We were using meddlr for training a recon + segmentation model using a sum of MAE (for recon) and the builtin functional dice_score. However, during training DSC was not improving at all. I took a peak at the function and wondered if the bool operator was somehow causing an issue in backdrop. I implemented a regular continuous dice and it seems to have fixed the issue.
Not sure if the dice_score was only meant for testing and not training? If so, maybe that should be documented? Happy to make a pull request if adding this is helpful.
ad12
commented
1 year ago
Hey,
We were using meddlr for training a recon + segmentation model using a sum of MAE (for recon) and the builtin functional dice_score. However, during training DSC was not improving at all. I took a peak at the function and wondered if the
booloperator was somehow causing an issue in backdrop. I implemented a regular continuous dice and it seems to have fixed the issue.Not sure if the dice_score was only meant for testing and not training? If so, maybe that should be documented? Happy to make a pull request if adding this is helpful.