This repo provides an example for a Generative Adversarial Network that is trained to create synthetic images for an segmentation algorithm. The used CycleGAN-algorithm learns an image-to-image transformation the tries to minimize the loss in the forward and backward conversion of the image.
The provided code follows the following medium-arcticle: https://medium.com/data-science-in-your-pocket/understanding-cyclegans-using-examples-codes-f5d6e1a47048
The repo is managed with poetry. To install the environment run
poetry install
from the root directory.
The algorithm is split into a training and a test part. Training requires template and target images to be located in a directory train. Additionally, a meta.csv file is required that contains the respective image urls to differentiate between template and target images. An example file is provided in the train-directory
Run train.py to start training the CylceGAN.