Python interface for ryg_rans entropy coding

This repository is based on CompressAI and ryg_rans with minor modifications to make ANS easier to use for PyTorch models.

• The C++ function accepts NumPy inputs (instead of Python lists)
• Supports CompressAI models with minimal changes
• Faster than CompressAI's encode_with_indexes() and decode_with_indexes()

Install

mkdir build
cd build

# Build the C++ extension
cmake ..
make

Then test-compressai.ipynb should work.

Speed comparison

End-to-end compression of Kodak images using compressai.zoo.mbt2018_mean

• Image size: 512x768
  • Speed averaged over 24 images
• GPU (neural networks forward): Nvidia RTX 4090
• CPU (entropy coding, using a single thread): Intel 10700k