/serving/image_http_client.py 9292 ./docs/images/logo.png +``` + +`9292`为发送请求的端口号，需要与服务启动时的端口号保持一致，`./docs/images/logo.png`为待识别的图像文件。最终返回Top1识别结果的类别ID以及概率值。 + +* 更多的服务部署类型，如`RPC预测服务`等，可以参考Serving的github官网：[https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/imagenet](https://github.com/PaddlePaddle/Serving/tree/develop/python/examples/imagenet) diff --git a/docs/zh_CN/faq.md b/docs/zh_CN/faq.md index fe83d31ba03bf649aab1bff097d78874310fb4e6..36aff3ec2458c80802a9da2c2c9244efa3387dd9 100644 --- a/docs/zh_CN/faq.md +++ b/docs/zh_CN/faq.md @@ -1,10 +1,5 @@ # FAQ ->> -* Q: 启动训练后，为什么当前终端中的输出信息一直没有更新？ -* A: 启动运行后，日志会实时输出到`mylog/workerlog.*`中，可以在这里查看实时的日志。 - - >> * Q: 多卡评估时，为什么每张卡输出的精度指标不相同？ * A: 目前PaddleClas基于fleet api使用多卡，在多卡评估时，每张卡都是单独读取各自part的数据，不同卡中计算的图片是不同的，因此最终指标也会有微量差异，如果希望得到准确的评估指标，可以使用单卡评估。 @@ -18,3 +13,36 @@ >> * Q: 评估和预测时，已经指定了预训练模型所在文件夹的地址，但是仍然无法导入参数，这么为什么呢？ * A: 加载预训练模型时，需要指定预训练模型的前缀，例如预训练模型参数所在的文件夹为`output/ResNet50_vd/19`，预训练模型参数的名称为`output/ResNet50_vd/19/ppcls.pdparams`，则`pretrained_model`参数需要指定为`output/ResNet50_vd/19/ppcls`，PaddleClas会自动补齐`.pdparams`的后缀。 + + +>> +* Q: 在评测`EfficientNetB0_small`模型时，为什么最终的精度始终比官网的低0.3%左右？ +* A: `EfficientNet`系列的网络在进行resize的时候，是使用`cubic插值方式`(resize参数的interpolation值设置为2)，而其他模型默认情况下为None，因此在训练和评估的时候需要显式地指定resiz的interpolation值。具体地，可以参考以下配置中预处理过程中ResizeImage的参数。 +``` +VALID: + batch_size: 16 + num_workers: 4 + file_list: "./dataset/ILSVRC2012/val_list.txt" + data_dir: "./dataset/ILSVRC2012/" + shuffle_seed: 0 + transforms: + - DecodeImage: + to_rgb: True + to_np: False + channel_first: False + - ResizeImage: + resize_short: 256 + interpolation: 2 + - CropImage: + size: 224 + - NormalizeImage: + scale: 1.0/255.0 + mean: [0.485, 0.456, 0.406] + std: [0.229, 0.224, 0.225] + order: '' + - ToCHWImage: +``` + +>> +* Q: 如果想将保存的`pdparams`模型参数文件转换为早期版本(Paddle1.7.0之前)的零碎文件(每个文件均为一个单独的模型参数)，该怎么实现呢？ +* A: 可以首先导入`pdparams`模型，之后使用`fluid.io.save_vars`函数将模型保存为零散的碎文件。 diff --git a/docs/zh_CN/models/DPN_DenseNet.md b/docs/zh_CN/models/DPN_DenseNet.md index 53092bd74014c42fae872cd20e0ea24ff858989c..25f61476e43d752f2426c000885166963de7ebc4 100644 --- a/docs/zh_CN/models/DPN_DenseNet.md +++ b/docs/zh_CN/models/DPN_DenseNet.md @@ -4,13 +4,15 @@ DenseNet是2017年CVPR best paper提出的一种新的网络结构，该网络设计了一种新的跨层连接的block，即dense-block。相比ResNet中的bottleneck，dense-block设计了一个更激进的密集连接机制，即互相连接所有的层，每个层都会接受其前面所有层作为其额外的输入。DenseNet将所有的dense-block堆叠，组合成了一个密集连接型网络。密集的连接方式使得DenseNe更容易进行梯度的反向传播，使得网络更容易训练。 DPN的全称是Dual Path Networks，即双通道网络。该网络是由DenseNet和ResNeXt结合的一个网络，其证明了DenseNet能从靠前的层级中提取到新的特征，而ResNeXt本质上是对之前层级中已提取特征的复用。作者进一步分析发现，ResNeXt对特征有高复用率，但冗余度低，DenseNet能创造新特征，但冗余度高。结合二者结构的优势，作者设计了DPN网络。最终DPN网络在同样FLOPS和参数量下，取得了比ResNeXt与DenseNet更好的结果。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -![](../../images/models/DPN.png.flops.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.DPN.flops.png) -![](../../images/models/DPN.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.DPN.params.png) -![](../../images/models/DPN.png.fp32.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.DPN.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs4.DPN.png) 目前PaddleClas开源的这两类模型的预训练模型一共有10个，其指标如上图所示，可以看到，在相同的FLOPS和参数量下，相比DenseNet，DPN拥有更高的精度。但是由于DPN有更多的分支，所以其推理速度要慢于DenseNet。由于DenseNet264的网络层数最深，所以该网络是DenseNet系列模型中参数量最大的网络，DenseNet161的网络的宽度最大，导致其是该系列中网络中计算量最大、精度最高的网络。从推理速度来看，计算量大且精度高的的DenseNet161比DenseNet264具有更快的速度，所以其比DenseNet264具有更大的优势。 @@ -34,9 +36,9 @@ DPN的全称是Dual Path Networks，即双通道网络。该网络是由DenseNet -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |-------------|-----------|-------------------|--------------------------| | DenseNet121 | 224 | 256 | 4.371 | | DenseNet161 | 224 | 256 | 8.863 | @@ -48,3 +50,20 @@ DPN的全称是Dual Path Networks，即双通道网络。该网络是由DenseNet | DPN98 | 224 | 256 | 21.057 | | DPN107 | 224 | 256 | 28.685 | | DPN131 | 224 | 256 | 28.083 | + + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| DenseNet121 | 224 | 256 | 4.16436 | 7.2126 | 10.50221 | 4.40447 | 9.32623 | 15.25175 | +| DenseNet161 | 224 | 256 | 9.27249 | 14.25326 | 20.19849 | 10.39152 | 22.15555 | 35.78443 | +| DenseNet169 | 224 | 256 | 6.11395 | 10.28747 | 13.68717 | 6.43598 | 12.98832 | 20.41964 | +| DenseNet201 | 224 | 256 | 7.9617 | 13.4171 | 17.41949 | 8.20652 | 17.45838 | 27.06309 | +| DenseNet264 | 224 | 256 | 11.70074 | 19.69375 | 24.79545 | 12.14722 | 26.27707 | 40.01905 | +| DPN68 | 224 | 256 | 11.7827 | 13.12652 | 16.19213 | 11.64915 | 12.82807 | 18.57113 | +| DPN92 | 224 | 256 | 18.56026 | 20.35983 | 29.89544 | 18.15746 | 23.87545 | 38.68821 | +| DPN98 | 224 | 256 | 21.70508 | 24.7755 | 40.93595 | 21.18196 | 33.23925 | 62.77751 | +| DPN107 | 224 | 256 | 27.84462 | 34.83217 | 60.67903 | 27.62046 | 52.65353 | 100.11721 | +| DPN131 | 224 | 256 | 28.58941 | 33.01078 | 55.65146 | 28.33119 | 46.19439 | 89.24904 | diff --git a/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md b/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md index eadd17683d0219b4ee72d7aba71dc04743896ea5..3cbe4009dfd17cbc88ead5f3bb91d4ed9cc7470d 100644 --- a/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md +++ b/docs/zh_CN/models/EfficientNet_and_ResNeXt101_wsl.md @@ -6,13 +6,15 @@ EfficientNet是Google于2019年发布的一个基于NAS的轻量级网络，其 ResNeXt是facebook于2016年提出的一种对ResNet的改进版网络。在2019年，facebook通过弱监督学习研究了该系列网络在ImageNet上的精度上限，为了区别之前的ResNeXt网络，该系列网络的后缀为wsl，其中wsl是弱监督学习（weakly-supervised-learning）的简称。为了能有更强的特征提取能力，研究者将其网络宽度进一步放大，其中最大的ResNeXt101_32x48d_wsl拥有8亿个参数，将其在9.4亿的弱标签图片下训练并在ImageNet-1k上做finetune，最终在ImageNet-1k的top-1达到了85.4%，这也是迄今为止在ImageNet-1k的数据集上以224x224的分辨率下精度最高的网络。Fix-ResNeXt中，作者使用了更大的图像分辨率，针对训练图片和验证图片数据预处理不一致的情况下做了专门的Fix策略，并使得ResNeXt101_32x48d_wsl拥有了更高的精度，由于其用到了Fix策略，故命名为Fix-ResNeXt101_32x48d_wsl。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -![](../../images/models/EfficientNet.png.flops.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.flops.png) -![](../../images/models/EfficientNet.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.EfficientNet.params.png) -![](../../images/models/EfficientNet.png.fp32.png) +![](../../images/models/T4_benchmark/t4.fp32.bs1.EfficientNet.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs1.EfficientNet.png) 目前PaddleClas开源的这两类模型的预训练模型一共有14个。从上图中可以看出EfficientNet系列网络优势非常明显，ResNeXt101_wsl系列模型由于用到了更多的数据，最终的精度也更高。EfficientNet_B0_Small是去掉了SE_block的EfficientNet_B0，其具有更快的推理速度。 @@ -36,9 +38,9 @@ ResNeXt是facebook于2016年提出的一种对ResNet的改进版网络。在2019 | EfficientNetB0_
small | 0.758 | 0.926 | | | 0.720 | 4.650 | -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |-------------------------------|-----------|-------------------|--------------------------| | ResNeXt101_
32x8d_wsl | 224 | 256 | 19.127 | | ResNeXt101_
32x16d_wsl | 224 | 256 | 23.629 | @@ -54,3 +56,24 @@ ResNeXt是facebook于2016年提出的一种对ResNet的改进版网络。在2019 | EfficientNetB6 | 528 | 560 | 18.381 | | EfficientNetB7 | 600 | 632 | 27.817 | | EfficientNetB0_
small | 224 | 256 | 1.692 | + + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|---------------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| ResNeXt101_
32x8d_wsl | 224 | 256 | 18.19374 | 21.93529 | 34.67802 | 18.52528 | 34.25319 | 67.2283 | +| ResNeXt101_
32x16d_wsl | 224 | 256 | 18.52609 | 36.8288 | 62.79947 | 25.60395 | 71.88384 | 137.62327 | +| ResNeXt101_
32x32d_wsl | 224 | 256 | 33.51391 | 70.09682 | 125.81884 | 54.87396 | 160.04337 | 316.17718 | +| ResNeXt101_
32x48d_wsl | 224 | 256 | 50.97681 | 137.60926 | 190.82628 | 99.01698256 | 315.91261 | 551.83695 | +| Fix_ResNeXt101_
