CycleGAN_Tensorlayer

Re-implement CycleGAN in TensorLayer

• Original CycleGAN
• Improved CycleGAN with resize-convolution

Prerequisites:

• TensorLayer
• TensorFlow
• Python

Run:

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:

The discriminator process:

Result Improvement

• Data augmentation
• Resize convolution[4]
• Instance normalization[5]

data augmentation:

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

Resize convolution (Remove Checkerboard Artifacts):

Final Results:

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