Variational Autoencoder in PyTorch and Fastai V1

An implementation of the VAE in pytorch with the fastai data api, applied on MNIST TINY (only contains 3 and 7). The notebook is the most comprehensive, but the script is runnable on its own as well. Results from sampling are saved in the results directory.

Script usage:

usage: vae.py [-h] [--batch-size N] [--epochs N] [--no-cuda]
              [--emb-size EMB_SIZE]

VAE MNIST Example

optional arguments:
  -h, --help           show this help message and exit
  --batch-size N       input batch size for training (default: 128)
  --epochs N           number of epochs to train (default: 10)
  --no-cuda            enables CUDA training
  --emb-size EMB_SIZE  size of embedding (default 10)

Results from sampling latent space