WGAN-GP model trained on the MNIST dataset using JAX in Colab.

Real Images	Generated Images

Training Progression

<video width="50%" controls> <source src="https://cdn-uploads.huggingface.co/production/uploads/649f9483d76ca0fe679011c2/nX7L6xkjvAvaca5pHyTp0.mp4" type="video/mp4"> </video>

Details

This model is based on WGAN-GP.

The model was trained for ~9h40m on a GCE VM instance (n1-standard-4, 1 x NVIDIA T4).

The Critic consists of 4 Convolutional Layers with strides for downsampling, and Leaky ReLU activation. The critic does not use Batch Normalization or Dropout.

The Generator consists of 4 Transposed Convolutional Layers with ReLU activation and Batch Normalization.

The learning rate was kept constant at 1e-4 for the first 50,000 steps, which was followed by cosine annealing cycles with a peak LR of 1e-3.

The Lambda (gradient penalty coefficient) used was 10 (same as the original paper).

For more details, please refer to the Colab Notebook.