vislearn / GIN

Code for the paper "Disentanglement by Nonlinear ICA with General Incompressible-flow Networks (GIN)" (2020)

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

GIN

Code for the paper "Disentanglement by Nonlinear ICA with General Incompressible-flow Networks (GIN)" (2020)

Prerequisites

Make sure you have numpy and pytorch installed (recommended versions below). Then install FrEIA:

pip install git+https://github.com/VLL-HD/FrEIA.git

Recommended package versions

The scripts in this repository were tested with the following package versions (may also work with earlier versions, eg python 3.7):

  • python 3.8.3
  • numpy 1.18.1
  • matplotlib 3.1.3
  • pytorch 1.5.0
  • torchvision 0.6.0
  • cudatoolkit 10.2.89

Tests were made with both CPU (artificial data only) and GPU (artificial data and EMNIST).

Usage

Clone the repository:

git clone https://github.com/VLL-HD/GIN.git
cd GIN

Artificial Data

To see the available options:

python artificial_data.py -h

Reconstructions are saved in ./artificial_data_save/{timestamp}/figures. Eight reconstructions are plotted, each corresponding to a different orientation of the reconstructed latent space.

Example reconstruction plot:

artificial_data_reconstruction_plot

EMNIST

To see the available options:

python emnist.py -h

Model checkpoints (.pt files) are saved in ./emnist_save/{timestamp}/model_save with the specified save frequency. Figures are saved in ./emnist_save/{timestamp}/figures whenever a checkpoint is made (including at the end of training).

Example plots:

emnist_spectrum emnist_first_dim

For further details please refer to the paper.

About

Code for the paper "Disentanglement by Nonlinear ICA with General Incompressible-flow Networks (GIN)" (2020)

License:MIT License


Languages

Language:Python 100.0%