CAVIA

Code for "Fast Context Adaptation via Meta-Learning" - Luisa M Zintgraf, Kyriacos Shiarlis, Vitaly Kurin, Katja Hofmann, Shimon Whiteson (ICML 2019).

I used Python 3.7 and PyTorch 1.0.1 for these experiments.

Regression

• Running experiments:

  To run the experiment with default settings, execute

  python3 regression/main.py
  

  This will run the sine curve experiment. To run the CelebA image completion experiment, run

  python3 regression/main --task celeba --num_context_params 128 --num_hidden_layers 128 128 128 128 128 --k_meta_test 1024
  

  To change the number of context parameters, use the flag --num_context_params.

  To run MAML with the default settings, run python3 regression/main.py --maml --num_context_params 0 --lr_inner 0.1

  For default settings and other argument options, see classification/arguments.py

• CelebA dataset:

  If you want to use the code for the CelebA dataset, you have to download it (http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) and change the path in tasks_celeba.py.

Classification

• Running experiments:

  To run the experiment with default settings, execute in your command line:

  python3 classification/cavia.py
  

  Use the --num_filters flag to set the number of filters. For default settings and other argument options, see arguments.py.

• Retrieving Mini-Imagenet:

  You need the Mini-Imagenet dataset to run these experiments. See e.g. https://github.com/y2l/mini-imagenet-tools for how to retrieve it. Put them in the folder classification/data/miniimagenet/images/ (the label files are already in there).

Reinforcement Learning

This code is an extended version of Tristan Deleu's PyTorch MAML implementation: https://github.com/tristandeleu/pytorch-maml-rl.

• Prerequisites:

  For the MuJoCo experiments you need mujoco-py and OpenAI gym.

• Running experiments:

  To run an experiment on the 2D navigation, use the following command:

  python3 main.py --env-name 2DNavigation-v0 --fast-lr 1.0 --phi-size 5 0  --output-folder results
  

Acknowledgements

Special thanks to Chelsea Finn, Jackie Loong and Tristan Deleu for their open-sourced MAML implementations. This was of great help to us, and parts of our implementation are based on the PyTorch code from:

• Jackie Loong's implementation of MAML, https://github.com/dragen1860/MAML-Pytorch
• Tristan Deleu's implementation of MAML-RL, https://github.com/tristandeleu/pytorch-maml-rl

BibTex

@article{zintgraf2018cavia,
  title={Fast Context Adaptation via Meta-Learning},
  author={Zintgraf, Luisa M and Shiarlis, Kyriacos and Kurin, Vitaly and Hofmann, Katja and Whiteson, Shimon},
  conference={Thirty-sixth International Conference on Machine Learning (ICML 2019)},
  year={2019}
}