Awesome

Few-Shot Meta-Baseline

This repository contains the code for Meta-Baseline: Exploring Simple Meta-Learning for Few-Shot Learning.

Citation

@inproceedings{chen2021meta,
  title={Meta-Baseline: Exploring Simple Meta-Learning for Few-Shot Learning},
  author={Chen, Yinbo and Liu, Zhuang and Xu, Huijuan and Darrell, Trevor and Wang, Xiaolong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={9062--9071},
  year={2021}
}

Main Results

The models on miniImageNet and tieredImageNet use ResNet-12 as backbone, the channels in each block are 64-128-256-512, the backbone does NOT introduce any additional trick (e.g. DropBlock or wider channel in some recent work).

5-way accuracy (%) on miniImageNet

5-way accuracy (%) on tieredImageNet

5-way accuracy (%) on ImageNet-800

Experiments on Meta-Dataset are in meta-dataset folder.

Running the code

Preliminaries

Environment

• Python 3.7.3
• Pytorch 1.2.0
• tensorboardX

Datasets

• miniImageNet (courtesy of Spyros Gidaris)
• tieredImageNet (courtesy of Kwonjoon Lee)
• ImageNet-800

Download the datasets and link the folders into materials/ with names mini-imagenet, tiered-imagenet and imagenet. Note imagenet refers to ILSVRC-2012 1K dataset with two directories train and val with class folders.

When running python programs, use --gpu to specify the GPUs for running the code (e.g. --gpu 0,1). For Classifier-Baseline, we train with 4 GPUs on miniImageNet and tieredImageNet and with 8 GPUs on ImageNet-800. Meta-Baseline uses half of the GPUs correspondingly.

In following we take miniImageNet as an example. For other datasets, replace mini with tiered or im800. By default it is 1-shot, modify shot in config file for other shots. Models are saved in save/.

1. Training Classifier-Baseline

python train_classifier.py --config configs/train_classifier_mini.yaml

(The pretrained Classifier-Baselines can be downloaded here)

2. Training Meta-Baseline

python train_meta.py --config configs/train_meta_mini.yaml

3. Test

To test the performance, modify configs/test_few_shot.yaml by setting load_encoder to the saving file of Classifier-Baseline, or setting load to the saving file of Meta-Baseline.

E.g., load: ./save/meta_mini-imagenet-1shot_meta-baseline-resnet12/max-va.pth

Then run

python test_few_shot.py --shot 1

Advanced instructions

Configs

A dataset/model is constructed by its name and args in a config file.

For a dataset, if root_path is not specified, it is materials/{DATASET_NAME} by default.

For a model, to load it from a specific saving file, change load_encoder or load to the corresponding path. load_encoder refers to only loading its .encoder part.

In configs for train_classifier.py, fs_dataset refers to the dataset for evaluating few-shot performance.

In configs for train_meta.py, both tval_dataset and val_dataset are validation datasets, while max-va.pth refers to the one with best performance in val_dataset.

Single-class AUC

To evaluate the single-class AUC, add --sauc when running test_few_shot.py.