Compression with Flows via Local Bits-Back Coding

Jonathan Ho, Evan Lohn, Pieter Abbeel

Neural Information Processing Systems, 2019

https://arxiv.org/abs/1905.08500

Contains a PyTorch implementation of Flow++.

Models available here.

Dependencies

• Python 3.6.7
• PyTorch 1.1.0 (CUDA 10.0)

Installation instructions

1. Install Anaconda with Python 3.6.7
2. conda install pytorch torchvision cudatoolkit=10.0 -c pytorch
3. pip install tqdm
4. Extract this codebase into a directory called compression
5. Build the underlying C++ library (need a compiler with OpenMP support):

cd compression/ans
mkdir build
cd build
cmake ..
make