merge devevelop

2022-05-05 14:42:06 +08:00 · 2022-05-05 14:42:06 +08:00 · 0b1481402b
parent faa4c08762 00bf84c43a
commit 0b1481402b
56 changed files with 298 additions and 231 deletions
--- a/README_ch.md
+++ b/README_ch.md
@ -7,7 +7,7 @@
 飞桨图像识别套件PaddleClas是飞桨为工业界和学术界所准备的一个图像识别任务的工具集，助力使用者训练出更好的视觉模型和应用落地。

 **近期更新**
- 2022.4.21 新增 CVPR2022 oral论文 [MixFormmer](https://arxiv.org/pdf/2204.02557.pdf) 相关[代码](https://github.com/PaddlePaddle/PaddleClas/pull/1820/files)。
+- 2022.4.21 新增 CVPR2022 oral论文 [MixFormer](https://arxiv.org/pdf/2204.02557.pdf) 相关[代码](https://github.com/PaddlePaddle/PaddleClas/pull/1820/files)。
 - 2022.1.27 全面升级文档；新增[PaddleServing C++ pipeline部署方式](./deploy/paddleserving)和[18M图像识别安卓部署Demo](./deploy/lite_shitu)。
 - 2021.11.1 发布[PP-ShiTu技术报告](https://arxiv.org/pdf/2111.00775.pdf)，新增饮料识别demo
 - 2021.10.23 发布轻量级图像识别系统PP-ShiTu，CPU上0.2s即可完成在10w+库的图像识别。
--- a/README_en.md
+++ b/README_en.md
@ -8,7 +8,7 @@ PaddleClas is an image recognition toolset for industry and academia, helping us

 **Recent updates**

- 2022.4.21 Added the related [code](https://github.com/PaddlePaddle/PaddleClas/pull/1820/files) of the CVPR2022 oral paper [MixFormmer](https://arxiv.org/pdf/2204.02557.pdf).
+- 2022.4.21 Added the related [code](https://github.com/PaddlePaddle/PaddleClas/pull/1820/files) of the CVPR2022 oral paper [MixFormer](https://arxiv.org/pdf/2204.02557.pdf).

 - 2021.09.17 Add PP-LCNet series model developed by PaddleClas, these models show strong competitiveness on Intel CPUs.
 For the introduction of PP-LCNet, please refer to [paper](https://arxiv.org/pdf/2109.15099.pdf) or [PP-LCNet model introduction](docs/en/models/PP-LCNet_en.md). The metrics and pretrained model are available [here](docs/en/ImageNet_models_en.md).
--- a/deploy/lite_shitu/README.md
+++ b/deploy/lite_shitu/README.md
@ -92,9 +92,9 @@ PaddleClas 提供了转换并优化后的推理模型，可以直接参考下方
 ```shell
 # 进入lite_ppshitu目录
 cd $PaddleClas/deploy/lite_shitu
-wget https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/lite/ppshitu_lite_models_v1.0.tar
-tar -xf ppshitu_lite_models_v1.0.tar
-rm -f ppshitu_lite_models_v1.0.tar
+wget https://paddle-imagenet-models-name.bj.bcebos.com/dygraph/lite/ppshitu_lite_models_v1.1.tar
+tar -xf ppshitu_lite_models_v1.1.tar
+rm -f ppshitu_lite_models_v1.1.tar
 ```

 #### 2.1.2 使用其他模型
@ -173,15 +173,11 @@ cp $code_path/PaddleDetection/inference/picodet_lcnet_x2_5_640_mainbody/mainbody

 2. 转换识别模型

-请先参考 [识别模型转分类模型](../../docs/zh_CN/advanced_tutorials/gallery2fc.md) 完成识别模型到分类模型的转换。
-在得到 inference 推理模型（后缀名为 `.pdmodel`、`.pdiparams`）以及 `label.txt` 后，再使用 PaddleLite opt 工具完成模型优化，命令如下：
-
 ```shell
 # 转换为Paddle-Lite模型
 paddle_lite_opt --model_file=inference/inference.pdmodel --param_file=inference/inference.pdiparams --optimize_out=inference/rec
-# 将模型、label文件拷贝到lite_shitu下
+# 将模型文件拷贝到lite_shitu下
 cp inference/rec.nb deploy/lite_shitu/models/
-cp inference/label.txt deploy/lite_shitu/models/
 cd deploy/lite_shitu
 ```

