Software Engineering Courses (SECourses)•3y ago

@Dr. Furkan Gözükara sorry if I bother you, but can I ask you if we can do what is in this video wit

@Dr. Furkan Gözükara sorry if I bother you, but can I ask you if we can do what is in this video without your patreon?
https://www.youtube.com/watch?v=16-b1AjvyBE

YouTubeSECourses

How To Do Stable Diffusion XL (SDXL) DreamBooth Training For Free -...

Master Stable Diffusion XL Training on Kaggle for Free!

Welcome to this comprehensive tutorial where I'll be guiding you through the exciting world of setting up and training Stable Diffusion XL (SDXL) with Kohya on a free Kaggle account. This video is your one-stop resource for learning everything from initiating a Kaggle session with dual ...

VVexmachina So I'm using all the latest instructions and settings, and I'm running training ...

Furkan Gözükara SECourses•12/23/23, 12:41 AM

you have 9600 steps

Furkan Gözükara SECourses•12/23/23, 12:41 AM

the training speed is accurate

MMax @Dr. Furkan Gözükara sorry if I bother you, but can I ask you if we can do what ...

Furkan Gözükara SECourses•12/23/23, 12:42 AM

you need to figure out hard to do

Furkan Gözükara SECourses•12/23/23, 12:42 AM

i can give you hint. you need to tunnel gui

Furkan Gözükara SECourses•12/23/23, 12:42 AM

or without gui you can just execute the command

VVexmachina This is with 15 example images, scaled to 1024x1024, and using the reg set of me...

Furkan Gözükara SECourses•12/23/23, 12:42 AM

if you have only 15 images

Furkan Gözükara SECourses•12/23/23, 12:42 AM

your repeat or epoch too many

FFurkan Gözükara SECourses or without gui you can just execute the command

MaxOP•12/23/23, 12:44 AM

so it would be easier with patreon?

Digital [Starburst]•12/23/23, 12:45 AM

Fooocus has Python 3.10.9 built-in to the install.

PPflanzen die Tanzen In the vast expanse of the whimsical world of Widdendream, where every creature ...

Furkan Gözükara SECourses•12/23/23, 12:46 AM

nice

DDigital [Starburst]Fooocus has Python 3.10.9 built-in to the install.

Furkan Gözükara SECourses•12/23/23, 12:46 AM

it has

Furkan Gözükara SECourses•12/23/23, 12:46 AM

i have auto installer too

MMax so it would be easier with patreon?

Furkan Gözükara SECourses•12/23/23, 12:46 AM

100%

Digital [Starburst]•12/23/23, 12:47 AM

I found an auto installer on Patreon for Kaggle and Runpod, but try as I might I couldn't find the one for straight-up Windows.

Digital [Starburst]•12/23/23, 12:48 AM

I know you made one because I have it installed on my other machine.

FFurkan Gözükara SECourses 100%

MaxOP•12/23/23, 12:48 AM

An information. With the $5 patreon, do I unlock all content, including what interests me?

MMax An information. With the $5 patreon, do I unlock all content, including what int...

Digital [Starburst]•12/23/23, 12:49 AM

I believe you do and it is well worth the investment.

Digital [Starburst]•12/23/23, 12:49 AM

He is releasing new scripts regularly as well for different purposes.

Digital [Starburst]•12/23/23, 12:50 AM

You can see a list of all his Patreon content on his Github repository.

Digital [Starburst]•12/23/23, 12:50 AM

https://github.com/FurkanGozukara/Stable-Diffusion/blob/main/Patreon-Posts-Index.md

GitHub

Stable-Diffusion/Patreon-Posts-Index.md at main · FurkanGozukara/St...

Stable Diffusion, SDXL, LoRA Training, DreamBooth Training, Automatic1111 Web UI, DeepFake, Deep Fakes, TTS, Animation, Text To Video, Tutorials, Guides, Lectures, Courses, ComfyUI, Google Colab, R...

Digital [Starburst]•12/23/23, 12:51 AM

Every post is updated anytime it is necessary and posted here on Discord to let us know we need to update.

Furkan Gözükara SECourses•12/23/23, 12:53 AM

new amazing script

Furkan Gözükara SECourses•12/23/23, 12:53 AM

https://www.patreon.com/posts/95143823?pr=true

Furkan Gözükara SECourses•12/23/23, 12:53 AM

DDigital [Starburst]I believe you do and it is well worth the investment.

