Awesome

Worst Case Matters for Few-Shot Recognition

This repository is the official implementation of Worst Case Matters for Few-Shot Recognition, which is accepted by ECCV2022.

This paper boost the worst-case performance in few-shot learning by reducing the standard deviation and increasing the average accuracy simultaneously from the perspective of bias-variance trade-off. To achieving so, a simple yet effective stability regularization (SR) loss together with model ensemble to reduce variance during fine-tuning, and an adaptability calibration (AC) mechanism to reduce the bias are proposed.

Environment

• Python 3.8

• Pytorch 1.10.0

• Pandas

Running the code

Datasets

Download datasets from this link, and organize these datasets as follows

YOUR_DATA_PATH
	miniImagenet/
		base/
		val/
		novel/
	CUB/
		base/
		val/
		novel/
	CIFAR-FS/
		base/
		val/
		novel/

After that, modify the parameter 'DATA_PATH' in config.py with the path of YOUR_DATA_PATH.

Checkpoints

Download file checkpoint.tar.gz (link), unpack it into a directory named checkpoint and place it in the root path of the project. The checkpoint contains everything else you need to run the code, including pre-trained models, pre-sampled tasks/episodes and features of base set extracted by pre-trained models.

We highly encourage researchers to conduct experiments on our pre-sampled episodes to further explore the way to boost few-shot learning, especially for worst-case performance.

Train

For AC+SR

python meta_test_ACSR.py --dataset [DATASET] --k [SHOT] 

For AC+EnSR, thanks to the parallel computing capabilities of pytorch, please run the code as

CUDA_VISIBLE_DEVICES=[CUDA_INDICES, e.g. 0,1,2,3] python -m torch.distributed.launch --nproc_per_node [NUM_OF_NODES, e.g. 4] --master_port [PORT]   meta_test_EnACSR.py --dataset [DATASET] --k [SHOT] 

which naturally allows the ensemble methods to run in parallel.

Important Arguments Some important arguments for our code.

• dataset: choices=['miniImagenet', 'CUB', 'CIFAR-FS']
• arch: choices=['wrn_28_10']
• lr: learning rate of fine-tuning on each task.
• tune_mode : adaptation calibration (AC) strategy used to the model. By default, the last convolutional block and final classification layer is set to be learnable.
• beta: trade-off scalar for classification loss and stability regularization (SR) loss.
• aux_bs: batch size to compute SR loss in one step.
• epochs_finetune: epochs for fine-tuning each few-shot episode.
• n : how many ways per few-shot episode.
• k : how many shots per few-shot episode.
• q : how many queries per few-shot episode.
• tasks : the number of episodes for evaluation..

Citation

If you find our code useful, please consider citing our work using the bibtex:

@article{fu2022ACSR,
  title={Worst Case Matters for Few-Shot Recognition},
  author={Fu, Minghao and Cao, Yun-Hao and Wu, Jianxin},
  journal={The European Conference on Computer Vision},
  year={2022}
}