MAML implementation in PyTorch

Introduction

This is a simple MAML implementation based on original paper and previous open-source implementations (listed in the references). This can be used as a reference for MAML beginners, which is well annotated and easy to fine-tune.

Running Results

You can take a look at some running results of this implementation at the following pages.

• Omniglot 5way-5shot
• Omniglot 5way-1shot

Usage

To run this code, you can follow the instructions below.

1. Clone this repo.

git clone https://github.com/Yifei20/MAML-PyTorch.git
cd MAML-PyTorch

2. Download the Omniglot dataset.

mkdir data
cd data
wget https://github.com/Yifei20/MAML-PyTorch/releases/download/dataset/Omniglot.zip
unzip Omniglot.zip

3. Run the code directly in the notebook, you can adjust the hyperparameter if you want.

References

Paper

• Original Paper: Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

Github Repo

• cbfnn/maml: Code of the Original Paper
• Runist/torch_maml

Course Materials

• NTU ML (Hung-Yi Lee), Spring 2023, HW15

• Standford CS 330: Deep Multi-Task and Meta Learning, Fall 2022, HW1