@ -191,10 +187,10 @@ cd deploy/lite_shitu

 ```shell
 # 如果测试单张图像
-python generate_json_config.py --det_model_path ppshitu_lite_models_v1.0/mainbody_PPLCNet_x2_5_640_quant_v1.0_lite.nb  --rec_model_path ppshitu_lite_models_v1.0/general_PPLCNet_x2_5_lite_v1.0_infer.nb --img_path images/demo.jpg
+python generate_json_config.py --det_model_path ppshitu_lite_models_v1.1/mainbody_PPLCNet_x2_5_640_quant_v1.1_lite.nb  --rec_model_path ppshitu_lite_models_v1.1/general_PPLCNet_x2_5_lite_v1.1_infer.nb --img_path images/demo.jpg
 # or
 # 如果测试多张图像
-python generate_json_config.py --det_model_path ppshitu_lite_models_v1.0/mainbody_PPLCNet_x2_5_640_quant_v1.0_lite.nb  --rec_model_path ppshitu_lite_models_v1.0/general_PPLCNet_x2_5_lite_v1.0_infer.nb --img_dir images
+python generate_json_config.py --det_model_path ppshitu_lite_models_v1.1/mainbody_PPLCNet_x2_5_640_quant_v1.1_lite.nb  --rec_model_path ppshitu_lite_models_v1.1/general_PPLCNet_x2_5_lite_v1.1_infer.nb --img_dir images
 # 执行完成后，会在lit_shitu下生成shitu_config.json配置文件
 ```

@ -263,7 +259,7 @@ make ARM_ABI=arm8

 ```shell
 mkdir deploy
-mv ppshitu_lite_models_v1.0 deploy/
+mv ppshitu_lite_models_v1.1 deploy/
 mv drink_dataset_v1.0 deploy/
 mv images deploy/
 mv shitu_config.json deploy/
@ -277,12 +273,12 @@ cp ../../../cxx/lib/libpaddle_light_api_shared.so deploy/

 ```shell
 deploy/
-|-- ppshitu_lite_models_v1.0/
-|   |--mainbody_PPLCNet_x2_5_lite_v1.0_infer.nb        优化后的主体检测模型文件
-|   |--general_PPLCNet_x2_5_quant_v1.0_lite.nb         优化后的识别模型文件
+|-- ppshitu_lite_models_v1.1/
+|   |--mainbody_PPLCNet_x2_5_640_quant_v1.1_lite.nb    优化后的主体检测模型文件
+|   |--general_PPLCNet_x2_5_lite_v1.1_infer.nb         优化后的识别模型文件
 |-- images/
 |   |--demo.jpg                                      图片文件
-|-- drink_dataset_v1.0/                                瓶装饮料demo数据
+|-- drink_dataset_v1.0/                              瓶装饮料demo数据
 |   |--index                                         检索index目录
 |-- pp_shitu                                         生成的移动端执行文件
 |-- shitu_config.json                                执行时参数配置文件
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_224.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_224.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1.25e-4
-    eta_min: 1.25e-6
+    learning_rate: 2.5e-4
+    eta_min: 2.5e-6
    warmup_epoch: 20
-    warmup_start_lr: 1.25e-7
+    warmup_start_lr: 2.5e-7


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_384.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_base_384.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 6.25e-5
-    eta_min: 6.25e-7
+    learning_rate: 1.25e-4
+    eta_min: 1.25e-6
    warmup_epoch: 20
-    warmup_start_lr: 6.25e-8
+    warmup_start_lr: 1.25e-7


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_224.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_224.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1.25e-4
-    eta_min: 1.25e-6
+    learning_rate: 2.5e-4
+    eta_min: 2.5e-6
    warmup_epoch: 20
-    warmup_start_lr: 1.25e-7
+    warmup_start_lr: 2.5e-7


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_384.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_large_384.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 3.125e-5
-    eta_min: 3.125e-7
+    learning_rate: 6.25e-5
+    eta_min: 6.25e-7
    warmup_epoch: 20
-    warmup_start_lr: 3.125e-8
+    warmup_start_lr: 6.25e-8


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_small_224.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_small_224.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 2.5e-4
-    eta_min: 2.5e-6
+    learning_rate: 5e-4
+    eta_min: 5e-6
    warmup_epoch: 20
-    warmup_start_lr: 2.5e-7
+    warmup_start_lr: 5e-7


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml
+++ b/ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml
@ -42,11 +42,12 @@ Optimizer:
  no_weight_decay_name: pos_embed cls_token .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_224.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_384.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_base_distilled_patch16_384.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_224.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_384.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_base_patch16_384.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_small_distilled_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_small_distilled_patch16_224.yaml
@ -41,10 +41,10 @@ Optimizer:
  one_dim_param_no_weight_decay: True
  lr:
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_small_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_small_patch16_224.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_tiny_distilled_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_tiny_distilled_patch16_224.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/DeiT/DeiT_tiny_patch16_224.yaml
+++ b/ppcls/configs/ImageNet/DeiT/DeiT_tiny_patch16_224.yaml
@ -40,11 +40,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token pos_embed dist_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 1e-3
-    eta_min: 1e-5
+    learning_rate: 2e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6

