hongbo-sun / AttriCLIP

CVPR2023: AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning

AttriCLIP

This is the pytorch version code of "AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning" in CVPR2023.

Content

• AttriCLIP
  • Content
  • Introduce of AttriCLIP
  • Datasets
  • Pretrained CLIP
  • Insturction

Introduce of AttriCLIP

AttriCLIP is introduced from《AttriCLIP: A Non-Incremental Learner for Incremental Knowledge Learning》

Paper ：Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lv, Baochang Zhang. "Few-Shot Learning with Visual Distribution Calibration and Cross-Modal Distribution Alignment". In CVPR, 2023.

Datasets

The test Datasets：CIFAR100, Download link
Datasets size：100 classes and 32*32 pixels for each image.

The test Datasets：a subset of ImageNet, Download link
Datasets size：100 classes. The chosen classes are shown in supplementary materials of the paper and dataset/imagenet100.py of this project. The dataset can be generated by running generate_ima100/generate_ima100.py.

Pretrained CLIP

We use the pretrained CLIP model from here

Insturction

pip install -r requirements.txt
python main_incremental_submit.py --root data_path #for cifar100
#or
python main_incremental_ima.py --root data_path #for imagenet100