MedSegDiff The poor performances of the multi-labels BraTs dataset

The poor performances of the multi-labels BraTs dataset

Open smallboy-code opened this issue 1 year ago • 8 comments

The raw images:

visdom_image (3) visdom_image (2) visdom_image (1)

The Ground Truth:

The output sample of 5 ensembles:

So why these outputs contain the brain boundary? And how to set 0,1, 2, 4?

Oct 16 '23 12:10 smallboy-code

I would like to know your batch and the number of training steps

Oct 24 '23 08:10 princerice

batch_size: 32 and training steps : 30000

Oct 24 '23 08:10 smallboy-code

Thank you for your answer. My batch is set to 8 and I use A5000GPU. I would like to know the size of the GPU memory you use

Oct 24 '23 11:10 princerice

I using the four A100 GPUs. And I get the normal mask without brain tissue boundary, but the Dice is lower, about 0.4. So do you have this issue?

Oct 24 '23 11:10 smallboy-code

I trained 40,000 steps on a gpu and the results are not good and we may need to adjust and improve ourselves

Oct 24 '23 11:10 princerice

I'm also confused that the loss has leveled off but the results are far from satisfactory

Oct 24 '23 11:10 princerice

Yes, and I want to know the clamp range in the final "process_xstart" function. I am use (0,3).

Oct 24 '23 11:10 smallboy-code

This part is not clear to me yet, I am just trying to reproduce and learn the code

Oct 24 '23 13:10 princerice