 # data loader for train and eval
 DataLoader:
--- a/ppcls/configs/ImageNet/Distillation/mv3_large_x1_0_distill_mv3_small_x1_0.yaml
+++ b/ppcls/configs/ImageNet/Distillation/mv3_large_x1_0_distill_mv3_small_x1_0.yaml
@ -49,9 +49,8 @@ Loss:
        model_name_pairs:
        - ["Student", "Teacher"]
  Eval:
-    - DistillationGTCELoss:
+    - CELoss:
        weight: 1.0
-        model_names: ["Student"]
        

 Optimizer:
--- a/ppcls/configs/ImageNet/Distillation/resnet34_distill_resnet18_afd.yaml
+++ b/ppcls/configs/ImageNet/Distillation/resnet34_distill_resnet18_afd.yaml
@ -88,10 +88,8 @@ Loss:
        s_shapes: *s_shapes
        t_shapes: *t_shapes
  Eval:
-    - DistillationGTCELoss:
+    - CELoss:
        weight: 1.0
-        model_names: ["Student"]
-        

 Optimizer:
  name: Momentum
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B0.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B0.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B1.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B1.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B2.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B2.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B3.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B3.yaml
@ -44,11 +44,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B4.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B4.yaml
@ -44,11 +44,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/PVTV2/PVT_V2_B5.yaml
+++ b/ppcls/configs/ImageNet/PVTV2/PVT_V2_B5.yaml
@ -44,11 +44,12 @@ Optimizer:
  no_weight_decay_name: pos_embed1 pos_embed2 pos_embed3 pos_embed4 cls_token
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 5e-6
+    learning_rate: 1e-3
+    eta_min: 1e-5
    warmup_epoch: 20
-    warmup_start_lr: 5e-7
+    warmup_start_lr: 1e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window12_384.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window12_384.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window7_224.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_base_patch4_window7_224.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window12_384.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window12_384.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window7_224.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_large_patch4_window7_224.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_small_patch4_window7_224.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_small_patch4_window7_224.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_tiny_patch4_window7_224.yaml
+++ b/ppcls/configs/ImageNet/SwinTransformer/SwinTransformer_tiny_patch4_window7_224.yaml
@ -41,11 +41,12 @@ Optimizer:
  no_weight_decay_name: absolute_pos_embed relative_position_bias_table .bias norm 
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 20
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/alt_gvt_base.yaml
+++ b/ppcls/configs/ImageNet/Twins/alt_gvt_base.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/alt_gvt_large.yaml
+++ b/ppcls/configs/ImageNet/Twins/alt_gvt_large.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/alt_gvt_small.yaml
+++ b/ppcls/configs/ImageNet/Twins/alt_gvt_small.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/pcpvt_base.yaml
+++ b/ppcls/configs/ImageNet/Twins/pcpvt_base.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/pcpvt_large.yaml
+++ b/ppcls/configs/ImageNet/Twins/pcpvt_large.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/configs/ImageNet/Twins/pcpvt_small.yaml
+++ b/ppcls/configs/ImageNet/Twins/pcpvt_small.yaml
@ -43,11 +43,12 @@ Optimizer:
  no_weight_decay_name: norm cls_token proj.0.weight proj.1.weight proj.2.weight proj.3.weight pos_block
  one_dim_param_no_weight_decay: True
  lr:
+    # for 8 cards
    name: Cosine
-    learning_rate: 5e-4
-    eta_min: 1e-5
+    learning_rate: 1e-3
+    eta_min: 2e-5
    warmup_epoch: 5
-    warmup_start_lr: 1e-6
+    warmup_start_lr: 2e-6


 # data loader for train and eval
--- a/ppcls/engine/engine.py
+++ b/ppcls/engine/engine.py
@ -98,8 +98,8 @@ class Engine(object):
        logger.info('train with paddle {} and device {}'.format(
            paddle.__version__, self.device))

-        # AMP training
-        self.amp = True if "AMP" in self.config and self.mode == "train" else False
+        # AMP training and evaluating
+        self.amp = "AMP" in self.config
        if self.amp and self.config["AMP"] is not None:
            self.scale_loss = self.config["AMP"].get("scale_loss", 1.0)
            self.use_dynamic_loss_scaling = self.config["AMP"].get(
@ -250,12 +250,17 @@ class Engine(object):
                    level=amp_level,
                    save_dtype='float32')

