https://github.com/luoxier/CycleGAN_Tensorlayer

Re-implement CycleGAN in Tensorlayer
https://github.com/luoxier/CycleGAN_Tensorlayer

cyclegan gan resize-convolution tensorlayer

Last synced: 18 days ago
JSON representation

Re-implement CycleGAN in Tensorlayer

• Host: GitHub
• URL: https://github.com/luoxier/CycleGAN_Tensorlayer
• Owner: luoxier
• License: mit
• Created: 2017-09-22T15:26:11.000Z (over 7 years ago)
• Default Branch: master
• Last Pushed: 2018-07-16T08:30:50.000Z (over 6 years ago)
• Last Synced: 2024-10-30T16:41:22.663Z (6 months ago)
• Topics: cyclegan, gan, resize-convolution, tensorlayer
• Language: Python
• Size: 86.3 MB
• Stars: 89
• Watchers: 7
• Forks: 25
• Open Issues: 6
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• awesome-tensorflow - Improved CycleGAN - Unpaired Image to Image Translation (Models/Projects)
• awesome-tensorflow - Improved CycleGAN - Unpaired Image to Image Translation (Models/Projects)
• fucking-awesome-tensorflow - Improved CycleGAN - Unpaired Image to Image Translation (Models/Projects)

README

        
# CycleGAN_Tensorlayer

Re-implement CycleGAN in TensorLayer

- Original CycleGAN

- Improved CycleGAN with resize-convolution

### Prerequisites:

* TensorLayer

* TensorFlow

* Python

### Run:

```Bash

CUDA_VISIBLE_DEVICES=0 python main.py 

```

(if datasets are collected by yourself, you can use dataset_clean.py or dataset_crop.py to pre-process images)

### Theory:

The generator process:

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/generator.png "generator")  

The discriminator process:

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/discriminator.png "discriminator")  

 

### Result Improvement

* Data augmentation

* Resize convolution[4]

* Instance normalization[5]

#### data augmentation:

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/data_augmentation.png) 

#### Instance normalization（comparision by original paper https://arxiv.org/abs/1607.08022）:

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/instance_norm.png) 

#### Resize convolution (Remove Checkerboard Artifacts):

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/compare1.png) 

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/compare2.png) 

### Final Results:

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/result.png)  

![Image text](https://github.com/luoxier/CycleGAN_Tensorlayer/blob/master/figures/result2.png) 

### Reference:

* [1] Original Paper: https://arxiv.org/pdf/1703.10593.pdf

* [2] Original implement in Torch: https://github.com/junyanz/CycleGAN/

* [3] TensorLayer by HaoDong: https://github.com/zsdonghao/tensorlayer

* [4] Resize Convolution: https://distill.pub/2016/deconv-checkerboard/

* [5] Instance Normalization: https://arxiv.org/abs/1607.08022