Repository navigation
[Tracker] modularize inferencing during and after training in the example scripts #6545
Description
We provide support for running validation inference during and after training in our officially maintained training examples. This is very helpful to keep track of the training progress.
We could modularize some bits in the example to reduce the LoC.
The train_lcm_distill_lora_sdxl.py script already does this:
It would be nice to follow something similar for the rest of the scripts too. Here's a handy list of the scripts where we'd like to incorporate this change:
Feel free to comment here if you're interested.
Also, when opening PRs, please target one example at a time and please tag me in the PRs.
Reacted by Charchit Sharma
Activity
charchit7 commentedon Jan 12, 2024
Hey @sayakpaul for text-to-image, SD
log-validationis present :diffusers/examples/text_to_image/train_text_to_image.py
Line 141 in 33d2b5b
haven't gone through the code, will check more if we need to modify it.
sayakpaul commentedon Jan 12, 2024
Ideally, this should be delegated to
log_validation()too:diffusers/examples/text_to_image/train_text_to_image.py
Line 1047 in 33d2b5b
As done in
train_lcm_distill_lora_sdxl.py.charchit7 commentedon Jan 12, 2024
Got it. @sayakpaul I missed that part. Thanks!
kim-sangyeon commentedon Jan 14, 2024
Hey @sayakpaul, @charchit7
I'm interested in incorporating the change to InstructPix2Pix SD
coolyashas commentedon Jan 14, 2024
Hey @sang-k, I have been looking into InstructPix2Pix SD too. Would be happy to collaborate if necessary.
kim-sangyeon commentedon Jan 14, 2024
@coolyashas
Oh, I see.
I will look into InstructPix2Pix SDXL then!
coolyashas commentedon Jan 21, 2024
Working on InstructPix2Pix SD
[-][Examples] modularize inferencing during and after training[/-][+][Tracker] modularize inferencing during and after training in the example scripts[/+]ParagEkbote commentedon Mar 12, 2025
Has this issue been resolved or are there any pending scripts I could contribute to modularize?
cc: @sayakpaul
sayakpaul commentedon Mar 12, 2025
Yeah I think this can be closed now :)