Re:NNCP: Lossless Data Compression with Neural Networks
by raphlinus on Saturday April 06, 2019 @11:40AM
(#19591257)
Fair enough. cmix is awesome work, obviously, and I'm looking forward to the ideas becoming production-ready. Probably the most notable contribution by Fabrice Bellard is having something relatively simple.
No, it's 10x faster than an experimental, completely performance-unoptimized LSTM model that is far from state-of-the-art on its own. From the paper:
"Regarding the speed, our results show that our implementation is much faster than lstm-compress[8] with a similar model and gain. The speed improvement comes from the more optimized matrix multiplication implementation and the larger batch size (16 instead of 1)."
Of most interest here is the library used, LibNC, which might allow us to speed up cmix a bit (even making it 2x faster would already help a lot in testing).
Disclaimer: I've worked on cmix with its author