Inverse Autoregressive Flows

A Pytorch implementation of this paper

Main work taken from the official implementation

Running the following command gets ~ 3.35 BPD (which is still some ways from the 3.11 in the original paper)

python main.py --batch_size 128 --depth 1 --n_blocks 20 --free_bits 0.001 --lr 0.002 --z_size 32 --h_size 160 --iaf 1

As comparison, the baseline (IAF disabled) gets ~ 3.55 BPD. You can test with the following command

python main.py --batch_size 128 --depth 1 --n_blocks 20 --free_bits 0.001 --lr 0.002 --z_size 32 --h_size 160 --iaf 0

Images

Here are some test set reconstructions, and samples

Judging by how good the reconstructions are, maybe the free bits constraint is too loose. More investigation required.

Contribute

I'm having trouble closing the performance gap with the offical code, as Conda does not support anymore the required tensorflow version to run it. Therefore, all contributions / comments / remarks are highly welcomed.