@Dr. Furkan Gözükara i got the block swapping to work by installing some more ram but the 12gb confi
@Dr. Furkan Gözükara i got the block swapping to work by installing some more ram but the 12gb config is running at 40s/it. Any thoughts?
EHow much Ram do you have? Physical or virtual ram? It did not work for me
EDoes anyone know the best way to train the model to merge two persons? I have datasets of two different persons and want to create a unique mix. What would be the best way?
N"I'm training on 4090D with 500 images. The speed is 4.4s/it (using Rank_3_18950MB_9_05_Second_IT). When I'm training on A800 (80 GB VRAM), the speed is 3.3s/it (using Rank_1_29500MB_8_85_Second_IT). I think the speed difference is abnormal. What could be causing this?"
Fi think you can deploy any pipeline to replicate
Ftotally normal i got same speed
Fi had to purchase 2x32 gb today, my 4 stick 16 gb keep causing problems
Fi wasted 3 hours today
Fmin 48 gb phyiscal ram
Frank 3 is slow
Fyour speed totally good
Fi suggest fine tuning though way better results
OFor training styles, how important is it to have the training images in 1024x. It's a bit more difficult to get those in uniform resolution vs training a subject.
Fwell it is really hard to test
Fthus i suggest people to make 1 training with 1024x1024
Fand 1 training with bucketing
Oto compare? or as a step 2 on the same model/lora?
Fcompare 2 different trainings
Fand see which one works better
Fso you need to train twicew
OI know you've provided scripts/tools to id faces and crop etc, is there something similar that will just expand the canvas of an image to make it square?
Fit does that
Fif image resolution is big
Fand if can do without cropping subject
Fso that is why crop first
Fresize second
Oso if there's a series of portrait style images, it will expand the canvas to make it square, and I guess fill it with something?
Fif that image
Fhas person in distant
Fso that canvas can be expanded
Fif not it will crop
Fi think i added expand with white background too
N
NShouldn't larger VRAM bring more speed? The A800 (using Rank 1) has more than 3 times the VRAM of the 4090D, but the speed is barely increased. Why is this? Is it because I haven't adjusted the batch_size?
Fno such thing. vram has 0 impact on speed if both fitting into vram
Fit is like case of a truck
Fbatch size can certainly make impact in your case
Fbut bigger batch size = different LR
F
EYeah the A800 is ampere generation so better to compare to a 3090, while the 4090 is the newer Ada generation. Ada has more CUDA and tensor cores so more raw compute power.
FA6000 speed almost same as RTX 3090
Fas long as model fits into VRAM
D@Dr. Furkan Gözükara I have a question on regularization images, you recommend 20 to 1 repetitions always or it depends on the number of images on my dataset per subjet?
Dor also I have to set the repetition based on the number of reg images I have?
Dif I set my subjects repetitions too low the reg images will have more impact on the training than my subjets?
V@Dr. Furkan Gözükara Does SwarmUI have something like this node for flux?
Ffor every 1 step
F1 reg image
Fso 15 images 200 epochs = 3000 steps
Fyou need 3000 unique reg images
Fi dont know