MaxOP•12/23/23, 12:54 AM

Of course I will! I just wanted to know if anything changed between the subscription plans

FFurkan Gözükara SECourses your repeat or epoch too many

Vexmachina•12/23/23, 12:58 AM

Ahh, so I should divide the default of 40 repeats by 6 if using 6 GPUs? Will experiment, thank you

Furkan Gözükara SECourses•12/23/23, 1:02 AM

https://twitter.com/GozukaraFurkan/status/1738363318245437890

MMax Of course I will! I just wanted to know if anything changed between the subscrip...

Furkan Gözükara SECourses•12/23/23, 1:03 AM

all plans have access to all scripts

Furkan Gözükara SECourses•12/23/23, 1:03 AM

but if you higher i appreicate

VVexmachina Ahh, so I should divide the default of 40 repeats by 6 if using 6 GPUs? Will exp...

Furkan Gözükara SECourses•12/23/23, 1:03 AM

if you are using 6 gpus

Furkan Gözükara SECourses•12/23/23, 1:03 AM

then you need to divide it to 6

Furkan Gözükara SECourses•12/23/23, 1:03 AM

because batch size will be 6

FFurkan Gözükara SECourses all plans have access to all scripts

MaxOP•12/23/23, 1:06 AM

Okay, Just signed up. Thank you so much!

MMax Okay, Just signed up. Thank you so much!

Furkan Gözükara SECourses•12/23/23, 1:07 AM

awesome welcome

AAntek Looking for the best workflow for a apparell mockup task: I have a t-shirt desi...

Furkan Gözükara SECourses•12/23/23, 1:07 AM

this is really hard

Sstillmotion Anyone messed around with training TI embeddings? Trying to figure out why my TI...

Furkan Gözükara SECourses•12/23/23, 1:14 AM

i am yet to investigate them further

stillmotion•12/23/23, 1:15 AM

Just found the issue, looks related to the scheduler https://github.com/huggingface/diffusers/issues/5433

GitHub

Visual artifacts when using `DPM++` schedulers and SDXL without the...

Describe the bug All DPM++ schedulers are showing visual artifacts out of the base model when denoising_end=1 (skipping the refiner). This effect is most notable with DPM++ 2M SDE configured using ...

Furkan Gözükara SECourses•12/23/23, 1:18 AM

these fooocus styles are amazing :d

VVexmachina Hmm, could this be related? It seems like with 6 processes, batch size of 2 and ...

Furkan Gözükara SECourses•12/23/23, 1:19 AM

the kohya GUI doesnt show multi GPU. so when you use 6 GPUs your actual batch size is being 6. it duplicates training on all

Furkan Gözükara SECourses•12/23/23, 1:19 AM

so you need to setup your number of epochs or repeating accordingly

FFurkan Gözükara SECourses the kohya GUI doesnt show multi GPU. so when you use 6 GPUs your actual batch si...

Vexmachina•12/23/23, 1:38 AM

I stopped using the kohya GUI because I couldn't see how to enable multi-gpu, so I've just been using a command line argument. This is the argument:

accelerate launch --num_cpu_threads_per_process=16 --num_processes=6 --multi_gpu --num_machines=1 --gpu_ids=0,1,2,3,4,5 "./sdxl_train.py" --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" --train_data_dir="training/img" --reg_data_dir="training/reg" --resolution="1024,1024" --output_dir="training/model" --logging_dir="training/log" --save_model_as=safetensors --full_bf16 --vae="stabilityai/sdxl-vae" --output_name="TESTsuperxl" --lr_scheduler_num_cycles="8" --max_data_loader_n_workers="0" --learning_rate_te1="3e-06" --learning_rate_te2="0.0" --learning_rate="1e-05" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="9600" --save_every_n_epochs="1" --mixed_precision="bf16" --save_precision="bf16" --cache_latents --cache_latents_to_disk --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False weight_decay=0.01 --max_data_loader_n_workers="0" --bucket_reso_steps=64 --gradient_checkpointing --bucket_no_upscale --noise_offset=0.0 --max_grad_norm=0.0 --no_half_vae --train_text_encoder

