Test speedup miou
resolve #32 (closed)
Add two functions in commons/metrics.py
- get_confusion_matrix_torch
- get_iou_metrics_torch
Use theses new functions in training_engine in place of old (non torch) function.
Also introduce two class attributes in training_engine , self.multilabel , self.micro_iou to configure the mIou computation. By default set to micro=True and multilabel=False to be coherent with previous code.
On lightunet I've observed a speedup from 3min to 20s for validation step.
Edited by samy khelifi