search

med-air / 3DSAM-adapter

Holistic Adaptation of SAM from 2D to 3D for Promptable Medical Image Segmentation
134 stars 12 forks source link

Couldn't reach the effectiveness. #12

Open hxp2396 opened 11 months ago

hxp2396 commented 11 months ago

I downloaded the processed kit dataset and followed the instruction to train, but the training 40epoch test dice is very low (0.002), don't know if this is normal. How many epochs did you train to achieve convergence? And the batchsize in the code seems to be unadjustable, which will cause size mismatch. How to solve it?

peterant330 commented 11 months ago

I downloaded the processed kit dataset and followed the instruction to train, but the training 40epoch test dice is very low (0.002), don't know if this is normal. How many epochs did you train to achieve convergence? And the batch size in the code seems to be unadjustable, which will cause a size mismatch. How to solve it?

It normally takes around 200 epochs to converge, but your reported dice is too low so there must be some bug. Sorry that we haven't made the code flexible enough to support multiple batch sizes yet.

I already received your email, we can discuss in detail about what modification you did and how to solve the problem.

liryy commented 11 months ago

I also encountered the same problem. Have you solved this problem? How to solve it? Thank you for your reply.

peterant330 commented 11 months ago

I also encountered the same problem. Have you solved this problem? How to solve it? Thank you for your reply.

Hi,

Could you download the latest code and try again? We made a few modifications in the past few weeks.