Commit Graph

34 Commits

Author SHA1 Message Date
mrq
6676c89c0e I sucked off the hyptothetical wizard again, just using BNB's ADAM optimizer nets HUGE savings, but I don't know the output costs, will need to test 2023-02-23 02:42:17 +00:00
mrq
4427d7fb84 initial conversion (errors out) 2023-02-22 23:07:05 +00:00
James Betker
f7d237a50a train quantizer with diffusion 2022-05-30 16:25:33 -06:00
James Betker
6b43915eb8 support projecting to vectors 2022-05-28 22:27:45 -06:00
James Betker
48aab2babe ressurect ctc code gen with some cool new ideas 2022-05-24 14:02:33 -06:00
James Betker
1e1bbe1a27 whoops 2022-05-23 12:28:36 -06:00
James Betker
560b83e770 default to residual encoder 2022-05-23 12:24:00 -06:00
James Betker
f432bdf7ae deeper resblock encoder 2022-05-23 11:46:40 -06:00
James Betker
dc471f5c6d residual features 2022-05-23 09:58:30 -06:00
James Betker
1f521d6a1d add reconstruction loss to m2v 2022-05-23 09:28:41 -06:00
James Betker
2270c89fdc . 2022-05-23 08:47:15 -06:00
James Betker
40f844657b tolong 2022-05-23 08:27:54 -06:00
James Betker
10f4a742bd reintroduce attention masks 2022-05-23 08:16:04 -06:00
James Betker
68c0afcbcc m2v frequency masking 2022-05-23 07:04:12 -06:00
James Betker
8f28404645 another fix 2022-05-22 21:32:43 -06:00
James Betker
41809a6330 Add 8x dim reductor 2022-05-22 20:23:16 -06:00
James Betker
3121bc4e43 flat diffusion 2022-05-20 11:01:48 -06:00
James Betker
c9c16e3b01 misc updates 2022-05-19 13:39:32 -06:00
James Betker
10378fc37f make codebooks specifiable 2022-05-18 11:07:12 -06:00
James Betker
efc2657b48 fiddle with init 2022-05-18 10:56:01 -06:00
James Betker
208a703080 use gelu act 2022-05-18 09:34:01 -06:00
James Betker
b2b37453df make the codebook bigger 2022-05-17 20:58:56 -06:00
James Betker
9a9c3cafba Make feature encoder a bit more descriptive 2022-05-17 18:14:52 -06:00
James Betker
ee364f4eeb just take the mean... 2022-05-17 18:09:23 -06:00
James Betker
6130391a85 fix div 2022-05-17 18:04:20 -06:00
James Betker
7213ad2b89 Do grad reduction 2022-05-17 17:59:40 -06:00
James Betker
7c82e18c6c darn mpi 2022-05-17 17:16:09 -06:00
James Betker
88ec0512f7 Scale losses 2022-05-17 17:12:20 -06:00
James Betker
a6397ce84a Fix incorrect projections 2022-05-17 16:53:52 -06:00
James Betker
c37fc3b4ed m2v grad norm groups 2022-05-17 16:29:36 -06:00
James Betker
c1bdb4f9a1 degrade gumbel softmax over time 2022-05-17 16:23:04 -06:00
James Betker
3853f37257 stable layernorm 2022-05-17 16:07:03 -06:00
James Betker
519151d83f m2v 2022-05-17 15:37:59 -06:00
James Betker
d1de94d75c Stash mel2vec work (gonna throw it all away..) 2022-05-17 12:35:01 -06:00