Check references at docs/references.
I am using RWKV for compression. It is pretty good, the layers are stackable.
N Layers | kb/min | # Params | gpu allocation |
---|---|---|---|
RWKV 1 | "1202,36" | 0.06M | 1.09G |
RWKV 2 | "714,29" | 0.11M | 1.42G |
RWKV 4 | "380,08" | 0.22M | 1.84G |
RWKV 8 | "202,10" | 0.42M | 3.43G |
"Tensorflow-compress (Mao, Y. et. al, 2022)" | "30,72~163,8 " | 258.3M | 10.64G |
"NNCP (Mao, Y. et. al, 2022)" | "28,6~122,9" | 151M-224M | 7.75G |
"Dzip (Mao, Y. et. al, 2022)" | "399,4" | 1M | 6.39G |
"TRACE (Mao, Y. et. al, 2022)" | "952,3" | 2.40M | 2.02G |
"TRACE+BP controller (Mao, Y. et. al, 2022)" | "1228,8" | 2.40M | 2.02G |
TRACE+BP is not open source, and they use a better graphics card, I have to test because I think in my computer it would be a lot slower.
Name of document | Ratio | File size | Type |
---|---|---|---|
enwik8 | 1 | 100000000 | Original |
enwik8_1_1024.compressed | 3,43 | 29120871 | RWKV |
enwik8_2_1024.compressed | 3,71 | 26958823 | RWKV |
enwik8_4_1024.compressed | 3,97 | 25215016 | RWKV |
enwik8_8_1024.compressed | 4,15 | 24114950 | RWKV |
enwik8.zip | 2,74 | 36445475 | Traditional |
gzip -9 enwik8.gz | 2,74 | 36445248 | Traditional |
7z enwik | 3,83 | 26080305 | Traditional |