#kohya_ss Trainer Colab

2103 messages · Page 3 of 3 (latest)

deep geyser
#

I'm not sure actually

celest crater
#

where do you handle the downloading logic? i might just add a console log there to see if its stuck or not

deep geyser
#

Before downloading there's a message, so you would know

celest crater
#

while you're here i might as well ask if you have some math behind locon dim's since this is an style lora with so many images idk how big it should be

deep geyser
#

With that many images it might get stuck saying it's more than 10000 steps?

celest crater
#

oh yeah i removed the check

deep geyser
#

Cool

celest crater
#

im trying prdigy to see how it goes

deep geyser
#

Prodigy is good for few images though, dunno how that will go

celest crater
#

oh well

deep geyser
#

alpha being half

celest crater
#

16/8 16/8?

deep geyser
#

yeah

celest crater
#

nice ill try then after this bake, colab has been a bitch today

#

had to make some changes to the screencap colab you linked on your dataset maker

#

turns out even if you're not using the gpu your time runs out

#

it handles tv series now so its cool

deep geyser
tacit zinc
#

been a long time since i last watched this space

#

where should i ask training loras on only 4gb vram or just give up

deep geyser
tacit zinc
#

man the trainer is really fast with batch size 6 now

#

it took less time to actually train the lora than myself preparing the dataset + fiddling with the options there

#

(kinda assumed lora training is also banned on colab after their policy change until finding out rn)

deep geyser
tacit zinc
#

anyway good job keeping the colab up

tacit zinc
#

how should i set ss_lr_warmup_steps? is it a ratio like .05, 5 (for 5%) or the actual no of warmup steps needed (like total steps x 5%)?

deep geyser
#

lora trainer working again

deep geyser
tacit zinc
#

i was using the linaqruf one

#

so apparently the linaqruf colab required me to input the exact no of steps so i need to multiply it

deep geyser
#

I suppose it's modular and all-in-one
Few more options, though I tried to put all the relevant ones in mine
I made mine based on linaqruf's you see, because I didn't like the "ergonomics" of it

tacit zinc
#

that was the one i bookmarked a while ago

#

which has more advanced option listed in the cells (i have to edit the values in the code for yours)

deep geyser
#

Which options would you like?

tacit zinc
#

just tried to run your colab and it got a 500 server error from github lmao

deep geyser
tacit zinc
#

alright i got this in your dataset maker (tagging cell)

#

env: PYTHONPATH=/env/python

🚶‍♂️ Launching program...

env: PYTHONPATH=/content/kohya-trainer
Traceback (most recent call last):
File "/content/kohya-trainer/finetune/tag_images_by_wd14_tagger.py", line 10, in <module>
from tensorflow.keras.models import load_model
File "/usr/local/lib/python3.10/dist-packages/keras/api/_v2/keras/init.py", line 3, in <module>
from keras.api._v2.keras import internal
File "/usr/local/lib/python3.10/dist-packages/keras/api/_v2/keras/init.py", line 3, in <module>
from keras.api._v2.keras import internal
File "/usr/local/lib/python3.10/dist-packages/keras/api/_v2/keras/internal/init.py", line 3, in <module>
from keras.api._v2.keras.internal import backend
File "/usr/local/lib/python3.10/dist-packages/keras/api/_v2/keras/internal/backend/init.py", line 3, in <module>
from keras.src.backend import _initialize_variables as initialize_variables
ImportError: cannot import name '_initialize_variables' from 'keras.src.backend' (/usr/local/lib/python3.10/dist-packages/keras/src/backend/init.py)

deep geyser
tacit zinc
#

change the resolution slider interval to 64px so i can do 704^2

deep geyser
#

you can edit it yourself

tacit zinc
#

ofc i know

#

realizing how colab might terminate my free session before it ends (its 768px on 1.5e-4)

deep geyser
tacit zinc
#

my lora would be ready 80% in so i might barely make it

tacit zinc
#

been talking to kohaku and they said they can add an update new post function to this

#

would be good for dataset maker

#

(am i really the only one still using your colab notebooks?)

deep geyser
#

Plenty of people using them just not in this channel

#

Though less people than before

faint bronze
tacit zinc
#

basically a danbooru mirror

deep geyser
tacit zinc
#

but yeah thats not your issue to solve, sorry for bothering you

tacit zinc
deep geyser
#

Most people I know train locally

tacit zinc
#

ok i will when i get my new pc/laptop

tacit zinc
#

i would still dl your dataset maker notebook to run it locally

deep geyser
tacit zinc
deep geyser
wispy lodge
#

Does anyone know of a suitable Colab link for clothing LoRA training (including subfolders)?

tacit zinc
#

the basic lora trainer is fine, it has a cell use custom subfolders

#

what you need to do is to do some magic with your dataset

#

theres a tutorial on civitai that basically tells you to add face censor to your dataset + crop out details of the outfit + add other naked ocs to compensate for the face

wispy lodge
tacit zinc
#

not fixed

tacit zinc
# tacit zinc

still unfixed @deep geyser can you check if the code is working

deep geyser
tacit zinc
#

really

#

it might be on my end

#

let me check again

tacit zinc
#

is this a dependencies issue?

deep geyser
#

Probably

tacit zinc
#

seems like im the only one using it & reporting back

#

it is killing my runtime midway when i have a dataset thats mixed sfw/nsfw when other datasets that is more nsfw were fine

#

(or is it just regular free timeout)

crude apex
#

Doesn't seem to work anymore

deep geyser
#

yeah that one's dead

crude apex
#

so is there one that still gets updated for sd 1.5

deep geyser
#

mine still works

crude apex
#

could you relink it ?

deep geyser
crude apex
stone warren
#

@deep geyser it appears the dataset maker is having trouble executing the 4th step to generate tags.


env: PYTHONPATH=/content/kohya-trainer
2024-09-29 09:52:46.679427: E external/local_xla/xla/stream_executor/cuda/cuda_dnn.cc:9261] Unable to register cuDNN factory: Attempting to register factory for plugin cuDNN when one has already been registered
2024-09-29 09:52:46.679478: E external/local_xla/xla/stream_executor/cuda/cuda_fft.cc:607] Unable to register cuFFT factory: Attempting to register factory for plugin cuFFT when one has already been registered
2024-09-29 09:52:46.681524: E external/local_xla/xla/stream_executor/cuda/cuda_blas.cc:1515] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered```

Trainers are still working though. Any ideas what happened to the dataset maker?
deep geyser
tacit zinc
#

does free colab t4 (16gb) have enough vram for sdxl trainer at batch size 2?

#

at least it wont cut running cells short in case times up for free users

tacit zinc
#

ok its doable (14gb)

#

going above batch size 3 is no good anyway so the only main optimization for pro users would be train on safetensors

tacit zinc
#

fuck free colab they cut my session off and stopped my cell right before i have a good epoch