-        # for distributed
+        # check the gpu num
        world_size = dist.get_world_size()
        self.config["Global"]["distributed"] = world_size != 1
-        if world_size != 4 and self.mode == "train":
-            msg = f"The training strategy in config files provided by PaddleClas is based on 4 gpus. But the number of gpus is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use config files in PaddleClas to train."
-            logger.warning(msg)
+        if self.mode == "train":
+            std_gpu_num = 8 if self.config["Optimizer"][
+                "name"] == "AdamW" else 4
+            if world_size != std_gpu_num:
+                msg = f"The training strategy provided by PaddleClas is based on {std_gpu_num} gpus. But the number of gpu is {world_size} in current training. Please modify the stategy (learning rate, batch size and so on) if use this config to train."
+                logger.warning(msg)
+
+        # for distributed
        if self.config["Global"]["distributed"]:
            dist.init_parallel_env()
            self.model = paddle.DataParallel(self.model)
--- a/ppcls/engine/evaluation/classification.py
+++ b/ppcls/engine/evaluation/classification.py
@ -73,68 +73,66 @@ def classification_eval(engine, epoch_id=0):
                    },
                    level=amp_level):
                out = engine.model(batch[0])
-                # calc loss
-                if engine.eval_loss_func is not None:
-                    loss_dict = engine.eval_loss_func(out, batch[1])
-                    for key in loss_dict:
-                        if key not in output_info:
-                            output_info[key] = AverageMeter(key, '7.5f')
-                        output_info[key].update(loss_dict[key].numpy()[0],
-                                                batch_size)
        else:
            out = engine.model(batch[0])
-            # calc loss
-            if engine.eval_loss_func is not None:
-                loss_dict = engine.eval_loss_func(out, batch[1])
-                for key in loss_dict:
-                    if key not in output_info:
-                        output_info[key] = AverageMeter(key, '7.5f')
-                    output_info[key].update(loss_dict[key].numpy()[0],
-                                            batch_size)

        # just for DistributedBatchSampler issue: repeat sampling
        current_samples = batch_size * paddle.distributed.get_world_size()
        accum_samples += current_samples

-        # calc metric
-        if engine.eval_metric_func is not None:
-            if paddle.distributed.get_world_size() > 1:
-                label_list = []
-                paddle.distributed.all_gather(label_list, batch[1])
-                labels = paddle.concat(label_list, 0)
+        if isinstance(out, dict) and "Student" in out:
+            out = out["Student"]
+        if isinstance(out, dict) and "logits" in out:
+            out = out["logits"]

-                if isinstance(out, dict):
-                    if "Student" in out:
-                        out = out["Student"]
-                        if isinstance(out, dict):
-                            out = out["logits"]
-                    elif "logits" in out:
-                        out = out["logits"]
-                    else:
-                        msg = "Error: Wrong key in out!"
-                        raise Exception(msg)
-                if isinstance(out, list):
-                    pred = []
-                    for x in out:
-                        pred_list = []
-                        paddle.distributed.all_gather(pred_list, x)
-                        pred_x = paddle.concat(pred_list, 0)
-                        pred.append(pred_x)
-                else:
+        # gather Tensor when distributed
+        if paddle.distributed.get_world_size() > 1:
+            label_list = []
+            paddle.distributed.all_gather(label_list, batch[1])
+            labels = paddle.concat(label_list, 0)
+
+            if isinstance(out, list):
+                preds = []
+                for x in out:
                    pred_list = []
-                    paddle.distributed.all_gather(pred_list, out)
-                    pred = paddle.concat(pred_list, 0)
-
-                if accum_samples > total_samples and not engine.use_dali:
-                    pred = pred[:total_samples + current_samples -
-                                accum_samples]
-                    labels = labels[:total_samples + current_samples -
-                                    accum_samples]
-                    current_samples = total_samples + current_samples - accum_samples
-                metric_dict = engine.eval_metric_func(pred, labels)
+                    paddle.distributed.all_gather(pred_list, x)
+                    pred_x = paddle.concat(pred_list, 0)
+                    preds.append(pred_x)
            else:
-                metric_dict = engine.eval_metric_func(out, batch[1])
+                pred_list = []
+                paddle.distributed.all_gather(pred_list, out)
+                preds = paddle.concat(pred_list, 0)

+            if accum_samples > total_samples and not engine.use_dali:
+                preds = preds[:total_samples + current_samples - accum_samples]
+                labels = labels[:total_samples + current_samples -
+                                accum_samples]
+                current_samples = total_samples + current_samples - accum_samples
+        else:
+            labels = batch[1]
+            preds = out
+
+        # calc loss
+        if engine.eval_loss_func is not None:
+            if engine.amp and engine.config["AMP"].get("use_fp16_test", False):
+                amp_level = engine.config['AMP'].get("level", "O1").upper()
+                with paddle.amp.auto_cast(
+                        custom_black_list={
+                            "flatten_contiguous_range", "greater_than"
+                        },
+                        level=amp_level):
+                    loss_dict = engine.eval_loss_func(preds, labels)
+            else:
+                loss_dict = engine.eval_loss_func(preds, labels)
+
+            for key in loss_dict:
+                if key not in output_info:
+                    output_info[key] = AverageMeter(key, '7.5f')
+                output_info[key].update(loss_dict[key].numpy()[0],
+                                        current_samples)
+        #  calc metric
+        if engine.eval_metric_func is not None:
+            metric_dict = engine.eval_metric_func(preds, labels)
            for key in metric_dict:
                if metric_key is None:
                    metric_key = key
--- a/ppcls/engine/evaluation/retrieval.py
+++ b/ppcls/engine/evaluation/retrieval.py
@ -259,10 +259,8 @@ def eval_func(distmat, q_pids, g_pids, q_camids, g_camids, max_rank=50):


 def cal_feature(engine, name='gallery'):
-    all_feas = None
-    all_image_id = None
-    all_unique_id = None
    has_unique_id = False
+    all_unique_id = None

    if name == 'gallery':
        dataloader = engine.gallery_dataloader
@ -273,6 +271,9 @@ def cal_feature(engine, name='gallery'):
    else:
        raise RuntimeError("Only support gallery or query dataset")

