Model-Agnostic Meta-Learning (MAML)

Paper

Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks [1]

Framework(s)
../_images/pytorch.png

PyTorch

API Reference

garage.torch.algos.MAML

Code

garage/torch/algos/maml.py

Examples

maml_ppo_half_cheetah_dir, maml_trpo_half_cheetah_dir, maml_trpo_metaworld_ml1_push, maml_trpo_metaworld_ml10. maml_trpo_metaworld_ml45

MAML is a meta-learning algorithm that trains the parameters of a policy such that they generalize well to unseen tasks. In essence, this technique produces models that are good few shot learners and easy to fine-tune.

Default Parameters

meta_batch_size=40,
inner_lr=0.1,
outer_lr=1e-3,
num_grad_updates=1,
meta_evaluator=None,
evaluate_every_n_epochs=1

Examples

maml_ppo_half_cheetah_dir

../_images/pytorch.png

maml_trpo_half_cheetah_dir

../_images/pytorch.png

maml_trpo_metaworld_ml1_push

../_images/pytorch.png

maml_trpo_metaworld_ml10

../_images/pytorch.png

maml_trpo_metaworld_ml45

../_images/pytorch.png

References

1

Chelsea Finn, Pieter Abbeel, and Sergey Levine. Model-agnostic meta-learning for fast adaptation of deep networks. 2017. arXiv:1703.03400.

This page was authored by Mishari Aliesa (@maliesa96).