--ddp_gradient_as_bucket_view
--ddp_gradient_as_bucket_view
DIs not working in kaggle
DNeed too much memory
TI'm having trouble with kaggle. I tried the original Lora training video and got into the Kohya gui without any issues. But, the json file was a no go.
TTried the updated tut and now the notebook keeps timing out? Any ideas?
TOk, so I got back into the gui. But loading the json file gives me this : Error
Unexpected token '<', " <!DOCTYPE "... is not valid JSON
Unexpected token '<', " <!DOCTYPE "... is not valid JSON
TAm I missing a step somewhere?
Blooks like the json file is not correct
Bsome error when you downloaded the file
THmm. Tried to redownload the fille. Then noticed kaggle stopped the session again: ^C
Keyboard interruption in main thread... closing server.
Killing tunnel 127.0.0.1:7860
Keyboard interruption in main thread... closing server.
Killing tunnel 127.0.0.1:7860
TThis is the file I'm using Kaggle_SDXL_DreamBooth_Best.json. I shows up fine. Copy and paste folder link in gui. But the gui won't even let me create a file like "test1.json
Bby
Unexpected token '<', " <!DOCTYPE "... is not valid JSON it's saying that the file starts with <!DOCTYPE and this isn't the case for a valid jsonBare you using the latest kaggle template? normally it is only killing the session with the old kaggle templates that don't use ngrok
DI have this problem in kaggle
Dupdate to kaggle dreambooth best .json is not working in kaggle for me
TSame. I think maybe kaggle updated something somewhere and errors are popping up. Maybe give it a few days. or ... sigh ... might have to find a gpu. I'm going to muddle around on colab.
Tgetting frustrated with kaggle at the moment
S@Dr. Furkan Gözükara I have been trying to do SDXL dreambooth training on Kaggle. I'm getting below issue. Pls assist
SMy bad.. i figured out localserver was not running. Its resolved. Thanks
DYeah they fixing the problem first there was catche issue
DYou can go with it
Fsometimes bugged
Fi would suggest this. do dreambooth on SDXL . then extract lora. are you really need sd 1.5?
Fplease do screen record hard to understand from there
Fresizer will resize
Fcrop will not
Fhello. you cant do lora of sdxl with 8 gb vram gpu
Fyou can only do sd 1.5
Fi am still yet to explore this. sadly couldnt find time yet
Fi updated the patreon post
Fupdated the atreon post
F
DHi @Dr. Furkan Gözükara Is there a good formula on setting repeats/epochs when you're using a larger number of images for a DreamBooth training? For example, I have a subject that has 90 images. If I use the default DreamBooth SDXL preset you created for 16GB textencoding, it defaults to doing 18000 steps which takes 7 hours on my mobile 4090. I can reduce the image count, but even if I cut the number down to 30 images, it still takes between 3.5-4 hours. I am using all of the default settings in your preset as well as 20 repeats for my images and 1 repeat for the regularization images. The default epochs in the preset is 8.
Fsadly no formula. that is why checkpointing is important to compare
Fbut usually gets overtrained quickly
Fif images are not diverse
DWhat would you recommend in a case like this? Fewer repeats? Fewer epochs? Setting a max number of steps?
Fwhat are you training
Fa person?
Dyes
Fi would discard similar images. if still very variety of images then try 100 repeat and save every n step
Flike every 1801 steps
Factually make it half
F50 repeat
F901 steps
Fsince our reg images count 5200
D100 repeats will create even more steps though. I'm currently only using 20
F20 * 90 * 2
Fit is 3600 steps
Fmake epoch 1
F1 epoch 50 repeat
