When Does Self-supervision Improve Few-shot Learning?

This repo contains the source code of the paper:

"When Does Self-supervision Improve Few-shot Learning?", Jong-Chyi Su, Subhransu Maji, Bharath Hariharan, ECCV, 2020.

Project page, arXiv, slides

The code is based on the repo of "A Closer Look at Few-shot Classification", ICLR, 2019.

Citation

@inproceedings{Su2020When,
	title = {When Does Self-supervision Improve Few-shot Learning?},
	author = {Jong-Chyi Su and Subhransu Maji and Bharath Hariharan},
	year = {2020},
	booktitle = {ECCV}
}

Enviroment

• Python3
• PyTorch (tested on > 1.0.0)

Getting started

Prepare datasets

• Please download images of Caltech-UCSD birds, Stanford cars, fgvc-aircraft, Stanford dogs, and Oxford flowers, and put them under filelists/${dset}/images.

Download mini- and tiered-ImageNet

• Change directory to filelists/miniImagenet
• run source download_miniImagenet.sh

(WARNING: This would download the 155G ImageNet dataset. You can comment out correponded line 5-6 in download_miniImagenet.sh if you already have one.)

Base/Val/Novel splits

• Require three data split json file: 'base.json', 'val.json', 'novel.json' for each dataset.
• Splits are included in this repo.

Training (including test after training is done)

For baseline, run python train.py --dataset CUB --train_aug

For jigsaw, run python train.py --dataset CUB --train_aug --jigsaw

For rotation, run python train.py --dataset CUB --train_aug --rotation

To use a separate unlabeled dataset for SSL, first make a json file for the unlabeled dataset (see original repo for details). Next, set --dataset_unlabel and --base_unlabel to the name of the json file. For example, to use 20% CUB dataset for supervised training (CUB/base_20.json) and 100% CUB for SSL (jigsaw) (CUB/base.json), run

python train_separate.py --dataset CUB --dataset_unlabel CUB --base base_20 --base_unlabel base --jigsaw --lbda 0.5 --lr 0.001 --train_aug --n_query 5 --stop_epoch 600 --bs 64

Note: For Mini-ImageNet and Tiered-Imagenet, please train the model for 600 epochs. For other datasets, 400 epochs are enough.

Author

Jong-Chyi Su (UMass Amherst) jcsu@cs.umass.edu