maml-pytorch

This repository is an implementation of Model Agnostic Meta-Learning by Finn et al. in PyTorch.

Currently, the code supports first-order MAML (FOMAML) & the Omniglot dataset.

Requirements

• Python 3.5.2
• Pytorch 1.3.1
• Torchmeta 1.2.1

Acknowledgments

This blog post by Oscar Knagg helped me immensely.

Usage

Setup hyperparameters in config.yml.

dataset: omniglot
num_ways: 5
num_shots: 1
meta:
  training_iterations: 10000
  batch_size: 16
  lr: 1.0e-3
inner:
  batch_size: 1
  training_iterations: 1
  lr: 0.4
logs:
  val_interval: 100
  writer_interval: 100
device: 'cpu'               # or 'gpu'

Run main.py.

python main.py

[RECOMMENDED]

To check tensorboard, open up a command prompt and enter,

tensorboard --logdir=tensorboard-runs/

You can select the run ID from the list on the bottom left.