accelerate launch --num_cpu_threads_per_process=16 --num_processes=6 --multi_gpu --num_machines=1 --gpu_ids=0,1,2,3,4,5 "./sdxl_train.py" --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" --train_data_dir="training/img" --reg_data_dir="training/reg" --resolution="1024,1024" --output_dir="training/model" --logging_dir="training/log" --save_model_as=safetensors --full_bf16 --vae="stabilityai/sdxl-vae" --output_name="TESTsuperxl" --lr_scheduler_num_cycles="8" --max_data_loader_n_workers="0" --learning_rate_te1="3e-06" --learning_rate_te2="0.0" --learning_rate="1e-05" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="9600" --save_every_n_epochs="1" --mixed_precision="bf16" --save_precision="bf16" --cache_latents --cache_latents_to_disk --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False weight_decay=0.01 --max_data_loader_n_workers="0" --bucket_reso_steps=64 --gradient_checkpointing --bucket_no_upscale --noise_offset=0.0 --max_grad_norm=0.0 --no_half_vae --train_text_encoder

I've tried this with a batch size of 1 and 2. 2 is actually slower than 1, and anything higher than 2 gives me an out of memory error.

I've gone ahead and tried doing a repeat of 40 on the original images, and now 7 (close enough to 40 / 6).

Changing that didn't change the total number of optimization steps, which was always 9600, or the time to complete training. I think it just wanted to do more epochs.

After changing the repeats to 7 and removing the flag for --max_train_steps="9600"--max_train_steps="9600" Now it's trying to do 46 epochs for a total of 1600 optimizations steps. It still says it's going to take over 4 hours.

I tried using the --ddp_gradient_as_bucket_view--ddp_gradient_as_bucket_view flag as specified in the updated sd-scripts repo, but that made it 5-6x slower.

Vexmachina•12/23/23, 1:39 AM

After all this, using just a single 4090 with your default settings gives me a training time of 3 hours @ 8 epochs and 9600 steps versus 4+h on 6 4090s.

Sorry for being a mess! I just don't understand what I'm doing wrong to mess up multi-gpu training so badly. Thank you for all that you're doing to help us though!

VVexmachina I stopped using the kohya GUI because I couldn't see how to enable multi-gpu, so...

Furkan Gözükara SECourses•12/23/23, 1:44 AM

kohya gui only prepares this command

Furkan Gözükara SECourses•12/23/23, 1:44 AM

accelerate is the one that does distributed training

Furkan Gözükara SECourses•12/23/23, 1:44 AM

i can give you private consultation if you need

MaxOP•12/23/23, 2:18 AM

Anyway, I'm following the tutorial. It's amazing!

AIGambino•12/23/23, 3:25 AM

How come img2img doesn't work?

AIGambino•12/23/23, 3:25 AM

it just barely changes the color and makes my character ruffled up

AIGambino•12/23/23, 3:25 AM

txt2img works fine for a realistic mouse

shortdwarf•12/23/23, 4:53 AM

@Dr. Furkan Gözükara sorry for the delay. Heres screenshots of my adetailer settings. I have also used the photo of ohwx man prompt. Also could DM photos if you want to see examples of with/without adetailer

accelerate launch --num_cpu_threads_per_process=16 --num_processes=6 --multi_gpu --num_machines=1 --gpu_ids=0,1,2,3,4,5 "./sdxl_train.py" --pretrained_model_name_or_path="stabilityai/stable-diffusion-xl-base-1.0" --train_data_dir="training/img" --reg_data_dir="training/reg" --resolution="1024,1024" --output_dir="training/model" --logging_dir="training/log" --save_model_as=safetensors --full_bf16 --vae="stabilityai/sdxl-vae" --output_name="TESTsuperxl" --lr_scheduler_num_cycles="8" --max_data_loader_n_workers="0" --learning_rate_te1="3e-06" --learning_rate_te2="0.0" --learning_rate="1e-05" --lr_scheduler="constant" --train_batch_size="1" --max_train_steps="9600" --save_every_n_epochs="1" --mixed_precision="bf16" --save_precision="bf16" --cache_latents --cache_latents_to_disk --optimizer_type="Adafactor" --optimizer_args scale_parameter=False relative_step=False warmup_init=False weight_decay=0.01 --max_data_loader_n_workers="0" --bucket_reso_steps=64 --gradient_checkpointing --bucket_no_upscale --noise_offset=0.0 --max_grad_norm=0.0 --no_half_vae --train_text_encoder

@Dr. Furkan Gözükara sorry if I bother you, but can I ask you if we can do what is in this video wit

Similar Threads