[P] 图片到图片转换 Improved CycleGAN with Resize-Convolution

Re-implement CycleGAN in Tensorlayer

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 arxiv.org/abs/1607.08…:

Resize convolution (Remove Checkerboard Artifacts):

Final Results:

Reference:

• [1] Original Paper: arxiv.org/pdf/1703.10…
• [2] Original implement in Torch: github.com/junyanz/Cyc…
• [3] TensorLayer by HaoDong: github.com/zsdonghao/t…
• [4] Resize Convolution: distill.pub/2016/deconv…
• [5] Instance Normalization: arxiv.org/abs/1607.08…