PyTorch implementation of MADE for MNIST
This is implementation of Masked Autoencoder for Distribution Estimation(MADE). To facilitate sampling numbers, I used one-hot encoding of numbers as additional input. The shape of input is batch size * 794 (not 784). This implementation includes direct connection between input and output and connectivity-agnostic training. But i didn't implement order-agonostic training.
The following results were obtained with the default setting. (command: python made.py)
| Reconstruction | Sampling |
 |
 |
The implementation is based on:
[1] https://arxiv.org/abs/1502.03509