Software Engineering Courses (SECourses)•3y ago

Hello guys. I just got approval for the Sagemaker Studio account and soon after I started I am conti

Hello guys. I just got approval for the Sagemaker Studio account and soon after I started I am continually being logged out. Any help?

Benjamin•10/12/23, 8:25 AM

@Dr. Furkan Gözükara this might also be interesting for precise captioning https://llava-vl.github.io/

LLaVA

Visual Instruction Tuning

Benjamin•10/12/23, 8:27 AM

sigh, just trained via dreambooth extension just to notice that it's broken in it's latest version... https://github.com/d8ahazard/sd_dreambooth_extension/issues/1363

GitHub

[Bug]: AttributeError: 'AttnProcessor2_0' object has no attribute '...

Is there an existing issue for this? I have searched the existing issues and checked the recent builds/commits of both this extension and the webui What happened? Training a standard 768 model the ...

mikemenders•10/12/23, 8:30 AM

How meaningful is the Tensorboard loss value? Because out of curiosity, I just looked at the Lora model in the training set that you described as having the lowest loss, and it doesn't really look like the one I'm training. However, for the similarity and this one it says big loss.

Benjamin•10/12/23, 9:21 AM

are you also planning to experiment with SDXL LoRAs @mikemenders?

MMetaphysix Is there a way to save the huggingface cache on another drive? It takes up all m...

Furkan Gözükara SECourses•10/12/23, 10:29 AM

ye it is taking huge space let me look for

BBenjamin @Dr. Furkan Gözükara this might also be interesting for precise captioning https...

Furkan Gözükara SECourses•10/12/23, 10:30 AM

ye i tried. told them not working on windows. they said they will add windows support

FFurkan Gözükara SECourses ye i tried. told them not working on windows. they said they will add windows su...

Benjamin•10/12/23, 10:30 AM

ah I see

MMetaphysix Is there a way to save the huggingface cache on another drive? It takes up all m...

Furkan Gözükara SECourses•10/12/23, 10:32 AM

https://chat.openai.com/share/8f80a664-b69f-48b3-be22-2b9a56cb4fa6

ChatGPT

TTomTekSol Hello guys. I just got approval for the Sagemaker Studio account and soon after ...

Furkan Gözükara SECourses•10/12/23, 10:33 AM

sadly i dont know

BBenjamin sigh, just trained via dreambooth extension just to notice that it's broken in i...

Furkan Gözükara SECourses•10/12/23, 10:34 AM

https://github.com/d8ahazard/sd_dreambooth_extension/issues/1358

GitHub

[Bug]: train_dreambooth.py", line 1582, in inner_loop not enough va...

Here full log I even did skip install of auto1111 and manually installed Here config file attached test1.txt venv "G:\temp 1\stable-diffusion-webui\venv\Scripts\Python.exe" Python 3.10.11...

FFurkan Gözükara SECourses https://github.com/d8ahazard/sd_dreambooth_extension/issues/1358

Benjamin•10/12/23, 10:35 AM

thanks for the hint

FFurkan Gözükara SECourses https://chat.openai.com/share/8f80a664-b69f-48b3-be22-2b9a56cb4fa6

Metaphysix•10/12/23, 11:53 AM

chatGDP useless as always, but I found the correct solution: https://huggingface.co/docs/huggingface_hub/guides/manage-cache#manage-huggingfacehub-cache-system
HUGGINGFACE_HUB_CACHE environment variable points to the new dir

Manage huggingface_hub cache-system

BBenjamin are you also planning to experiment with SDXL LoRAs @mikemenders?

mikemenders•10/12/23, 11:53 AM

Not for the moment, because it would require a more powerful machine or an online machine. I try a lot, so it would cost a lot of money to do the latter. However, I've been experimenting with SD 1.5 and have had some nice results recently. I seem to have the Prodigy's best setup for RealVision models (one of the best SD 1.5 models), so I'm using that to make more Lora models now to see how usable the setup is.

virginia•10/12/23, 12:46 PM

hello Furkan, the notebook does not work

virginia•10/12/23, 12:46 PM

───────────────────── Traceback (most recent call last) ──────────────────────╮
│ /kaggle/working/kohya_ss/kohya_gui.py:13 in <module> │
│ │
│ 12 from library.custom_logging import setup_logging │
│ ❱ 13 from localization_ext import add_javascript │
│ 14 │
╰──────────────────────────────────────────────────────────

virginia•10/12/23, 12:46 PM

i tried several times

land23•10/12/23, 12:48 PM

Hi @Dr. Furkan Gözükara any updates on the SD1.5 Realistic Vision model testing?

