Well I find this scrolling up

FFurkan Gözükara SECourses it depends several factors

V

Vinmi•9/16/24, 3:07 PM

I trained a 200 epoch LoRA (81 images dataset), with the following settings:

V

Vinmi•9/16/24, 3:07 PM

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

{
  "engine": "kohya",
  "unetLR": 0.0005,
  "clipSkip": 1,
  "loraType": "lora",
  "keepTokens": 0,
  "networkDim": 2,
  "numRepeats": 1,
  "resolution": 1024,
  "lrScheduler": "cosine_with_restarts",
  "minSnrGamma": 5,
  "noiseOffset": 0.1,
  "targetSteps": 4050,
  "enableBucket": true,
  "networkAlpha": 16,
  "optimizerType": "AdamW8Bit",
  "textEncoderLR": 0,
  "maxTrainEpochs": 200,
  "shuffleCaption": false,
  "trainBatchSize": 4,
  "flipAugmentation": false,
  "lrSchedulerNumCycles": 3
}

V

Vinmi•9/16/24, 3:08 PM

Got a 18mb LoRA out of it

V

Vinmi•9/16/24, 3:08 PM

It works, but is not incredible

VVinmi I trained a 200 epoch LoRA (81 images dataset), with the following settings:

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

i explained in video the factors

KKallamamran Well I find this scrolling up

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

set accelerator and gpu id 0

KKallamamran Kohya gets stuck here

F

Furkan Gözükara SECourses•9/16/24, 3:14 PM

it is using cpu that is the error

K

KallamamranOP•9/16/24, 3:14 PM

Oh shit

K

KallamamranOP•9/16/24, 3:16 PM

May I ask where? command line?

F

Furkan Gözükara SECourses•9/16/24, 3:17 PM

read

K

KallamamranOP•9/16/24, 3:17 PM

KKallamamran Click to see attachment

F

Furkan Gözükara SECourses•9/16/24, 3:18 PM

ok then set accelerator

F

Furkan Gözükara SECourses•9/16/24, 3:18 PM

you have multiple gpus?

K

KallamamranOP•9/16/24, 3:18 PM

I do

K

KallamamranOP•9/16/24, 3:18 PM

I shitty internal GPU that gets listed

K

KallamamranOP•9/16/24, 3:19 PM

Aha

K

KallamamranOP•9/16/24, 3:19 PM

K

KallamamranOP•9/16/24, 3:19 PM

accelerator device: cpu :/

K

KallamamranOP•9/16/24, 3:19 PM

Well have to eat. Hopefully I can solve this later

K

KallamamranOP•9/16/24, 3:19 PM

Thanks!

F

Furkan Gözükara SECourses•9/16/24, 3:20 PM

ok try gpu id 1

K

Koudkunstje (Roelof)•9/16/24, 3:57 PM

To test a Lora which guidance value is best? Like 1 or 3.5 or even higher?

K

Koudkunstje (Roelof)•9/16/24, 4:08 PM

I have a pretty hard time to find the sweet spot. Sometimes it changes the image quite a bit like on epoch 200 strenght 1.0 and on epoch 150 at strenght 1 .1

K

Koudkunstje (Roelof)•9/16/24, 4:22 PM

ok this image isn't very useful as discord resizes it

KKoudkunstje (Roelof)To test a Lora which guidance value is best? Like 1 or 3.5 or even higher?

F

Furkan Gözükara SECourses•9/16/24, 4:23 PM

cfg 1

F

Furkan Gözükara SECourses•9/16/24, 4:23 PM

flux guidance 4

F

Furkan Gözükara SECourses•9/16/24, 4:24 PM

i got a better config for fine tuning

F

Furkan Gözükara SECourses•9/16/24, 4:24 PM

it is nice

K

Koudkunstje (Roelof)•9/16/24, 4:27 PM

default shift (base 0.5 / max 1.15) ?

KKoudkunstje (Roelof)default shift (base 0.5 / max 1.15) ?

F

Furkan Gözükara SECourses•9/16/24, 4:35 PM

ah from your image i dont know what are they

K

Koudkunstje (Roelof)•9/16/24, 4:39 PM

if discord doesn't resize the image maybe this is more useful. Seem pretty spot on with epoch 200 and lora_strenght 1.0 at 1.1 if starts mutilate hands.

KKoudkunstje (Roelof)if discord doesn't resize the image maybe this is more useful. Seem pretty spot ...

F

Furkan Gözükara SECourses•9/16/24, 4:40 PM

ye i use lora str 1

F

Furkan Gözükara SECourses•9/16/24, 4:40 PM

v3 test results added : https://www.patreon.com/posts/112099700

Patreon

Kohya FLUX Fine Tuning (Full Checkpoints) Training Full Tutorial Fo...

Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon

K

Koudkunstje (Roelof)•9/16/24, 4:44 PM

Okey, but that is not for lora. check that out another time as it was a 48hr run with 100 images in a data set. for this lora.

K

Koudkunstje (Roelof)•9/16/24, 4:46 PM

on a RTX 3090 ti

KKoudkunstje (Roelof)Okey, but that is not for lora. check that out another time as it was a 48hr run...

F

Furkan Gözükara SECourses•9/16/24, 4:46 PM

ye 100 images 100 epochs would take a lot of time

K

Koudkunstje (Roelof)•9/16/24, 4:47 PM