This repository is the official implementation for Class-Specific Channel Attention for Few-Shot Learning.
Pytorch 1.8.0 is used for the experiments in the paper.
All pretrained weights and extracted features for 5-way 5-shot expriments in the paper can be downloaded from the PT-MAP repository.
Create directories "./pretrained_models_features/[miniImagenet/Tiered_ImageNet/CIFAR_FS/CUB]", and place the plk file in the corresponding directory.
5-way 5-shot
python main.py --dataset [miniImagenet/Tiered_ImageNet/CIFAR_FS/CUB] --meta_train_epoch [10/15/20/25]
5-way 1-shot
Work in progress...
| Dataset | 5-Way 1-Shot | 5-Way 5-Shot |
|---|---|---|
| miniImageNet | 96.68% | 99.96% |
| Tiered-ImageNet | 96.58% | 99.37% |
| CIFAR-FS | 98.85% | 99.82% |
| CUB | 97.43% | 99.09% |
Channel Importance Matters in Few-Shot Image Classification
Charting the Right Manifold: Manifold Mixup for Few-shot Learning
Manifold Mixup: Better Representations by Interpolating Hidden States
Leveraging the Feature Distribution in Transfer-based Few-Shot Learning