Software Engineering Courses (SECourses)•3y ago

Hello @Dr. Furkan Gözükara I have a question about Kohya. If I want to use the scheduler cosine_with

Hello @Dr. Furkan Gözükara I have a question about Kohya. If I want to use the scheduler cosine_with_restarts, how do I set the restarts? I is something like this? "lrSchedulerNumCycles": 3

iloveh8•10/13/23, 4:49 PM

can we create a video by 'looping' GPT4V and DALL-3 'recursively'

Zono50•10/13/23, 5:32 PM

anyone use comfyui and reactor? some reason I'm always getting blurry faces

Zono50•10/13/23, 5:33 PM

and i can get good looking faces in automatica1111, but not sure why I can't in comfy

virginia•10/13/23, 6:20 PM

Furkan's kaggle notebook for training is pretty good.

FFurkan Gözükara SECourses yes

Stefanoid9343•10/13/23, 9:14 PM

How much the training improve with the caption?

Stefanoid9343•10/13/23, 9:14 PM

I’m planning to train logos

Ttita6070 Hi, I have a issue with my automatic1111, it says that Torch is not able to use ...

Furkan Gözükara SECourses•10/13/23, 10:45 PM

what is your gpu

Tthesilvermoth Hello @Dr. Furkan Gözükara I have a question about Kohya. If I want to use the s...

Furkan Gözükara SECourses•10/13/23, 10:46 PM

i am not sure never used

Furkan Gözükara SECourses•10/13/23, 10:46 PM

but if you do that

Furkan Gözükara SECourses•10/13/23, 10:46 PM

follow LR during training

Furkan Gözükara SECourses•10/13/23, 10:46 PM

use wandb

Iiloveh8 can we create a video by 'looping' GPT4V and DALL-3 'recursively'

Furkan Gözükara SECourses•10/13/23, 10:46 PM

probably not. not very accurate for that

SStefanoid9343 How much the training improve with the caption?

Furkan Gözükara SECourses•10/13/23, 10:47 PM

i am in still research

Furkan Gözükara SECourses•10/13/23, 10:47 PM

for a very poor game character

Furkan Gözükara SECourses•10/13/23, 10:47 PM

i got quite decent

Furkan Gözükara SECourses•10/13/23, 10:47 PM

but i will test on myself later

Furkan Gözükara SECourses•10/13/23, 10:47 PM

this is for a client currently i am testing

vp•10/14/23, 12:54 AM

I’m seeing your comment on Github re LLaVA req for venv installer. Did you succeed?

FFurkan Gözükara SECourses i am in still research

Stefanoid9343•10/14/23, 2:57 AM

Thank you!

mayna•10/14/23, 10:26 AM

Hi all, Anyone know how this can be solved please while training on KOhya SS

wandb: ERROR The API key you provided is either invalid or missing. If the WANDB_API_KEYWANDB_API_KEY environment variable is set, make sure it is correct. Otherwise, to resolve this issue, you may try running the 'wandb login --relogin' command. If you are using a local server, make sure that you're using the correct hostname. If you're not sure, you can try logging in again using the 'wandb login --relogin --host [hostname]' command.(Error 401: Unauthorized)

mayna•10/14/23, 10:27 AM

where to run wandb login --relogin ?

FFurkan Gözükara SECourses what is your gpu

tita6070•10/14/23, 10:36 AM

RTX 3060

mayna•10/14/23, 1:32 PM

anyone know how to pause resume a trainign ? been running for 3h still 3h ..

mayna•10/14/23, 1:32 PM

my machine is burning

FFurkan Gözükara SECourses yes

Elisoli•10/14/23, 1:37 PM

This only use RAM and not gpu? when using the model you use on the video (the advance one) - it takes about 1 hour per image, does that make sense?

Elisoli•10/14/23, 1:37 PM

3060 gpu, 64gm ram

iloveh8•10/14/23, 4:07 PM

havent seen any real example that produces good video, any suggestion for very specific use case that has higher chance of success - video looks really good? this is for ecommerce use case / product showcasing

Mmikemenders I have now tried the Prodigy Lora training with another parameter that is also r...

Constipated•10/14/23, 4:20 PM

Can you share your full training settings?

CConstipated Can you share your full training settings?

mikemenders•10/14/23, 5:04 PM

With Prodigy, I find that it's very sensitive to the face, so you definitely need a Compreface in front of it to get images with a score of 0.99-1 only, smaller ones will distort the face badly.
I use two types of parameters: a strong one that will take most of the body but needs a strong prompt, and a flexible weaker one that will take the face but less the body.

