Merge pull request #171 from littletomatodonkey/fix_main

add faq and competition en doc

README.md

@@ -19,10 +19,10 @@
 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)。
 
 <div align="center">
-    <img src="./docs/images/models/V100_benchmark/v100.fp32.bs1.main_fps_top1_s.jpg" width="700">
+    <img src="./docs/images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png" width="700">
 </div>
 
-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)。
+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83.0%的ResNet50_vd_ssld_v2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，其中v2表示在训练时添加了AutoAugment数据增广策略。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)。
 
 <div align="center">
 <img

README_en.md

@@ -4,9 +4,9 @@
 
 > **English version of PaddleClas. Updating...**
 
-**Book**: https://paddleclas.readthedocs.io
+**Book**: https://paddleclas-en.readthedocs.io/en/latest/
 
-**Quick start PaddleClas in 30 minutes**: https://paddleclas.readthedocs.io/zh_CN/latest/tutorials/quick_start.html
+**Quick start PaddleClas in 30 minutes**: https://paddleclas-en.readthedocs.io/en/latest/tutorials/quick_start_en.html
 
 ## Introduction
 
@@ -18,14 +18,14 @@ PaddleClas is a toolset for image classification tasks prepared for the industry
 
 ## Rich model zoo
 
-Based on ImageNet1k dataset, PaddleClas provides 23 series of image classification networks such as ResNet, ResNet_vd, Res2Net, HRNet, and MobileNetV3 with brief introductions, reproduction configurations and training tricks. At the same time, the corresponding 117 image classification pretrained models are also available. The GPU inference time of the server-side models are evaluated based on TensorRT. The CPU inference time and storage size of the mobile-side models are evaluated on the Snapdragon 855 (SD855). For more detailed information on the supported pretrained models and their download links, please refer to [**models introduction tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html).
+Based on ImageNet1k dataset, PaddleClas provides 23 series of image classification networks such as ResNet, ResNet_vd, Res2Net, HRNet, and MobileNetV3 with brief introductions, reproduction configurations and training tricks. At the same time, the corresponding 117 image classification pretrained models are also available. The GPU inference time of the server-side models are evaluated based on TensorRT. The CPU inference time and storage size of the mobile-side models are evaluated on the Snapdragon 855 (SD855). For more detailed information on the supported pretrained models and their download links, please refer to [**models introduction tutorial**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html).
 
 
 <div align="center">
-    <img src="./docs/images/models/V100_benchmark/v100.fp32.bs1.main_fps_top1_s.jpg" width="700">
+    <img src="./docs/images/models/T4_benchmark/t4.fp32.bs4.main_fps_top1.png" width="700">
 </div>
 
-The above figure shows some of the latest server-side pretrained models. It can be seen from the figure that when using V100 GPU with FP32 and TensorRT, the `Top1` accuracy of the ResNet50_vd_ssld pretrained model on ImageNet1k-val dataset is **82.4%** and that of ResNet101_vd_ssld pretrained model is 83.7%. These pretained models are obtained from  SSLD knowledge distillation solution provided by PaddleClas. The marks of the same color and symbol in the figure represent models of different model sizes in the same series. For the introduction of different models, FLOPS, Params and detailed GPU inference time (including the inference speed of T4 GPU with different batch size), please refer to the documentation tutorial for more details: [https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)
+The above figure shows some of the latest server-side pretrained models. It can be seen from the figure that when using V100 GPU with FP32 and TensorRT, the `Top1` accuracy of the ResNet50_vd_ssld pretrained model on ImageNet1k-val dataset is **83.0%** and that of ResNet101_vd_ssld pretrained model is 83.7%. These pretained models are obtained from  SSLD knowledge distillation solution provided by PaddleClas. The marks of the same color and symbol in the figure represent models of different model sizes in the same series. For the introduction of different models, FLOPS, Params and detailed GPU inference time (including the inference speed of T4 GPU with different batch size), please refer to the documentation tutorial for more details: [https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html)
 
 
 <div align="center">
@@ -34,7 +34,7 @@ src="./docs/images/models/mobile_arm_top1.png" width="700">
 </div>
 
