Software Engineering Courses (SECourses)•6mo ago

I noticed you are using adafactor, what is your take on prodigy? the learning rate will adapt accord

I noticed you are using adafactor, what is your take on prodigy? the learning rate will adapt accordingly right? it might consume more vram but what do you think the result quality would be with prodigy?

I might be wrong but because the learning rate adapts, I think we might eliminate the risk of overfitting. Your advise would be helpful.

Uunderdog I noticed you are using adafactor, what is your take on prodigy? the learning ra...

Furkan Gözükara SECourses•7/11/25, 9:14 PM

i tested them and none of the others yielded as realistic as adafactor

Furkan Gözükara SECourses•7/11/25, 9:14 PM

i did a big test recently

underdogOP•7/11/25, 9:15 PM

Thank you

for 15 image dataset, were you able to figure what the unet learning rate was ideal?

Uunderdog Thank you 🙂 for 15 image dataset, were you able to figure what the unet learnin...

Furkan Gözükara SECourses•7/11/25, 9:15 PM

yes our rate is ideal

Furkan Gözükara SECourses•7/11/25, 9:15 PM

just do at least 200 epoch

Furkan Gözükara SECourses•7/11/25, 9:15 PM

but you can do up to like 3000 and compare

FFurkan Gözükara SECourses i did a big test recently

underdogOP•7/11/25, 9:15 PM

are you able to share the results for learning purposes?

underdogOP•7/11/25, 9:21 PM

also for batch size, what number do you recommend? I'm renting A100 gpu. I'm testing with a batch size of 1 for 15 images. I want to increase it but I'm afraid the model might overfit. I can test them myself but I'm paying by the hour so testing all of them is going to burn me out

Uunderdog also for batch size, what number do you recommend? I'm renting A100 gpu. I'm tes...

Furkan Gözükara SECourses•7/11/25, 9:22 PM

i recommend bs 1

Uunderdog are you able to share the results for learning purposes?

Furkan Gözükara SECourses•7/11/25, 9:23 PM

i planned the test but later didnt have time

underdogOP•7/11/25, 9:24 PM

Thank you very much. Your testing evaluation is saving my time

underdogOP•7/11/25, 9:28 PM

one more question
you used constant LR scheduler, why not cosine or cosine with restarts?

With the constant scheduler, the model doesn't learn. The training loss doesn't decrease. Is this okay?

Uunderdog one more question you used constant LR scheduler, why not cosine or cosine with ...

Furkan Gözükara SECourses•7/11/25, 9:33 PM

training loss useless

Furkan Gözükara SECourses•7/11/25, 9:33 PM

constant also gives you chance to continue perfectly fine

Konstantin•7/12/25, 5:59 PM

What could be the problem, I tried the kohya-ss/sd-scripts branch for lumina 2 (feature-lumina-image) for lumina_train.py. There were a couple of typos img instead of noisy_model_input, but after correcting them, the training started, the loss even goes down, but at the first generation of samples I get noise. If I add "sample_at_first = true" then I see that before training the image is generated normally, but from the very beginning of training there is noise. Now I have
max_timestep = 1000
guidance_scale = 4
timestep_sampling = "nextdit_shift"
discrete_flow_shift = 6.0
model_prediction_type = "raw"
optimizer_args = [ "weight_decay=0.00",]
optimizer_type = "PagedLion8bit"

I tried changing the optimizer, timestep_sampling, discrete_flow_shift, all the same, no results... Already at the 30th step of training there is noise.

KKonstantin What could be the problem, I tried the kohya-ss/sd-scripts branch for lumina 2 (...

Furkan Gözükara SECourses•7/12/25, 9:03 PM

never tried lumina 2

K4bronsky•7/14/25, 2:17 AM

@Furkan Gözükara SECourses is theres a way to understand how to improve speed? i followed all the steps from the videos on and rtx 5090

KK4bronsky <@205854764540362752> is theres a way to understand how to improve speed? i foll...

Furkan Gözükara SECourses•7/14/25, 2:50 AM

which config?

FFurkan Gözükara SECourses which config?

K4bronsky•7/14/25, 2:53 AM

test_24GB_GPU_22900MB_7.3_second_it_Tier_1.json

KK4bronsky test_24GB_GPU_22900MB_7.3_second_it_Tier_1.json

Furkan Gözükara SECourses•7/14/25, 2:56 AM

use 48 gb config

Furkan Gözükara SECourses•7/14/25, 2:56 AM

rtx 5090 can do it since full model uses like 28.5 gb

K4bronsky•7/14/25, 2:56 AM

ok thanks

Furkan Gözükara SECourses•7/14/25, 2:59 AM

you are welcome i got to sleep

FFurkan Gözükara SECourses you are welcome i got to sleep

K4bronsky•7/14/25, 11:38 AM

i changed to 48 but stills runs at the same speed.
ae = "X:/AI/Kohya/ae.safetensors" do you see anything wrong with the toml?
I really appreciate your help
blocks_to_swap = 0
bucket_no_upscale = true
bucket_reso_steps = 64
cache_latents = true
cache_latents_to_disk = true
cache_text_encoder_outputs = true
cache_text_encoder_outputs_to_disk = true
caption_dropout_every_n_epochs = 0
caption_dropout_rate = 0

K4bronsky•7/14/25, 11:38 AM

single_blocks_to_swap = 0
t5xxl = "X:/AI/Kohya/t5xxl_fp16.safetensors"
t5xxl_max_token_length = 512
timestep_sampling = "sigmoid"
train_batch_size = 1
train_blocks = "all"
train_data_dir = "X:\AI\Kohya\SwarmUI_Model_Downloader_v54\SwarmUI\Models\diffusion_models\img"
vae_batch_size = 4
wandb_run_name = "miakillamodel"
xformers = true

KK4bronsky i changed to 48 but stills runs at the same speed. ae = "X:/AI/Kohya/ae.safetens...

Furkan Gözükara SECourses•7/14/25, 11:39 AM

wait at least 100 steps

Furkan Gözükara SECourses•7/14/25, 11:39 AM

also show me nvitop output

FFurkan Gözükara SECourses wait at least 100 steps

K4bronsky•7/14/25, 11:42 AM