3 ms·
The point of the challenge is to actually create a model that fits in 15MB together with its compressed data. Not just to prove its existence. One of the best
by klipt 7y ago
The point of the challenge is to actually create a model that fits in 15MB together with its compressed data. Not just to prove its existence.
One of the best contenders is the open source cmix which, among other tricks, uses LSTM neural networks to predict the next character/bit: http://www.byronknoll.com/cmix.html http://www.byronknoll.com/cmix.html
- kzrdude 7y agoSeems like cmix compresses better than the winners, but uses far too much memory to be eligible for the prize.
- Dylan16807 7y agoI had to search out the size of the decompressor on a different page, but it proves you right. The compressed file is ~400KB smaller than the record holder and the cmix decompressor is only ~200KB. I wonder what the memory/speed/efficiency tradeoffs looks like for that kind of design.