Cycle GAN

Tensorflow implementation of Cycle GAN.

• Dataset are converted to TFRecord format
• Tensorboard visualization (metrics, output generated image)

Get started

Clone repository

git clone https://github.com/asahi417/CycleGAN
cd CycleGAN
pip install .

setup dataset

mkdir datasets
bash bin/download_cyclegan_dataset.sh monet2photo

You can choose dataset from following list: monet2photo, horse2zebra, vangogh2photo, ukiyoe2photo, cezanne2photo

convert dataset to TFRecord format

python bin/build_tfrecord.py --data monet2photo 

train cycle GAN

python bin/train.py --data monet2photo -e 200

You can tune hyperparameter by modifying this file.

visualize by tensorboard

tensorboard logdir=./checkpoint --port 555

Result

Here, some generated examples are shown with different identity map regularizer (lambda_id in the paper).

with 0.0 identity regularizer

Fig 1: Horse <-> zebra ([Original, Generated, Cycled, Identity])

Fig 2: Monet <-> photo ([Original, Generated, Cycled, Identity])

with 0.1 identity regularizer

Fig 3: Horse <-> zebra ([Original, Generated, Cycled, Identity])

Fig 4: Monet <-> photo ([Original, Generated, Cycled, Identity])

with 0.5 identity regularizer

Fig 5: Horse <-> zebra ([Original, Generated, Cycled, Identity])

Fig 6: Monet <-> photo ([Original, Generated, Cycled, Identity])