Software Engineering Courses (SECourses)•3y ago

hello Furkan, I only use a few images, about 10 approx, see, out of memory,

JM•10/7/23, 3:22 AM

@Dr. Furkan Gözükara I want to know how to train. lora use multi gpu in Ubuntu ? How to set accelerate config ?

virginiaOP•10/7/23, 3:53 AM

morning all, possible to make models or sth under temp folder to gain more ram, for the training notebook? i have tried two notebooks, only 10 pics, all failed ./sdxl_train_network.py FAILED
-----------------------------------------------------
Failures:
<NO_OTHER_FAILURES>
-----------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-10-07_02:51:09
host : 08833fa205ae
rank : 0 (local_rank: 0)
exitcode : -9 (pid: 1097)
error_file: <N/A>
traceback : Signal 9 (SIGKILL) received by PID 1097 i tried, two notebooks, possible to make everything in temp folders? otherwise it is too big, out of ram

virginiaOP•10/7/23, 3:53 AM

accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" --train_data_dir="/kaggle/working/results/img" --resolution="1024,1024" --output_dir="/kaggle/working/results/model" --logging_dir="/kaggle/working/results/log" --network_alpha="1" --save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=0.0004 --unet_lr=0.0004 --network_dim=32 --output_name="kaggle_flower" --lr_scheduler_num_cycles="8" --no_half_vae --learning_rate="0.0004" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="1000" --save_every_n_epochs="1" --mixed_precision="fp16" --save_precision="fp16" --cache_latents --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False --max_data_loader_n_workers="0" --bucket_reso_steps=64 --gradient_checkpointing --xformers --bucket_no_upscale --noise_offset=0.0 --lowram, are these settings too big?

killermoth•10/7/23, 4:04 AM

FFurkan Gözükara SECourses i will also research this for a company

Benjamin•10/7/23, 8:07 AM

will you share those results in patreon?

Vvirginia accelerate launch --num_cpu_threads_per_process=2 "./sdxl_train_network.py" --pr...

Deleted User•10/7/23, 9:59 AM

I have the solution for this

virginiaOP•10/7/23, 10:36 AM

hello Ihsas, good morning

virginiaOP•10/7/23, 10:36 AM

does it mean my settings are too big?

virginiaOP•10/7/23, 10:50 AM

because of no half vae?

virginiaOP•10/7/23, 10:50 AM

if so, i do not notice it

Stefanoid9343•10/7/23, 11:39 AM

@Dr. Furkan Gözükara Good morning I follow step by step the lora training that you prepare on youtube until the end I have 17 images and I use the regularization woman images but when I try to use the lora nothing happen on the generated images the 17 images are pictures of a woman

Stefanoid9343•10/7/23, 11:40 AM

any advice? with dreamboth is a better way to train? what is the difference

Zono50•10/7/23, 12:37 PM

anyone got a good starting point for me getting a runpod so i can train an sdxl lora of my face?

JM•10/7/23, 12:39 PM

Who knows how to do it?Make the picture clear without losing details?

land23•10/7/23, 2:42 PM

Anyone been able to do a face swap and also swap the skin color??

EEyeSpyBekoAI Interesting. Thanks for the heads up I'll give it a shot.

mikemenders•10/7/23, 3:08 PM

This could also be an interesting solution, although I haven't played with it much: https://www.reddit.com/r/StableDiffusion/comments/16vkhrt/how_to_use_ipadapter_controlnets_for_consistent/
It also says here that you should use the IP adapter + face swap (which is ReActor). This way you can create more accurate real reproductions.:
https://www.reddit.com/r/StableDiffusion/comments/16vkhrt/comment/k2tuyfb/?utm_source=reddit&utm_medium=web2x&context=3

From the StableDiffusion community on Reddit: How to use IP-adapter...

Explore this post and more from the StableDiffusion community

CeraRalaz's comment on "How to use IP-adapter controlnets for consi...

Explore this conversation and more from the StableDiffusion community

land23•10/7/23, 3:17 PM

man it's quite annoying the majority of tutorials, including those here, are done with Automatic1111 GUI instead of with simple jupyter notebooks

GrayFox•10/7/23, 3:17 PM

Hi guys, I encountered an error when trying to train with optimizer adam8bit, does anyone have any suggestions how to solve it? Have kohya 21.8.9, Python 30.10.9

error_2.txt11.66KB

Mmikemenders This could also be an interesting solution, although I haven't played with it mu...

EyeSpyBekoAI•10/7/23, 5:32 PM

Hey @mikemenders , thanks very much for sharing! I'm going to try to test these different options out this week. On that note, it's actually really interesting. Adetailer can be really good if your prompts are a bit simpler, and with certain shape faces and skin tones. If you look at some of my pictures in the "strictly-sfw-image-share channel", with me, it worked very well in most cases. But the more complicated prompts get, or if the person trained has certain features, it starts faltering. Still not terrible, it's definitely better having it than not, but you do start seeing some cracks. I'd be curious to see if the other solutions solve for that.

