Software Engineering Courses (SECourses)•2y ago

7950x3d+4090+64gb ddr5 .. for SDXL i use your config dreambooth 24GB_NoTextEncoder , with 60 Trainin

7950x3d+4090+64gb ddr5 .. for SDXL i use your config dreambooth 24GB_NoTextEncoder , with 60 Training images my face 40 repeats ,your Regularisation images 1024x1024_2734_imgs 1 Repeats .....9 hours of training and good result of high quality , kohya-ss v22.4.1 without any error messages

thawadiooOP•1/19/24, 11:20 PM

1.5 Based Models Training Config me_sd_7e7_v2.json same 60 Training images of my face 768x768 ,40 repeats and Regularisation images man_5200_imgs_768x768 1 Repeats = 5 hours of bad result getting all same picture

thawadiooOP•1/19/24, 11:21 PM

what im doing wrong ?

Tthawadioo 7950x3d+4090+64gb ddr5 .. for SDXL i use your config dreambooth 24GB_NoTextEncod...

Furkan Gözükara SECourses•1/19/24, 11:46 PM

60 train 40 repeat

Furkan Gözükara SECourses•1/19/24, 11:46 PM

how many epochs?

Furkan Gözükara SECourses•1/19/24, 11:46 PM

it makes 4800 steps with reg images

Furkan Gözükara SECourses•1/19/24, 11:46 PM

if 1 epoch

thawadiooOP•1/19/24, 11:47 PM

8 , i didnt change anything just put the images folders

Tthawadioo 8 , i didnt change anything just put the images folders

Furkan Gözükara SECourses•1/19/24, 11:47 PM

ok 8 is too many

Furkan Gözükara SECourses•1/19/24, 11:47 PM

what is your speed how many second / it

thawadiooOP•1/19/24, 11:47 PM

1.25

Furkan Gözükara SECourses•1/19/24, 11:47 PM

that is normal speed

Furkan Gözükara SECourses•1/19/24, 11:47 PM

so you need to reduce number of epochs

Furkan Gözükara SECourses•1/19/24, 11:47 PM

i suggest this

Furkan Gözükara SECourses•1/19/24, 11:47 PM

100 repeat

Furkan Gözükara SECourses•1/19/24, 11:47 PM

1 epoch

Furkan Gözükara SECourses•1/19/24, 11:47 PM

and save checkpoints

Furkan Gözükara SECourses•1/19/24, 11:48 PM

100 * 60 *2 = 12000 steps

Furkan Gözükara SECourses•1/19/24, 11:48 PM

so save every 1201 steps

Furkan Gözükara SECourses•1/19/24, 11:48 PM

to see which one will perform best

thawadiooOP•1/19/24, 11:48 PM

OK thank you i will test it

thawadiooOP•1/20/24, 12:16 AM

train unet: True, text_encoder1: False, text_encoder2: False
number of models: 1
number of trainable parameters: 2567463684
prepare optimizer, data loader etc.
use Adafactor optimizer | {'scale_parameter': False, 'relative_step': False, 'warmup_init': False, 'weight_decay': 0.01}
constant_with_warmup will be good / スケジューラはconstant_with_warmupが良いかもしれません
enable full bf16 training.
running training / 学習開始
num examples / サンプル数: 6000
num batches per epoch / 1epochのバッチ数: 12000
num epochs / epoch数: 1
batch size per device / バッチサイズ: 1
gradient accumulation steps / 勾配を合計するステップ数 = 1
total optimization steps / 学習ステップ数: 12000
steps: 0%| | 0/12000 [00:00<?, ?it/s]
epoch 1/1
steps: 3%|█▊ | 416/12000 [07:49<3:38:06, 1.13s/it, avr_loss=0.105]

thawadiooOP•1/20/24, 12:19 AM

100% gpu utilization ,dedicated gpu memory 17.2/24.0 GB , 20.6 gb ram in use

thawadiooOP•1/20/24, 12:28 AM

i turn off Gradient checkpointing ,

thawadiooOP•1/20/24, 12:28 AM

steps: 0%| | 20/12000 [06:13<62:08:29, 18.67s/it, avr_loss=0.129]

Tthawadioo i turn off Gradient checkpointing ,

Furkan Gözükara SECourses•1/20/24, 12:38 PM

it starts using shared VRAM

Furkan Gözükara SECourses•1/20/24, 12:38 PM

therefore it will become too slow

Furkan Gözükara SECourses•1/20/24, 12:39 PM

OneTrainer is fitting into 24 GB so it brings speed up there

Dazzastrous [4090]•1/21/24, 12:35 PM

https://github.com/chaojie/ComfyUI-Moore-AnimateAnyone

GitHub

GitHub - chaojie/ComfyUI-Moore-AnimateAnyone

Contribute to chaojie/ComfyUI-Moore-AnimateAnyone development by creating an account on GitHub.

DDazzastrous [4090]https://github.com/chaojie/ComfyUI-Moore-AnimateAnyone

Furkan Gözükara SECourses•1/21/24, 5:36 PM

thanks

bert•1/21/24, 5:37 PM

I want to create a space on my mac where i can install and run packages and experiment around but I should be able to delete this space/folder and after that none of what i did or installed during my experimentation should remain on my pc. how can I do this?

Bbert I want to create a space on my mac where i can install and run packages and expe...

Furkan Gözükara SECourses•1/21/24, 5:38 PM

sadly i dont know mac. but best would be installing a windows or linux on mac if possible

bert•1/21/24, 5:39 PM

does docker do this kind of thing

Bbert does docker do this kind of thing

Furkan Gözükara SECourses•1/21/24, 5:46 PM

are you able to run dockers on mac?

Furkan Gözükara SECourses•1/21/24, 5:46 PM

yes it would do

Furkan Gözükara SECourses•1/21/24, 5:47 PM

but it is important if docker will be able to use your bare metal hardware

ciphermonk•1/22/24, 7:41 PM

@Furkan Gözükara SECourses have you done a tutorial on stable video diffusion

ciphermonk•1/22/24, 7:42 PM

also comfi.ui?

Cciphermonk <@205854764540362752> have you done a tutorial on stable video diffusion

Furkan Gözükara SECourses•1/22/24, 11:11 PM

not yet but will do hopefully

Furkan Gözükara SECourses•1/22/24, 11:11 PM

so many stuff recently

Partha•1/25/24, 3:56 PM

Is anyone online?

PPartha Is anyone online?

Furkan Gözükara SECourses•1/25/24, 6:24 PM

yes

Dazzastrous [4090]•1/26/24, 1:58 PM

https://huggingface.co/spaces/HumanAIGC/OutfitAnyone This shows much promise impressive

OutfitAnyone - a Hugging Face Space by HumanAIGC

DDazzastrous [4090]https://huggingface.co/spaces/HumanAIGC/OutfitAnyone This shows much promise i...

Furkan Gözükara SECourses•1/26/24, 10:24 PM

i agree

DDazzastrous [4090]https://huggingface.co/spaces/HumanAIGC/OutfitAnyone This shows much promise i...

EyeSpyBekoAI•1/27/24, 7:17 PM

very interesting. Is it stable diffusion based? The fact that they only allow the preset model images to be put in there makes me think there is a pre-prepared controlnet going on but it is very smooth none the less. Thanks for sharing.

EEyeSpyBekoAI very interesting. Is it stable diffusion based? The fact that they only allow th...

Dazzastrous [4090]•1/29/24, 2:45 PM

Original message was deleted

Sipriyani•1/30/24, 12:42 AM

im from NJ. what exactly you need?

Tthawadioo OK thank you i will test it 🌹

Chip•1/30/24, 3:40 AM

did you test it? What was your result?

Jillani SoftTech•2/1/24, 3:31 AM

Hey Everyone today I am recently sharing my new updated blog post please check and share your feedback.
https://jillanisofttech.medium.com/redefining-excellence-in-mlops-a-guide-to-optimal-machine-learning-experimentation-environments-0c4ab61905d1

ai

Medium

Redefining Excellence in MLOps: A Guide to Optimal Machine Learning...

Muhammad Ghulam Jillani(Jillani SoftTech), Senior Data Scientist and Machine Learning Engineer

JJillani SoftTech Hey Everyone today I am recently sharing my new updated blog post please check a...

Furkan Gözükara SECourses•2/1/24, 10:53 AM

nice thanks

Evan•2/2/24, 3:14 AM

Hey everyone! After doing a lot of work in the generative AI space I want to start investing in training an LLM to process input - like a chatbot as an example

7950x3d+4090+64gb ddr5 .. for SDXL i use your config dreambooth 24GB_NoTextEncoder , with 60 Trainin

Similar Threads