32x48d_wsl | 320 | 320 | 78.62869 | 191.76039 | 317.15436 | 160.0838242 | 595.99296 | 1151.47384 | +| EfficientNetB0 | 224 | 256 | 3.40122 | 5.95851 | 9.10801 | 3.442 | 6.11476 | 9.3304 | +| EfficientNetB1 | 240 | 272 | 5.25172 | 9.10233 | 14.11319 | 5.3322 | 9.41795 | 14.60388 | +| EfficientNetB2 | 260 | 292 | 5.91052 | 10.5898 | 17.38106 | 6.29351 | 10.95702 | 17.75308 | +| EfficientNetB3 | 300 | 332 | 7.69582 | 16.02548 | 27.4447 | 7.67749 | 16.53288 | 28.5939 | +| EfficientNetB4 | 380 | 412 | 11.55585 | 29.44261 | 53.97363 | 12.15894 | 30.94567 | 57.38511 | +| EfficientNetB5 | 456 | 488 | 19.63083 | 56.52299 | - | 20.48571 | 61.60252 | - | +| EfficientNetB6 | 528 | 560 | 30.05911 | - | - | 32.62402 | - | - | +| EfficientNetB7 | 600 | 632 | 47.86087 | - | - | 53.93823 | - | - | +| EfficientNetB0_small | 224 | 256 | 2.39166 | 4.36748 | 6.96002 | 2.3076 | 4.71886 | 7.21888 | diff --git a/docs/zh_CN/models/HRNet.md b/docs/zh_CN/models/HRNet.md index c33fb0fa025fd9b7f1993186d622c2acfdd22acb..f694f7b0c1d6d6c9b195fa61aa0cc9544564859d 100644 --- a/docs/zh_CN/models/HRNet.md +++ b/docs/zh_CN/models/HRNet.md @@ -3,13 +3,16 @@ ## 概述 HRNet是2019年由微软亚洲研究院提出的一种全新的神经网络，不同于以往的卷积神经网络，该网络在网络深层仍然可以保持高分辨率，因此预测的关键点热图更准确，在空间上也更精确。此外，该网络在对分辨率敏感的其他视觉任务中，如检测、分割等，表现尤为优异。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -![](../../images/models/HRNet.png.flops.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.HRNet.flops.png) -![](../../images/models/HRNet.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.HRNet.params.png) + +![](../../images/models/T4_benchmark/t4.fp32.bs4.HRNet.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs4.HRNet.png) -![](../../images/models/HRNet.png.fp32.png) 目前PaddleClas开源的这类模型的预训练模型一共有7个，其指标如图所示，其中HRNet_W48_C指标精度异常的原因可能是因为网络训练的正常波动。 @@ -26,9 +29,9 @@ HRNet是2019年由微软亚洲研究院提出的一种全新的神经网络， | HRNet_W64_C | 0.793 | 0.946 | 0.795 | 0.946 | 57.830 | 128.060 | -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |-------------|-----------|-------------------|--------------------------| | HRNet_W18_C | 224 | 256 | 7.368 | | HRNet_W30_C | 224 | 256 | 9.402 | @@ -37,3 +40,18 @@ HRNet是2019年由微软亚洲研究院提出的一种全新的神经网络， | HRNet_W44_C | 224 | 256 | 11.497 | | HRNet_W48_C | 224 | 256 | 12.165 | | HRNet_W64_C | 224 | 256 | 15.003 | + + + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| HRNet_W18_C | 224 | 256 | 6.79093 | 11.50986 | 17.67244 | 7.40636 | 13.29752 | 23.33445 | +| HRNet_W30_C | 224 | 256 | 8.98077 | 14.08082 | 21.23527 | 9.57594 | 17.35485 | 32.6933 | +| HRNet_W32_C | 224 | 256 | 8.82415 | 14.21462 | 21.19804 | 9.49807 | 17.72921 | 32.96305 | +| HRNet_W40_C | 224 | 256 | 11.4229 | 19.1595 | 30.47984 | 12.12202 | 25.68184 | 48.90623 | +| HRNet_W44_C | 224 | 256 | 12.25778 | 22.75456 | 32.61275 | 13.19858 | 32.25202 | 59.09871 | +| HRNet_W48_C | 224 | 256 | 12.65015 | 23.12886 | 33.37859 | 13.70761 | 34.43572 | 63.01219 | +| HRNet_W64_C | 224 | 256 | 15.10428 | 27.68901 | 40.4198 | 17.57527 | 47.9533 | 97.11228 | diff --git a/docs/zh_CN/models/Inception.md b/docs/zh_CN/models/Inception.md index 8c9333656b1e78199b1f29feff17ef5d15c593b8..b85c2bf1b5068936daa3091215540b866d4d31b3 100644 --- a/docs/zh_CN/models/Inception.md +++ b/docs/zh_CN/models/Inception.md @@ -9,13 +9,15 @@ Xception 是 Google 继 Inception 后提出的对 InceptionV3 的另一种改进 InceptionV4是2016年由Google设计的新的神经网络，当时残差结构风靡一时，但是作者认为仅使用Inception 结构也可以达到很高的性能。InceptionV4使用了更多的Inception module，在ImageNet上的精度再创新高。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -![](../../images/models/Inception.png.flops.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.Inception.flops.png) -![](../../images/models/Inception.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.Inception.params.png) -![](../../images/models/Inception.png.fp32.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.Inception.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs4.Inception.png) 上图反映了Xception系列和InceptionV4的精度和其他指标的关系。其中Xception_deeplab与论文结构保持一致，Xception是PaddleClas的改进模型，在预测速度基本不变的情况下，精度提升约0.6%。关于该改进模型的详细介绍正在持续更新中，敬请期待。 @@ -35,14 +37,28 @@ InceptionV4是2016年由Google设计的新的神经网络，当时残差结构 -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |------------------------|-----------|-------------------|--------------------------| | GoogLeNet | 224 | 256 | 1.807 | | Xception41 | 299 | 320 | 3.972 | -| Xception41
_deeplab | 299 | 320 | 4.408 | +| Xception41_
deeplab | 299 | 320 | 4.408 | | Xception65 | 299 | 320 | 6.174 | -| Xception65
_deeplab | 299 | 320 | 6.464 | +| Xception65_
deeplab | 299 | 320 | 6.464 | | Xception71 | 299 | 320 | 6.782 | | InceptionV4 | 299 | 320 | 11.141 | + + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|--------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| GoogLeNet | 299 | 320 | 1.75451 | 3.39931 | 4.71909 | 1.88038 | 4.48882 | 6.94035 | +| Xception41 | 299 | 320 | 2.91192 | 7.86878 | 15.53685 | 4.96939 | 17.01361 | 32.67831 | +| Xception41_
deeplab | 299 | 320 | 2.85934 | 7.2075 | 14.01406 | 5.33541 | 17.55938 | 33.76232 | +| Xception65 | 299 | 320 | 4.30126 | 11.58371 | 23.22213 | 7.26158 | 25.88778 | 53.45426 | +| Xception65_
deeplab | 299 | 320 | 4.06803 | 9.72694 | 19.477 | 7.60208 | 26.03699 | 54.74724 | +| Xception71 | 299 | 320 | 4.80889 | 13.5624 | 27.18822 | 8.72457 | 31.55549 | 69.31018 | +| InceptionV4 | 299 | 320 | 9.50821 | 13.72104 | 20.27447 | 12.99342 | 25.23416 | 43.56121 | diff --git a/docs/zh_CN/models/Mobile.md b/docs/zh_CN/models/Mobile.md index 6cf4ed90241ca959e7c5b66f9a14a8c415e7dd87..3c0ebe37693b5564fbc654d99948d69b04b97a85 100644 --- a/docs/zh_CN/models/Mobile.md +++ b/docs/zh_CN/models/Mobile.md @@ -8,10 +8,16 @@ MobileNetV2是Google继MobileNetV1提出的一种轻量级网络。相比MobileN ShuffleNet系列网络是旷视提出的轻量化网络结构，到目前为止，该系列网络一共有两种典型的结构，即ShuffleNetV1与ShuffleNetV2。ShuffleNet中的Channel Shuffle操作可以将组间的信息进行交换，并且可以实现端到端的训练。在ShuffleNetV2的论文中，作者提出了设计轻量级网络的四大准则，并且根据四大准则与ShuffleNetV1的不足，设计了ShuffleNetV2网络。 MobileNetV3是Google于2019年提出的一种基于NAS的新的轻量级网络，为了进一步提升效果，将relu和sigmoid激活函数分别替换为hard_swish与hard_sigmoid激活函数，同时引入了一些专门减小网络计算量的改进策略。 + ![](../../images/models/mobile_arm_top1.png) + ![](../../images/models/mobile_arm_storage.png) -![](../../images/models/mobile_trt.png.flops.png) -![](../../images/models/mobile_trt.png.params.png) + +![](../../images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.flops.png) + +![](../../images/models/T4_benchmark/t4.fp32.bs4.mobile_trt.params.png) + + 目前PaddleClas开源的的移动端系列的预训练模型一共有32个，其指标如图所示。从图片可以看出，越新的轻量级模型往往有更优的表现，MobileNetV3代表了目前最新的轻量级神经网络结构。在MobileNetV3中，作者为了获得更高的精度，在global-avg-pooling后使用了1x1的卷积。该操作大幅提升了参数量但对计算量影响不大，所以如果从存储角度评价模型的优异程度，MobileNetV3优势不是很大，但由于其更小的计算量，使得其有更快的推理速度。此外，我们模型库中的ssld蒸馏模型表现优异，从各个考量角度下，都刷新了当前轻量级模型的精度。由于MobileNetV3模型结构复杂，分支较多，对GPU并不友好，GPU预测速度不如MobileNetV1。 @@ -53,9 +59,9 @@ MobileNetV3是Google于2019年提出的一种基于NAS的新的轻量级网络 | ShuffleNetV2_swish | 0.700 | 0.892 | | | 0.290 | 2.260 | -## CPU预测速度和存储大小 +## 基于SD855的预测速度和存储大小 -| Models | batch_size=1(ms) | Storage Size(M) | +| Models | Batch Size=1(ms) | Storage Size(M) | |:--:|:--:|:--:| | MobileNetV1_x0_25 | 3.220 | 1.900 | | MobileNetV1_x0_5 | 9.580 | 5.200 | @@ -89,3 +95,40 @@ MobileNetV3是Google于2019年提出的一种基于NAS的新的轻量级网络 | ShuffleNetV2_x1_5 | 19.352 | 14.000 | | ShuffleNetV2_x2_0 | 34.770 | 28.000 | | ShuffleNetV2_swish | 16.023 | 9.100 | + + +## 基于T4 GPU的预测速度 + +| Models | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-----------------------------|-----------------------|-----------------------|-----------------------|-----------------------|-----------------------|-----------------------| +| MobileNetV1_x0_25 | 0.68422 | 1.13021 | 1.72095 | 0.67274 | 1.226 | 1.84096 | +| MobileNetV1_x0_5 | 0.69326 | 1.09027 | 1.84746 | 0.69947 | 1.43045 | 2.39353 | +| MobileNetV1_x0_75 | 0.6793 | 1.29524 | 2.15495 | 0.79844 | 1.86205 | 3.064 | +| MobileNetV1 | 0.71942 | 1.45018 | 2.47953 | 0.91164 | 2.26871 | 3.90797 | +| MobileNetV1_ssld | 0.71942 | 1.45018 | 2.47953 | 0.91164 | 2.26871 | 3.90797 | +| MobileNetV2_x0_25 | 2.85399 | 3.62405 | 4.29952 | 2.81989 | 3.52695 | 4.2432 | +| MobileNetV2_x0_5 | 2.84258 | 3.1511 | 4.10267 | 2.80264 | 3.65284 | 4.31737 | +| MobileNetV2_x0_75 | 2.82183 | 3.27622 | 4.98161 | 2.86538 | 3.55198 | 5.10678 | +| MobileNetV2 | 2.78603 | 3.71982 | 6.27879 | 2.62398 | 3.54429 | 6.41178 | +| MobileNetV2_x1_5 | 2.81852 | 4.87434 | 8.97934 | 2.79398 | 5.30149 | 9.30899 | +| MobileNetV2_x2_0 | 3.65197 | 6.32329 | 11.644 | 3.29788 | 7.08644 | 12.45375 | +| MobileNetV2_ssld | 2.78603 | 3.71982 | 6.27879 | 2.62398 | 3.54429 | 6.41178 | +| MobileNetV3_large_x1_25 | 2.34387 | 3.16103 | 4.79742 | 2.35117 | 3.44903 | 5.45658 | +| MobileNetV3_large_x1_0 | 2.20149 | 3.08423 | 4.07779 | 2.04296 | 2.9322 | 4.53184 | +| MobileNetV3_large_x0_75 | 2.1058 | 2.61426 | 3.61021 | 2.0006 | 2.56987 | 3.78005 | +| MobileNetV3_large_x0_5 | 2.06934 | 2.77341 | 3.35313 | 2.11199 | 2.88172 | 3.19029 | +| MobileNetV3_large_x0_35 | 2.14965 | 2.7868 | 3.36145 | 1.9041 | 2.62951 | 3.26036 | +| MobileNetV3_small_x1_25 | 2.06817 | 2.90193 | 3.5245 | 2.02916 | 2.91866 | 3.34528 | +| MobileNetV3_small_x1_0 | 1.73933 | 2.59478 | 3.40276 | 1.74527 | 2.63565 | 3.28124 | +| MobileNetV3_small_x0_75 | 1.80617 | 2.64646 | 3.24513 | 1.93697 | 2.64285 | 3.32797 | +| MobileNetV3_small_x0_5 | 1.95001 | 2.74014 | 3.39485 | 1.88406 | 2.99601 | 3.3908 | +| MobileNetV3_small_x0_35 | 2.10683 | 2.94267 | 3.44254 | 1.94427 | 2.94116 | 3.41082 | +| MobileNetV3_large_x1_0_ssld | 2.20149 | 3.08423 | 4.07779 | 2.04296 | 2.9322 | 4.53184 | +| MobileNetV3_small_x1_0_ssld | 1.73933 | 2.59478 | 3.40276 | 1.74527 | 2.63565 | 3.28124 | +| ShuffleNetV2 | 1.95064 | 2.15928 | 2.97169 | 1.89436 | 2.26339 | 3.17615 | +| ShuffleNetV2_x0_25 | 1.43242 | 2.38172 | 2.96768 | 1.48698 | 2.29085 | 2.90284 | +| ShuffleNetV2_x0_33 | 1.69008 | 2.65706 | 2.97373 | 1.75526 | 2.85557 | 3.09688 | +| ShuffleNetV2_x0_5 | 1.48073 | 2.28174 | 2.85436 | 1.59055 | 2.18708 | 3.09141 | +| ShuffleNetV2_x1_5 | 1.51054 | 2.4565 | 3.41738 | 1.45389 | 2.5203 | 3.99872 | +| ShuffleNetV2_x2_0 | 1.95616 | 2.44751 | 4.19173 | 2.15654 | 3.18247 | 5.46893 | +| ShuffleNetV2_swish | 2.50213 | 2.92881 | 3.474 | 2.5129 | 2.97422 | 3.69357 | diff --git a/docs/zh_CN/models/Others.md b/docs/zh_CN/models/Others.md index 35cabf68d5e2d05ec336d7e5cdfac3d363477971..c24f76652bc5e322df2533bbf8c59889bb420910 100644 --- a/docs/zh_CN/models/Others.md +++ b/docs/zh_CN/models/Others.md @@ -27,10 +27,10 @@ DarkNet53是YOLO作者在论文设计的用于目标检测的backbone，该网 -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |---------------------------|-----------|-------------------|----------------------| | AlexNet | 224 | 256 | 1.176 | | SqueezeNet1_0 | 224 | 256 | 0.860 | @@ -41,3 +41,20 @@ DarkNet53是YOLO作者在论文设计的用于目标检测的backbone，该网 | VGG19 | 224 | 256 | 3.076 | | DarkNet53 | 256 | 256 | 3.139 | | ResNet50_ACNet
_deploy | 224 | 256 | 5.626 | + + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-----------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| AlexNet | 224 | 256 | 1.06447 | 1.70435 | 2.38402 | 1.44993 | 2.46696 | 3.72085 | +| SqueezeNet1_0 | 224 | 256 | 0.97162 | 2.06719 | 3.67499 | 0.96736 | 2.53221 | 4.54047 | +| SqueezeNet1_1 | 224 | 256 | 0.81378 | 1.62919 | 2.68044 | 0.76032 | 1.877 | 3.15298 | +| VGG11 | 224 | 256 | 2.24408 | 4.67794 | 7.6568 | 3.90412 | 9.51147 | 17.14168 | +| VGG13 | 224 | 256 | 2.58589 | 5.82708 | 10.03591 | 4.64684 | 12.61558 | 23.70015 | +| VGG16 | 224 | 256 | 3.13237 | 7.19257 | 12.50913 | 5.61769 | 16.40064 | 32.03939 | +| VGG19 | 224 | 256 | 3.69987 | 8.59168 | 15.07866 | 6.65221 | 20.4334 | 41.55902 | +| DarkNet53 | 256 | 256 | 3.18101 | 5.88419 | 10.14964 | 4.10829 | 12.1714 | 22.15266 | +| ResNet50_ACNet | 256 | 256 | 3.89002 | 4.58195 | 9.01095 | 5.33395 | 10.96843 | 18.70368 | +| ResNet50_ACNet_deploy | 224 | 256 | 2.6823 | 5.944 | 7.16655 | 3.49161 | 7.78374 | 13.94361 | diff --git a/docs/zh_CN/models/ResNet_and_vd.md b/docs/zh_CN/models/ResNet_and_vd.md index bc2946e99bc270c084c146808188da53c475ebae..ea045f12ca545ca0cea2229fcfc1993fd50ec77b 100644 --- a/docs/zh_CN/models/ResNet_and_vd.md +++ b/docs/zh_CN/models/ResNet_and_vd.md @@ -10,18 +10,19 @@ ResNet系列模型是在2015年提出的，一举在ILSVRC2015比赛中取得冠 其中，ResNet50_vd_v2与ResNet50_vd_ssld采用了知识蒸馏，保证模型结构不变的情况下，进一步提升了模型的精度，具体地，ResNet50_vd_v2的teacher模型是ResNet152_vd（top1准确率80.59%），数据选用的是ImageNet-1k的训练集，ResNet50_vd_ssld的teacher模型是ResNeXt101_32x16d_wsl（top1准确率84.2%），数据选用结合了ImageNet-1k的训练集和ImageNet-22k挖掘的400万数据。知识蒸馏的具体方法正在持续更新中。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +![](../../images/models/T4_benchmark/t4.fp32.bs4.ResNet.flops.png) -![](../../images/models/ResNet.png.flops.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.ResNet.params.png) -![](../../images/models/ResNet.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.ResNet.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs4.ResNet.png) -![](../../images/models/ResNet.png.fp32.png) 通过上述曲线可以看出，层数越多，准确率越高，但是相应的参数量、计算量和延时都会增加。ResNet50_vd_ssld通过用更强的teacher和更多的数据，将其在ImageNet-1k上的验证集top-1精度进一步提高，达到了82.39%，刷新了ResNet50系列模型的精度。 -**注意**：所有模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。 ## 精度、FLOPS和参数量 @@ -46,9 +47,9 @@ ResNet系列模型是在2015年提出的，一举在ILSVRC2015比赛中取得冠 -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |------------------|-----------|-------------------|--------------------------| | ResNet18 | 224 | 256 | 1.499 | | ResNet18_vd | 224 | 256 | 1.603 | @@ -65,3 +66,24 @@ ResNet系列模型是在2015年提出的，一举在ILSVRC2015比赛中取得冠 | ResNet200_vd | 224 | 256 | 8.885 | | ResNet50_vd_ssld | 224 | 256 | 3.165 | | ResNet101_vd_ssld | 224 | 256 | 5.252 | + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| ResNet18 | 224 | 256 | 1.3568 | 2.5225 | 3.61904 | 1.45606 | 3.56305 | 6.28798 | +| ResNet18_vd | 224 | 256 | 1.39593 | 2.69063 | 3.88267 | 1.54557 | 3.85363 | 6.88121 | +| ResNet34 | 224 | 256 | 2.23092 | 4.10205 | 5.54904 | 2.34957 | 5.89821 | 10.73451 | +| ResNet34_vd | 224 | 256 | 2.23992 | 4.22246 | 5.79534 | 2.43427 | 6.22257 | 11.44906 | +| ResNet50 | 224 | 256 | 2.63824 | 4.63802 | 7.02444 | 3.47712 | 7.84421 | 13.90633 | +| ResNet50_vc | 224 | 256 | 2.67064 | 4.72372 | 7.17204 | 3.52346 | 8.10725 | 14.45577 | +| ResNet50_vd | 224 | 256 | 2.65164 | 4.84109 | 7.46225 | 3.53131 | 8.09057 | 14.45965 | +| ResNet50_vd_v2 | 224 | 256 | 2.65164 | 4.84109 | 7.46225 | 3.53131 | 8.09057 | 14.45965 | +| ResNet101 | 224 | 256 | 5.04037 | 7.73673 | 10.8936 | 6.07125 | 13.40573 | 24.3597 | +| ResNet101_vd | 224 | 256 | 5.05972 | 7.83685 | 11.34235 | 6.11704 | 13.76222 | 25.11071 | +| ResNet152 | 224 | 256 | 7.28665 | 10.62001 | 14.90317 | 8.50198 | 19.17073 | 35.78384 | +| ResNet152_vd | 224 | 256 | 7.29127 | 10.86137 | 15.32444 | 8.54376 | 19.52157 | 36.64445 | +| ResNet200_vd | 224 | 256 | 9.36026 | 13.5474 | 19.0725 | 10.80619 | 25.01731 | 48.81399 | +| ResNet50_vd_ssld | 224 | 256 | 2.65164 | 4.84109 | 7.46225 | 3.53131 | 8.09057 | 14.45965 | +| ResNet101_vd_ssld | 224 | 256 | 5.05972 | 7.83685 | 11.34235 | 6.11704 | 13.76222 | 25.11071 | diff --git a/docs/zh_CN/models/SEResNext_and_Res2Net.md b/docs/zh_CN/models/SEResNext_and_Res2Net.md index 90955354aece52a3770c57c6d387c4bdf8238453..1a8c125ee931a2bc036834d122d529c26264bb66 100644 --- a/docs/zh_CN/models/SEResNext_and_Res2Net.md +++ b/docs/zh_CN/models/SEResNext_and_Res2Net.md @@ -7,18 +7,20 @@ SENet是2017年ImageNet分类比赛的冠军方案，其提出了一个全新的 Res2Net是2019年提出的一种全新的对ResNet的改进方案，该方案可以和现有其他优秀模块轻松整合，在不增加计算负载量的情况下，在ImageNet、CIFAR-100等数据集上的测试性能超过了ResNet。Res2Net结构简单，性能优越，进一步探索了CNN在更细粒度级别的多尺度表示能力。Res2Net揭示了一个新的提升模型精度的维度，即scale，其是除了深度、宽度和基数的现有维度之外另外一个必不可少的更有效的因素。该网络在其他视觉任务如目标检测、图像分割等也有相当不错的表现。 -该系列模型的FLOPS、参数量以及FP32预测耗时如下图所示。 +该系列模型的FLOPS、参数量以及T4 GPU上的预测耗时如下图所示。 -![](../../images/models/SeResNeXt.png.flops.png) -![](../../images/models/SeResNeXt.png.params.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.flops.png) -![](../../images/models/SeResNeXt.png.fp32.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.params.png) + +![](../../images/models/T4_benchmark/t4.fp32.bs4.SeResNeXt.png) + +![](../../images/models/T4_benchmark/t4.fp16.bs4.SeResNeXt.png) -目前PaddleClas开源的这三类的预训练模型一共有24个，其指标如图所示，从图中可以看出，在同样Flops和Params下，改进版的模型往往有更高的精度，但是推理速度往往不如ResNet系列。另一方面，Res2Net表现也较为优秀，相比ResNeXt中的group操作、SEResNet中的SE结构操作，Res2Net在相同Flops、Params和推理速度下往往精度更佳。 +目前PaddleClas开源的这三类的预训练模型一共有24个，其指标如图所示，从图中可以看出，在同样Flops和Params下，改进版的模型往往有更高的精度，但是推理速度往往不如ResNet系列。另一方面，Res2Net表现也较为优秀，相比ResNeXt中的group操作、SEResNet中的SE结构操作，Res2Net在相同Flops、Params和推理速度下往往精度更佳。 -**注意**：所有模型在预测时，图像的crop_size设置为224，resize_short_size设置为256。 ## 精度、FLOPS和参数量 @@ -52,9 +54,9 @@ Res2Net是2019年提出的一种全新的对ResNet的改进方案，该方案可 -## FP32预测速度 +## 基于V100 GPU的预测速度 -| Models | Crop Size | Resize Short Size | Batch Size=1
