https://github.com/AllanYangZhou/metalearning-symmetries

Experiments for Meta-Learning Symmetries by Reparameterization
https://github.com/AllanYangZhou/metalearning-symmetries

Last synced: 4 months ago
JSON representation

Experiments for Meta-Learning Symmetries by Reparameterization

• Host: GitHub
• URL: https://github.com/AllanYangZhou/metalearning-symmetries
• Owner: AllanYangZhou
• Created: 2020-07-02T01:05:46.000Z (almost 4 years ago)
• Default Branch: master
• Last Pushed: 2021-04-11T20:13:51.000Z (about 3 years ago)
• Last Synced: 2024-01-16T10:46:18.766Z (5 months ago)
• Language: Python
• Size: 12.7 KB
• Stars: 51
• Watchers: 13
• Forks: 7
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Lists

• awesome-stars - AllanYangZhou/metalearning-symmetries - Experiments for Meta-Learning Symmetries by Reparameterization (Python)

README

        
# Meta-learning Symmetries by Reparameterization (MSR)

Code and weights corresponding to Meta-learning Symmetries by Reparameterization, found at this [arXiv link](https://arxiv.org/abs/2007.02933).

## Installation and requirements

All experiments were run in a conda environment with python 3.7.5, using pytorch. The conda environment we used is exported in `environment.yml`, though it likely contains more packages than are strictly necessary.

## Synthetic experiments

### Generate data

The file `generate_synthetic_data.py` contains the data generation code.

You must specify the name of the problem we wish to generate data for. Options are:

* `rank1`: 1-D data generated by a convolution layer

* `rank2`: 1-D data generated by a locally connected layer, with rank 2 weight factorization.

* `rank5`: 1-D data generated by a locally connected layer, with rank 5 weight factorization.

* `2d_rot8`: 2-D data that is equivariant to 45-degree rotations.

* `2d_rot8_flip`: 2-D data that is equivariant to 45-degree rotations and flips.

Example:

```sh

python generate_synthetic_data.py --problem rank1

```

### Train

The file `train_synthetic.py` contains training and evaluation code. Specifying the argument `model` controls whether to use MSR, plain MAML, or something else.

Some examples are given below:

```sh

python train_synthetic.py --problem rank1 --model share_fc  # MSR+FC on rank1 problem.

python train_synthetic.py --problem rank1 --model fc  # MAML+FC on rank1 problem.

python train_synthetic.py --problem rank1 --model conv  # MAML+Conv model on rank1 problem.

python train_synthetic.py --problem 2d_rot8 --model share_conv  # MSR+Conv on 2d_rot8 problem.

python train_synthetic.py --problem 2d_rot8 --model conv  # MAML+Conv on 2d_rot8 problem.

```

## Augmented-(Omniglot/Miniimagenet)

WIP