https://github.com/rosinality/glow-pytorch

PyTorch implementation of Glow
https://github.com/rosinality/glow-pytorch

Last synced: 3 months ago
JSON representation

PyTorch implementation of Glow

• Host: GitHub
• URL: https://github.com/rosinality/glow-pytorch
• Owner: rosinality
• License: mit
• Created: 2018-07-12T06:34:34.000Z (over 6 years ago)
• Default Branch: master
• Last Pushed: 2021-11-20T23:21:48.000Z (almost 3 years ago)
• Last Synced: 2024-04-20T12:39:17.530Z (7 months ago)
• Language: Python
• Size: 11.7 MB
• Stars: 494
• Watchers: 9
• Forks: 97
• Open Issues: 30
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# glow-pytorch

PyTorch implementation of Glow, Generative Flow with Invertible 1x1 Convolutions (https://arxiv.org/abs/1807.03039)

Usage:

> python train.py PATH

as trainer uses ImageFolder of torchvision, input directory should be structured like this even when there are only 1 classes. (Currently this implementation does not incorporate class classification loss.)

> PATH/class1 


> PATH/class2 


> ...

## Notes

![Sample](sample.png)

I have trained model on vanilla celebA dataset. Seems like works well. I found that learning rate (I have used 1e-4 without scheduling), learnt prior, number of bits (in this cases, 5), and using sigmoid function at the affine coupling layer instead of exponential function is beneficial to training a model.

In my cases, LU decomposed invertible convolution was much faster than plain version. So I made it default to use LU decomposed version.

![Progression of samples](progression.gif)

Progression of samples during training. Sampled once per 100 iterations during training.