CPU Training

[ADH369]

hi,
I updated dreambooth_extension and after that i cant see CPU Training option anymore.
is the CPU Training option removed from the extension ?

I was doing CPU training for my models.

Thanks

[runner22k]

How log it used to take per model? How many images you used, per training?

[ADH369]

I have AMD 8320 FX CPU with 4 CORE that takes 9 hours to train 4 images of 512x512 Res.

[runner22k]

My goodness! too much for 4 images! my be try Google colab at some point, and use LoRA. I was also looking for dreambooth options for 8GBVRAM. Later I got convinced to relay on LoRA instead of entire model tuning. If we have a 16GB VRAM card, right now it will take 8 minutes to traing 72 images, (6 images per batch), through Google colab, if we get a faster GPU, I tried 3 times and I got lucky once, else it takes 2 hours on google colab.

I agree that dreambooth training is easy as LoRA needs captions (that is the headache I need to deal with to get the best out of LoRA training.)

[MaKraMc]

I have the same problem. Sadly my GPU is not powerful enough but my CPU could do the training over night.

[ADH369]

yes i was doing my models on CPU overnight. it is not fast as GPU but works really well.

but i am not sure if it is me only or CPU training options is being removed ?

[MaKraMc]

I didn't find the option as well.

[MaKraMc]

I really don't know why it has been removed. I did some digging in an older version and reimplemented it into the code. Will be doing some testing tomorrow but it should be pretty straightforward. :)

[ADH369]

oh that would be great. i am waiting for the result.

[make23l]

I've been looking for this option for two days! It's good to know that it still exists even though it has been lost. I'm waiting! thx!

[gsgoldma]

please post the code, if you have got it figured out.

[d8ahazard]

Add --force-cpu to your command line arguments and restart. This will force CPU only training.

…

On Sun, Jan 22, 2023, 10:49 AM gsgoldma ***@***.***> wrote: please post the code, if you have got it figured out. — Reply to this email directly, view it on GitHub <#773 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAMO4NCLMDFW6BBNKTF3453WTVQKHANCNFSM6AAAAAAT2ORSLE> . You are receiving this because you are subscribed to this thread.Message ID: <d8ahazard/sd_dreambooth_extension/repo-discussions/773/comments/4750377@ github.com>

[ADH369]

yes it is possible if you add --force-cpu to the dreambooth bat.Rather than using GPU and 4Gb VRam it will use CPU and RAM for Training.This process is really slow.

[econundrum]

I know it's slow I trained a few things on the older version, I don't seem to be able to get --force-cpu to work CUDA is still trying to allocate vRAM from the error I'm getting.

[ADH369]

I got same problem too but after one or two restarts it worked properly.

only just to make sure did you add that line in weui-user.bat file inside sd-dreambooth folder?

please check attached image for more information

.

[gsgoldma]

I got same problem too but after one or two restarts it worked properly.

only just to make sure did you add that line in weui-user.bat file inside sd-dreambooth folder?

please check attached image for more information .

which one should it be added to? I added to the extension dream booths folder commandline args, is that wrong?

[econundrum]

Thank you I'd added it in the wrong place I'll give that a try.

[econundrum]

Have to say colab which I've tried now is a much better option than cpu training generally, still the option is nice.