(ms) | +| Models | Crop Size | Resize Short Size | FP32
Batch Size=1
(ms) | |-----------------------|-----------|-------------------|--------------------------| | Res2Net50_26w_4s | 224 | 256 | 4.148 | | Res2Net50_vd_26w_4s | 224 | 256 | 4.172 | @@ -80,3 +82,33 @@ Res2Net是2019年提出的一种全新的对ResNet的改进方案，该方案可 | SE_ResNeXt50_vd_32x4d | 224 | 256 | 9.011 | | SE_ResNeXt101_32x4d | 224 | 256 | 19.204 | | SENet154_vd | 224 | 256 | 50.406 | + + +## 基于T4 GPU的预测速度 + +| Models | Crop Size | Resize Short Size | FP16
Batch Size=1
(ms) | FP16
Batch Size=4
(ms) | FP16
Batch Size=8
(ms) | FP32
Batch Size=1
(ms) | FP32
Batch Size=4
(ms) | FP32
Batch Size=8
(ms) | +|-----------------------|-----------|-------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------|------------------------------| +| Res2Net50_26w_4s | 224 | 256 | 3.56067 | 6.61827 | 11.41566 | 4.47188 | 9.65722 | 17.54535 | +| Res2Net50_vd_26w_4s | 224 | 256 | 3.69221 | 6.94419 | 11.92441 | 4.52712 | 9.93247 | 18.16928 | +| Res2Net50_14w_8s | 224 | 256 | 4.45745 | 7.69847 | 12.30935 | 5.4026 | 10.60273 | 18.01234 | +| Res2Net101_vd_26w_4s | 224 | 256 | 6.53122 | 10.81895 | 18.94395 | 8.08729 | 17.31208 | 31.95762 | +| Res2Net200_vd_26w_4s | 224 | 256 | 11.66671 | 18.93953 | 33.19188 | 14.67806 | 32.35032 | 63.65899 | +| ResNeXt50_32x4d | 224 | 256 | 7.61087 | 8.88918 | 12.99674 | 7.56327 | 10.6134 | 18.46915 | +| ResNeXt50_vd_32x4d | 224 | 256 | 7.69065 | 8.94014 | 13.4088 | 7.62044 | 11.03385 | 19.15339 | +| ResNeXt50_64x4d | 224 | 256 | 13.78688 | 15.84655 | 21.79537 | 13.80962 | 18.4712 | 33.49843 | +| ResNeXt50_vd_64x4d | 224 | 256 | 13.79538 | 15.22201 | 22.27045 | 13.94449 | 18.88759 | 34.28889 | +| ResNeXt101_32x4d | 224 | 256 | 16.59777 | 17.93153 | 21.36541 | 16.21503 | 19.96568 | 33.76831 | +| ResNeXt101_vd_32x4d | 224 | 256 | 16.36909 | 17.45681 | 22.10216 | 16.28103 | 20.25611 | 34.37152 | +| ResNeXt101_64x4d | 224 | 256 | 30.12355 | 32.46823 | 38.41901 | 30.4788 | 36.29801 | 68.85559 | +| ResNeXt101_vd_64x4d | 224 | 256 | 30.34022 | 32.27869 | 38.72523 | 30.40456 | 36.77324 | 69.66021 | +| ResNeXt152_32x4d | 224 | 256 | 25.26417 | 26.57001 | 30.67834 | 24.86299 | 29.36764 | 52.09426 | +| ResNeXt152_vd_32x4d | 224 | 256 | 25.11196 | 26.70515 | 31.72636 | 25.03258 | 30.08987 | 52.64429 | +| ResNeXt152_64x4d | 224 | 256 | 46.58293 | 48.34563 | 56.97961 | 46.7564 | 56.34108 | 106.11736 | +| ResNeXt152_vd_64x4d | 224 | 256 | 47.68447 | 48.91406 | 57.29329 | 47.18638 | 57.16257 | 107.26288 | +| SE_ResNet18_vd | 224 | 256 | 1.61823 | 3.1391 | 4.60282 | 1.7691 | 4.19877 | 7.5331 | +| SE_ResNet34_vd | 224 | 256 | 2.67518 | 5.04694 | 7.18946 | 2.88559 | 7.03291 | 12.73502 | +| SE_ResNet50_vd | 224 | 256 | 3.65394 | 7.568 | 12.52793 | 4.28393 | 10.38846 | 18.33154 | +| SE_ResNeXt50_32x4d | 224 | 256 | 9.06957 | 11.37898 | 18.86282 | 8.74121 | 13.563 | 23.01954 | +| SE_ResNeXt50_vd_32x4d | 224 | 256 | 9.25016 | 11.85045 | 25.57004 | 9.17134 | 14.76192 | 19.914 | +| SE_ResNeXt101_32x4d | 224 | 256 | 19.34455 | 20.6104 | 32.20432 | 18.82604 | 25.31814 | 41.97758 | +| SENet154_vd | 224 | 256 | 49.85733 | 54.37267 | 74.70447 | 53.79794 | 66.31684 | 121.59885 | diff --git a/docs/zh_CN/models/models_intro.md b/docs/zh_CN/models/models_intro.md index da653ae30194cfbd05baa4051998b0fe916b9035..935db13a158db7466986ec08ab09f8bb1a0750f9 100644 --- a/docs/zh_CN/models/models_intro.md +++ b/docs/zh_CN/models/models_intro.md @@ -23,7 +23,8 @@ python tools/infer/predict.py \ --batch_size=1 ``` -![](../../images/models/main_fps_top1.png) +![](../../images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png) + ![](../../images/models/mobile_arm_top1.png) @@ -193,7 +194,7 @@ python tools/infer/predict.py \ - [VGG16](https://paddle-imagenet-models-name.bj.bcebos.com/VGG16_pretrained.tar) - [VGG19](https://paddle-imagenet-models-name.bj.bcebos.com/VGG19_pretrained.tar) - DarkNet系列[[21](#ref21)]([论文地址](https://arxiv.org/abs/1506.02640)) - - [DarkNet53](https://paddle-imagenet-models-name.bj.bcebos.com/DarkNet53_pretrained.tar) + - [DarkNet53](https://paddle-imagenet-models-name.bj.bcebos.com/DarkNet53_ImageNet1k_pretrained.tar) - ACNet系列[[22](#ref22)]([论文地址](https://arxiv.org/abs/1908.03930)) - [ResNet50_ACNet_deploy](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_ACNet_deploy_pretrained.tar) diff --git a/docs/zh_CN/update_history.md b/docs/zh_CN/update_history.md index a3ba148b2be262e920e8a8f06c87c6b429b30413..b2ab286109354b9a75defc3ec5b29fbf85596349 100644 --- a/docs/zh_CN/update_history.md +++ b/docs/zh_CN/update_history.md @@ -1,3 +1,12 @@ # 更新日志 -* 2020.04.10: 第一次提交 +* 2020.05.17 + * 添加混合精度训练。 + +* 2020.05.09 + * 添加Paddle Serving使用文档。 + * 添加Paddle-Lite使用文档。 + * 添加T4 GPU的FP32/FP16预测速度benchmark。 + +* 2020.04.10: + * 第一次提交。 diff --git a/ppcls/data/imaug/operators.py b/ppcls/data/imaug/operators.py index 98e89ca8307d44ee156deaaeb0bf64ffbd6db58a..a8454740f4f6047f16104b32c5a06b6f5cc1e327 100644 --- a/ppcls/data/imaug/operators.py +++ b/ppcls/data/imaug/operators.py @@ -22,7 +22,6 @@ from __future__ import unicode_literals import six import math import random -import functools import cv2 import numpy as np @@ -38,8 +37,8 @@ class DecodeImage(object): def __init__(self, to_rgb=True, to_np=False, channel_first=False): self.to_rgb = to_rgb - self.to_np = to_np #to numpy - self.channel_first = channel_first #only enabled when to_np is True + self.to_np = to_np # to numpy + self.channel_first = channel_first # only enabled when to_np is True def __call__(self, img): if six.PY2: @@ -64,7 +63,8 @@ class DecodeImage(object): class ResizeImage(object): """ resize image """ - def __init__(self, size=None, resize_short=None): + def __init__(self, size=None, resize_short=None, interpolation=-1): + self.interpolation = interpolation if interpolation >= 0 else None if resize_short is not None and resize_short > 0: self.resize_short = resize_short self.w = None @@ -86,8 +86,10 @@ class ResizeImage(object): else: w = self.w h = self.h - - return cv2.resize(img, (w, h)) + if self.interpolation is None: + return cv2.resize(img, (w, h)) + else: + return cv2.resize(img, (w, h), interpolation=self.interpolation) class CropImage(object): @@ -138,8 +140,7 @@ class RandCropImage(object): scale_max = min(scale[1], bound) scale_min = min(scale[0], bound) - target_area = img_w * img_h * random.uniform(\ - scale_min, scale_max) + target_area = img_w * img_h * random.uniform(scale_min, scale_max) target_size = math.sqrt(target_area) w = int(target_size * w) h = int(target_size * h) @@ -176,7 +177,8 @@ class NormalizeImage(object): """ def __init__(self, scale=None, mean=None, std=None, order='chw'): - if isinstance(scale, str): scale = eval(scale) + if isinstance(scale, str): + scale = eval(scale) self.scale = np.float32(scale if scale is not None else 1.0 / 255.0) mean = mean if mean is not None else [0.485, 0.456, 0.406] std = std if std is not None else [0.229, 0.224, 0.225] diff --git a/ppcls/data/reader.py b/ppcls/data/reader.py index 5bf83c2170319fa13c6c86cddc42b04ec48da7e7..20072db90923055818f2b148732e78915f4d73e0 100755 --- a/ppcls/data/reader.py +++ b/ppcls/data/reader.py @@ -1,27 +1,25 @@ -#copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. +# copyright (c) 2020 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 +# 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. +# 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 cv2 import numpy as np import os import signal - +import imghdr import paddle from . import imaug from .imaug import transform -from .imaug import MixupOperator from ppcls.utils import logger trainers_num = int(os.environ.get('PADDLE_TRAINERS_NUM', 1)) @@ -35,7 +33,7 @@ class ModeException(Exception): def __init__(self, message='', mode=''): message += "\nOnly the following 3 modes are supported: " \ - "train, valid, test. Given mode is {}".format(mode) + "train, valid, test. Given mode is {}".format(mode) super(ModeException, self).