3cefb57fc6
fp32 Update
2022-12-03 19:42:50 +09:00
18ff256be5
Implement Text Encoder Training
2022-12-03 12:47:40 +09:00
c4b17e41f1
Add np seed
...
Just in case!
2022-12-02 01:49:12 +09:00
d225cd57e7
Fix Seed
2022-12-02 01:41:22 +09:00
b96eabfe2a
Fix
2022-12-02 01:24:00 +09:00
138cb7bbed
Merge pull request #54 from harubaru/extended-mode
...
Adding Extended Mode Functionality
2022-11-30 22:55:30 -07:00
55a555850d
Remove dangling barrier
2022-11-30 22:52:48 -07:00
7102d313ac
Synchronize ranks for DDP
2022-11-30 19:05:02 -07:00
1074bd6f3b
Bug Fix
2022-12-01 05:10:21 +09:00
fb75cbe029
Provide Tokens for Inference
2022-12-01 04:46:01 +09:00
981c6ca41a
Cleanup
2022-12-01 04:32:10 +09:00
ee281badcd
Extended Mode Updates
2022-12-01 04:31:30 +09:00
4572617ff9
use ddp for everything
2022-11-30 10:54:30 -07:00
b0cec788be
Get DDP to work
2022-11-29 22:06:21 -07:00
8decb0bc7d
Fix distributed training
2022-11-29 18:01:17 -07:00
5787f7d080
Update diffusers_trainer.py
2022-11-21 13:57:26 +09:00
f2cfe65d09
Move the movel to device BEFORE creating the optimizer
...
>It shouldn’t matter, as the optimizer should hold the references to the parameter (even after moving them). However, the “safer” approach would be to move the model to the device first and create the optimizer afterwards.
https://discuss.pytorch.org/t/should-i-create-optimizer-after-sending-the-model-to-gpu/133418/2
https://discuss.pytorch.org/t/effect-of-calling-model-cuda-after-constructing-an-optimizer/15165
At least in my experience with hivemind, if you initialize the optimizer and move the model afterwards, it will throw errors about finding some data in CPU and other on GPU. This shouldn't affect performance or anything I believe.
2022-11-20 00:09:35 -05:00
9916294de1
fix wandb init mode, don't log hf token
...
correct value for mode ('enabled' is invalid)
clear hf_token passed to wandb to avoid logging it
2022-11-16 22:28:16 +01:00
dc5849b235
Merge branch 'main' into inference-option
2022-11-16 16:20:57 -05:00
a2772fc668
fixes
2022-11-16 10:55:38 -05:00
fed3431f03
Revert "sync trainer with main branch"
...
This reverts commit 80e2422967
2022-11-16 10:44:39 -05:00
80e2422967
sync trainer with main branch
2022-11-16 10:39:20 -05:00
015eeae274
Documentation consistency.
2022-11-15 10:34:55 -07:00
29ffbd645e
Fix noise scheduler
2022-11-15 11:08:38 -05:00
6c5b2e7149
Fix of fix
2022-11-15 07:15:18 -07:00
2c18d29613
Fix from upstream merge.
2022-11-15 06:42:14 -07:00
d600078008
Add options and local inference
...
Added options to:
- Disable Inference (it consumes about 2gb of VRAM even when not active)
- Disable wandb
and:
- if no hftoken is provided it just fills it with nothing so it doesn't argues
- if wandb is not enabled, save the inference outputs to a local folder along with information about it
2022-11-14 22:08:16 -05:00
773e65f324
Merge origin:main into remote:staging-migration
2022-11-14 19:59:45 -07:00
95b9407a3e
Add+config .gitignore (bring back git stage) and fix up documentation.
2022-11-12 18:48:16 -07:00
6bd6c6a4ef
Fixed/flipped help text.
2022-11-12 16:11:30 -07:00
189f621a1e
Here, let's fix this while we're at it.
2022-11-12 15:47:17 -07:00
cd0910e82d
Parse booleans in argument parser
...
true, yes or 1 correspond to True, else False.
2022-11-12 11:14:48 +01:00
d1eb3ace3f
I lied.
2022-11-11 18:17:29 -07:00
6c2d5d8066
Final cleanup.
2022-11-11 18:09:09 -07:00
de221ea42e
Derp. ImageStore.__init__ already iterates fully :)
2022-11-11 17:59:32 -07:00
925eacf374
Cleanup
2022-11-11 17:50:23 -07:00
c12cbfced3
Fixed ref typo.
2022-11-11 17:43:09 -07:00
6480336d2c
Cleanup test code.
2022-11-11 17:17:50 -07:00
120d406355
Implementation of validation/resize classes.
2022-11-11 17:14:46 -07:00
624f0f14af
correctly set args
2022-11-11 08:25:27 -07:00
1f5b671b67
relicense
2022-11-10 12:59:53 -07:00
cafeai