+    batch_feas_list = []
+    img_id_list = []
+    unique_id_list = []
    max_iter = len(dataloader) - 1 if platform.system() == "Windows" else len(
        dataloader)
    for idx, batch in enumerate(dataloader):  # load is very time-consuming
@ -317,32 +318,39 @@ def cal_feature(engine, name='gallery'):
        if engine.config["Global"].get("feature_binarize") == "sign":
            batch_feas = paddle.sign(batch_feas).astype("float32")

-        if all_feas is None:
-            all_feas = batch_feas
+        if paddle.distributed.get_world_size() > 1:
+            batch_feas_gather = []
+            img_id_gather = []
+            unique_id_gather = []
+            paddle.distributed.all_gather(batch_feas_gather, batch_feas)
+            paddle.distributed.all_gather(img_id_gather, batch[1])
+            batch_feas_list.append(paddle.concat(batch_feas_gather))
+            img_id_list.append(paddle.concat(img_id_gather))
            if has_unique_id:
-                all_unique_id = batch[2]
-            all_image_id = batch[1]
+                paddle.distributed.all_gather(unique_id_gather, batch[2])
+                unique_id_list.append(paddle.concat(unique_id_gather))
        else:
-            all_feas = paddle.concat([all_feas, batch_feas])
-            all_image_id = paddle.concat([all_image_id, batch[1]])
+            batch_feas_list.append(batch_feas)
+            img_id_list.append(batch[1])
            if has_unique_id:
-                all_unique_id = paddle.concat([all_unique_id, batch[2]])
+                unique_id_list.append(batch[2])

    if engine.use_dali:
        dataloader.reset()

-    if paddle.distributed.get_world_size() > 1:
-        feat_list = []
-        img_id_list = []
-        unique_id_list = []
-        paddle.distributed.all_gather(feat_list, all_feas)
-        paddle.distributed.all_gather(img_id_list, all_image_id)
-        all_feas = paddle.concat(feat_list, axis=0)
-        all_image_id = paddle.concat(img_id_list, axis=0)
-        if has_unique_id:
-            paddle.distributed.all_gather(unique_id_list, all_unique_id)
-            all_unique_id = paddle.concat(unique_id_list, axis=0)
+    all_feas = paddle.concat(batch_feas_list)
+    all_img_id = paddle.concat(img_id_list)
+    if has_unique_id:
+        all_unique_id = paddle.concat(unique_id_list)
+
+    # just for DistributedBatchSampler issue: repeat sampling
+    total_samples = len(
+        dataloader.dataset) if not engine.use_dali else dataloader.size
+    all_feas = all_feas[:total_samples]
+    all_img_id = all_img_id[:total_samples]
+    if has_unique_id:
+        all_unique_id = all_unique_id[:total_samples]

    logger.info("Build {} done, all feat shape: {}, begin to eval..".format(
        name, all_feas.shape))
-    return all_feas, all_image_id, all_unique_id
+    return all_feas, all_img_id, all_unique_id
--- a/ppcls/loss/deephashloss.py
+++ b/ppcls/loss/deephashloss.py
@ -20,6 +20,7 @@ class DSHSDLoss(nn.Layer):
    """
    # DSHSD(IEEE ACCESS 2019)
    # paper [Deep Supervised Hashing Based on Stable Distribution](https://ieeexplore.ieee.org/document/8648432/)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/DSHSD.py
    """

    def __init__(self, alpha, multi_label=False):
@ -62,6 +63,7 @@ class DSHSDLoss(nn.Layer):
 class LCDSHLoss(nn.Layer):
    """
    # paper [Locality-Constrained Deep Supervised Hashing for Image Retrieval](https://www.ijcai.org/Proceedings/2017/0499.pdf)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/LCDSH.py
    """

    def __init__(self, n_class, _lambda):
@ -100,6 +102,7 @@ class DCHLoss(paddle.nn.Layer):
    """
    # paper [Deep Cauchy Hashing for Hamming Space Retrieval]
    URL:(http://ise.thss.tsinghua.edu.cn/~mlong/doc/deep-cauchy-hashing-cvpr18.pdf)
+    # code reference: https://github.com/swuxyj/DeepHash-pytorch/blob/master/DCH.py
    """

    def __init__(self, gamma, _lambda, n_class):
--- a/ppcls/loss/emlloss.py
+++ b/ppcls/loss/emlloss.py
@ -23,6 +23,11 @@ from .comfunc import rerange_index