 
-The above figure shows the performance of some commonly used mobile-side models, including MobileNetV1, MobileNetV2, MobileNetV3 and ShuffleNetV2 series. The inference time is tested on Snapdragon 855 (SD855) with the batch size set as 1. The `Top1` accuracy of the MV3_large_x1_0_ssld, MV3_small_x1_0_ssld, MV1_ssld and MV2_ssld pretrained model on ImageNet1k-val dataset are 79%, 71.3%, 76.74%, 77.89%, respectively (M is short for MobileNet). MV3_large_x1_0_ssld_int8 is a quantizatied pretrained model for MV3_large_x1_0. More details about the mobile-side models can be seen in [**models introduction tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/models/models_intro.html)
+The above figure shows the performance of some commonly used mobile-side models, including MobileNetV1, MobileNetV2, MobileNetV3 and ShuffleNetV2 series. The inference time is tested on Snapdragon 855 (SD855) with the batch size set as 1. The `Top1` accuracy of the MV3_large_x1_0_ssld, MV3_small_x1_0_ssld, MV1_ssld and MV2_ssld pretrained model on ImageNet1k-val dataset are 79%, 71.3%, 76.74%, 77.89%, respectively (M is short for MobileNet). MV3_large_x1_0_ssld_int8 is a quantizatied pretrained model for MV3_large_x1_0. More details about the mobile-side models can be seen in [**models introduction tutorial**](https://paddleclas-en.readthedocs.io/en/latest/models/models_intro_en.html)
 
 - TODO
 - [ ] Reproduction and performance evaluation of EfficientLite, GhostNet, RegNet and ResNeSt.
@@ -54,7 +54,7 @@ Knowledge distillation refers to using the teacher model to guide the student mo
 src="./docs/images/distillation/distillation_perform_s.jpg" width="700">
 </div>
 
-Taking the ImageNet1k dataset as an example, the following figure shows the SSLD knowledge distillation method framework. The key points of the method include the choice of teacher model, loss calculation method, iteration number, use of unlabeled data, and ImageNet1k dataset finetune. For detailed introduction and experiments, please refer to [**knowledge distillation tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/advanced_tutorials/distillation/index.html)
+Taking the ImageNet1k dataset as an example, the following figure shows the SSLD knowledge distillation method framework. The key points of the method include the choice of teacher model, loss calculation method, iteration number, use of unlabeled data, and ImageNet1k dataset finetune. For detailed introduction and experiments, please refer to [**knowledge distillation tutorial**](https://paddleclas-en.readthedocs.io/en/latest/advanced_tutorials/distillation/distillation_en.html)
 
 <div align="center">
 <img
@@ -72,7 +72,7 @@ src="./docs/images/image_aug/image_aug_samples_s_en.jpg" width="800">
 </div>
 
 
-PaddleClas provides the reproduction of the above 8 data augmentation algorithms and the evaluation of the effect in a unified environment. The following figure shows the performance of different data augmentation methods based on ResNet50. Compared with the standard transformation, using data augmentation, the recognition accuracy  can be increased by up to 1%. For more detailed introduction of data augmentation methods, please refer to the [**data augmentation tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/advanced_tutorials/image_augmentation/index.html).
+PaddleClas provides the reproduction of the above 8 data augmentation algorithms and the evaluation of the effect in a unified environment. The following figure shows the performance of different data augmentation methods based on ResNet50. Compared with the standard transformation, using data augmentation, the recognition accuracy  can be increased by up to 1%. For more detailed introduction of data augmentation methods, please refer to the [**data augmentation tutorial**](https://paddleclas-en.readthedocs.io/en/latest/advanced_tutorials/image_augmentation/ImageAugment_en.html).
 
 
 <div align="center">
@@ -83,12 +83,12 @@ src="./docs/images/image_aug/main_image_aug_s.jpg" width="600">
 
 ## Quick start
 
-Based on flowers102 dataset, one can easily experience different networks, pretrained models and SSLD knowledge distillation method in PaddleClas. More details can be seen in [**Quick start PaddleClas in 30 minutes**](https://paddleclas.readthedocs.io/zh_CN/latest/tutorials/quick_start.html).
+Based on flowers102 dataset, one can easily experience different networks, pretrained models and SSLD knowledge distillation method in PaddleClas. More details can be seen in [**Quick start PaddleClas in 30 minutes**](https://paddleclas-en.readthedocs.io/en/latest/tutorials/quick_start_en.html).
 
 
 ## Getting started
 
-For installation, model training, inference, evaluation and finetune in PaddleClas, you can refer to [**gettting started tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/tutorials/index.html).
+For installation, model training, inference, evaluation and finetune in PaddleClas, you can refer to [**gettting started tutorial**](https://paddleclas-en.readthedocs.io/en/latest/tutorials/index.html).
 