__init__(message) @@ -46,10 +44,10 @@ class SampleNumException(Exception): def __init__(self, message='', sample_num=0, batch_size=1): message += "\nError: The number of the whole data ({}) " \ - "is smaller than the batch_size ({}), and drop_last " \ - "is turnning on, so nothing will feed in program, " \ - "Terminated now. Please reset batch_size to a smaller " \ - "number or feed more data!".format(sample_num, batch_size) + "is smaller than the batch_size ({}), and drop_last " \ + "is turnning on, so nothing will feed in program, " \ + "Terminated now. Please reset batch_size to a smaller " \ + "number or feed more data!".format(sample_num, batch_size) super(SampleNumException, self).__init__(message) @@ -80,12 +78,12 @@ def check_params(params): data_dir = params.get('data_dir', '') assert os.path.isdir(data_dir), \ - "{} doesn't exist, please check datadir path".format(data_dir) + "{} doesn't exist, please check datadir path".format(data_dir) if params['mode'] != 'test': file_list = params.get('file_list', '') assert os.path.isfile(file_list), \ - "{} doesn't exist, please check file list path".format(file_list) + "{} doesn't exist, please check file list path".format(file_list) def create_file_list(params): @@ -176,8 +174,8 @@ def partial_reader(params, full_lines, part_id=0, part_num=1): part_id(int): part index of the current partial data part_num(int): part num of the dataset """ - assert part_id < part_num, ("part_num: {} should be larger " \ - "than part_id: {}".format(part_num, part_id)) + assert part_id < part_num, ("part_num: {} should be larger " + "than part_id: {}".format(part_num, part_id)) full_lines = full_lines[part_id::part_num] @@ -187,8 +185,9 @@ def partial_reader(params, full_lines, part_id=0, part_num=1): def reader(): ops = create_operators(params['transforms']) + delimiter = params.get('delimiter', ' ') for line in full_lines: - img_path, label = line.split() + img_path, label = line.split(delimiter) img_path = os.path.join(params['data_dir'], img_path) with open(img_path, 'rb') as f: img = f.read() @@ -220,7 +219,7 @@ def mp_reader(params): def term_mp(sig_num, frame): - """ kill all child processes + """ kill all child processes """ pid = os.getpid() pgid = os.getpgid(os.getpid()) diff --git a/ppcls/modeling/architectures/__init__.py b/ppcls/modeling/architectures/__init__.py index 34be3ce2addc0a8320e4b3606e1d87f0873d2a47..b6b69a370b4bb56752715f6faed51501609609b5 100644 --- a/ppcls/modeling/architectures/__init__.py +++ b/ppcls/modeling/architectures/__init__.py @@ -36,7 +36,7 @@ from .densenet import DenseNet121, DenseNet161, DenseNet169, DenseNet201, DenseN from .squeezenet import SqueezeNet1_0, SqueezeNet1_1 from .darknet import DarkNet53 from .resnext101_wsl import ResNeXt101_32x8d_wsl, ResNeXt101_32x16d_wsl, ResNeXt101_32x32d_wsl, ResNeXt101_32x48d_wsl, Fix_ResNeXt101_32x48d_wsl -from .efficientnet import EfficientNet, EfficientNetB0, EfficientNetB1, EfficientNetB2, EfficientNetB3, EfficientNetB4, EfficientNetB5, EfficientNetB6, EfficientNetB7 +from .efficientnet import EfficientNet, EfficientNetB0, EfficientNetB0_small, EfficientNetB1, EfficientNetB2, EfficientNetB3, EfficientNetB4, EfficientNetB5, EfficientNetB6, EfficientNetB7 from .res2net import Res2Net50_48w_2s, Res2Net50_26w_4s, Res2Net50_14w_8s, Res2Net50_26w_6s, Res2Net50_26w_8s, Res2Net101_26w_4s, Res2Net152_26w_4s from .res2net_vd import Res2Net50_vd_48w_2s, Res2Net50_vd_26w_4s, Res2Net50_vd_14w_8s, Res2Net50_vd_26w_6s, Res2Net50_vd_26w_8s, Res2Net101_vd_26w_4s, Res2Net152_vd_26w_4s, Res2Net200_vd_26w_4s from .hrnet import HRNet_W18_C, HRNet_W30_C, HRNet_W32_C, HRNet_W40_C, HRNet_W44_C, HRNet_W48_C, HRNet_W60_C, HRNet_W64_C, SE_HRNet_W18_C, SE_HRNet_W30_C, SE_HRNet_W32_C, SE_HRNet_W40_C, SE_HRNet_W44_C, SE_HRNet_W48_C, SE_HRNet_W60_C, SE_HRNet_W64_C diff --git a/ppcls/modeling/architectures/efficientnet.py b/ppcls/modeling/architectures/efficientnet.py index 33135839d8ff3247b8f228946111c97ce1f31e68..d6bac79bd8674b6bcb315d512fc095577fc6c97a 100644 --- a/ppcls/modeling/architectures/efficientnet.py +++ b/ppcls/modeling/architectures/efficientnet.py @@ -288,7 +288,7 @@ class EfficientNet(): name=conv_name, use_bias=use_bias) - if use_bn == False: + if use_bn is False: return conv else: bn_name = name + bn_name @@ -350,8 +350,11 @@ class EfficientNet(): conv = self._project_conv_norm(conv, block_args, is_test, name) # Skip connection and drop connect - input_filters, output_filters = block_args.input_filters, block_args.output_filters - if id_skip and block_args.stride == 1 and input_filters == output_filters: + input_filters = block_args.input_filters + output_filters = block_args.output_filters + if id_skip and \ + block_args.stride == 1 and \ + input_filters == output_filters: if drop_connect_rate: conv = self._drop_connect(conv, drop_connect_rate, self.is_test) @@ -416,7 +419,8 @@ class EfficientNet(): num_repeat=round_repeats(block_args.num_repeat, self._global_params)) - # The first block needs to take care of stride and filter size increase. + # The first block needs to take care of stride, + # and filter size increase. drop_connect_rate = self._global_params.drop_connect_rate if drop_connect_rate: drop_connect_rate *= float(idx) / block_size @@ -444,7 +448,9 @@ class EfficientNet(): class BlockDecoder(object): - """ Block Decoder for readability, straight from the official TensorFlow repository """ + """ + Block Decoder, straight from the official TensorFlow repository. + """ @staticmethod def _decode_block_string(block_string): @@ -460,9 +466,10 @@ class BlockDecoder(object): options[key] = value # Check stride - assert ( - ('s' in options and len(options['s']) == 1) or - (len(options['s']) == 2 and options['s'][0] == options['s'][1])) + cond_1 = ('s' in options and len(options['s']) == 1) + cond_2 = ((len(options['s']) == 2) + and (options['s'][0] == options['s'][1])) + assert (cond_1 or cond_2) return BlockArgs( kernel_size=int(options['k']), @@ -491,10 +498,11 @@ class BlockDecoder(object): @staticmethod def decode(string_list): """ - Decodes a list of string notations to specify blocks inside the network. + Decode a list of string notations to specify blocks in the network. - :param string_list: a list of strings, each string is a notation of block - :return: a list of BlockArgs namedtuples of block args + string_list: list of strings, each string is a notation of block + return + list of BlockArgs namedtuples of block args """ assert isinstance(string_list, list) blocks_args = [] @@ -529,6 +537,19 @@ def EfficientNetB0(is_test=False, return model +def EfficientNetB0_small(is_test=False, + padding_type='DYNAMIC', + override_params=None, + use_se=False): + model = EfficientNet( + name='b0', + is_test=is_test, + padding_type=padding_type, + override_params=override_params, + use_se=use_se) + return model + + def EfficientNetB1(is_test=False, padding_type='SAME', override_params=None, diff --git a/ppcls/utils/config.py b/ppcls/utils/config.py index b1c1be4ef616d504fce196e51d3ea5316e5b8c67..93b11569e7490ba1c7d3576e254dbfabf29f9344 100644 --- a/ppcls/utils/config.py +++ b/ppcls/utils/config.py @@ -64,14 +64,14 @@ def print_dict(d, delimiter=0): placeholder = "-" * 60 for k, v in sorted(d.items()): if isinstance(v, dict): - logger.info("{}{} : ".format(delimiter * " ", k)) + logger.info("{}{} : ".format(delimiter * " ", logger.coloring(k, "HEADER"))) print_dict(v, delimiter + 4) elif isinstance(v, list) and len(v) >= 1 and isinstance(v[0], dict): - logger.info("{}{} : ".format(delimiter * " ", k)) + logger.info("{}{} : ".format(delimiter * " ", logger.coloring(str(k),"HEADER"))) for value in v: print_dict(value, delimiter + 4) else: - logger.info("{}{} : {}".format(delimiter * " ", k, v)) + logger.info("{}{} : {}".format(delimiter * " ", logger.coloring(k,"HEADER"), logger.coloring(v,"OKGREEN"))) if k.isupper(): logger.info(placeholder) diff --git a/ppcls/utils/logger.py b/ppcls/utils/logger.py index 5b4ae2ca0b12cad7b68a4f76603ae8bc4fd60d25..5b192c61b46d82cc99e88269c5118c3e9e182b66 100644 --- a/ppcls/utils/logger.py +++ b/ppcls/utils/logger.py @@ -14,15 +14,48 @@ import logging import os +import datetime -logging.basicConfig(level=logging.INFO) +from imp import reload +reload(logging) + +logging.basicConfig(level=logging.INFO, + format="%(asctime)s %(levelname)s: %(message)s", + datefmt = "%Y-%m-%d %H:%M:%S") + + +def time_zone(sec, fmt): + real_time = datetime.datetime.now() + datetime.timedelta(hours=8) + return real_time.timetuple() + + +logging.Formatter.converter = time_zone _logger = logging.getLogger(__name__) +Color= { + 'RED' : '\033[31m' , + 'HEADER' : '\033[35m' , # deep purple + 'PURPLE' : '\033[95m' ,# purple + 'OKBLUE' : '\033[94m' , + 'OKGREEN' : '\033[92m' , + 'WARNING' : '\033[93m' , + 'FAIL' : '\033[91m' , + 'ENDC' : '\033[0m' } + + +def coloring(message, color="OKGREEN"): + assert color in Color.keys() + if os.environ.get('PADDLECLAS_COLORING', False): + return Color[color]+str(message)+Color["ENDC"] + else: + return message + + def anti_fleet(log): """ - Because of the fucking Fleet, logs will print multi-times. - So we only display one of them and ignore the others. + logs will print multi-times when calling Fleet API. + Only display single log and ignore the others. """ def wrapper(fmt, *args): @@ -39,12 +72,12 @@ def info(fmt, *args): @anti_fleet def warning(fmt, *args): - _logger.warning(fmt, *args) + _logger.warning(coloring(fmt, "RED"), *args) @anti_fleet def error(fmt, *args): - _logger.error(fmt, *args) + _logger.error(coloring(fmt, "FAIL"), *args) def advertise(): @@ -66,7 +99,7 @@ def advertise(): website = "https://github.com/PaddlePaddle/PaddleClas" AD_LEN = 6 + len(max([copyright, ad, website], key=len)) - info("\n{0}\n{1}\n{2}\n{3}\n{4}\n{5}\n{6}\n{7}\n".format( + info(coloring("\n{0}\n{1}\n{2}\n{3}\n{4}\n{5}\n{6}\n{7}\n".format( "=" * (AD_LEN + 4), "=={}==".format(copyright.center(AD_LEN)), "=" * (AD_LEN + 4), @@ -74,4 +107,4 @@ def advertise(): "=={}==".format(ad.center(AD_LEN)), "=={}==".format(' ' * AD_LEN), "=={}==".format(website.center(AD_LEN)), - "=" * (AD_LEN + 4), )) + "=" * (AD_LEN + 4), ),"RED")) diff --git a/ppcls/utils/model_zoo.py b/ppcls/utils/model_zoo.py index 8a154a90de99432c44ec004c08f0b5cb2a61d26e..dd65e921f3e258b75c911599ee234359f3fd7c51 100644 --- a/ppcls/utils/model_zoo.py +++ b/ppcls/utils/model_zoo.py @@ -24,7 +24,6 @@ import tqdm import zipfile from ppcls.modeling import similar_architectures -from ppcls.utils.check import check_architecture from ppcls.utils import logger __all__ = ['get'] @@ -168,11 +167,16 @@ def _decompress(fname): os.remove(fname) +def _get_pretrained(): + with open('./ppcls/utils/pretrained.list') as flist: + pretrained = [line.strip() for line in flist] + return pretrained + + def _check_pretrained_name(architecture): assert isinstance(architecture, str), \ - ("the type of architecture({}) should be str". format(architecture)) - with open('./configs/pretrained.list') as flist: - pretrained = [line.strip() for line in flist] + ("the type of architecture({}) should be str". format(architecture)) + pretrained = _get_pretrained() similar_names = similar_architectures(architecture, pretrained) model_list = ', '.join(similar_names) err = "{} is not exist! Maybe you want: [{}]" \ @@ -181,6 +185,14 @@ def _check_pretrained_name(architecture): raise ModelNameError(err) +def list_models(): + pretrained = _get_pretrained() + msg = "All avialable pretrained models are as follows: {}".format( + pretrained) + logger.info(msg) + return + + def get(architecture, path, decompress=True): """ Get the pretrained model. @@ -188,5 +200,6 @@ def get(architecture, path, decompress=True): _check_pretrained_name(architecture) url = _get_url(architecture) fname = _download(url, path) - if decompress: _decompress(fname) + if decompress: + _decompress(fname) logger.info("download {} finished ".format(fname)) diff --git a/configs/pretrained.list b/ppcls/utils/pretrained.list similarity index 98% rename from configs/pretrained.list rename to ppcls/utils/pretrained.list index b01272ea2bc7ed0d9587f486001a1bdb3d7697d9..633cafd921d7390f434b2b5f82dad70129349658 100644 --- a/configs/pretrained.list +++ b/ppcls/utils/pretrained.list @@ -114,5 +114,5 @@ VGG11 VGG13 VGG16 VGG19 -DarkNet53 +DarkNet53_ImageNet1k ResNet50_ACNet_deploy diff --git a/ppcls/utils/save_load.py b/ppcls/utils/save_load.py index 27c680432b878799ef7551f4aaabeb3f077a383d..e310166ece236cadffa0fc28d3bef6558b4c43e4 100644 --- a/ppcls/utils/save_load.py +++ b/ppcls/utils/save_load.py @@ -74,7 +74,7 @@ def load_params(exe, prog, path, ignore_params=None): raise ValueError("Model pretrain path {} does not " "exists.".format(path)) - logger.info('Loading parameters from {}...'.format(path)) + logger.info(logger.coloring('Loading parameters from {}...'.format(path), 'HEADER')) ignore_set = set() state = _load_state(path) @@ -100,7 +100,7 @@ def load_params(exe, prog, path, ignore_params=None): if len(ignore_set) > 0: for k in ignore_set: if k in state: - logger.warning('variable {} not used'.format(k)) + logger.warning('variable {} is already excluded automatically'.format(k)) del state[k] fluid.io.set_program_state(prog, state) @@ -113,7 +113,7 @@ def init_model(config, program, exe): checkpoints = config.get('checkpoints') if checkpoints: fluid.load(program, checkpoints, exe) - logger.info("Finish initing model from {}".format(checkpoints)) + logger.info(logger.coloring("Finish initing model from {}".format(checkpoints),"HEADER")) return pretrained_model = config.get('pretrained_model') @@ -122,7 +122,7 @@ def init_model(config, program, exe): pretrained_model = [pretrained_model] for pretrain in pretrained_model: load_params(exe, program, pretrain) - logger.info("Finish initing model from {}".format(pretrained_model)) + logger.info(logger.coloring("Finish initing model from {}".format(pretrained_model),"HEADER")) def save_model(program, model_path, epoch_id, prefix='ppcls'): @@ -133,4 +133,4 @@ def save_model(program, model_path, epoch_id, prefix='ppcls'): _mkdir_if_not_exist(model_path) model_prefix = os.path.join(model_path, prefix) fluid.save(program, model_prefix) - logger.info("Already save model in {}".format(model_path)) + logger.info(logger.coloring("Already save model in {}".format(model_path),"HEADER")) diff --git a/tools/download.py b/tools/download.py index 157bebccdaa26bbb40a46a5bf065b4e092527af7..d9fe1a8ee04a14b31cf4917e60f9348ae51b8d20 100644 --- a/tools/download.py +++ b/tools/download.py @@ -1,18 +1,17 @@ # Copyright (c) 2020 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 +# 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. +# 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 import argparse from ppcls import model_zoo @@ -26,6 +25,7 @@ def parse_args(): parser.add_argument('-a', '--architecture', type=str, default='ResNet50') parser.add_argument('-p', '--path', type=str, default='./pretrained/') parser.add_argument('-d', '--decompress', type=str2bool, default=True) + parser.add_argument('-l', '--list', type=str2bool, default=False) args = parser.parse_args() return args @@ -33,7 +33,10 @@ def parse_args(): def main(): args = parse_args() - model_zoo.get(args.architecture, args.path, args.decompress) + if args.list: + model_zoo.list_models() + else: + model_zoo.get(args.architecture, args.path, args.decompress) if __name__ == '__main__': diff --git a/tools/export_model.py b/tools/export_model.py index 2cc32b3e52242652e9dcb53a1ea94c6b90ca80f9..763ce473c45d60ba577d65782f5d4737f38a10bc 