VAE + (Conditioned) Quantile Networks for MNIST

PyTorch VAE example extended with Autoregressive Quantile Networks (though, they aren't autoregressive when applied to VAE) https://arxiv.org/abs/1806.05575

By default, writes tensorboard logs to ./tensorboard, change with --log-folder /log/path. Saving just png images without tensorboard is not implemented.

With --conditioned argument AIQN network also receives class labels as inputs. VAE encoder / decoder does not have access to class labels.

Requirements

PyTorch 0.4.0

tensorboard-pytorch https://github.com/lanpa/tensorboard-pytorch

Running

Conditioned AIQN with KL regularizer main.py --conditioned --kl-scale 1 --log-folder ./tensorboard

No AIQN with KL regularizer (same as PyTorch example) main.py --no-aiqn --kl-scale 1 --log-folder ./tensorboard

Examples

AIQN / no AIQN with KL regularizer. AIQN gives some improvements.

AIQN / no AIQN without KL regularizer. Without AIQN and KL penalty generated images are mess. AIQN gives noticible improvement.

Conditioned AIQN with / without KL regularizer.