Software Engineering Courses (SECourses)•3y ago

Personally cosine and polynomial have given me good results. Seems to be a matter of just finding se

Personally cosine and polynomial have given me good results. Seems to be a matter of just finding settings. There's a write up in discussions for dreambooth on GitHub that has a very thorough fleshing out of how to do your settings according to a formula.
https://github.com/d8ahazard/sd_dreambooth_extension/discussions/547

GitHub

[Pre-Final] A large research of the main training parameters *Updat...

NEW UPDATE Updated formulas for calculating frame resolutions. Also updates with permissions were carried out in the table with mathematics. The custom lr_scheduler has been moved to a separate doc...

Uundelirium Personally cosine and polynomial have given me good results. Seems to be a matte...

ashley•5/26/23, 9:50 AM

Thanks, any chance you could share your settings with me?

Furkan Gözükara SECourses•5/26/23, 10:32 AM

newest xformers is compatiable with torch 2.0.1

Uundelirium Personally cosine and polynomial have given me good results. Seems to be a matte...

Furkan Gözükara SECourses•5/26/23, 10:33 AM

i tested a lot in this video : https://youtu.be/sRdtVanSRl4

YouTubeSECourses

Epic Web UI DreamBooth Update - New Best Settings - 10 Stable Diffu...

RunPod: https://bit.ly/RunPodIO. Discord : https://bit.ly/SECoursesDiscord. Web UI DreamBooth got epic update and we tested all new features to find the best settings (10). If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of...

FFurkan Gözükara SECourses i tested a lot in this video : https://youtu.be/sRdtVanSRl4

ashley•5/26/23, 10:35 AM

Buy these are all constant_with_warmup aren't they? and polynomial and linear both work better.

undeliriumOP•5/26/23, 10:41 AM

@ashleyk I would not presume to know better than the good doctor.

I am mobile at this time and cannot access my dreambooth settings. The advantage I like of any scheduler on a curve is that it greatly helps with avoiding over fitting. Lots of ways to preserve the model you train on.

undeliriumOP•5/26/23, 10:41 AM

But I cannot argue with the doctor's results.

FFurkan Gözükara SECourses i tested a lot in this video : https://youtu.be/sRdtVanSRl4

undeliriumOP•5/26/23, 10:42 AM

It's 6am here and bed time, but as I need to get back into training this weekend, ty for the link. I will definitely watch. I've been bookmarking several of your courses for references.

Uundelirium It's 6am here and bed time, but as I need to get back into training this weekend...

ashley•5/26/23, 10:45 AM

This video is awesome, you definitely need to watch it if you have not already done so

undeliriumOP•5/26/23, 10:46 AM

I plan to. His courses are fantastic. This weekend I have a lot of work I need to get done, so I'll be training to his spec and looking forward to the results

undeliriumOP•5/26/23, 10:47 AM

I am really putting my 3090 TI through the ringer! Wish I had a 4090 but I can't be unhappy with what I do have.

ashley•5/26/23, 10:55 AM

Yeah you are fortunate to have your own GPU, I don't have a GPU so use the cloud for training

Diego•5/26/23, 11:13 AM

@Furkan Gözükara SECourses I'm trying to do dreambooth training using the Inkpunk model to get the same style but I'm having a few errors. Using 1 x A100 80GB
24 vCPU 125 GB RAM on Runpod.

I installed dreambooth manually, changed the requirements, but when training I get this error:
AssertionError:
from user code:
File "/workspace/venv/lib/python3.10/site-packages/torch/random.py", line 23, in get_rng_state
return default_generator.get_state()
Set torch._dynamo.config.verbose=True for more information

So I tried installing torch torch 1.13 following the steps in https://www.youtube.com/watch?v=zA4LksIVas8&ab_channel=SECourses

pip3 install torch==1.13.1 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117

pip uninstall xformers

pip install https://huggingface.co/MonsterMMORPG/SECourses/resolve/main/xformers-0.0.19-cp310-cp310-manylinux2014_x86_64.whl

In dreambooth extension folder: pip install -r requirements.txt

Installed VAE file

Restart pod

fuser -k 3000/tcp

But now when I type Python relauncher.py it won't start:( I get this error: TypeError: expected string or bytes-like object

I've stayed awake many hours trying to solve this but can't. Please help

YouTubeSECourses

RunPod Fix For DreamBooth & xFormers - How To Use Automatic1111 Web...

RunPod: https://bit.ly/RunPodIO. Discord : https://bit.ly/SECoursesDiscord. How to install latest version of Automatic1111 Web UI and #DreamBooth on #RunPod. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

Playlist of #StableDiffusi...

Uundelirium I plan to. His courses are fantastic. This weekend I have a lot of work I need t...

