CycleGAN-TensorFlow

CycleGAN implementation in tensorflow using keras layers

Supports up to 2 GPUs

Features

• Supports up to 2 GPUs
• Outputs a TensorBoard log file and the model checkpoints in the 'runs' directory
• The model checkpoints are at the Keras hdf5 saved model format. To load a model, use keras.models.load_model(path_to_model)

Configuration

See example_config.py

Usage

Training the models from scratch

python train.py config_file dataA_path dataB_path

Resuming training

python src/train.py config_file dataA_path dataB_path path_to_checkpoints epoch

Generating samples

python src/generate.py checkpoint_path files_path output_path

Dependencies

• python3
• numpy
• tensorflow >= 1.10
• python-opencv
• progressbar2

Models

• Vanilla CycleGAN generator and discriminator
• PatchGAN discriminator
• UNet generator