 class EmlLoss(paddle.nn.Layer):
+    """Ensemble Metric Learning Loss
+    paper: [Large Scale Strongly Supervised Ensemble Metric Learning, with Applications to Face Verification and Retrieval](https://arxiv.org/pdf/1212.6094.pdf)
+    code reference: https://github.com/PaddlePaddle/models/blob/develop/PaddleCV/metric_learning/losses/emlloss.py
+    """
+
    def __init__(self, batch_size=40, samples_each_class=2):
        super(EmlLoss, self).__init__()
        assert (batch_size % samples_each_class == 0)
--- a/ppcls/loss/googlenetloss.py
+++ b/ppcls/loss/googlenetloss.py
@ -18,11 +18,13 @@ import paddle.nn.functional as F
 class GoogLeNetLoss(nn.Layer):
    """
    Cross entropy loss used after googlenet
+    reference paper: [https://arxiv.org/pdf/1409.4842v1.pdf](Going Deeper with Convolutions)
    """
+
    def __init__(self, epsilon=None):
        super().__init__()
-        assert (epsilon is None or epsilon <= 0 or epsilon >= 1), "googlenet is not support label_smooth"
-            
+        assert (epsilon is None or epsilon <= 0 or
+                epsilon >= 1), "googlenet is not support label_smooth"

    def forward(self, inputs, label):
        input0, input1, input2 = inputs
--- a/ppcls/loss/msmloss.py
+++ b/ppcls/loss/msmloss.py
@ -21,10 +21,12 @@ from .comfunc import rerange_index

 class MSMLoss(paddle.nn.Layer):
    """
-    MSMLoss Loss, based on triplet loss. USE P * K samples.
+    paper : [Margin Sample Mining Loss: A Deep Learning Based Method for Person Re-identification](https://arxiv.org/pdf/1710.00478.pdf)
+    code reference: https://github.com/michuanhaohao/keras_reid/blob/master/reid_tripletcls.py
+    Margin Sample Mining Loss, based on triplet loss. USE P * K samples.
    the batch size is fixed. Batch_size = P * K;  but the K may vary between batches.
    same label gather together
-    
+
            supported_metrics = [
            'euclidean',
            'sqeuclidean',
@ -41,7 +43,7 @@ class MSMLoss(paddle.nn.Layer):
        self.rerange_index = rerange_index(batch_size, samples_each_class)

    def forward(self, input, target=None):
-        #normalization 
+        #normalization
        features = input["features"]
        features = self._nomalize(features)
        samples_each_class = self.samples_each_class
@ -53,7 +55,7 @@ class MSMLoss(paddle.nn.Layer):
                features, axis=0)
        similary_matrix = paddle.sum(paddle.square(diffs), axis=-1)

-        #rerange 
+        #rerange
        tmp = paddle.reshape(similary_matrix, shape=[-1, 1])
        tmp = paddle.gather(tmp, index=rerange_index)
        similary_matrix = paddle.reshape(tmp, shape=[-1, self.batch_size])
--- a/ppcls/loss/npairsloss.py
+++ b/ppcls/loss/npairsloss.py
@ -5,6 +5,11 @@ import paddle


 class NpairsLoss(paddle.nn.Layer):
+    """Npair_loss_
+    paper [Improved deep metric learning with multi-class N-pair loss objective](https://dl.acm.org/doi/10.5555/3157096.3157304)
+    code reference: https://www.tensorflow.org/versions/r1.15/api_docs/python/tf/contrib/losses/metric_learning/npairs_loss
+    """
+
    def __init__(self, reg_lambda=0.01):
        super(NpairsLoss, self).__init__()
        self.reg_lambda = reg_lambda
--- a/ppcls/loss/pairwisecosface.py
+++ b/ppcls/loss/pairwisecosface.py
@ -23,6 +23,11 @@ import paddle.nn.functional as F


 class PairwiseCosface(nn.Layer):
+    """
+    paper: Circle Loss: A Unified Perspective of Pair Similarity Optimization
+    code reference: https://github.com/leoluopy/circle-loss-demonstration/blob/main/circle_loss.py
+    """
+
    def __init__(self, margin, gamma):
        super(PairwiseCosface, self).__init__()
        self.margin = margin
@ -36,8 +41,10 @@ class PairwiseCosface(nn.Layer):
        dist_mat = paddle.matmul(embedding, embedding, transpose_y=True)

        N = dist_mat.shape[0]
-        is_pos = targets.reshape([N,1]).expand([N,N]).equal(paddle.t(targets.reshape([N,1]).expand([N,N]))).astype('float')
-        is_neg = targets.reshape([N,1]).expand([N,N]).not_equal(paddle.t(targets.reshape([N,1]).expand([N,N]))).astype('float')
+        is_pos = targets.reshape([N, 1]).expand([N, N]).equal(
+            paddle.t(targets.reshape([N, 1]).expand([N, N]))).astype('float')
+        is_neg = targets.reshape([N, 1]).expand([N, N]).not_equal(
+            paddle.t(targets.reshape([N, 1]).expand([N, N]))).astype('float')

