Software Engineering Courses (SECourses)•3y ago

still a little new here. I am trying to wrap my head around why epochs are halved if using classific

still a little new here. I am trying to wrap my head around why epochs are halved if using classification images in kohya lora training (shown here https://youtu.be/TpuDOsuKIBo?t=1456) should i just not worry the "why" about it? lol also, im doing a training right now, training myself as the subject. 22 images, did 40 repeats, 10 epochs, so i thought that should equal out to 8,800 steps, but its actually showing a total of 17,600 (double) is that because i have class images?

YouTubeSECourses

Generate Studio Quality Realistic Photos By Kohya LoRA Stable Diffu...

#Kohya SS web GUI DreamBooth #LoRA training full tutorial. You don't need technical knowledge to follow this tutorial. In this tutorial I have explained how to generate professional photo studio quality portrait / self images for free with Stable Diffusion training.

Our Discord server

https://bit.ly/SECoursesDiscord

If I have been of assis...

Generate Studio Quality Realistic Photos By Kohya LoRA Stable Diffu...

Mmccoypauley here's a question: I'm trying to create a lora of a type of humanoid (Disney gar...

Furkan Gözükara SECourses•5/28/23, 11:24 PM

ye you may need to try different classes

HHeistergand Hi. Can I run stable-diffusion with dreambooth on python 3.11 with torch 2.0.1+...

Furkan Gözükara SECourses•5/28/23, 11:24 PM

no python 3.11. use python 3.10

Furkan Gözükara SECourses•5/28/23, 11:24 PM

with 8 gb you can do lora training

EEggo still a little new here. I am trying to wrap my head around why epochs are halve...

Furkan Gözükara SECourses•5/28/23, 11:24 PM

i think they might have fixed it.

Furkan Gözükara SECourses•5/28/23, 11:25 PM

double steps are correct

Furkan Gözükara SECourses•5/28/23, 11:25 PM

that is a fix

FFurkan Gözükara SECourses double steps are correct

EggoOP•5/28/23, 11:29 PM

I want to do no more than 6,000 ish steps for Lora training a face, right? So does that mean 12,000, or 6,000

EEggo I want to do no more than 6,000 ish steps for Lora training a face, right? So do...

Furkan Gözükara SECourses•5/28/23, 11:29 PM

you should aim epoch

Furkan Gözükara SECourses•5/28/23, 11:29 PM

not steps

flowwolf•5/28/23, 11:30 PM

I got kohya scritps working today. very successful with them. I used fp16 instead of bf16

Fflowwolf I got kohya scritps working today. very successful with them. I used fp16 instea...

Yume_Origins•5/28/23, 11:31 PM

i would be glad to see a tutorial about the training on kohya ss guy in case automatic 1111 broke again

Fflowwolf I got kohya scritps working today. very successful with them. I used fp16 instea...

Furkan Gözükara SECourses•5/28/23, 11:33 PM

ye bf16 not available on all cards

YYume_Origins i would be glad to see a tutorial about the training on kohya ss guy in case aut...

Furkan Gözükara SECourses•5/28/23, 11:33 PM

i already have tutorial :d

Furkan Gözükara SECourses•5/28/23, 11:33 PM

this is kohya ss : https://www.youtube.com/watch?v=TpuDOsuKIBo

YouTubeSECourses

Generate Studio Quality Realistic Photos By Kohya LoRA Stable Diffu...

https://bit.ly/SECoursesDiscord

If I have been of assis...

FFurkan Gözükara SECourses i already have tutorial :d

Yume_Origins•5/28/23, 11:34 PM

lol thx i will see it

Furkan Gözükara SECourses•5/28/23, 11:38 PM

i have planned so many tutorials but so many things taking so much time including family :d

FFurkan Gözükara SECourses ye bf16 not available on all cards

flowwolf•5/28/23, 11:38 PM

rtx 4080 it should work

YYume_Origins i would be glad to see a tutorial about the training on kohya ss guy in case aut...

flowwolf•5/28/23, 11:39 PM

yes i used the one linked on SECourses and a few others here and there. kohya is very widely supported . most tutorials i've noticed are for it, rather than the webui

Fflowwolf rtx 4080 it should work

Furkan Gözükara SECourses•5/28/23, 11:41 PM

yes

flowwolf•5/28/23, 11:46 PM

i also managed to get pytorch 2 working so i trained a style using lion. got some nice results but i might've set the lr way to low

Furkan Gözükara SECourses•5/28/23, 11:49 PM

i use 1e-7 for lion

Furkan Gözükara SECourses•5/28/23, 11:50 PM

torch 2 works on windows

Furkan Gözükara SECourses•5/28/23, 11:50 PM

but on unix i couldnt yet

Yume_Origins•5/28/23, 11:50 PM

simple question : i've trained a dreambooth that work fine as long i don't put " oil painting in the prompt : in this case the face is destroyed

FFurkan Gözükara SECourses you should aim epoch

EggoOP•5/28/23, 11:50 PM

I figured epoch is more important, but then how would I decid how many repeats? :/ ugh I feel that dream booth extension is simpler, but it always gives us trouble.

