GauGAN

Unofficial Pytorch implementation of GauGAN, from Semantic Image Synthesis with Spatially-Adaptive Normalization (Park et al. 2019). Implementation for Generative Adversarial Networks (GANs) Specialization course material.

Usage

1. Download the Cityscapes dataset, unzip the gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip folders and move them to data directory.
2. All Python requirements can be found in requirements.txt. Support for Python>=3.7.
3. Default config for can be found in config.yml. All defaults are as per the configurations described in the original paper and code.

Training

By default, all checkpoints will be stored in logs/YYYY-MM-DD_hh_mm_ss, but this can be edited via the train.log_dir field in the config files.

1. To train GauGAN, run python train.py.

Inference

1. Edit the resume_checkpoint field config.yml to reflect the desired checkpoint from training and run python infer.py --encode. The --encode flag generates Gaussian statistics from the input image via the encoder. If not specified, noise will be sampled from a standard Gaussian.

You can edit the number of test images to show with the flag --n_show. Defaults to 5.