crashmoon / MAML-Pytorch-Multi-GPUs

pytorch maml with Multi-GPUs, fast and simplest implementation

MAML-Pytorch-Multi-GPUs

It is a reproduced version of maml, which is implemented with PyTorch 1.6.0 and support Multi-GPUs both in Meta-training phase and Meta-testing phase.

All the hyper-parameters and tricks, e.g. gradient clip, are strictly consistent with the original paper Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks HERE and its Tensorflow Implementation HERE.

Platform

• python: 3.7
• Pytorch: 1.6.0

Howto

1. Downloading MiniImagenet dataset
2. Changing dataset_dir in run_meta.py with your own root of the MiniImagenet dataset
3. python run_meta.py

Comparison to original MAML implementation for miniImageNet

	5-way 1-shot	5-way 5-shot
MAML	48.7%	63.1%
Ours	48.9%	#TODO#

reference

https://github.com/cbfinn/maml

https://github.com/dragen1860/MAML-Pytorch

https://github.com/jik0730/MAML-in-pytorch

https://arxiv.org/abs/1703.03400