/PURER-Plus

PURER-Plus: An Extension of PURER (CVPR-2023)

Primary LanguagePython

PURER-Plus: An Extension of PURER (CVPR-2023)

PURER-Plus can:

  1. work well with ProtoNet (metric-based) and ANIL/MAML (optimization-based);
  2. work well with a larger number of pre-trained models with overlapping label spece.

Paper & Framework:

For our paper accepted by CVPR-2023, please refer to

[camera-ready], [arxiv], [poster], [slide], [video].

pipeline

Prerequisites:

  • python == 3.8.15
  • torch == 1.13.1

Datasets & Pre-trained Modes:

Datasets:

  • CIFAR-FS:

    • Please manually download the CIFAR-FS dataset from here to obtain "cifar100.zip".

    • Unzip ".zip". The directory structure is presented as follows:

      cifar100
├─mete_train
	├─apple (label_directory)
		└─ ***.png (image_file)
	...
├─mete_val
	├─ ...
		├─ ...
└─mete_test
	├─ ...
		├─ ...

    • Place it in "./DFL2Ldata/".

  • Mini-Imagenet: Please manually download it here. Unzip and then place it in "./DFL2Ldata/".

  • CUB: Please manually download it here. Unzip and then place it in "./DFL2Ldata/".

  • Flower: Please manually download it here. Unzip and then place it in "./DFL2Ldata/".

Pre-trained models:

  • You can pre-train the models following the instructions below (Step 3).

Quick Start:

  1. Make sure that the root directory is "./PURER-Plus".

  2. Prepare the dataset files.

    • For CIFAR-FS:

      python ./write_cifar100_filelist.py

      After running, you will obtain "meta_train.csv", "meta_val.csv", and "meta_test.csv" files under "./DFL2Ldata/cifar100/split/".

    • For MiniImageNet:

      python ./write_miniimagenet_filelist.py

      After running, you will obtain "meta_train.csv", "meta_val.csv", and "meta_test.csv" files under "./DFL2Ldata/Miniimagenet/split/".

    • For CUB:

      python ./write_CUB_filelist.py

      After running, you will obtain "meta_train.csv", "meta_val.csv", and "meta_test.csv" files under "./DFL2Ldata/CUB_200_2011/split/".

    • For Flower:

      python ./write_flower_filelist.py

      After running, you will obtain "meta_train.csv", "meta_val.csv", and "meta_test.csv" files under "./DFL2Ldata/flower/split/".

  3. Prepare the pre-trained models.

    bash ./scripts/pretrain.sh

    Some options you may change:

    Option Help
    --dataset cifar100/miniimagenet/cub/flower
    --pre_backbone conv4/resnet10/resnet18

  4. Meta training

     bash ./scripts/metatrain.sh

    Some options you may change:

    Option Help
    --dataset & --testdataset cifar100/miniimagenet/cub/flower
    --num_sup_train 1 for 1-shot, 5 for 5-shot
    --backbone conv4, the architecture of meta model
    --Glr learning rate of pueudo data generation
    --teacherMethod Proponent/anil/maml
    --APInum number of pre-trained models

Citation:

If you find this code is useful to your research, please consider to cite our paper.

@inproceedings{hu2023architecture,
  title={Architecture, Dataset and Model-Scale Agnostic Data-free Meta-Learning},
  author={Hu, Zixuan and Shen, Li and Wang, Zhenyi and Liu, Tongliang and Yuan, Chun and Tao, Dacheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={7736--7745},
  year={2023}
}

Acknowledgements:

Some codes are inspired from DeepInversion and CloserLookFewShot.

Contact: