Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks

This is a PyTorch implementation of the Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks.

Dataset

You can find CycleGAN datasets from this, and chose what kind of dataset you want to download.

To train a model on your own datasets, you need to create a data folder as the following (apple2orange is dataset name).

apple2orange/
	train/
		X/...jpg
		Y/...jpg
	test/
		X/...jpg
		Y/...jpg

Usage

python cyclegan.py ARGS

Possible ARGS are:

• --img_size size of image, default is 128;
• --batch_size size of minibatch, default is 8;
• --dataset_name name of the dataset, default is 'apple2orange';
• --n_residual_blocks number of residual blocks in the generator, default is 6;
• --max_epoch numbers of epoch to train, default is 200;
• --lr the learning rate of Adam, default is 0.0002;
• --decay_start_epoch start of epoch to decay lr, default is 100;
• --sample_interval interval between saving generator outputs, default is 100;
• --checkpoint_interval interval between saving models, default is 100;
• --save_images directory of saving generator outputs, default is 'images';
• --save_models directory of saving models, default is 'save_models'.

An example:

python cyclegan.py --dataset_name "apple2orange"

Result