Furkan Gözükara SECourses•5/26/23, 11:41 AM

i also have 3090 ti

Furkan Gözükara SECourses•5/26/23, 11:41 AM

purchased 13900k so i will also compare cpu effect

Furkan Gözükara SECourses•5/26/23, 11:41 AM

hopefully a video coming for that too

DDiego @Furkan Gözükara SECourses I'm trying to do dreambooth training using the Inkpu...

Furkan Gözükara SECourses•5/26/23, 11:42 AM

so you get error at Python relauncher.py ?

FFurkan Gözükara SECourses so you get error at Python relauncher.py ?

Diego•5/26/23, 11:50 AM

Yes:/

DDiego Yes:/

Furkan Gözükara SECourses•5/26/23, 12:04 PM

if you are our patreon supporter i can do a fresh install for you

FFurkan Gözükara SECourses if you are our patreon supporter i can do a fresh install for you

Diego•5/26/23, 12:12 PM

Ok I'm patreon supporter now:)

DDiego Ok I'm patreon supporter now:)

Furkan Gözükara SECourses•5/26/23, 12:16 PM

ty please open google remote desktop

Furkan Gözükara SECourses•5/26/23, 12:16 PM

i will quickly install

Furkan Gözükara SECourses•5/26/23, 12:16 PM

also start a fresh runpod instance

felixsanz•5/26/23, 12:34 PM

i train a lot of models, using different dreambooth options and when i want to compare models to get the best one..... it's an impossible task

felixsanz•5/26/23, 12:35 PM

depending on seed, cfg, etc... sometimes a model is better at 3.000 steps, other times at 20.000 steps, other times using filewords, other times not using it...

felixsanz•5/26/23, 12:35 PM

damn it

felixsanz•5/26/23, 12:35 PM

what a waste of time

felixsanz•5/26/23, 12:35 PM

i have like 40 models of the same thing and can't decide which ones to delete

felixsanz•5/26/23, 12:36 PM

i also tried using deepface to get the scores but it also depends on many things!

Furkan Gözükara SECourses•5/26/23, 12:39 PM

what is your ultimate aim?

felixsanz•5/26/23, 12:40 PM

my results are not even consistent, even if i'm using the deterministic option

FFurkan Gözükara SECourses what is your ultimate aim?

felixsanz•5/26/23, 12:40 PM

trying to get a model or two, one for best txt2img, flexible and good quality, and if needed, another for inpainting with insane quality

felixsanz•5/26/23, 12:42 PM

maybe i shoauldn't be using euler_a to compare? when comparing checkpoints at different steps... they produce very different stuff and it's hard to compare

felixsanz•5/26/23, 12:42 PM

for example model a 13920 steps produces a totally different image at 17400 steps

felixsanz•5/26/23, 12:46 PM

so it's impossible to compare at what epoch is better, because results are so different

Furkan Gözükara SECourses•5/26/23, 2:17 PM

You are using same training dataset?

FFurkan Gözükara SECourses You are using same training dataset?

undeliriumOP•5/26/23, 2:52 PM

Yeah there's no way you're guessing what cpu I have....

undeliriumOP•5/26/23, 2:53 PM

And 64gb more ram in the mail

undeliriumOP•5/26/23, 3:00 PM

(totally replied to the wrong sentence, oops)

FFurkan Gözükara SECourses You are using same training dataset?

felixsanz•5/26/23, 3:05 PM

yes

Uundelirium Yeah there's no way you're guessing what cpu I have....

Furkan Gözükara SECourses•5/26/23, 4:40 PM

nice

Ffelixsanz yes

Furkan Gözükara SECourses•5/26/23, 4:40 PM

with deterministic it is suppsed to be same but as you say it is not

Furkan Gözükara SECourses•5/26/23, 4:41 PM

there are random parts of training and i think they cant make it deterministic

Furkan Gözükara SECourses•5/26/23, 4:41 PM

especially with optimizer if you use

felixsanz•5/26/23, 5:03 PM

it's very hard to compare stuff

felixsanz•5/26/23, 5:03 PM

i tried 10 class images vs 30 vs 50

felixsanz•5/26/23, 5:03 PM

and i can't see a difference

undeliriumOP•5/26/23, 5:54 PM

Class images? If using class images, aren't you supposed to have significantly more? Last time I used class images in training it was smthg like 100 per training image, and then 100 or 150 steps per epoch. Oof I'm a little rusty off the top of my head.

FFurkan Gözükara SECourses ChatGPT has no idead about stable diffusion yet. but i am planning an in depth ...

thebean•5/26/23, 6:12 PM

interestingly it actually does in chatgpt4 since you can give it internet access. i didn't choose to use it in this case though

undeliriumOP•5/26/23, 6:28 PM

ChatGPT 3.5 and earlier only has knowledge up to September 2021.

Uundelirium ChatGPT 3.5 and earlier only has knowledge up to September 2021.

thebean•5/26/23, 9:08 PM

ah yep true

Personally cosine and polynomial have given me good results. Seems to be a matter of just finding se

Similar Threads