Online Label Smoothing

The code for the paper "Delving Deep into Label Smoothing"
I have only cleaned the code on the fine-grained datasets. Since I am not currently in school, I have not tested it. So if there are any bugs, please feel easy to contact me (zhangchbin AATT gmail Ddot com).

Citation

@ARTICLE{zhang2021delving,
  author={Zhang, Chang-Bin and Jiang, Peng-Tao and Hou, Qibin and Wei, Yunchao and Han, Qi and Li, Zhen and Cheng, Ming-Ming},
  journal={IEEE Transactions on Image Processing}, 
  title={Delving Deep into Label Smoothing}, 
  year={2021},
  volume={30},
  number={},
  pages={5984-5996},
  doi={10.1109/TIP.2021.3089942}}

Performance

Classification for fine-grained datasets

Model ensemble on CIFAR-100

TODO

• supporting for Jittor
• training scripts for more network and datasets
• ~~EfficientNet for classification~~
• ~~SAN network for classification~~
• training code on the ImageNet
• training code on the Noisy-CIFAR
• adversarial attack code on the ImageNet and CIFAR
• ~~training code on the CIFAR~~
• ~~the download link of CUB-200-2011, Flowers, Cars and dogs~~

Requirements

pytorch >= 1.0
torchvision
numpy
tensorboardX
apex
tqdm
efficientnet_pytorch
SAN_network

efficientnet_pytorch
SAN network

Data Preparation

Download all datasets to the data directory, note that we modify the division for datasets as shown in files in the data directory:

• You can download the images from Cars. And the lists for training and validation are in the data directory.
• You can download the images from Aircrafts. And the lists are in the data directory.
• You can download the images from Flowers with 102 classes. And the lists are in the data directory.
• You can download the images from Standford-dogs. And the lists are in the data directory.

Train and Validate

1. download the ImageNet pretrained model to checkpoint
   MobileNet-v2, ResNet-50, Res2Net

2. train the model with online label smoothing:

   CUDA_VISIBLE_DEVICES=1 python main.py \
   --mode train \
   --pretrained_model ./checkpoint/mobilenet_v2-b0353104.pth \
   --epochs 100 \
   --lr 0.01 \
   --arch mobilenetv2 \
   --dataset cub \
   --method ols \
   --batch_size 64 \
   

   (optional) test the model ensemble performance:

   python main.py \
   --mode ensemble \
   --ensemble 'runs/mobilenetv2_cub_ols/20.pth' 'runs/mobilenetv2_cub_ols/60.pth' \
   --epochs 100 \
   --lr 0.01 \
   --arch mobilenetv2 \
   --dataset aircraft \
   --method ols \
   --batch_size 64 \
   

Train on CIFAR

```
cd cifar
sh train_cifar_imagenetresnet34.sh
sh train_cifar_resnext29_2.sh
````

Other realated implementation

Thanks to the re-implementation in Kurumi233 and ankandrew