100644 --- a/tools/export_model.py +++ b/tools/export_model.py @@ -13,7 +13,6 @@ # limitations under the License. import argparse -import numpy as np from ppcls.modeling import architectures import paddle.fluid as fluid @@ -25,13 +24,14 @@ def parse_args(): parser.add_argument("-p", "--pretrained_model", type=str) parser.add_argument("-o", "--output_path", type=str) parser.add_argument("--class_dim", type=int, default=1000) + parser.add_argument("--img_size", type=int, default=224) return parser.parse_args() -def create_input(): +def create_input(img_size=224): image = fluid.data( - name='image', shape=[None, 3, 224, 224], dtype='float32') + name='image', shape=[None, 3, img_size, img_size], dtype='float32') return image @@ -57,7 +57,7 @@ def main(): with fluid.program_guard(infer_prog, startup_prog): with fluid.unique_name.guard(): - image = create_input() + image = create_input(args.img_size) out = create_model(args, model, image, class_dim=args.class_dim) infer_prog = infer_prog.clone(for_test=True) diff --git a/tools/export_serving_model.py b/tools/export_serving_model.py new file mode 100644 index 0000000000000000000000000000000000000000..e6e4472cdbf8dfd1738dede98b5aa61121f8191a --- /dev/null +++ b/tools/export_serving_model.py @@ -0,0 +1,76 @@ +# Copyright (c) 2020 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. + +import argparse +import os +from ppcls.modeling import architectures + +import paddle.fluid as fluid +import paddle_serving_client.io as serving_io + + +def parse_args(): + parser = argparse.ArgumentParser() + parser.add_argument("-m", "--model", type=str) + parser.add_argument("-p", "--pretrained_model", type=str) + parser.add_argument("-o", "--output_path", type=str, default="") + parser.add_argument("--class_dim", type=int, default=1000) + parser.add_argument("--img_size", type=int, default=224) + + return parser.parse_args() + + +def create_input(img_size=224): + image = fluid.data( + name='image', shape=[None, 3, img_size, img_size], dtype='float32') + return image + + +def create_model(args, model, input, class_dim=1000): + if args.model == "GoogLeNet": + out, _, _ = model.net(input=input, class_dim=class_dim) + else: + out = model.net(input=input, class_dim=class_dim) + out = fluid.layers.softmax(out) + return out + + +def main(): + args = parse_args() + + model = architectures.__dict__[args.model]() + + place = fluid.CPUPlace() + exe = fluid.Executor(place) + + startup_prog = fluid.Program() + infer_prog = fluid.Program() + + with fluid.program_guard(infer_prog, startup_prog): + with fluid.unique_name.guard(): + image = create_input(args.img_size) + out = create_model(args, model, image, class_dim=args.class_dim) + + infer_prog = infer_prog.clone(for_test=True) + fluid.load( + program=infer_prog, model_path=args.pretrained_model, executor=exe) + + model_path = os.path.join(args.output_path, "ppcls_model") + conf_path = os.path.join(args.output_path, "ppcls_client_conf") + serving_io.save_model(model_path, conf_path, {"image": image}, + {"prediction": out}, infer_prog) + + +if __name__ == "__main__": + main() diff --git a/tools/lite/benchmark.sh b/tools/lite/benchmark.sh new file mode 100644 index 0000000000000000000000000000000000000000..591331e42a4ff7dc184a9873b3aaf05889d2aa02 --- /dev/null +++ b/tools/lite/benchmark.sh @@ -0,0 +1,73 @@ +#!/bin/bash +# ref1: https://github.com/PaddlePaddle/Paddle-Lite/blob/58b2d7dd89/lite/api/benchmark.cc +# ref2: https://paddle-inference-dist.bj.bcebos.com/PaddleLite/benchmark_0/benchmark.sh + +set -e + +# Check input +if [ $# -lt 3 ]; +then + echo "Input error" + echo "Usage:" + echo " sh benchmark.sh " + echo " sh benchmark.sh " + exit +fi + +# Set benchmark params +ANDROID_DIR=/data/local/tmp +BENCHMARK_BIN=$1 +MODELS_DIR=$2 +RESULT_FILENAME=$3 + +WARMUP=10 +REPEATS=30 +IS_RUN_MODEL_OPTIMIZE=false +IS_RUN_QUANTIZED_MODEL=false +NUM_THREADS_LIST=(1 2 4) +MODELS_LIST=$(ls $MODELS_DIR) + +# Check input +if [ $# -gt 3 ]; +then + IS_RUN_MODEL_OPTIMIZE=$4 +fi + +# Adb push benchmark_bin, models +adb push $BENCHMARK_BIN $ANDROID_DIR/benchmark_bin +adb shell chmod +x $ANDROID_DIR/benchmark_bin +adb push $MODELS_DIR $ANDROID_DIR + +# Run benchmark +adb shell "echo 'PaddleLite Benchmark' > $ANDROID_DIR/$RESULT_FILENAME" +for threads in ${NUM_THREADS_LIST[@]}; do + adb shell "echo Threads=$threads Warmup=$WARMUP Repeats=$REPEATS >> $ANDROID_DIR/$RESULT_FILENAME" + for model_name in ${MODELS_LIST[@]}; do + echo "Model=$model_name Threads=$threads" + if [ "$IS_RUN_MODEL_OPTIMIZE" = true ]; + then + adb shell "$ANDROID_DIR/benchmark_bin \ + --model_dir=$ANDROID_DIR/${MODELS_DIR}/$model_name \ + --model_filename=model \ + --param_filename=params \ + --warmup=$WARMUP \ + --repeats=$REPEATS \ + --threads=$threads \ + --result_filename=$ANDROID_DIR/$RESULT_FILENAME" + else + adb shell "$ANDROID_DIR/benchmark_bin \ + --optimized_model_path=$ANDROID_DIR/${MODELS_DIR}/$model_name \ + --warmup=$WARMUP \ + --repeats=$REPEATS \ + --threads=$threads \ + --result_filename=$ANDROID_DIR/$RESULT_FILENAME" + fi + done + adb shell "echo >> $ANDROID_DIR/$RESULT_FILENAME" +done + +# Adb pull benchmark result, show result +adb pull $ANDROID_DIR/$RESULT_FILENAME . +echo "\n--------------------------------------" +cat $RESULT_FILENAME +echo "--------------------------------------" diff --git a/tools/program.py b/tools/program.py index 50c609de17186b234f8ce3632a94cf7f59454dcd..b73f1064284a7f1929f99d892bee43798a5d52fb 100644 --- a/tools/program.py +++ b/tools/program.py @@ -297,6 +297,19 @@ def dist_optimizer(config, optimizer): return optimizer +def mixed_precision_optimizer(config, optimizer): + use_fp16 = config.get('use_fp16', False) + amp_scale_loss = config.get('amp_scale_loss', 1.0) + use_dynamic_loss_scaling = config.get('use_dynamic_loss_scaling', False) + if use_fp16: + optimizer = fluid.contrib.mixed_precision.decorate( + optimizer, + init_loss_scaling=amp_scale_loss, + use_dynamic_loss_scaling=use_dynamic_loss_scaling) + + return optimizer + + def build(config, main_prog, startup_prog, is_train=True): """ Build a program using a model and an optimizer @@ -337,6 +350,8 @@ def build(config, main_prog, startup_prog, is_train=True): optimizer = create_optimizer(config) lr = optimizer._global_learning_rate() fetchs['lr'] = (lr, AverageMeter('lr', 'f', need_avg=False)) + + optimizer = mixed_precision_optimizer(config, optimizer) optimizer = dist_optimizer(config, optimizer) optimizer.minimize(fetchs['loss'][0]) @@ -396,19 +411,30 @@ def run(dataloader, exe, program, fetchs, epoch=0, mode='train'): for i, m in enumerate(metrics): metric_list[i].update(m[0], len(batch[0])) fetchs_str = ''.join([str(m.value) + ' ' - for m in metric_list] + [batch_time.value]) + for m in metric_list] + [batch_time.value]) + 's' if mode == 'eval': logger.info("{:s} step:{:<4d} {:s}s".format(mode, idx, fetchs_str)) else: - logger.info("epoch:{:<3d} {:s} step:{:<4d} {:s}s".format( - epoch, mode, idx, fetchs_str)) + epoch_str = "epoch:{:<3d}".format(epoch) + step_str = "{:s} step:{:<4d}".format(mode, idx) + + logger.info("{:s} {:s} {:s}".format( + logger.coloring(epoch_str, "HEADER") + if idx == 0 else epoch_str, + logger.coloring(step_str, "PURPLE"), + logger.coloring(fetchs_str, 'OKGREEN'))) end_str = ''.join([str(m.mean) + ' ' - for m in metric_list] + [batch_time.total]) + for m in metric_list] + [batch_time.total]) + 's' if mode == 'eval': logger.info("END {:s} {:s}s".format(mode, end_str)) else: - logger.info("END epoch:{:<3d} {:s} {:s}s".format(epoch, mode, end_str)) + end_epoch_str = "END epoch:{:<3d}".format(epoch) + + logger.info("{:s} {:s} {:s}".format( + logger.coloring(end_epoch_str, "RED"), + logger.coloring(mode, "PURPLE"), + logger.coloring(end_str, "OKGREEN"))) # return top1_acc in order to save the best model if mode == 'valid': diff --git a/tools/serving/image_http_client.py b/tools/serving/image_http_client.py