Platform to create syntetic image datasets using CycleGAN. It transforms images from a domain A with some features to a domain B with different features (for example horses to horses, day to night, etc). It uses CycleGAN, which allows to use unsupervised learning to transform images between two domains. Apart from CycleGAN, other image-to-image translation algorithms are being added, such as MixerGAN. Better parametrizations and more models are going to be added in the future.
In order to do so, it is possible to add the the image dataset to the images folder as follows:
Where train_A contains the images to train from the domain A and train_B contains images to train from the domain B. The same happens for the testing dataset. The datasets must have the previous format in order to be used by the platform.
If your dataset is divided in two folder containing the images in two domains, it is possible to use the scripts/split_dataset.py script to obtain the required format.
python scripts/split_datasets.py ./images/image_folder_A ./images/image_folder_B imagesA2imagesBThe images dataset is going to be stored in the images/ folder. Use the the format images_a2images_b as folder name.
Clone the repository
git clone https://github.com/GonzaloRomeroR/agro_cycle_gan.gitFor pip users, install the required packages: pip install environments/requirements.txt
To train the models use:
python -u train.py horse2zebra --batch_size 10 --num_epochs 20 --image_resize 64 64 --plot_image_epochWhere horse2zebra is the name of the folder with the training images.
In order to generate images run:
python generate.py ./images/horse2zebra/test_A/A ./images_gen/horse2zebra/ --generator_name horse2zebra --dest_domain BWhere ./images/horse2zebra/test_A/A is the folder containing the images to transform, ./images_gen/horse2zebra/ is the destination folder where the transformed images will be stored. The generator name must be equal to the dataset name. It is possible to select the desination domain with --dest_domain (it can be A or B).