Software Engineering Courses (SECourses)•2y ago

but usually gets overtrained quickly

Furkan Gözükara SECoursesOP•12/31/23, 2:30 PM

if images are not diverse

FFurkan Gözükara SECourses but usually gets overtrained quickly

DJRich•12/31/23, 2:31 PM

What would you recommend in a case like this? Fewer repeats? Fewer epochs? Setting a max number of steps?

DDJRich What would you recommend in a case like this? Fewer repeats? Fewer epochs? Setti...

Furkan Gözükara SECoursesOP•12/31/23, 2:31 PM

what are you training

Furkan Gözükara SECoursesOP•12/31/23, 2:31 PM

a person?

DJRich•12/31/23, 2:31 PM

yes

Furkan Gözükara SECoursesOP•12/31/23, 2:32 PM

i would discard similar images. if still very variety of images then try 100 repeat and save every n step

Furkan Gözükara SECoursesOP•12/31/23, 2:32 PM

like every 1801 steps

Furkan Gözükara SECoursesOP•12/31/23, 2:32 PM

actually make it half

Furkan Gözükara SECoursesOP•12/31/23, 2:32 PM

50 repeat

Furkan Gözükara SECoursesOP•12/31/23, 2:32 PM

901 steps

Furkan Gözükara SECoursesOP•12/31/23, 2:33 PM

since our reg images count 5200

FFurkan Gözükara SECourses i would discard similar images. if still very variety of images then try 100 rep...

DJRich•12/31/23, 2:33 PM

100 repeats will create even more steps though. I'm currently only using 20

Furkan Gözükara SECoursesOP•12/31/23, 2:33 PM

20 * 90 * 2

Furkan Gözükara SECoursesOP•12/31/23, 2:33 PM

it is 3600 steps

Furkan Gözükara SECoursesOP•12/31/23, 2:33 PM

make epoch 1

Furkan Gözükara SECoursesOP•12/31/23, 2:33 PM

1 epoch 50 repeat

FFurkan Gözükara SECourses 20 * 90 * 2

DJRich•12/31/23, 2:34 PM

Ok I'm a bit lost now

DJRich•12/31/23, 2:34 PM

20 steps, 90 images, x2 because of regularization.

DJRich•12/31/23, 2:34 PM

I get that part

FFurkan Gözükara SECourses resizer will resize

Kallamamran•12/31/23, 2:34 PM

I don't have a resizer

FFurkan Gözükara SECourses 1 epoch 50 repeat

DJRich•12/31/23, 2:35 PM

Perhaps you can show me via a screenshot what this would look like in Kohya?

KKallamamran I don't have a resizer

Furkan Gözükara SECoursesOP•12/31/23, 2:58 PM

download it from this post

Furkan Gözükara SECoursesOP•12/31/23, 2:58 PM

https://www.patreon.com/posts/sota-subject-and-88391247

DDJRich Perhaps you can show me via a screenshot what this would look like in Kohya?

Furkan Gözükara SECoursesOP•12/31/23, 2:58 PM

you set 1 epoch in gui

Furkan Gözükara SECoursesOP•12/31/23, 2:58 PM

and 50 repeat name in folder

Furkan Gözükara SECoursesOP•12/31/23, 2:59 PM

Kaggle SDXL training works. after cache it will give vram error. just click training again

FFurkan Gözükara SECourses and 50 repeat name in folder

DJRich•12/31/23, 3:00 PM

And set the save every # of steps on the parameters tab?

FFurkan Gözükara SECourses Kaggle SDXL training works. after cache it will give vram error. just click trai...

DJRich•12/31/23, 3:02 PM

Not sure if that was for me, but I'm doing local training. I stopped using Kaggle.

DDJRich And set the save every # of steps on the parameters tab?

Furkan Gözükara SECoursesOP•12/31/23, 3:15 PM

yes

DDJRich Not sure if that was for me, but I'm doing local training. I stopped using Kaggl...

