My code for various GANS related tasks using wgan with gradient penalty. This adversarial model (generator and critic models) were trained on a single NIDIA RTX 3070 8gb GPU. The trained generator requires about 2gb of memory in production(evaluation + no_grad) mode. train.py contains the code for training the generator and critic using wgan-gp. A docker container will be provided in the future if I can get my hands on better hardware and retrain the models with more parameters.
Fig. Output of generator during traing using celeb face dataset(Left)Click the image to download the WebM video., using smaller model on cat dataset(Right).
.png)
Fig. Outputs of trained generators.
Fig. Training of celeb and cat generators and critics.