        # Mask scores related to itself
        is_pos = is_pos - paddle.eye(N, N)
@ -46,10 +53,12 @@ class PairwiseCosface(nn.Layer):
        s_n = dist_mat * is_neg

        logit_p = -self.gamma * s_p + (-99999999.) * (1 - is_pos)
-        logit_n = self.gamma * (s_n + self.margin) + (-99999999.) * (1 - is_neg)
+        logit_n = self.gamma * (s_n + self.margin) + (-99999999.) * (1 - is_neg
+                                                                     )
+
+        loss = F.softplus(
+            paddle.logsumexp(
+                logit_p, axis=1) + paddle.logsumexp(
+                    logit_n, axis=1)).mean()

-        loss = F.softplus(paddle.logsumexp(logit_p, axis=1) + paddle.logsumexp(logit_n, axis=1)).mean()
-  
        return {"PairwiseCosface": loss}
-
-
--- a/ppcls/loss/rkdloss.py
+++ b/ppcls/loss/rkdloss.py
@ -29,6 +29,7 @@ def pdist(e, squared=False, eps=1e-12):


 class RKdAngle(nn.Layer):
+    # paper : [Relational Knowledge Distillation](https://arxiv.org/abs/1904.05068?context=cs.LG)
    # reference: https://github.com/lenscloth/RKD/blob/master/metric/loss.py
    def __init__(self, target_size=None):
        super().__init__()
@ -64,6 +65,7 @@ class RKdAngle(nn.Layer):


 class RkdDistance(nn.Layer):
+    # paper : [Relational Knowledge Distillation](https://arxiv.org/abs/1904.05068?context=cs.LG)
    # reference: https://github.com/lenscloth/RKD/blob/master/metric/loss.py
    def __init__(self, eps=1e-12, target_size=1):
        super().__init__()
--- a/ppcls/loss/supconloss.py
+++ b/ppcls/loss/supconloss.py
@ -4,6 +4,7 @@ from paddle import nn

 class SupConLoss(nn.Layer):
    """Supervised Contrastive Learning: https://arxiv.org/pdf/2004.11362.pdf.
+    code reference: https://github.com/HobbitLong/SupContrast/blob/master/losses.py
    It also supports the unsupervised contrastive loss in SimCLR"""

    def __init__(self,
--- a/ppcls/loss/trihardloss.py
+++ b/ppcls/loss/trihardloss.py
@ -22,10 +22,12 @@ from .comfunc import rerange_index

 class TriHardLoss(paddle.nn.Layer):
    """
+    paper: In Defense of the Triplet Loss for Person Re-Identification
+    code reference: https://github.com/VisualComputingInstitute/triplet-reid/blob/master/loss.py
    TriHard Loss, based on triplet loss. USE P * K samples.
    the batch size is fixed. Batch_size = P * K;  but the K may vary between batches.
    same label gather together
-    
+
            supported_metrics = [
            'euclidean',
            'sqeuclidean',
@ -45,7 +47,7 @@ class TriHardLoss(paddle.nn.Layer):
        features = input["features"]
        assert (self.batch_size == features.shape[0])

-        #normalization 
+        #normalization
        features = self._nomalize(features)
        samples_each_class = self.samples_each_class
        rerange_index = paddle.to_tensor(self.rerange_index)
@ -56,7 +58,7 @@ class TriHardLoss(paddle.nn.Layer):
                features, axis=0)
        similary_matrix = paddle.sum(paddle.square(diffs), axis=-1)

-        #rerange 
+        #rerange
        tmp = paddle.reshape(similary_matrix, shape=[-1, 1])
        tmp = paddle.gather(tmp, index=rerange_index)
        similary_matrix = paddle.reshape(tmp, shape=[-1, self.batch_size])
--- a/ppcls/loss/triplet.py
+++ b/ppcls/loss/triplet.py
@ -1,10 +1,10 @@
-# copyright (c) 2021 PaddlePaddle Authors. All Rights Reserve.
+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
 # You may obtain a copy of the License at
 #
-#    http://www.apache.org/licenses/LICENSE-2.0
+#     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,
@ -22,6 +22,8 @@ import paddle.nn as nn

 class TripletLossV2(nn.Layer):
    """Triplet loss with hard positive/negative mining.
+    paper : [Facenet: A unified embedding for face recognition and clustering](https://arxiv.org/pdf/1503.03832.pdf)
+    code reference: https://github.com/okzhili/Cartoon-face-recognition/blob/master/loss/triplet_loss.py
    Args:
        margin (float): margin for triplet.
    """
--- a/ppcls/optimizer/init.py
+++ b/ppcls/optimizer/init.py
@ -120,8 +120,6 @@ def build_optimizer(config, epochs, step_each_epoch, model_list=None):
                    elif hasattr(model_list[i], optim_scope):
                        optim_model.append(getattr(model_list[i], optim_scope))