Furkan Gözükara SECoursesOP•12/31/23, 3:15 PM

that was for kaggle users

Furkan Gözükara SECoursesOP•12/31/23, 3:17 PM

i am testing 2100 and 4200 steps on kaggle

Furkan Gözükara SECoursesOP•12/31/23, 3:17 PM

since batch size is 2 by default

FFurkan Gözükara SECourses yes

DJRich•12/31/23, 3:19 PM

Doing another test now. Keeping the 90 images and following your advice 4590 latents to cache. It's going to take 30 minutes just to do that part and likely another 4 hours to do the training since it is showing to be around 9000 steps

Mmikemenders Can you share your training script?

FiveSAI•12/31/23, 4:02 PM

I finetuned my own base model with 10000 images, i train on that. Dreambooth, same config from his patreon. Except i use blip captions, 40 repeats, and noise ofset

FiveSAI•12/31/23, 4:04 PM

But you can get similar results training on cyberrealistic

mikemenders•12/31/23, 4:58 PM

10000 images with captions... Nice! A huge job.

FiveSAI•12/31/23, 5:12 PM

Yep i trained my own base model, I probably will release it soon when its done, there are still some tweaking needed. I still have another 5-6k images left to train. I captioned those images with llava.

then i train my own models, celebs etc on it. Comes out quite nice

DJRich•12/31/23, 5:54 PM

Has anyone seen this before? I have a laptop with a 4090 and 16GB of VRAM. Out of the blue I am starting to get out of memory errors during a hires fix in A1111 where the upscale is only 1300x1700 approximately. The same images worked a couple of days ago without issue. Not running any updated versions of any software and even switched from sdp to xformers because that seemed to help upscale some images that were failing. According to Windows 11 my GPU is using 8.4GB of VRAM just starting A1111. I've tried disabling all extensions and no change. Without A1111 running VRAM is 0.3/16GB.

FFiveSAI Yep i trained my own base model, I probably will release it soon when its done, ...

Lorelai AI•12/31/23, 6:07 PM

do you have any good resources on llava I have been looking at decent captioning. Blip I find not that great nor is git tremendously accurate

DDJRich Has anyone seen this before? I have a laptop with a 4090 and 16GB of VRAM. Out o...

Furkan Gözükara SECoursesOP•12/31/23, 6:10 PM

interesting

LLorelai AI do you have any good resources on llava I have been looking at decent captioning...

Furkan Gözükara SECoursesOP•12/31/23, 6:10 PM

i have auto llava installer

Furkan Gözükara SECoursesOP•12/31/23, 6:10 PM

batch caption as well

Furkan Gözükara SECoursesOP•12/31/23, 6:10 PM

https://www.patreon.com/posts/sota-very-best-90744385

Patreon

SOTA Image Captioning Scripts For Stable Diffusion: LLaVA, BLIP-2, ...

Patreon is empowering a new generation of creators.
Support and engage with artists and creators as they live out their passions!

LLorelai AI do you have any good resources on llava I have been looking at decent captioning...

FiveSAI•12/31/23, 6:10 PM

I used the same one he posted on his Patreon for the auto installer. I had used a previous one before.

I only use llava for fine tuning base models.

I like blip for training people. With an instance prompt

FFurkan Gözükara SECourses i have auto llava installer

Lorelai AI•12/31/23, 6:10 PM

Awesome

FiveSAI•12/31/23, 6:11 PM

If you caption wrong you’ll get bad results

FFiveSAI I used the same one he posted on his Patreon for the auto installer. I had used ...

Lorelai AI•12/31/23, 6:11 PM

I don't understand blip entirely but I found that the higher the beams the worse it got

FiveSAI•12/31/23, 6:12 PM

I usually edit them. Yeah I stopped using higher beams, I just use 1 now. And I change the other setting from 0.9 to 0.7

FFurkan Gözükara SECourses interesting

DJRich•12/31/23, 6:12 PM

I just installed the A1111 package from StabilityMatrix with all of the same extensions and it only takes up 4GB of VRAM.

FiveSAI•12/31/23, 6:13 PM

Things I’ll usually caption is just my instance prompt, their hair style, and length, and background.

but usually gets overtrained quickly

Similar Threads