Software Engineering Courses (SECourses)•3y ago

@Furkan Gözükara SECourses in your last video, when you set class images per instance to 48, and 200

@Furkan Gözükara SECourses in your last video, when you set class images per instance to 48, and 200 epochs - how many total training steps did you get?

IVGOP•1/24/23, 1:39 PM

and how long was your training process?

IIVG @Furkan Gözükara SECourses in your last video, when you set class images per ins...

Furkan Gözükara SECourses•1/24/23, 1:40 PM

class images count don't affect your training process speed or steps count

Furkan Gözükara SECourses•1/24/23, 1:40 PM

it depends on your cpu but it would get calculated like this

Furkan Gözükara SECourses•1/24/23, 1:41 PM

lets say you have 10 training images

Furkan Gözükara SECourses•1/24/23, 1:41 PM

1 epoch would mean 10 + 10 (this comes from classiication images) now in same bucket so 20 steps

Furkan Gözükara SECourses•1/24/23, 1:41 PM

lets say your gpu can process 1 it second so 1 step second

Furkan Gözükara SECourses•1/24/23, 1:41 PM

so 200 epoch would mean 200 * 20 = 4000 steps = 4000 seconds

IVGOP•1/24/23, 1:42 PM

so what settings would you use for 34 instance images?

Colored Tonalities•1/24/23, 1:43 PM

Instance images: 10 = I
Epoch: I+I(CI)
36+36(

IIVG so what settings would you use for 34 instance images?

Furkan Gözükara SECourses•1/24/23, 1:44 PM

for dreambooth i think you can test same. just save every 5 epoch. also try lower training rate if it gets too fast overtrained

Furkan Gözükara SECourses•1/24/23, 1:44 PM

so like

Furkan Gözükara SECourses•1/24/23, 1:44 PM

50 classification for example for each instance

Furkan Gözükara SECourses•1/24/23, 1:44 PM

1 batch 1 gradient default 0.000002 learn rate

IVGOP•1/24/23, 1:44 PM

and 200 epochs?

Furkan Gözükara SECourses•1/24/23, 1:44 PM

in my last case i got best results with just 30 epoch

IIVG and 200 epochs?

Furkan Gözükara SECourses•1/24/23, 1:44 PM

yes

Furkan Gözükara SECourses•1/24/23, 1:44 PM

but save checkpoints

Furkan Gözükara SECourses•1/24/23, 1:45 PM

because you cant know which one will be best

Colored Tonalities•1/24/23, 1:45 PM

IVGOP•1/24/23, 1:45 PM

ok thats a long training - like 13600 stepaa

Colored Tonalities•1/24/23, 1:46 PM

Colored Tonalities•1/24/23, 1:46 PM

Thats how I am calc in excel

IVGOP•1/24/23, 1:47 PM

well the calculation done by DB is that with 34 instance images, and 200 epochs its 13600 steps...

IVGOP•1/24/23, 1:47 PM

@thirdman are u using warmup steps in your training?

Colored Tonalities•1/24/23, 1:48 PM

I have a calculation, but no I do not follow it

Colored Tonalities•1/24/23, 1:48 PM

I am however trying it this run

IVGOP•1/24/23, 1:49 PM

it'd be interesting to see how the warmup steps are affecting the model

Colored Tonalities•1/24/23, 1:51 PM

CColored Tonalities Click to see attachment

Furkan Gözükara SECourses•1/24/23, 2:00 PM

i think for loss graph analyze, the paper authors could be email to get their opinions

Colored Tonalities•1/24/23, 2:00 PM

oh? I want to email them

Colored Tonalities•1/24/23, 2:01 PM

Its very interesting question

Colored Tonalities•1/24/23, 2:01 PM

dont you think?

Seagreen•1/24/23, 2:01 PM

Guys do you use background removal for instance images?

IVGOP•1/24/23, 2:03 PM

I don't

Colored Tonalities•1/24/23, 2:03 PM

you should

IVGOP•1/24/23, 2:03 PM

background removal? you do it?

Colored Tonalities•1/24/23, 2:03 PM

Is white background diffusable?

Colored Tonalities•1/24/23, 2:04 PM

And is there a difference between whatever happens when the background is gone?

Colored Tonalities•1/24/23, 2:04 PM

when ai diffuse?

Colored Tonalities•1/24/23, 2:04 PM

I dont know what the PNG says

IVGOP•1/24/23, 2:05 PM

i haven't tried it without the background

CColored Tonalities dont you think?

Furkan Gözükara SECourses•1/24/23, 2:05 PM

yep email them

Furkan Gözükara SECourses•1/24/23, 2:05 PM

they would know best

Furkan Gözükara SECourses•1/24/23, 2:05 PM

loss meaning

Colored Tonalities•1/24/23, 2:06 PM

https://arxiv.org/pdf/2208.01618.pdf

Colored Tonalities•1/24/23, 2:06 PM

This paper?

Furkan Gözükara SECourses•1/24/23, 2:07 PM

this is dreambooth : https://arxiv.org/pdf/2208.12242.pdf

Furkan Gözükara SECourses•1/24/23, 2:08 PM

when you register you can see author email i presume : https://arxiv.org/abs/2208.12242

arXiv.org

DreamBooth: Fine Tuning Text-to-Image Diffusion Models for...

Large text-to-image models achieved a remarkable leap in the evolution of AI,
enabling high-quality and diverse synthesis of images from a given text prompt.
However, these models lack the ability...

Furkan Gözükara SECourses•1/24/23, 2:08 PM

also searching their names and other papers probably contains their email

Colored Tonalities•1/24/23, 2:08 PM

I have not read this paper

@Furkan Gözükara SECourses in your last video, when you set class images per instance to 48, and 200

Similar Threads