virginia•10/12/23, 12:49 PM

@nateland the notebook on my side does not work, you should check it on your side

land23•10/12/23, 12:50 PM

wait link me up to the notebook you're talking about?

land23•10/12/23, 12:50 PM

I'm using my own scripts with @Dr. Furkan Gözükara 's configurations

virginia•10/12/23, 12:50 PM

land23•10/12/23, 12:53 PM

which notebook are you referring to, @virginia

baloooo•10/12/23, 12:54 PM

hey, it looks like everyone is doing Dreambooth full model. I am still on LoRA. It seems that Kohya has updated and the old sdxl.json config does not fit it.

virginia•10/12/23, 12:58 PM

you are right, maybe

baloooo•10/12/23, 1:27 PM

How much difference between LoRA and regular dreambooth?
I read an article about Low Rank ...(areas?) and I decided that this was just the right way of storing weights. I concluded that the amount of info lost is insignificant. So is it worth to do the full model?

Vvirginia ───────────────────── Traceback (most recent call last) ──────────────────────╮ ...

Benjamin•10/12/23, 1:27 PM

yes the kohya ss gui file that @Dr. Furkan Gözükara replaced with his own modifications does not fit anymore to the latest kohya ss version

︻

︻╦╤─꧁⚜ 𝓢𝓲𝓭𝓷𝓮𝔂 ⚜꧂─╤╦︻•10/12/23, 1:52 PM

I tried to run dreambooth locally and it keeps giving me a cuda error. Not memory related. Any way to get that working?

virginia•10/12/23, 3:04 PM

hi, i have a question about dataset, you see, my datasets are like this, i have tried many times to improve image quality, but always get training results like this regardless of 1.5 or sdxl. i do not know the reason, very curious, because of the white and black? because of the white body and white background?

virginia•10/12/23, 3:05 PM

i tried with another image , results regardless of 1.5 and sdlx, are pretty good, like

virginia•10/12/23, 3:06 PM

really puzzled, i have lots troubles with the panda training for a month. because of the dataset white and black?

MMetaphysix chatGDP useless as always, but I found the correct solution: https://huggingface...

Furkan Gözükara SECourses•10/12/23, 5:21 PM

i didnt surprise

Vvirginia hello Furkan, the notebook does not work

Furkan Gözükara SECourses•10/12/23, 5:21 PM

hello fixed with v4 update

Lland23 Hi @Dr. Furkan Gözükara any updates on the SD1.5 Realistic Vision model testing?

Furkan Gözükara SECourses•10/12/23, 5:21 PM

sadly currently focused on sdxl - experimenting captioning

Bbaloooo hey, it looks like everyone is doing Dreambooth full model. I am still on LoRA. ...

Furkan Gözükara SECourses•10/12/23, 5:21 PM

use my shared config on patreon

Furkan Gözükara SECourses•10/12/23, 5:21 PM

the very best

Furkan Gözükara SECourses•10/12/23, 5:21 PM

found with 100+ trainings

Bbaloooo How much difference between LoRA and regular dreambooth? I read an article abou...

Furkan Gözükara SECourses•10/12/23, 5:21 PM

huge

Furkan Gözükara SECourses•10/12/23, 5:21 PM

i am sharing on twitter comparison

BBenjamin yes the kohya ss gui file that @Dr. Furkan Gözükara replaced with his own modifi...

Furkan Gözükara SECourses•10/12/23, 5:22 PM

fixed it

Vvirginia hi, i have a question about dataset, you see, my datasets are like this, i have...

Furkan Gözükara SECourses•10/12/23, 5:22 PM

your dataset is very poor

Furkan Gözükara SECourses•10/12/23, 5:22 PM

make images full size

Furkan Gözükara SECourses•10/12/23, 5:22 PM

1024x1024

Furkan Gözükara SECourses•10/12/23, 5:22 PM

different backgrounds

Furkan Gözükara SECourses•10/12/23, 5:22 PM

or transparent background

Furkan Gözükara SECourses•10/12/23, 5:23 PM

check this out guys : https://twitter.com/GozukaraFurkan/status/1712518063256449224

Jas•10/12/23, 5:31 PM

I want to create a style LORA based on a 1.5 model, I tried to follow your advice here to get Koyha working, but I am using Runpod, do you know which file I need to edit to get a Gradiolink to come up?