search

marinbenc / medical-polar-training

41 stars 7 forks source link

Using the polyp dataset, run train_ hourglass. py, DSC is always kept at a value of -0.0032 #2

Closed CTGUzhangguangyi closed 2 years ago

CTGUzhangguangyi commented 2 years ago

Using the polyp dataset, run train_ hourglass. py, DSC is always kept at a value of -0.0032. How can I solve this problem.

marinbenc commented 2 years ago

Hello, the stacked hourglass network does not use the Dice coefficient as a metric, so I'm not really sure what the problem is, unless I misunderstood your question. The stacked hourglass uses the MSE from each stack's heatmap compared to the GT heatmap, as defined here:

https://github.com/marinbenc/medical-polar-training/blob/91f9540a07dc63138faa23c36031236d4e616b8c/models/stacked_hourglass/stacked_hourglass.py#L19