Brainstorming changes to training preprocessing #3802
enn-nafnlaus
started this conversation in
Ideas
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Not sure I'll actually have time for this, but in case I do... I'm thinking about making changes to the preprocessing tab. It's nice to have, but sorely deficient at present. I'd like to brainstorm some ideas about what would be a desirable approach.
Right now you're sort of operating blind - you just pick two directories and then do the same steps on every file in them, and have no individualized control.
General:
Structural:
Input tab:
Output tab:
What are people's thoughts on this? Obviously - if I have time to work on it at all (I've started browsing through the code), I wouldn't work on everything at once. But before one can start they have to know what's even a desirable direction to move. What do people think? Right now, making a dataset is tedious, and it'd be nice if it were a bit less tedious.
I figured, as a series of steps, it might go:
In the end, it could be amazing to have a workflow like "Specify a search string and autodownload the images, delete the ones you don't like, bulk-choose reasonable parameters for subsampling the images, tell it to generate thousands of training images from your dataset, and then bulk-edit the resultant captions."
I dunno whether I'll even have time - just thinking about this. What are everyone's thoughts? It's just been tedious generating and labeling datasets manually, both in terms of turning hi-res images into numerous smaller images, and in terms of Blip not labeling the very thing you want to train in. And I've been writing various bash and python scripts to help myself out, but I know that's not the right solution - the RIGHT solution is that all this would be in the interface itself.
Beta Was this translation helpful? Give feedback.
All reactions