Official Pytorch Implementation for "Learning to Learn from APIs: Black-Box Data-Free Meta-Learning" (ICML-2023)

Paper & Framework:

For our paper accepted by ICML-2023, please refer to [camera-ready] (on-going) or arxiv with the supplementary materials.

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/".

Pre-trained models:

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

Training:

1. Make sure that the root directory is "./BiDf-MKD".

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/".

3. Prepare the pre-trained models.

   bash ./scriptskd/pretrain.sh

   Some options you may change:

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

4. Meta training

   • For API-SS scenario:

      bash ./scriptskd/ss.sh

   • For API-SH scenario:

      bash ./scriptskd/sh.sh

   • For API-MH scenario:

      bash ./scriptskd/mh.sh

   Some options you may change:

   Option	Help
   --dataset	cifar100/miniimagenet/cub for API-SS and API-SH, mix for API-MH
   --num_sup_train	1 for 1-shot, 5 for 5-shot
   --backbone	conv4, the architecture of meta model
   --pre_backbone	conv4/resnet10/resnet18 for SS, mix for SH and MH
   --q	inference times of zero-order gradient estimation
   --numsplit	for parallel inference, reduce it for low memory cost

Citation:

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

@inproceedings{hu2023learning,
  title={Learning to Learn from APIs: Black-Box Data-Free Meta-Learning},
  author={Zixuan Hu, Li Shen, Zhenyi Wang, Baoyuan Wu, Chun Yuan, Dacheng Tao},
  booktitle={International Conference on Machine Learning},
  year={2023}
}

Acknowledgements:

Some codes are inspired from CMI.

Contact:

• Zixuan Hu: huzixuan21@mails.tsinghua.edu.cn