Software Engineering Courses (SECourses)•2y ago

do i need to change "dataloader threads?" i don't see it in your video

SSandstormleader do i need to change "dataloader threads?" i don't see it in your video

Furkan Gözükara SECourses•5/9/24, 11:09 PM

it is caching batch size

Furkan Gözükara SECourses•5/9/24, 11:09 PM

make like 2

SandstormleaderOP•5/9/24, 11:10 PM

thanks

SandstormleaderOP•5/9/24, 11:34 PM

do i really need regulation images?

SandstormleaderOP•5/9/24, 11:34 PM

nvm you just said it's not necessary

SandstormleaderOP•5/9/24, 11:34 PM

haha

Furkan Gözükara SECourses•5/9/24, 11:40 PM

but reg helps

SandstormleaderOP•5/9/24, 11:54 PM

OSError: [WinError 123] The filename, directory name, or volume label syntax is incorrect: 'workspace-cache/runL:'

SandstormleaderOP•5/10/24, 12:03 AM

my guess is that i have to run it from my C: drive

SandstormleaderOP•5/10/24, 12:14 AM

File "C:\1T\OneTrainer\modules\modelLoader\stableDiffusionXL\StableDiffusionXLModelLoader.py", line 232, in load
raise Exception("could not load model: " + model_names.base_model)
Exception: could not load model: C:/auto1111/New folder/wowXLV2_v2.safetensors

SandstormleaderOP•5/10/24, 12:15 AM

it won't load any or my models

xlads•5/10/24, 12:29 AM

@Dr. Furkan Gözükara do you have a link to a list of prompts?

SSandstormleader File "C:\1T\OneTrainer\modules\modelLoader\stableDiffusionXL\StableDiffusionXLMo...

Furkan Gözükara SECourses•5/10/24, 12:51 AM

please report to their discord

Furkan Gözükara SECourses•5/10/24, 12:51 AM

they help

Furkan Gözükara SECourses•5/10/24, 12:51 AM

can be app bug

Xxlads @Dr. Furkan Gözükara do you have a link to a list of prompts?

Furkan Gözükara SECourses•5/10/24, 12:51 AM

ytes

Furkan Gözükara SECourses•5/10/24, 12:52 AM

https://www.patreon.com/posts/amazing-prompt-1-90346033

Patreon

Amazing Prompt List For DreamBooth or LoRA Trained Stable Diffusion...

Get more from SECourses: Tutorials, Guides, Resources, Training, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözükara on Patreon

Furkan Gözükara SECourses•5/10/24, 12:52 AM

you can also see some prompts here : https://civitai.com/user/SECourses/images

SandstormleaderOP•5/10/24, 12:59 AM

do you have a link? @Dr. Furkan Gözükara

SSandstormleader do you have a link? @Dr. Furkan Gözükara

Furkan Gözükara SECourses•5/10/24, 1:44 AM

https://discord.gg/KwgcQd5scF

SandstormleaderOP•5/10/24, 2:54 AM

so is 1.4s/it pretty standard for sdxl 1024 training?

SandstormleaderOP•5/10/24, 2:54 AM

SSandstormleader so is 1.4s/it pretty standard for sdxl 1024 training?

Furkan Gözükara SECourses•5/10/24, 3:52 AM

it is better than mine :d

SandstormleaderOP•5/10/24, 9:41 AM

Oof ok, training just going to take 4x longer then, got it

kareem9000•5/10/24, 10:06 AM

if im trying to put my syelf in a movie poster, is the best way to train an SDXL model on my images and then create LORAs for the movie scenes?

kareem9000•5/10/24, 10:07 AM

like say for e.g the movie is avatar and i wanna put my self on the poster of avatar

Kkareem9000 like say for e.g the movie is avatar and i wanna put my self on the poster of av...

dxqb•5/10/24, 10:20 AM

if you are just trying to make 1 picture, inpainting is the best way

rafraf•5/10/24, 10:20 AM

hey @Dr. Furkan Gözükara i am switching my kohya_ss sd-scripts branch to a PR that passes args to wandb.
I use:
git fetch sd-scripts pull/1285/head:hyperparamsgit fetch sd-scripts pull/1285/head:hyperparams
how do i git pull the sd-scripts above to the remote. I am doing something wrong, when i git pull hyperparamsgit pull hyperparams it pulls the sd-scripts PR to the root of kohya_ss instead of updating sd-scripts with the PR. What would be the correct command?