 
 ## Featured extension and application
@@ -107,12 +107,12 @@ The models trained on ImageNet1K dataset are often used as pretrained models for
 | Geology    | class_num:4<br/>train/val:671/296         | 0.5719        | 0.6781        |
 
 
-The 100,000 categories' pretrained model can be downloaded here: [download link](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_10w_pretrained.tar). More details can be seen in [**Transfer learning tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/application/transfer_learning.html).
+The 100,000 categories' pretrained model can be downloaded here: [download link](https://paddle-imagenet-models-name.bj.bcebos.com/ResNet50_vd_10w_pretrained.tar). More details can be seen in [**Transfer learning tutorial**](https://paddleclas-en.readthedocs.io/en/latest/application/transfer_learning_en.html).
 
 
 ### Object detection
 
-In recent years, object detection tasks attract a lot of attention in academia and industry. The ImageNet classification model is often used for pretrained model in object detection, which can directly affect the effect of object detection. Based on 82.39% ResNet50_vd pretrained model, PaddleDetection provides a Practical Server-side Detection solution, PSS-DET. The solution contains many strategies that can effectively improve the performance while taking limited extra computation cost, such as model pruning, better pretrained model, deformable convolution, cascade rcnn, autoaugment, libra sampling and multi-scale training. Compared with the 79.12% ImageNet1k pretrained model, the 82.39% model can help improve the COCO mAP by 1.5% without any computation cost. Using PSS-DET, the inference speed on single V100 GPU can reach 20FPS when COCO mAP is 47.8%, and reach 61FPS when COCO mAP is 41.6%. For more details, please refer to [**Object Detection tutorial**](https://paddleclas.readthedocs.io/zh_CN/latest/application/object_detection.html).
+In recent years, object detection tasks attract a lot of attention in academia and industry. The ImageNet classification model is often used for pretrained model in object detection, which can directly affect the effect of object detection. Based on 82.39% ResNet50_vd pretrained model, PaddleDetection provides a Practical Server-side Detection solution, PSS-DET. The solution contains many strategies that can effectively improve the performance while taking limited extra computation cost, such as model pruning, better pretrained model, deformable convolution, cascade rcnn, autoaugment, libra sampling and multi-scale training. Compared with the 79.12% ImageNet1k pretrained model, the 82.39% model can help improve the COCO mAP by 1.5% without any computation cost. Using PSS-DET, the inference speed on single V100 GPU can reach 20FPS when COCO mAP is 47.8%, and reach 61FPS when COCO mAP is 41.6%. For more details, please refer to [**Object Detection tutorial**](https://paddleclas-en.readthedocs.io/en/latest/application/object_detection_en.html).
 
 
 - TODO

docs/en/advanced_tutorials/distillation/distillation_en.md

@@ -10,7 +10,7 @@ With enough training data, increasing parameters of the neural network by buildi
 Parameter redundancy exists in deep neural networks. There are several methods to compress the model suck as pruning ,quantization, knowledge distillation, etc. Knowledge distillation refers to using the teacher model to guide the student model to learn specific tasks, ensuring that the small model has a relatively large effect improvement with the computation cost unchanged, and even obtains similar accuracy with the large model [1]. Combining some of the existing distillation methods [2,3], PaddleClas provides a simple semi-supervised label knowledge distillation solution (SSLD). Top-1 Accuarcy on ImageNet1k dataset has an improvement of more than 3% based on ResNet_vd and MobileNet series, which can be shown as below.
 
 
-![](../../../images/distillation/distillation_perform.png)
+![](../../../images/distillation/distillation_perform_s.jpg)
 
 
 # 2. SSLD
@@ -97,6 +97,12 @@ Finetuning is carried out on ImageNet1k dataset to restore distribution between
 | ResNet50_vd | 60 | 7e-5 | 1024/32 | 0.004 | cosine_decay_warmup | 82.39% |
 | ResNet101_vd | 30 | 7e-5 | 1024/32 | 0.004 | cosine_decay_warmup | 83.73% |
 
+## 3.4 Data agmentation and Fix strategy
+
+* Based on experiments mentioned above, we add AutoAugment [4] during training process, and reduced l2_decay from 4e-5 t 2e-5. Finally, the Top-1 accuracy on ImageNet1k dataset can reach 82.99%, with 0.6% improvement compared to the standard SSLD distillation strategy.
+
+* For image classsification tasks, The model accuracy can be further improved when the test scale is 1.15 times that of training[5]. For the 82.99% ResNet50_vd pretrained model, it comes to 83.7% using 320x320 for the evaluation. We use Fix strategy to finetune the model with the training scale set as 320x320. During the process, the pre-preocessing pipeline is same for both training and test. All the weights except the fully connected layer are freezed. Finally the top-1 accuracy comes to **84.0%**.
+
 
 # 4. Application of the distillation model
 
@@ -225,3 +231,7 @@ python -m paddle.distributed.launch \
 [2] Bagherinezhad H, Horton M, Rastegari M, et al. Label refinery: Improving imagenet classification through label progression[J]. arXiv preprint arXiv:1805.02641, 2018.
 