new file mode 100644 index 0000000000000000000000000000000000000000..3b92091c659613c83e4423a3f22b0d4d20321f43 --- /dev/null +++ b/tools/serving/image_http_client.py @@ -0,0 +1,47 @@ +# Copyright (c) 2020 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. + +import requests +import base64 +import json +import sys +import numpy as np + +py_version = sys.version_info[0] + + +def predict(image_path, server): + if py_version == 2: + image = base64.b64encode(open(image_path).read()) + else: + image = base64.b64encode(open(image_path, "rb").read()).decode("utf-8") + req = json.dumps({"feed": [{"image": image}], "fetch": ["prediction"]}) + r = requests.post( + server, data=req, headers={"Content-Type": "application/json"}) + try: + pred = r.json()["result"]["prediction"][0] + cls_id = np.argmax(pred) + score = pred[cls_id] + pred = {"cls_id": cls_id, "score": score} + return pred + except ValueError: + print(r.text) + return r + + +if __name__ == "__main__": + server = "http://127.0.0.1:{}/image/prediction".format(sys.argv[1]) + image_file = sys.argv[2] + res = predict(image_file, server) + print("res:", res) diff --git a/tools/serving/image_service_cpu.py b/tools/serving/image_service_cpu.py new file mode 100644 index 0000000000000000000000000000000000000000..92f67d3220670ffa880ff663ed887984325a0723 --- /dev/null +++ b/tools/serving/image_service_cpu.py @@ -0,0 +1,60 @@ +# Copyright (c) 2020 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. + +import sys +import base64 +from paddle_serving_server.web_service import WebService +import utils + + +class ImageService(WebService): + def __init__(self, name): + super(ImageService, self).__init__(name=name) + self.operators = self.create_operators() + + def create_operators(self): + size = 224 + img_mean = [0.485, 0.456, 0.406] + img_std = [0.229, 0.224, 0.225] + img_scale = 1.0 / 255.0 + decode_op = utils.DecodeImage() + resize_op = utils.ResizeImage(resize_short=256) + crop_op = utils.CropImage(size=(size, size)) + normalize_op = utils.NormalizeImage( + scale=img_scale, mean=img_mean, std=img_std) + totensor_op = utils.ToTensor() + return [decode_op, resize_op, crop_op, normalize_op, totensor_op] + + def _process_image(self, data, ops): + for op in ops: + data = op(data) + return data + + def preprocess(self, feed={}, fetch=[]): + feed_batch = [] + for ins in feed: + if "image" not in ins: + raise ("feed data error!") + sample = base64.b64decode(ins["image"]) + img = self._process_image(sample, self.operators) + feed_batch.append({"image": img}) + return feed_batch, fetch + + +image_service = ImageService(name="image") +image_service.load_model_config(sys.argv[1]) +image_service.prepare_server( + workdir=sys.argv[2], port=int(sys.argv[3]), device="cpu") +image_service.run_server() +image_service.run_flask() diff --git a/tools/serving/image_service_gpu.py b/tools/serving/image_service_gpu.py new file mode 100644 index 0000000000000000000000000000000000000000..df61cdd60659713ac77176beb8c1ecfad1c8efd8 --- /dev/null +++ b/tools/serving/image_service_gpu.py @@ -0,0 +1,62 @@ +# Copyright (c) 2020 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. + +import sys +import base64 +from paddle_serving_server_gpu.web_service import WebService + +import utils + + +class ImageService(WebService): + def __init__(self, name): + super(ImageService, self).__init__(name=name) + self.operators = self.create_operators() + + def create_operators(self): + size = 224 + img_mean = [0.485, 0.456, 0.406] + img_std = [0.229, 0.224, 0.225] + img_scale = 1.0 / 255.0 + decode_op = utils.DecodeImage() + resize_op = utils.ResizeImage(resize_short=256) + crop_op = utils.CropImage(size=(size, size)) + normalize_op = utils.NormalizeImage( + scale=img_scale, mean=img_mean, std=img_std) + totensor_op = utils.ToTensor() + return [decode_op, resize_op, crop_op, normalize_op, totensor_op] + + def _process_image(self, data, ops): + for op in ops: + data = op(data) + return data + + def preprocess(self, feed={}, fetch=[]): + feed_batch = [] + for ins in feed: + if "image" not in ins: + raise ("feed data error!") + sample = base64.b64decode(ins["image"]) + img = self._process_image(sample, self.operators) + feed_batch.append({"image": img}) + return feed_batch, fetch + + +image_service = ImageService(name="image") +image_service.load_model_config(sys.argv[1]) +image_service.set_gpus("0") +image_service.prepare_server( + workdir=sys.argv[2], port=int(sys.argv[3]), device="gpu") +image_service.run_server() +image_service.run_flask() diff --git a/tools/serving/utils.py b/tools/serving/utils.py new file mode 100644 index 0000000000000000000000000000000000000000..6c4a75e1afe2fc1e1710c7e8213f8ac4de8ffcc2 --- /dev/null +++ b/tools/serving/utils.py @@ -0,0 +1,84 @@ +# Copyright (c) 2020 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. + +import cv2 +import numpy as np + + +class DecodeImage(object): + def __init__(self, to_rgb=True): + self.to_rgb = to_rgb + + def __call__(self, img): + data = np.frombuffer(img, dtype='uint8') + img = cv2.imdecode(data, 1) + if self.to_rgb: + assert img.shape[2] == 3, 'invalid shape of image[%s]' % ( + img.shape) + img = img[:, :, ::-1] + + return img + + +class ResizeImage(object): + def __init__(self, resize_short=None): + self.resize_short = resize_short + + def __call__(self, img): + img_h, img_w = img.shape[:2] + percent = float(self.resize_short) / min(img_w, img_h) + w = int(round(img_w * percent)) + h = int(round(img_h * percent)) + return cv2.resize(img, (w, h)) + + +class CropImage(object): + def __init__(self, size): + if type(size) is int: + self.size = (size, size) + else: + self.size = size + + def __call__(self, img): + w, h = self.size + img_h, img_w = img.shape[:2] + w_start = (img_w - w) // 2 + h_start = (img_h - h) // 2 + + w_end = w_start + w + h_end = h_start + h + return img[h_start:h_end, w_start:w_end, :] + + +class NormalizeImage(object): + def __init__(self, scale=None, mean=None, std=None): + self.scale = np.float32(scale if scale is not None else 1.0 / 255.0) + mean = mean if mean is not None else [0.485, 0.456, 0.406] + std = std if std is not None else [0.229, 0.224, 0.225] + + shape = (1, 1, 3) + self.mean = np.array(mean).reshape(shape).astype('float32') + self.std = np.array(std).reshape(shape).astype('float32') + + def __call__(self, img): + return (img.astype('float32') * self.scale - self.mean) / self.std + + +class ToTensor(object): + def __init__(self): + pass + + def __call__(self, img): + img = img.transpose((2, 0, 1)) + return img diff --git a/tools/train.py b/tools/train.py index ab7752fd17c6621c605365d5529bf20d91837c89..cd5b7d25b499751ccc9289fe96139e630b09fee7 100644 --- a/tools/train.py +++ b/tools/train.py @@ -62,7 +62,7 @@ def main(args): startup_prog = fluid.Program() train_prog = fluid.Program() - best_top1_acc_list = (0.0, -1) # (top1_acc, epoch_id) + best_top1_acc = 0.0 # best top1 acc record train_dataloader, train_fetchs = program.build( config, train_prog, startup_prog, is_train=True) @@ -101,13 +101,15 @@ def main(args): top1_acc = program.run(valid_dataloader, exe, compiled_valid_prog, valid_fetchs, epoch_id, 'valid') - if top1_acc > best_top1_acc_list[0]: - best_top1_acc_list = (top1_acc, epoch_id) - logger.info("Best top1 acc: {}, in epoch: {}".format( - *best_top1_acc_list)) - model_path = os.path.join(config.model_save_dir, + if top1_acc > best_top1_acc: + best_top1_acc = top1_acc + message = "The best top1 acc {:.5f}, in epoch: {:d}".format(best_top1_acc, epoch_id) + logger.info("{:s}".format(logger.coloring(message, "RED"))) + if epoch_id % config.save_interval==0: + + model_path = os.path.join(config.model_save_dir, config.ARCHITECTURE["name"]) - save_model(train_prog, model_path, "best_model") + save_model(train_prog, model_path, "best_model_in_epoch_"+str(epoch_id)) # 3. save the persistable model if epoch_id % config.save_interval == 0: