A minimum implementation of importance weighted autoencoder from Burda et al. (2016). https://arxiv.org/abs/1509.00519
In Pytorch.
Negative log-likelihood of test data using k importance samples after 50 epochs.
| k | Negative Log-Likelihood |
|---|---|
| 1 | 86.886 |
| 5 | 81.439 |
| 50 | 79.008 |
| 1000 | 77.317 |
- I used the original MNIST dataset. You may want to download the binarized version of MNIST referenced in the paper.
- At this time torchvision cannot automatically download MNIST dataset. Apparently this is a server side issue. You can however download it manually and then set the root arg of torchvision.datasets.MNIST function with the proper local dataset directory. Keep the download=True flag.
- I did not use the schedule for the Beta coefficient of ADAM optimizer used in the paper.