Official code for Closing the Dequantization Gap: PixelCNN as a Single-Layer Flow by Didrik Nielsen, Ole Winther (2020).

Code will soon be updated to include new experiments!

To run the code, we used python 3.6.8 with the following packages installed:

• pytorch==1.2.0
• torchvision==0.4.0
• numpy==1.16.4
• ptable==0.9.2

The code used for experiments have been collected in the python package pixelflow. Thus, we begin by installing this package.

In the folder containing setup.py, run

pip install --user -e .

The --user option ensures the library will only be installed for your user.
The -e option makes it possible to modify the library, and modifications will be loaded on the fly.

You should now be able to use it.

The code for training models can be found in experiments/train, while the code for evaluation can be found in experiments/eval.

The code for training models can be found in experiments/train.
The .py files contain the experiment code and the .sh files contain the commands that were run:

PixelCNN interpolation experiments:

sh interpolations.sh 0

Dequantization gap experiments:

sh dequant_gap.sh 0

Multilayer PixelCNN experiments:

sh multilayer.sh 0

The code for evaluating test performance can be found in experiments/eval/bpd, while the code for performing interpolations can be found in experiments/eval/interpolations.
The commands used to evaluate models can be found in the .sh files (Note that the MODEL_PATH variable need to be set to the path of the pretrained models):

PixelCNN interpolation experiments:

sh interp_pixelcnn.sh 0
sh interp_pixelcnn_pp.sh 0

Dequantization gap experiments:

sh dequant_gap_pixelcnn.sh 0
sh dequant_gap_pixelcnn_quad.sh 0
sh dequant_gap_pixelcnn_pp.sh 0

Multilayer PixelCNN experiments:

sh multilayer.sh 0

Parts of the code build on:

• https://github.com/bayesiains/nsf
• https://github.com/pclucas14/pixel-cnn-pp

Thanks to the authors of these repositories!