 [3] Yalniz I Z, Jégou H, Chen K, et al. Billion-scale semi-supervised learning for image classification[J]. arXiv preprint arXiv:1905.00546, 2019.
+
+[4] Cubuk E D, Zoph B, Mane D, et al. Autoaugment: Learning augmentation strategies from data[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2019: 113-123.
+
+[5] Touvron H, Vedaldi A, Douze M, et al. Fixing the train-test resolution discrepancy[C]//Advances in Neural Information Processing Systems. 2019: 8250-8260.

docs/en/competition_support_en.md

@@ -0,0 +1,21 @@
+### Competition Support
+
+PaddleClas stems from the Baidu's visual business applications and the exploration of frontier visual capabilities. It has helped us achieve leading results in many key events, and continues to promote more frontier visual solutions and landing applications.
+
+
+* 1st place in 2018 Kaggle Open Images V4 object detection challenge
+
+
+* 2nd place in 2019 Kaggle Open Images V5 object detection challenge
+    * The report is avaiable here: [https://arxiv.org/pdf/1911.07171.pdf](https://arxiv.org/pdf/1911.07171.pdf)
+    * The pretrained model and code is avaiable here: [source code](https://github.com/PaddlePaddle/PaddleDetection/blob/master/docs/featured_model/OIDV5_BASELINE_MODEL.md)
+
+* 2nd place in Kacggle Landmark Retrieval Challenge 2019
+    * The report is avaiable here: [https://arxiv.org/abs/1906.03990](https://arxiv.org/abs/1906.03990)
+    * The pretrained model and code is avaiable here: [source code](https://github.com/PaddlePaddle/Research/tree/master/CV/landmark)
+
+* 2nd place in Kaggle Landmark Recognition Challenge 2019
+    * The report is avaiable here: [https://arxiv.org/abs/1906.03990](https://arxiv.org/abs/1906.03990)
+    * The pretrained model and code is avaiable here: [source code](https://github.com/PaddlePaddle/Research/tree/master/CV/landmark)
+
+* A-level certificate of three tasks: printed text OCR, face recognition and landmark recognition in the first multimedia information recognition technology competition

docs/en/faq_en.md

@@ -0,0 +1,48 @@
+# FAQ
+
+>>
+* Why are the metrics different for different cards?
+* A: Fleet is the default option for the use of PaddleClas. Each GPU card is taken as a single trainer and deals with different images, which cause the final small difference. Single card evalution is suggested to get the accurate results if you use `tools/eval.py`. You can also use  `tools/eval_multi_platform.py` to evalute the models on multiple GPU cards, which is also supported on Windows and CPU.
+
+
+>>
+* Q: Why `Mixup` or `Cutmix` is not used even if I have already add the data operation in the configuration file?
+* A: When using `Mixup` or `Cutmix`, you also need to add `use_mix: True` in the configuration file to make it work properly.
+
+
+>>
+* Q: During evaluation and inference, pretrained model address is assgined, but the weights can not be imported. Why?
+* A: Prefix of the pretrained model is needed. For example, if the pretained weights are located in `output/ResNet50_vd/19`, with the filename `output/ResNet50_vd/19/ppcls.pdparams`, then `pretrained_model` in the configuration file needs to be `output/ResNet50_vd/19/ppcls`.
+
+>>
+* Q: Why are the metrics 0.3% lower than that shown in the model zoo for `EfficientNet` series of models?
+* A: Resize method is set as `Cubic` for `EfficientNet`(interpolation is set as 2 in OpenCV), while other models are set as `Bilinear`(interpolation is set as None in OpenCV). Therefore, you need to modify the interpolation explicitly in `ResizeImage`. Specifically, the following configuration is a demo for EfficientNet.
+
+```
+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: What should I do if I want to transform the weights' format from `pdparams` to an earlier version(before Paddle1.7.0), which consists of the scattered files?
+* A: You can use `fluid.load` to load the `pdparams` weights and use `fluid.io.save_vars` to save the weights as scattered files.

