A lightweight CycleGAN tensorflow implementation.

If you plan to use a CycleGAN model for real-world purposes, you should use the Torch CycleGAN implementation.

@eyyub_s

lion2leopard (cherry-picked)

More lion2leopard (each classes contain only 100 instances!)

horse2zebra

horse2zebra failure

zebra2horse

wtf

More zebra2horse

apple2orange

See more in readme_imgs/

• Download horse2zebra.zip from CycleGAN datasets
• Unzip it here .
• Run python build_dataset.py horse2zebra/trainA horse2zebra/trainB trainA trainB
• (make sure dataset_trainA.npy & dataset_trainB.npy are created)
• Then, run python example.py
• (If you want to stop and restart your training later you can do: python example.py restore <last_iter_number>)

• Python 3.5
• Tensorflow
• Matplotlib
• Pillow
• (Only tested on Windows so far)

• Training took me ~1day (GTX 1060 3g)
• Each 100 steps the script adds an image in the images/ folder
• Each 1000 steps the model is saved in models
• CycleGAN seems to be init-sensitive, if the generators only inverse colors: kill & re-try training

• Image Pool
• Add learning reate linear decay