Closed CharlesGaydon closed 2 months ago
Tested on a small patch, with 1 and 2 gpus. Had to add a small related fix in model.py that apparently was not dealt with in the previous bugfix MR.
Small diff since we are dealing with a single patch + we now use lightning's version instead of a manual computation from a confusion matrix.
The diff disapear on larger datasets :
Log more metrics. Keep the model.py clean. It is hard to unit-test that metrics are actually logged, I'll do some manual checks on known datasets, in mono and multi gpu.