log loss per image #7278
log loss per image #7278
Conversation
| @@ -19,6 +19,7 @@ | |||
| import logging | |||
| import math | |||
| import os | |||
| import os.path | |||
There was a problem hiding this comment.
We already have os module imported. I think we can reuse it.
| for fname in batch['filenames']: | ||
| accelerator.log({'loss_for_' + fname: loss}, step=global_step) |
There was a problem hiding this comment.
Should we accept a CLI argument for this? That will be nice no?
There was a problem hiding this comment.
I agree, it's nice to have.
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
@sayakpaul Which code quality issues? What should i fix? |
|
The CI says: 
To be more specific the following step in the CI fails:
You need to run |
| parser.add_argument( | ||
| "--debug-loss", | ||
| action="store_true", | ||
| default=False, |
There was a problem hiding this comment.
We don't need to set a default for boolean flag CLI args. If someone doesn't pass --debug-loss it will be defaulted to False. Also, we use "_" rather than "-" even for CLI args.
|
Great! Could you please merge the PR? |
|
Let's go! |
What does this PR do?
logging loss for each image in dataset. Given that Lora datasets are usually small this feature might help to one to spot low-quality images or bad captions etc.