dchiji-ntt/meta-ticket

Meta-ticket: Finding optimal subnetworks for few-shot learning within randomly initialized neural networks

Official repository for Meta-ticket paper. (ArXiv, OpenReview)

Presented at NeurIPS'22 (Poster): https://neurips.cc/virtual/2022/poster/53373

Requirements

• Python 3.8
• NVIDIA CUDA 11.3
• torch==1.9.0
• learn2learn==0.1.6
• torchmeta==1.8.0
• pyyaml==5.3.1

Usage

python3.8 main.py <command> <exp_name>

• <command> is one of meta_train, meta_test, meta_cross_test and parallel. The meta_train, meta_test and meta_cross_test commands can be used to meta-train/meta-test/cross-test a single model, and parallel can be used to reproduce our experiments with multiple seeds or to search hyperparameters.
• <exp_name> is one of the keys defined in the config file config.yaml.

Datasets

We can download datasets for meta-learning by the scripts in scripts directory.

Example:

python3.8 scripts/download_miniimagenet.py

Reproduce Experimental Results

For each experiment in our paper, we have the corresponding setting in config.yaml. For example, we can run the meta-training for MetaTicket with ResNet12 on miniImageNet (5-shot, 5-way) by parallel command:

python main.py parallel config.yaml miniimagenet_5s5w_resnet12_ticket

In the end of experiments, we can check the final results by:

python utils/test_info.py __outputs__/miniimagenet_5s5w_resnet12_ticket/