-        assert len(optim_model) == 1, \
-            "Invalid optim model for optim scope({}), number of optim_model={}".format(optim_scope, len(optim_model))
        optim = getattr(optimizer, optim_name)(
            learning_rate=lr, grad_clip=grad_clip,
            **optim_cfg)(model_list=optim_model)
--- a/test_tipc/config/CSWinTransformer/CSWinTransformer_tiny_224_train_infer_python.txt
+++ b/test_tipc/config/CSWinTransformer/CSWinTransformer_tiny_224_train_infer_python.txt
@ -13,14 +13,14 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null
 null:null
 null:null
 ##
-===========================eval_params=========================== 
+===========================eval_params===========================
 eval:tools/eval.py -c ppcls/configs/ImageNet/CSWinTransformer/CSWinTransformer_tiny_224.yaml
 null:null
 ##
--- a/test_tipc/config/MobileViT/MobileViT_S_train_infer_python.txt
+++ b/test_tipc/config/MobileViT/MobileViT_S_train_infer_python.txt
@ -13,14 +13,14 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViT/MobileViT_S.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/MobileViT/MobileViT_S.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null
 null:null
 null:null
 ##
-===========================eval_params=========================== 
+===========================eval_params===========================
 eval:tools/eval.py -c ppcls/configs/ImageNet/MobileViT/MobileViT_S.yaml
 null:null
 ##
--- a/test_tipc/config/PVTV2/PVT_V2_B2_Linear_train_infer_python.txt
+++ b/test_tipc/config/PVTV2/PVT_V2_B2_Linear_train_infer_python.txt
@ -13,7 +13,7 @@ train_infer_img_dir:./dataset/ILSVRC2012/val
 null:null
 ##
 trainer:norm_train
-norm_train:tools/train.py -c ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False
+norm_train:tools/train.py -c ppcls/configs/ImageNet/PVTV2/PVT_V2_B2_Linear.yaml -o Global.seed=1234 -o DataLoader.Train.sampler.shuffle=False -o DataLoader.Train.loader.num_workers=0 -o DataLoader.Train.loader.use_shared_memory=False -o Global.print_batch_step=1
 pact_train:null
 fpgm_train:null
 distill_train:null
--- a/test_tipc/prepare.sh
+++ b/test_tipc/prepare.sh
@ -1,7 +1,7 @@
 #!/bin/bash
 FILENAME=$1

-# MODE be one of ['lite_train_lite_infer' 'lite_train_whole_infer' 'whole_train_whole_infer',  
+# MODE be one of ['lite_train_lite_infer' 'lite_train_whole_infer' 'whole_train_whole_infer',
 #                 'whole_infer', 'klquant_whole_infer',
 #                 'cpp_infer', 'serving_infer',  'lite_infer']

@ -67,9 +67,9 @@ if [ ${MODE} = "cpp_infer" ];then
 	    model_dir=${tar_name%.*}
 	    eval "tar xf ${tar_name}"
 	    eval "mv ${model_dir} ${cls_inference_model_dir}"
-	    
+
 	    eval "wget -nc $det_inference_url"
-	    tar_name=$(func_get_url_file_name "$det_inference_url") 
+	    tar_name=$(func_get_url_file_name "$det_inference_url")
 	    model_dir=${tar_name%.*}
 	    eval "tar xf ${tar_name}"
 	    eval "mv ${model_dir} ${det_inference_model_dir}"
@ -120,7 +120,7 @@ if [ ${MODE} = "lite_train_lite_infer" ] || [ ${MODE} = "lite_train_whole_infer"
    wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/data/whole_chain/whole_chain_little_train.tar
    tar xf whole_chain_little_train.tar
    ln -s whole_chain_little_train ILSVRC2012
-    cd ILSVRC2012 
+    cd ILSVRC2012
    mv train.txt train_list.txt
    mv val.txt val_list.txt
    cp -r train/* val/
@ -132,7 +132,7 @@ elif [ ${MODE} = "whole_infer" ] || [ ${MODE} = "klquant_whole_infer" ];then
    wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/data/whole_chain/whole_chain_infer.tar
    tar xf whole_chain_infer.tar
    ln -s whole_chain_infer ILSVRC2012
-    cd ILSVRC2012 
+    cd ILSVRC2012
    mv val.txt val_list.txt
    ln -s val_list.txt train_list.txt
    cd ../../
@ -153,7 +153,7 @@ elif [ ${MODE} = "whole_train_whole_infer" ];then
    wget -nc https://paddle-imagenet-models-name.bj.bcebos.com/data/whole_chain/whole_chain_CIFAR100.tar
    tar xf whole_chain_CIFAR100.tar
    ln -s whole_chain_CIFAR100 ILSVRC2012
-    cd ILSVRC2012 
+    cd ILSVRC2012
    mv train.txt train_list.txt
    mv test.txt val_list.txt
    cd ../../