30% weakerr? Probably less because the upper you have to take into consideration warmup I think?
30% weakerr? Probably less because the upper you have to take into consideration warmup I think?
Fnot sure about that
Fmaybe opposite
Coh wow
CSeed before tuning
Cohwx woman
Cafter fine tuning ohwx woman ceteris paribus
C
Cbefore tuning, after tuning same seed
CBefore and after
I@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?
Iand how long was your training process?
Fclass images count don't affect your training process speed or steps count
Fit depends on your cpu but it would get calculated like this
Flets say you have 10 training images
F1 epoch would mean 10 + 10 (this comes from classiication images) now in same bucket so 20 steps
Flets say your gpu can process 1 it second so 1 step second
Fso 200 epoch would mean 200 * 20 = 4000 steps = 4000 seconds
Iso what settings would you use for 34 instance images?
CInstance images: 10 = I
Epoch: I+I(CI)
36+36(
Epoch: I+I(CI)
36+36(
Ffor dreambooth i think you can test same. just save every 5 epoch. also try lower training rate if it gets too fast overtrained
Fso like
F50 classification for example for each instance
F1 batch 1 gradient default 0.000002 learn rate
Iand 200 epochs?
Fin my last case i got best results with just 30 epoch
Fyes
Fbut save checkpoints
Fbecause you cant know which one will be best
C
Iok thats a long training - like 13600 stepaa
C
CThats how I am calc in excel 
Iwell the calculation done by DB is that with 34 instance images, and 200 epochs its 13600 steps...
I@thirdman are u using warmup steps in your training?
CI have a calculation, but no I do not follow it
CI am however trying it this run
Iit'd be interesting to see how the warmup steps are affecting the model
C
Fi think for loss graph analyze, the paper authors could be email to get their opinions
Coh? I want to email them
CIts very interesting question
Cdont you think?
SGuys do you use background removal for instance images?
II don't
Cyou should
Ibackground removal? you do it?
CIs white background diffusable?
CAnd is there a difference between whatever happens when the background is gone?
Cwhen ai diffuse?
CI dont know what the PNG says