docs/en/index.rst

@@ -11,7 +11,7 @@ Welcome to PaddleClas！
    advanced_tutorials/index
    application/index
    extension/index
-   competition_support.md
+   competition_support_en.md
    update_history_en.md
    faq_en.md
 

docs/zh_CN/advanced_tutorials/distillation/distillation.md

@@ -8,7 +8,7 @@
 深度神经网络一般有较多的参数冗余，目前有几种主要的方法对模型进行压缩，减小其参数量。如裁剪、量化、知识蒸馏等，其中知识蒸馏是指使用教师模型(teacher model)去指导学生模型(student model)学习特定任务，保证小模型在参数量不变的情况下，得到比较大的性能提升，甚至获得与大模型相似的精度指标[1]。PaddleClas融合已有的蒸馏方法[2,3]，提供了一种简单的半监督标签知识蒸馏方案（SSLD，Simple Semi-supervised Label Distillation），基于ImageNet1k分类数据集，在ResNet_vd以及MobileNet系列上的精度均有超过3%的绝对精度提升，具体指标如下图所示。
 
 
-![](../../../images/distillation/distillation_perform.png)
+![](../../../images/distillation/distillation_perform_s.jpg)
 
 
 # 二、SSLD 蒸馏策略
@@ -17,10 +17,8 @@
 
 SSLD的流程图如下图所示。
 
-
 ![](../../../images/distillation/ppcls_distillation.png)
 
-
 首先，我们从ImageNet22k中挖掘出了近400万张图片，同时与ImageNet-1k训练集整合在一起，得到了一个新的包含500万张图片的数据集。然后，我们将学生模型与教师模型组合成一个新的网络，该网络分别输出学生模型和教师模型的预测分布，与此同时，固定教师模型整个网络的梯度，而学生模型可以做正常的反向传播。最后，我们将两个模型的logits经过softmax激活函数转换为soft label，并将二者的soft label做JS散度作为损失函数，用于蒸馏模型训练。下面以MobileNetV3（该模型直接训练，精度为75.3%）的知识蒸馏为例，介绍该方案的核心关键点（baseline为79.12%的ResNet50_vd模型蒸馏MobileNetV3，训练集为ImageNet1k训练集，loss为cross entropy loss，迭代轮数为120epoch，精度指标为75.6%）。
 
 * 教师模型的选择。在进行知识蒸馏时，如果教师模型与学生模型的结构差异太大，蒸馏得到的结果反而不会有太大收益。相同结构下，精度更高的教师模型对结果也有很大影响。相比于79.12%的ResNet50_vd教师模型，使用82.4%的ResNet50_vd教师模型可以带来0.4%的绝对精度收益(`75.6%->76.0%`)。
@@ -103,6 +101,14 @@ SSLD的流程图如下图所示。
 | ResNet101_vd | 30 | 7e-5 | 1024/32 | 0.004 | cosine_decay_warmup | 83.73% |
 
 
+## 3.4 数据增广以及基于Fix策略的微调
+
+* 基于前文所述的实验结论，我们在训练的过程中加入自动增广(AutoAugment)[4]，同时进一步减小了l2_decay(4e-5->2e-5)，最终ResNet50_vd经过SSLD蒸馏策略，在ImageNet1k上的精度可以达到82.99%，相比之前不加数据增广的蒸馏策略再次增加了0.6%。
+
+
+* 对于图像分类任务，在测试的时候，测试尺度为训练尺度的1.15倍左右时，往往在不需要重新训练模型的情况下，模型的精度指标就可以进一步提升[5]，对于82.99%的ResNet50_vd在320x320的尺度下测试，精度可达83.7%，我们进一步使用Fix策略，即在320x320的尺度下进行训练，使用与预测时相同的数据预处理方法，同时固定除FC层以外的所有参数，最终在320x320的预测尺度下，精度可以达到**84.0%**。
+
+
 ## 3.4 实验过程中的一些问题
 
 ### 3.4.1 bn的计算方法
@@ -266,3 +272,7 @@ sh tools/run.sh
 [2] Bagherinezhad H, Horton M, Rastegari M, et al. Label refinery: Improving imagenet classification through label progression[J]. arXiv preprint arXiv:1805.02641, 2018.
 
 [3] Yalniz I Z, Jégou H, Chen K, et al. Billion-scale semi-supervised learning for image classification[J]. arXiv preprint arXiv:1905.00546, 2019.
+
+[4] Cubuk E D, Zoph B, Mane D, et al. Autoaugment: Learning augmentation strategies from data[C]//Proceedings of the IEEE conference on computer vision and pattern recognition. 2019: 113-123.
+
+[5] Touvron H, Vedaldi A, Douze M, et al. Fixing the train-test resolution discrepancy[C]//Advances in Neural Information Processing Systems. 2019: 8250-8260.