rafraf•5/10/24, 10:20 AM

I am following Bernard's readme:
https://github.com/bmaltais/kohya_ss/blob/master/examples/pull kohya_ss sd-scripts updates in.md#updating-a-local-branch-with-the-latest-sd-scripts-changes

rafraf•5/10/24, 10:22 AM

Currently my remote:

origin  https://github.com/bmaltais/kohya_ss.git (fetch)
origin  https://github.com/bmaltais/kohya_ss.git (push)
sd-scripts      https://github.com/kohya-ss/sd-scripts.git (fetch)
sd-scripts      https://github.com/kohya-ss/sd-scripts.git (push)

origin  https://github.com/bmaltais/kohya_ss.git (fetch)
origin  https://github.com/bmaltais/kohya_ss.git (push)
sd-scripts      https://github.com/kohya-ss/sd-scripts.git (fetch)
sd-scripts      https://github.com/kohya-ss/sd-scripts.git (push)

rafraf•5/10/24, 10:23 AM

this is the PR i am trying to switch to
https://github.com/kohya-ss/sd-scripts/pull/1285/commits/2c9db5d9f2f6b57f15b9312139d0410ae8ae4f3c#diff-36b34773120418762617b3bcd90be991f9ef2a620e5976a1b7e783b8d2460d09L310-R316

GitHub

Hyperparameter tracking by ccharest93 · Pull Request #1285 · kohya-...

Sensitive parameters are not really DRY since they are repeated in verify_command_line_training_args, but for now only two repeats is not too bad.
The setting of hyperparameters through config is h...

Ddxqb if you are just trying to make 1 picture, inpainting is the best way

kareem9000•5/10/24, 10:43 AM

so i use the checkpoint of myself -> generate a few images and then inpaint them onto the poster?

FFurkan Gözükara SECourses it is better than mine :d

VAMP•5/10/24, 11:19 AM

doc whic do you use? windows 10 or 11? 11 has started to irritate me a bit...buggy drivers and shitty ui updates...

Rrafraf this is the PR i am trying to switch to https://github.com/kohya-ss/sd-scripts/...

dxqb•5/10/24, 12:16 PM

sd-scripts and kohya_ss seem to be separate. go to sd-scripts and pull there?

Sebastianblood•5/10/24, 12:18 PM

I assume something must be wrong but don't know what it is exactly. Am training around 8,000 images on Dreambooth and the duration is around 400 hours. Is that even normal?

SSebastianblood I assume something must be wrong but don't know what it is exactly. Am training ...

dxqb•5/10/24, 12:19 PM

no, depending on what card you have I have seen 0,5 it/s up to 4 it/s. that's a max of about 9 hours for your 8000 images

dxqb•5/10/24, 12:19 PM

how many it/s do you see?

dxqb•5/10/24, 12:19 PM

sorry, I mean s/it above, not it/s

Ddxqb no, depending on what card you have I have seen 0,5 it/s up to 4 it/s. that's a ...

Sebastianblood•5/10/24, 12:20 PM

RTX 4090

dxqb•5/10/24, 12:21 PM

I have done some benchmarks on runpod when I was deciding which card to buy. I think 4090 was 0,5 s/it, == 2 it/s

dxqb•5/10/24, 12:21 PM

so 2 pictures per second

dxqb•5/10/24, 12:24 PM

anybody here has experience with how the loss function during dreambooth training works? I am trying to implement something into kohya_ss...

Ddxqb so 2 pictures per second

Sebastianblood•5/10/24, 12:25 PM

With this math, that means 8000 images are supposed to train for 1.11 hours

Sebastianblood•5/10/24, 12:26 PM

What could possibly be slowing down my training then?

SSebastianblood With this math, that means 8000 images are supposed to train for 1.11 hours

dxqb•5/10/24, 12:26 PM

yes. how many s/it do you see? maybe you have many repetitions, so you are training 8000 images 40 times each

Sebastianblood•5/10/24, 12:29 PM

I have separated the dataset into 14 different folders. 13 of them have 17 repeats each and one has 1 repeat

Ddxqb sd-scripts and kohya_ss seem to be separate. go to sd-scripts and pull there?

rafraf•5/10/24, 12:29 PM

Yes, but I’m setting a PR as-script as remote, so it’s tied in with kohya_ss own git structure

rafraf•5/10/24, 12:29 PM

I guess that could be a hack

SSebastianblood I have separated the dataset into 14 different folders. 13 of them have 17 repea...

dxqb•5/10/24, 12:31 PM

what training software do you use? all that I know display the steps. you can easily see how many steps you have in total. is that 8000?

do i need to change "dataloader threads?" i don't see it in your video

Similar Threads

Similar Threads

Similar Threads