Meta-Learning Models for Few-shot Classification

This repository contains several meta-learning models used for few-shot tasks. Also includes pretrain process i.e. training in a classical way and test on 5-way 5-shot task directly.

Combine the pretraining and meta-training would make the performance better.

Running

Environment

• python 3.9.3
• pytorch 1.14.0
• wandb

Datasets

• miniImageNet (courtesy of Spyros Gidaris)
• tieredImageNet (courtesy of Kwonjoon Lee)

Training

Pretrain

python pretrain.py --config config/pretrain_mini.yaml

Meta Training

python train.py --config/train_proto_mini.yaml

Testing

The trained model would be saved in logs/.../... .pth Modify the load path in the test.yaml to load the model

python test.py --config/test.yaml