Tested training / validating / inferencing (weight only trained for <5 epochs) with binary / multiclass / multilabel.
The only breaking change is with FocalLossSeg, from F.binary_cross_entropy_with_logits to F.binary_cross_entropy because we already have activation in last layer. Check this issue on segmentation_models.pytorch for more detail. Other than that, everything should be identical to how the repo was if the user doesn't switch seg_multilabel on in project.yml.
xoiga123
commented
2 years ago
Tested training / validating / inferencing (weight only trained for <5 epochs) with binary / multiclass / multilabel.
The only breaking change is with FocalLossSeg, from
F.binary_cross_entropy_with_logitstoF.binary_cross_entropybecause we already have activation in last layer. Check this issue on segmentation_models.pytorch for more detail. Other than that, everything should be identical to how the repo was if the user doesn't switchseg_multilabelon inproject.yml.