Paper link: https://arxiv.org/abs/2206.08126
Few-Shot Learning (FSL) requires vision models to quickly adapt to brand-new classification tasks with a shift in task distribution. Understanding the difficulties posed by this task distribution shift is central to FSL. In this paper, we show that a simple channel-wise feature transformation may be the key to unraveling this secret from a channel perspective. When facing novel few-shot tasks in the test-time datasets, this transformation can greatly improve the generalization ability of learned image representations, while being agnostic to the choice of training algorithms and datasets. Through an in-depth analysis of this transformation, we find that the difficulty of representation transfer in FSL stems from the severe channel bias problem of image representations: channels may have different importance in different tasks, while convolutional neural networks are likely to be insensitive, or respond incorrectly to such a shift. This points out a core problem of the generalization ability of modern vision systems and needs further attention in the future.
We have added the extended version of oracle transformation in each classifier. Compared to the version in our ICML paper, the extended version additionally supports features taking negative values. So any pretrained backbone can be tested, including ResNets using LeakyReLU or variants of ViT like SwinTransformer. We will include the extended version and far more analysis about feature importance of visual representions in an extended paper. If you have any problem, please open an issue or just contact the author Xu Luo through the email.
The single-line simple_transform function has been put inside each test-time classifier, e.g., see in architectures/classifier/proto_head.py.
A list of packages to install is available in requirements.txt, just install them using command:
pip install -r requirements.txt
All 11 pre-trained models in Table 1 of our paper can be found here. We use LightningFSL to train all backbones. For training the models, please refer to it.
After obtaining pre-trained models and downloading needed datasets (see below), one can evaluate the simple channel-wise feature transformation. Exemplar commands for running the code can be found in eval.sh.
Download any of the 19 test-time datasets in Table 1 of our paper for evaluation.
- miniImageNet: Download and extract miniImageNet.zip, and use the resultant folder as the root directory.
- CUB: Download and extract cub.tar, and use the
images/folder as the root directory. - Textures: Download and extract dtd-r1.0.1.tar.gz, and use the
images/folder as the root directory. - Traffic Signs: Download and extract GTSRB_Final_Training_Images.zip, and use the
Final_Training/Images/folder as the root directory. - Aircraft: Download and extract fgvc-aircraft-2013b.tar.gz, and use the
fgvc-aircraft-2013b/datafolder as the root directory. - Omniglot: Download images_background.zip and images_evaluation.zip, then extract them into the same directory as the root directory.
- VGG Flower: Download 102flowers.tgz and imagelabels.mat, then extract 102flowers.tgz, and put them into the same directory as the root directory.
- MSCOCO: Download val2017.zip and annotations_trainval2017.zip, then extract them into the same directory as the root directory.
- Quick Draw: Download and extract quickdraw.zip, and use the resultant folder as the root directory.
- Fungi: Download and extract fungi_train_val.tgz, and use the
images/folder as the root directory. - Plant Disease: Download from kaggle, and use the train split as the root directory.
- ISIC: Download ISIC2018_Task3_Training_Input.zip and ISIC2018_Task3_Training_GroundTruth.zip, then extract them into the same directory as the root directory.
- EuroSAT: Dowaload and extract EuroSAT.zip, and use the resultant folder as the root directory.
- ChestX: Download from kaggle, and use the resultant folder as the root directory.
- Real: Download and extract real.zip, and use the resultant folder as the root directory.
- Sketch: Download and extract sketch.zip, and use the resultant folder as the root directory.
- Infograph: Download and extract infograph.zip, and use the resultant folder as the root directory.
- Painting: Download and extract painting.zip, and use the resultant folder as the root directory.
- Clipart: Download and extract clipart.zip, and use the resultant folder as the root directory.
If you find our work useful in your research please consider citing:
@inproceedings{
Luo2022Channel,
title={Channel Importance Matters in Few-Shot Image Classification},
author={Luo, Xu and Xu, Jing and Xu, Zenglin},
booktitle={International Conference on Machine Learning},
year={2022},
}