0f04206aa2
added ability to toggle some settings with envvars for later testing without needing to manually edit this file (and some other things like disabling it when a user requests it in the future)
2023-02-24 23:08:56 +00:00
1433b7c0ea
working Embedding override
2023-02-23 07:28:27 +00:00
fd66c4104b
ugh
2023-02-23 07:18:07 +00:00
7bcedca771
I guess I can't easily toggle it outside of here, but it works
2023-02-23 07:02:06 +00:00
58600274ac
Disabling bitsandbytes optimization as default for now, in the off chance that it actually produces garbage (which shouldn't happen, there's no chance, if training at float16 from a model at float16 works fine, then this has to work)
2023-02-23 03:22:59 +00:00
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
4427d7fb84
initial conversion (errors out)
2023-02-22 23:07:05 +00:00
6c284ef8ec
oops
2023-02-18 03:27:04 +00:00
8db762fa17
thought I copied this over
2023-02-18 03:15:44 +00:00
73d9c3bd46
set output folder to be sane with the cwd as a reference point
2023-02-18 02:01:09 +00:00
5ecf7da881
Fix later
2023-02-17 20:49:29 +00:00
e3e8801e5f
Fix I thought wasn't needed since it literally worked without it earlier
2023-02-17 20:41:20 +00:00
535549c3f3
add some snark about the kludge I had to fix, and the kludge I used to fix it
2023-02-17 19:20:19 +00:00
a09cf98c7f
more cleanup, pip-ifying won't work, got an alternative
2023-02-17 15:47:55 +00:00
6afa2c299e
break if your dataset size is smaller than your batch size
2023-02-17 04:08:27 +00:00
94d0f16608
Necessary fixes to get it to work
2023-02-17 02:03:00 +00:00
49e23b226b
pip-ify
2023-02-17 00:33:50 +00:00
James Betker
f31a333c4f
more sampling fixes
2022-10-10 20:11:28 -06:00
James Betker
5d172fbf7e
Fix eval
2022-10-10 14:22:36 -06:00
James Betker
9502e0755e
ugh
2022-10-10 12:15:51 -06:00
James Betker
fce2c8f5db
and listify them
2022-10-10 12:13:49 -06:00
James Betker
3cf78e3c44
train mel head even when not
2022-10-10 12:10:56 -06:00
James Betker
cc74a43675
Checkin
2022-10-10 11:30:20 -06:00
James Betker
3cb14123bc
glc fix
2022-07-29 11:24:36 -06:00
James Betker
4ddd01a7fb
support generating cheaters from the new cheater network
2022-07-29 09:19:20 -06:00
James Betker
27a9b1b750
rename perplexity->log perplexity
2022-07-28 09:48:40 -06:00
James Betker
1d68624828
fix some imports..
2022-07-28 02:35:32 -06:00
James Betker
cfe907f13f
i like this better
2022-07-28 02:33:23 -06:00
James Betker
d44ed5d12d
probably too harsh on ninfs
2022-07-28 01:33:54 -06:00
James Betker
4509cfc705
track logperp for diffusion evals
2022-07-28 01:30:44 -06:00
James Betker
19eb939ccf
gd perplexity
...
# Conflicts:
# codes/trainer/eval/music_diffusion_fid.py
2022-07-28 00:25:05 -06:00
James Betker
a1bbde8a43
few things
2022-07-26 11:52:03 -06:00
James Betker
f8108cfdb2
update environment and fix a bunch of deps
2022-07-24 23:43:25 -06:00
James Betker
45afefabed
fix booboo
2022-07-24 18:00:14 -06:00
James Betker
cc62ba9cba
few more tfd13 things
2022-07-24 17:39:33 -06:00
James Betker
f3d967dbf5
remove eta from mdf
2022-07-24 17:21:20 -06:00
James Betker
76464ca063
some fixes to mdf to support new archs
2022-07-21 10:55:50 -06:00
James Betker
13c263e9fb
go all in on m2wv3
2022-07-21 00:51:27 -06:00
James Betker
24a78bd7d1
update tfd14 too
2022-07-21 00:45:33 -06:00
James Betker
02ebda42f2
#yolo
2022-07-21 00:43:03 -06:00
James Betker
b92ff8de78
misc
2022-07-20 23:59:32 -06:00
James Betker
a1743d26aa
Revert "Try to squeeze a bit more performance out of this arch"
...
This reverts commit 767f963392
2022-07-20 23:57:56 -06:00
James Betker
767f963392
Try to squeeze a bit more performance out of this arch
2022-07-20 23:51:11 -06:00
James Betker
b9d0f7e6de
simplify parameterization a bit
2022-07-20 23:41:54 -06:00
James Betker
ee8ceed6da
rework tfd13 further
...
- use a gated activation layer for both attention & convs
- add a relativistic learned position bias. I believe this is similar to the T5 position encodings but it is simpler and learned
- get rid of prepending to the attention matrix - this doesn't really work that well. the model eventually learns to attend one of its heads to these blocks but why not just concat if it is doing that?
2022-07-20 23:28:29 -06:00
James Betker
40427de8e3
update tfd13 for inference
2022-07-20 21:51:25 -06:00
James Betker
dbebe18602
Fix ts=0 with new formulation
2022-07-20 12:12:33 -06:00
James Betker
82bd62019f
diffuse the cascaded prior for continuous sr model
2022-07-20 11:54:09 -06:00
James Betker
b0e3be0a17
transition to nearest interpolation mode for downsampling
2022-07-20 10:56:17 -06:00
James Betker
7b3fc79737
iq checkin
2022-07-20 10:19:32 -06:00
