https://github.com/dragen1860/introvae-pytorch

Pytorch Implementation for paper: IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis
https://github.com/dragen1860/introvae-pytorch

Last synced: 6 months ago
JSON representation

Pytorch Implementation for paper: IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis

• Host: GitHub
• URL: https://github.com/dragen1860/introvae-pytorch
• Owner: dragen1860
• Created: 2018-12-05T08:34:26.000Z (almost 7 years ago)
• Default Branch: master
• Last Pushed: 2018-12-10T04:12:26.000Z (almost 7 years ago)
• Last Synced: 2025-03-30T17:11:10.290Z (7 months ago)
• Language: Python
• Homepage:
• Size: 1.69 MB
• Stars: 39
• Watchers: 4
• Forks: 11
• Open Issues: 2
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

          
# IntroVAE-Pytorch

Pytorch Implementation for NeuraIPS2018 paper: 

[IntroVAE: Introspective Variational Autoencoders for Photographic Image Synthesis](https://arxiv.org/abs/1807.06358).

The rep. contains a basic implementation for IntroVAE. However, due to no official implementation released, some hyperparameters can only be guessed and can not reach the performance as stated in paper.

![](assets/heart.gif)

# HowTo

1. Download [CelebA](http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html) dataset and extract it as:

```

├── /home/i/dbs/

	├──img_align_celeba # only one folder in this directory

		├── 050939.jpg

		├── 050940.jpg

		├── 050941.jpg

		├── 050942.jpg

		├── 050943.jpg

		├── 050944.jpg

		├── 050945.jpg

```

modify `/home/i/dbs` to your specific path, making sure that the `/home/i/dbs/` comtains only ONE folder since we use

`torchvision.datasets.ImageFolder` API to load dataset. 

```python

    argparser.add_argument('--root', type=str, default='/home/i/dbs/',

                           help='root/label/*.jpg')

```

2. run `python main.py  --epoch 750000` to train from strach, and use `python main.py --resume '' --epoch 1000000` to resume training from latest checkpoint.

# Training

only tested for CelebA 128x128 exp.

- training curves

![](assets/train.png)

- sampled x

![](assets/xr_750000.png)