---
license: cc-by-sa-3.0
datasets:
- mnist
---

[WGAN-GP](https://arxiv.org/abs/1704.00028) model trained on the [MNIST dataset](https://www.tensorflow.org/datasets/catalog/mnist) using [JAX in Colab](https://colab.research.google.com/drive/1RzQfrc4Xf_pvGJD2PaNJyaURLh0nO4Fp?usp=sharing).

| Real Images | Generated Images |
| ------- | -------- |
| ![image.png](https://cdn-uploads.huggingface.co/production/uploads/649f9483d76ca0fe679011c2/YlmgxAdyvJl-oy4Ae_fGB.png) | ![image.png](https://cdn-uploads.huggingface.co/production/uploads/649f9483d76ca0fe679011c2/sNDUja9lFPKiH8UDUqBvl.png) |

# 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](https://arxiv.org/abs/1704.00028).

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](https://colab.research.google.com/drive/1RzQfrc4Xf_pvGJD2PaNJyaURLh0nO4Fp?usp=sharing).