Vvirginia hello Furkan, I only use a few images, about 10 approx, see, out of memory,

Furkan Gözükara SECourses•10/7/23, 5:52 PM

i replied you from patreon

Furkan Gözükara SECourses•10/7/23, 5:52 PM

you are missing full fp16

JJM @Dr. Furkan Gözükara I want to know how to train. lora use multi gpu in Ubuntu ?...

Furkan Gözükara SECourses•10/7/23, 5:52 PM

you can use my kaggle video install settings

Furkan Gözükara SECourses•10/7/23, 5:52 PM

they auto use

Furkan Gözükara SECourses•10/7/23, 5:53 PM

https://youtu.be/JF2P7BIUpIU

YouTubeSECourses

How to Do SDXL Training For FREE with Kohya LoRA - Kaggle - NO GPU ...

If you don't have a strong GPU for Stable Diffusion XL training then this is the tutorial you are looking for. We will use Kaggle free notebook to do Kohya SDXL LoRA training. In this tutorial you will master Kohya SDXL with Kaggle!

Curious about training Kohya SDXL? Learn why Kaggle outshines Google Colab! We will uncover the power of free Ka...

BBenjamin will you share those results in patreon?

Furkan Gözükara SECourses•10/7/23, 5:53 PM

i am sharing on Twitter the results. batch size didn't reduce quality much really good. but uses more vram

Furkan Gözükara SECourses•10/7/23, 5:53 PM

I also updated best workflow files on Patreon for dreambooth : https://www.patreon.com/posts/89213064

Vvirginia morning all, possible to make models or sth under temp folder to gain more ram, ...

Furkan Gözükara SECourses•10/7/23, 5:55 PM

as i replied from patreon you are missing full fp16

SStefanoid9343 @Dr. Furkan Gözükara Good morning I follow step by step the lora training that y...

Furkan Gözükara SECourses•10/7/23, 5:56 PM

dreambooth much better

Furkan Gözükara SECourses•10/7/23, 5:56 PM

what gpu do you have?

Furkan Gözükara SECourses•10/7/23, 5:56 PM

here very best json files updated : https://www.patreon.com/posts/89213064

ZZono50 anyone got a good starting point for me getting a runpod so i can train an sdxl ...

Furkan Gözükara SECourses•10/7/23, 5:58 PM

https://youtu.be/-xEwaQ54DI4

YouTubeSECourses

How To Do SDXL LoRA Training On RunPod With Kohya SS GUI Trainer & ...

If you don't have a GPU, or have a strong GPU, or you are using Mac and your computer not supporting Stable Diffusion training, SDXL training, then this is the tutorial you are looking for. In this tutorial, we will use a cheap cloud GPU service provider RunPod to use both Stable Diffusion Web UI Automatic1111 and Stable Diffusion trainer Kohya ...

JJM Who knows how to do it?Make the picture clear without losing details?

Furkan Gözükara SECourses•10/7/23, 5:58 PM

you mean existing picture?

Lland23 man it's quite annoying the majority of tutorials, including those here, are don...

Furkan Gözükara SECourses•10/7/23, 5:59 PM

jupyter notebooks are harder

GGrayFox Hi guys, I encountered an error when trying to train with optimizer adam8bit, do...

Furkan Gözükara SECourses•10/7/23, 6:00 PM

disable caching caching Text Encoder outputs

Furkan Gözükara SECourses•10/7/23, 6:00 PM

https://twitter.com/GozukaraFurkan/status/1710715982975508708

FFurkan Gözükara SECourses you mean existing picture?

JM•10/7/23, 6:04 PM

No I typed wrong , I mean, how can I adjust the details of the enhanced image to make it sharper? Without losing details Refers to my face shape used for training

EEyeSpyBekoAI Hey @mikemenders , thanks very much for sharing! I'm going to try to test these ...

mikemenders•10/7/23, 6:04 PM

I liked the hires fix better. For the face trained loras, it definitely worked better with the hires fix than with adetailer. I tried the other adetailer models, but most of the time they made the original result worse. True, I just turned it on by default, not trying to get a better result, but I would think it would work fine by default. But I'll check later to see if I really need a pilot test.

JJM No I typed wrong , I mean, how can I adjust the details of the enhanced image to...

Furkan Gözükara SECourses•10/7/23, 6:09 PM

i dont know

Mmikemenders I liked the hires fix better. For the face trained loras, it definitely worked b...

EyeSpyBekoAI•10/7/23, 6:10 PM

You know I forgot all about hiresfix since I switched to SDXL. Gonna try that too.

Stefanoid9343•10/7/23, 6:16 PM

@Dr. Furkan Gözükara hi, I trying to do a training in dreamboth using runpod when I try to put the path of the source directory like "/workspace/768x1024" where my images to be trained is supposed to be I and the click on preprocess
I have an error
---
Error completing request

files = listfiles(src)
File "/workspace/stable-diffusion-webui/modules/textual_inversion/preprocess.py", line 30, in listfiles
return os.listdir(dirname)
FileNotFoundError: [Errno 2] No such file or directory: '/workspace/768x1024'

---