Awesome
Integer-Only Discrete Flows (ICML 2022) [arxiv]
This repository contains Pytorch implementation of experiments from the paper Fast Lossless Neural Compression with Integer-Only Discrete Flows. The implementation is based on Integer Discrete Flows. rANS entropy coding in C language is based on local bits back.
Main Dependency
- Python >= 3.7
- Pytorch 1.9.0
- TensorRT 8.2.0.6 + CUDA 10.2
Usage
<!-- Basic training IODF and coding with IODF: ``` python run_train.py --nn_type resnet --dataset imagenet32 --batchsize 256 python run_coding.py --nn_type resnet --dataset imagenet32 --batchsize 500 --resume base --no_decode ``` -->-
Follow training procedure described by Algorithm.1 in the paper. Refer to commands.sh for detailed scripts.
-
For TensorRT Implementation, switch to branch trt.
Contact
Please open an issue.
Cite
@inproceedings{wang2022fast,
title={Fast Lossless Neural Compression with Integer-Only Discrete Flows},
author={Wang, Siyu and Chen, Jianfei and Li, Chongxuan and Zhu, Jun and Zhang, Bo},
booktitle={International Conference on Machine Learning},
year={2022},
organization={PMLR}
}