Awesome

When Prompt-based Incremental Learning Does Not Meet Strong Pretraining (ICCV2023)

Official PyTorch implementation of our ICCV2023 paper “When Prompt-based Incremental Learning Does Not Meet Strong Pretraining”[paper]

Environments

• Python: 3.8
• PyTorch: 1.10
• See requirements.txt for more details

Data preparation

Download Datasets

• CIFAR100 dataset will be downloaded automatically to the directory specified by --data-path.

• ImageNet100 dataset cannot be downloaded automatically, please download it from image-net.org. Place the dataset the directory specified by --data-path.

  In order to conduct incremental training, we also need to put imagenet split file train_100.txt, val_100.txt into the data path. Symbolic link is recommended:

  ln -s ImageNet/imagenet_split/train_100.txt data_path/imagenet/train_100.txt
  ln -s ImageNet/imagenet_split/val_100.txt data_path/imagenet/val_100.txt
  

• ImageNet-R is fromThe Many Faces of Robustness. Since the official dataset didn't provide the train/test split, you can use our split in the datalink. Place the dataset the directory specified by --data-path.

• EuroSAT is from the official github repo, you can use the version in the datalink for convenience. Place the dataset the directory specified by --data-path.

• RESISC45 is from the paper (url), you can use the version in the datalink. Place the dataset the directory specified by --data-path.

Path structure

After downloading, the dataset should be organized like this

datasets
│  
│──imagenet
│   │
│   └───train
│       │   n01440764
│       │   n01443537 
│       │   ...
│   │
│   └───val
│       │   n01440764
│       │   n01443537
│       │   ...
│   │   
│   │ train_100.txt
│   │ train_900.txt
│   │ val_100.txt 
│
│──cifar-100-python
│   │ ...
│ 
│──my-imagenet-r 
│   │ ...
│ 
│──my-EuroSAT_RGB
│   │ ...
│ 
│──my-NWPU-RESISC45
│   │ ...
└

Training

Non-pretrained incremental learning

ImageNet-100

This setting is the default setting of class-incremental learning. In this setting, the network is trained from scratch.

• Step 1: Download weights from the following link and put them in 'chkpts' folder.

• Step 2: To reproduce the results of our method in Table 1 & 2:

  bash runs/non_pretrained/run_nonPretrained_imageNetSub_B50_T10.sh 0,1 10241
  

Notes for step1:

Since this setting is not using any pretrained weights, we treat the first task as the pretraining tasks. For ViT we use, it is difficult to train the model from scratch to match the first-task performance as the CNN (ResNets) . So we use an ResNet teacher to assist the first-task training. The resnet teacher can be trained under repos like PODNet or our previous work Imagine)

For simplicity, we provide the trained ViT weights on the first task which can be found in the link above. Running the script above will load such weights. If you want to train the ViT on the first task, you can run

bash runs/non_pretrained/imagnet_pretrain/run_nonPretrained_imageNetSub_B50_teacher.sh 0,1,2,3 10241

After training, please put the checkpoint on the chkpts folder and modify the path in the script above (line28).

CIFAR100

• Step 1: Download weights from the following link and put them in 'chkpts' folder.

• Step 2: To reproduce the results of our method in Table 3:

  bash runs/non_pretrained/run_nonPretrained_cifar100_B50_T10.sh 0,1 10241
  

Notes for step1:

Like the Imagenet, here is the script for training the first stage.

bash runs/non_pretrained/cifar_pretrain/run_nonPretrained_CIFAR100_B50_teacher.sh 0,1,2,3 10241

Pretrained incremental learning

In order to compare our method with other prompt-based methods, we also conduct experiments with pretrained weights (ImageNet21k & TinyImageNet pretrained).

Download the pretrained checkpoints

Download weights from the following link and put them in 'chkpts' folder. We use the pretrained-weight(deit_base_patch16_224-b5f2ef4d.pth) from the Deit repo. We use the tiny-ImageNet weight from this repo.

ImageNet-R

• To reproduce the results of our method in Table 5:

  (ImageNet-pretrained)

  bash runs/pretrained/ImageNetR_pretrained_vitbase_T10.sh 0,1 10241
  

• To reproduce the results of our method with tiny-imagenet pretrained in Table A4 (in the Appendix):

  (tinyImageNet-pretrained)

  bash runs/pretrained/ImageNetR_tinyImageNet_pretrained_vitbase_T10.sh 0,1 10241
  

CIFAR100

• To reproduce the results of our method in Table 4:

  (ImageNet-pretrained)

  bash runs/pretrained/CIFAR100_pretrained_vitbase_T10.sh 0,1 10241
  

• To reproduce the results of our method with tiny-imagenet pretrained in Table 4 (in the Appendix):

  (tinyImageNet-pretrained)

  bash runs/pretrained/CIFAR100_tinyImageNet_pretrained_vitbase_T10.sh 0,1 10241
  

EuroSAT and RESISC45

• To reproduce the results of our method with ImageNet pretrained in Table A2 (in the Appendix):

  (EuroSAT_RGB)

  bash runs/pretrained/EuroSAT_RGB_pretrained_T5.sh 0,1 10241
  

  (NWPU-RESISC45)

  bash runs/pretrained/NWPU-RESISC45_pretrained_T10.sh 0,1 10241
  

• Further, with tiny-imagenet pretrained:

  (EuroSAT_RGB)

  bash runs/pretrained/EuroSAT_RGB_tinyImageNet_pretrained_T5.sh 0,1 10241
  

  (NWPU-RESISC45)

  bash runs/pretrained/NWPU-RESISC45_tinyImageNet_pretrained_T10.sh 0,1 10241
  

Todo list

-[x] Non-pretrained incremental learning on CIFAR100

-[x] Pretrained incremental learning on ImageNetR

-[x] Pretrained incremental learning on EuroSAT and RESISC45

Acknowledgement

• This repository is heavily based on incremental_learning.pytorch by arthurdouillard.

• If you use this paper/code in your research, please consider citing us:

@inproceedings{tang2022learning,
  title={When Prompt-based Incremental Learning Does Not Meet Strong Pretraining},
  author={Tang, Yu-Ming and Peng, Yi-Xing and Zheng, Wei-Shi},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision (ICCV)},
  year={2023}
}