Ensemble Model with Batch Spectral Regularization and Data Blending for Cross-Domain Few-Shot Learning with Unlabeled Data

Code release for Track 2 in Cross-Domain Few-Shot Learning (CD-FSL) Challenge .

Enviroment

Python 2.7.16

Pytorch 1.0.0

Steps

1. Prepare the source dataset miniImageNet and four target datasets CropDisease, EuroSAT, ISIC and ChestX.

2. Modify the paths of the datasets in configs.py according to the real paths.

3. Train base models on miniImageNet (pre-train)

   • Train single model

   python ./train_bsr.py --model ResNet10 --train_aug

   • Generate projection matrices and train ensemble model

   python ./create_Pmatrix.py
   python ./train_Pbsr.py --model ResNet10 --train_aug

4. Fine-tune and test for the 5-shot task in CropDisease as an example (change n_shot parameter to 20, 50 for 20-shot and 50-shot evaluations and change dtarget parameter to EuroSAT, ISIC, ChestX for the other target domains)

   • Test the BSDB and BSDB+LP methods for single model

    python ./finetune_lp_bsdb.py --model ResNet10 --train_aug --use_saved --dtarget CropDisease --n_shot 5

   The use_saved flag is used to test with our saved models. You can close it to test with the reproduced models.

   Example output:

    BSDB: 600 Test Acc = 93.48% +- 0.42%
    BSDB+LP: 600 Test Acc = 95.31% +- 0.37%
   

   • Test the BSDB and BSDB+LP methods for ensemble model

    python ./finetune_P_lp_bsdb.py --model ResNet10 --train_aug --use_saved --dtarget CropDisease --n_shot 5

   Example output:

    BSDB (Ensemble): 600 Test Acc = 94.05% +- 0.41%
    BSDB+LP (Ensemble): 600 Test Acc = 95.93% +- 0.37%