YYume_Origins simple question : i've trained a dreambooth that work fine as long i don't put "...

Furkan Gözükara SECourses•5/28/23, 11:52 PM

that means either overtrained or undertrained

EEggo I figured epoch is more important, but then how would I decid how many repeats? ...

Furkan Gözükara SECourses•5/28/23, 11:52 PM

ye it depends on used script

Furkan Gözükara SECourses•5/28/23, 11:52 PM

normally 1 epoch means 1 time all training items are trained

FFurkan Gözükara SECourses that means either overtrained or undertrained

Yume_Origins•5/28/23, 11:53 PM

i have several steps ckpt of this model i will try them all

felixsanz•5/29/23, 12:13 AM

dreambooth learning rate 0.00001 vs 0.000002 anyone?

Ffelixsanz dreambooth learning rate 0.00001 vs 0.000002 anyone?

Furkan Gözükara SECourses•5/29/23, 12:19 AM

i prefer 1e-6 for adam, 1e-7 for lion

Ffelixsanz dreambooth learning rate 0.00001 vs 0.000002 anyone?

Yume_Origins•5/29/23, 12:19 AM

i use 0.000002

YYume_Origins i use 0.000002

felixsanz•5/29/23, 12:22 AM

but you tried 0.00001 ?

Ffelixsanz but you tried 0.00001 ?

Yume_Origins•5/29/23, 12:22 AM

not so far

mccoypauley•5/29/23, 12:25 AM

if we’re trying to train concepts (again I’m doing fantasy races), does it make sense to include a lot more training images?

flowwolf•5/29/23, 12:27 AM

quality over quantity i lean towards. add more but only if they meet a bar

flowwolf•5/29/23, 12:29 AM

the concept in a myriad of ways too . not more images that are similar to whats there already. unless they're high quality haha

EggoOP•5/29/23, 12:36 AM

do i need to upgrade xformers for 1111 dreambooth extension to work properly, or are these settings alright?

EEggo do i need to upgrade xformers for 1111 dreambooth extension to work properly, or...

Furkan Gözükara SECourses•5/29/23, 12:49 AM

i would update to latest

FFurkan Gözükara SECourses i would update to latest

EggoOP•5/29/23, 1:00 AM

Will do! You have a video on it? I’ll watch it. On how to update it

PickledJoi•5/29/23, 1:01 AM

wouldn't you just need to force pip to upgrade the xformers package

flowwolf•5/29/23, 1:02 AM

i got bf16 working. i'm not sure why training would just hang up the other day. came back to it with fresh eyes a day later and it just works this time. cool. style training is going well. i'm doing lion with cosine 3 restarts. te lr of 1e-6 and unet lr 2e-5. i guess i'm supposed to use less but the cosine schedulling keeps dipping it towards 1e-6

EEggo Will do! You have a video on it? I’ll watch it. On how to update it

Furkan Gözükara SECourses•5/29/23, 1:21 AM

in this video i show the logic : https://youtu.be/pom3nQejaTs

YouTubeSECourses

How To Install New DREAMBOOTH & Torch 2 On Automatic1111 Web UI PC ...

Discord : https://bit.ly/SECoursesDiscord. Torch 2 / Pytorch 2 is now supported along with new DreamBooth Automatic1111 Web UI extension. If I have been of assistance to you and you would like to show your support for my work, please consider becoming a patron on

https://www.patreon.com/SECourses

In this video, I am explaining how you can man...

How To Install New DREAMBOOTH & Torch 2 On Automatic1111 Web UI PC ...

Furkan Gözükara SECourses•5/29/23, 1:24 AM

i am trying to find best settings for speech training

Furkan Gözükara SECourses•5/29/23, 1:24 AM

i have a famous voice in my mind to train

Furkan Gözükara SECourses•5/29/23, 1:25 AM

496.wav1.88MB

Furkan Gözükara SECourses•5/29/23, 1:28 AM

i am going to sleep hopefully see you later

thebean•5/29/23, 2:15 AM

anyone had massive it/s slowdown on the new 1.3.0 a1111 commit? reverting commits didn't give me back my old speed either and i'm trying to figure out what might've caused this

FFurkan Gözükara SECourses but on unix i couldnt yet

ashley•5/29/23, 5:32 AM

I also couldn't get the latest version of torch working on Linux

Tthebean anyone had massive it/s slowdown on the new 1.3.0 a1111 commit? reverting commit...

ashley•5/29/23, 5:36 AM

It shouldn't make any difference to things like that. It's usually your GPU that affects things like that.

still a little new here. I am trying to wrap my head around why epochs are halved if using classific

Similar Threads