PaCL: Part-level Contrastive Learning for Fine-grained Few-shot Image Classification

Requirements

• PyTorch >= 1.8
• torchvision
• easydict
• numpy
• pyyaml
• prettytable
• tqdm

Folder Structure

pytorch-template/
│
├── train_exp.py - main script to start training
├── valid_exp.py - evaluation of trained model
│
├── config/ - configure files for the experiments
│   ├── conv-idt.yml - a basic config file
│
├── data/ - code for creating data loader
│   ├── source_data - scripts to create cached data
│   ├── collect_fn  - utils for a fast dataloader
│   ├── fs_dataset  - class file for few-shot dataset
│   ├── fs_sampler  - sampler for the dataloader to sample episodes
│   ├── transforms  - image transform functions
│
├── libs/ - some utils for the experiments
│   ├── checkpoint - save model and its weights
│   ├── count_params - calculate the number of parameters in the model
│   ├── init_exp  - create some basic info for the exp.
│   ├── lr_scheduler  - a learning rate scheduler with warm up
│
├── model/ - some utils for the experiments
│   ├── backbone - used backbone for the model
│   ├── modules - IDT, PaProCL, and some metric-based classifiers for PaCL
│   ├── pacl_net.py  - class file for the pacl-net
│   ├── pacl_model.py  - control the training process of pacl-net
│   ├── utils.py  - some utils for the model

Usage

1. create cached data

cd data/source_data/ucsd_cub_200
python generate_file.py generate_file --mode 'cl' --data_path 'path/to/cub-200-2011'

2. modify values in conv-idt.yml
3. run the Experiment

python train_exp.py --config ./config/conv-idt.yml

Citation

If you find this paper or our code useful in your research, please consider citing:

@inproceedings{DBLP:conf/mm/WangFM22,
  author    = {Chuanming Wang and
               Huiyuan Fu and
               Huadong Ma},
  editor    = {Jo{\~{a}}o Magalh{\~{a}}es and
               Alberto Del Bimbo and
               Shin'ichi Satoh and
               Nicu Sebe and
               Xavier Alameda{-}Pineda and
               Qin Jin and
               Vincent Oria and
               Laura Toni},
  title     = {PaCL: Part-level Contrastive Learning for Fine-grained Few-shot Image
               Classification},
  booktitle = {{MM} '22: The 30th {ACM} International Conference on Multimedia, Lisboa,
               Portugal, October 10 - 14, 2022},
  pages     = {6416--6424},
  publisher = {{ACM}},
  year      = {2022},
  doi       = {10.1145/3503161.3547997},
}