if you get 7 GPU SXM machine with A100
if you get 7 GPU SXM machine with A100
Fit will be like 7 times faster
F
Fand you can use same learning rate
Fdefinitely better
F15 images underperform
WSo I can't use A6000 right ? because based your named 48gb I though 7xA6000 are enough
CIs Torch 2.5 OK now or not? I can't figure it out from the Video Tutorial, this is for Dreambooth Flux
Fyou can use
Fbut not for fine tuning
Fbut for lora multi gpu
Fit is ok to use for windows
Whttps://1drv.ms/u/s!AoU1QUjSr6pJrbdRgnQNUMMyMe2qTg?e=xCOQuv
Hi, could you guide me through what's wrong with this training. My latest samples at 4400 steps are still far from imitating Justin Sweet's style. Thanks
Hi, could you guide me through what's wrong with this training. My latest samples at 4400 steps are still far from imitating Justin Sweet's style. Thanks
Fcrop all to 1024x1024
Fdont use captions
Ftoken can be like
Fohwx painting style
WWould you have a way to automate croping/upscaling please, with ComfyUI for example?
Fi have auto cropper for some subjects
Wfor the token it's here? like that?
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
Fyep
Wthank you
V@Furkan Gözükara SECourses Hi Dr. Can you please share which video teaches me how to use huggingface for kohya_ss save state uploads/sync etc. ?
M@Furkan Gözükara SECourses Bot attack again
Dhm, is Flux schnell trainable?
Fi couldnt make it work
Fi manually upload via notebook
Fmy results were bad
V@Furkan Gözükara SECourses What is the correct folder path for the image? I would need to enter this in the resume save-state field.
Fi never made it work
VI made it work in windows, but I don't know to get the proper folder path in ubuntu. In windows, if you select the two options in the picture, you can shutdown and resume based on the last save, with the same learning rate etc. Nothing is lost or needs to be done, which is preffered instead of just loading the last saved model file. You would then link to the folder of the laste save-state folder of the model you are training to resume. In onetrainer its much easier.
Fyou can use last checkpoint
Feasier
CQuick question, does fine tuning require less steps than LORA?
My 3,000 steps fine tuning from last night (15 pics x 200 epochs) is giving much less good results than my 5,800 steps Dim 16 Lora I did with Fluxgym.
Is that normal?
My 3,000 steps fine tuning from last night (15 pics x 200 epochs) is giving much less good results than my 5,800 steps Dim 16 Lora I did with Fluxgym.
Is that normal?
Walmost finished rendering but an error (
V@Furkan Gözükara SECourses The use of save-states is much better as its actually continuing the training of the model. Starting off from the last saved checkpoint is essentially a whole different training of the model and the results will be different.
VAnyway, I figured out how to use saved state and it worked, you just have to enable it at the start and it saves a special folder called saved-state. If for some reason you have to continue training due to crash or no HD space you can continue by inputting the folder patch into this field before you start. If you get an error then the path is not correct. If the folder exists with what is seen in the photo then it will definitely continue training with the same state it left off at as well as same result of training, just a continuation.
VI forgot to mention, you would also have to edit the number of epochs, to be what you did not initially finish. So if you have a training for 200 epochs, and the saved state is from the 150 epoch, you have to change the epoch to 50 in the settings, or else it will continue training from the 150th epoch save-state for another 200 epochs. You can increase the epochs what ever you like, but it will be start from epoch 1 to what you set, since you are continuing the training using the saved-state folder. Its confusing if you don't understand this part.
WIs there any point in modifying LR with adafactor? 22 hours of training and 4400 steps later, I have the impression that my final samples are still undertrained.
WIf I choose the last checkpoint, can I resume training from there for more steps?
Flearning rates
Fand step count depends on dataset
Fyour yolo file is not properly downloaded
Fnice but that is extra space :d
Fsure you can increase
Fto get faster
Fsure you can
