https://bellard.org/nncp/

NNCP: Lossless Data Compression with Neural Networks

NNCP is an experiment to build a practical lossless data compressor
with neural networks. The latest version uses a Transformer model.

The papers nncp_v2.1.pdf and nncp.pdf describe the algorithms and
results of previous releases of NNCP.

The current release of NNCP is implemented in C and uses LibNC to get
better performance than PyTorch.

Compression ratio

Result for enwik8:


     Program      Compr. size Ratio
                    (bytes)   (bpb)
gzip              36 445 248  2.92
xz                24 865 244  1.99
NNCP (2021-02-06) 15 020 691  1.20
CMIX (v18)        14 838 332  1.19


Result for enwik9:


    Program      Compr. size Ratio    Program size(zip,       Total
                   (bytes)   (bpb)         bytes)            (bytes)
gzip             322 591 995 2.58  38 801                  322 630 796
xz               197 331 816 1.58  36 752                  197 368 568
CMIX (v18)       115 714 367 0.926 208 961                 115 923 328
NNCP             110 034 293 0.880 197 491                 110 231 784
(2021-04-24)

* The results for the other programs are from the Large Text
Compression Benchmark.

Download

  * NNCP v3: Linux version (including CUDA support):
    nncp-2021-04-24.tar.gz (Changelog, readme.txt).
  * NNCP v3: Precompiled Windows version (no CUDA support):
    nncp-2021-04-24-win64.zip.
  * NNCP v2 (Python+PyTorch, GPU required):
    nncp_v2-2021-02-06-1.tar.gz

Related Links

  * LibNC: C Library for Tensor Manipulation.
  * gpt2tc: Text Completion and Compression using GPT-2.
  * NNCP thread on the encode.su forum.
  * CMIX lossless data compression program.
  * lstm-compress: lossless data compression with LSTM.
  * Large Text Compression Benchmark.

---------------------------------------------------------------------
Fabrice Bellard - https://bellard.org/