Mykee_-_Prodigy_-_RealisticVision_-_Person_Flexible.json2.9KB

Mykee_-_Prodigy_-_RealisticVision_-_Person.json2.88KB

mikemenders•10/14/23, 5:05 PM

I train for 25 pictures with 4 repetitions

mikemenders•10/14/23, 5:06 PM

I just made a too strong Lora, the Compreface gives a value of 1 even without the hires

mikemenders•10/14/23, 5:41 PM

The strong lora script can also be damped with safeguard_warmup=True in case it gets strong too soon (I use a test prompt)

mikemenders•10/14/23, 6:50 PM

Here is a different without (black) and with (blue) safeguard_warmup=True command. Blue is better now. Black is my strong model/script.

Mmikemenders With Prodigy, I find that it's very sensitive to the face, so you definitely nee...

Constipated•10/14/23, 6:59 PM

Thanks for sharing! Very helpful!

mikemenders•10/14/23, 7:30 PM

I also made a version with weight_decay=0.01, so with these parameters: use_bias_correction=True weight_decay=0.01 d_coef=2 safeguard_warmup=Trueuse_bias_correction=True weight_decay=0.01 d_coef=2 safeguard_warmup=True
The pink line is the new one. You can see that almost nothing has changed. I'm going to make a flexible version with d_coef set to 1 by default (that's why I didn't include it), weight_decay will be set to 0.3, because I read on a website that this is the best value.

mikemenders•10/14/23, 7:32 PM

For some reason, these three models became very strong (I really couldn't find a correlation why), and when generated or trained with Clip Skip 2, they become more flexible.

Mmikemenders With Prodigy, I find that it's very sensitive to the face, so you definitely nee...

Constipated•10/14/23, 7:48 PM

Giving it a try right now.

This works for SDXL right? (noticed that the json file was set to resolution 512x512)

mikemenders•10/14/23, 7:49 PM

Yes, work on SDXL too

Constipated•10/14/23, 7:49 PM

And what do you mean "needing a compreface in front of it"?

mikemenders•10/14/23, 7:49 PM

But I don't train on SDXL yet, because it also gives very nice results on RealVision.

mikemenders•10/14/23, 7:50 PM

Here is CompreFace:
https://github.com/exadel-inc/CompreFace

GitHub

GitHub - exadel-inc/CompreFace: Leading free and open-source face r...

Leading free and open-source face recognition system - GitHub - exadel-inc/CompreFace: Leading free and open-source face recognition system

Constipated•10/14/23, 7:51 PM

Yeah I googled that too

Constipated•10/14/23, 7:51 PM

Not sure what face recognition had to do with Lora training

mikemenders•10/14/23, 7:53 PM

I installed for TI with this guide: https://discord.com/channels/1010980909568245801/1096067465177464842/1102035161438240798

mikemenders•10/14/23, 7:54 PM

The essence of this is that you put in the original image and next to it the generated one, and it outputs how identical the two are

mikemenders•10/14/23, 7:55 PM

Finally, here is the fourth model (yellow) where I used these commands:
use_bias_correction=True safeguard_warmup=True weight_decay=0.3use_bias_correction=True safeguard_warmup=True weight_decay=0.3
You can see that if d_coef defaults to 1, it learns slower, the pattern is not fixed as much (the model becomes more flexible). That's why I use two types of scripts: for most of them the strong script is good, but where the model becomes too rigid, I use the flexible version, the other parameters are the same, I just change the arguments. With the strong model it not only takes the face but also the body very well, but on some image sets it gets stubborn and learns quickly, in those cases I need the flexible one. However, it is a fact that if any image on the CompreFace score is lower than 0.99, the face of the generated person is distorted to an extreme.

Constipated•10/14/23, 7:56 PM

Wow

Constipated•10/14/23, 7:56 PM

This is deep... Will need to study this

mikemenders•10/14/23, 8:00 PM

The flexible version is now flexible, but the compreface score is only 0.8 (above 0.9 would be good), so maybe a d_coef=1.5 would be better and flexible yet, but I won't bore you with that. So I hope you can see which switch has a stronger effect and which one doesn't.

mikemenders•10/14/23, 8:06 PM

What's very interesting is that the loss is always the same, even if I set the parameters, so it's not relevant. In the other graphs you see more variation.

Hello @Dr. Furkan Gözükara I have a question about Kohya. If I want to use the scheduler cosine_with

Similar Threads

Similar Threads

Similar Threads