This example tries to closely resemble the original pix2pix paper with the same architecture and hyper-parameters, but using Swift for TensorFlow. Results seem to be close to what the original authors claimed to achieve, this is what a generator outputs after 7 hours of training using single GTX 1080 card:
In order to run the project you have to just call
swift run -c release pix2pix- Trainer will emit a lot of intermediate results in a calling directory, you can easily comment out that part.
- You need to specify a different URL for
datasetor this will download the facades dataset for you. - This implementation only provides a data loader for the facades dataset, but the model is capable of learning on any pix2pix compatible dataset.