Difference between MetricReduction and LossReduction

[anais2390]

Hi everyone !

I am currently working on a 2D binary segmentation problem and I am using the dice coefficient to evaluate performance during training. To do so I use MeanDice() combined with a SupervisedEvaluator:

val_metric = {'mean_dice':MeanDice()}
evaluator= create_supervised_evaluator(net, val_metric , device, non_blocking=True, output_transform=lambda x, y, y_pred: ([postproc_pred(i) for i in decollate_batch(y_pred)], [postproc_label(i) for i in decollate_batch(y)]), prepare_batch=prepare_batch)

As a double check I also add the DiceLoss class to my evaluator such as :

val_metric = {'diceloss':1 - Loss(DiceLoss())}

However, the results return by the two metrics are highly different, for exemple :
INFO:trainer:Epoch[1]: mean_dice: 0.492 diceloss: 0.176

I suppose this difference comes from the Reduction step but i can't figure out why ?
Could you bring me some highlights ?

Thanks !

[Nic-Ma]

Hi @anais2390 ,

Thanks for your feedback.
I think that because we use different reduction computations for Nan in metrics and loss:
https://github.com/Project-MONAI/MONAI/blob/dev/monai/losses/dice.py#L182
https://github.com/Project-MONAI/MONAI/blob/dev/monai/metrics/utils.py#L74
And it should be expected as loss and metrics have different meanings for Nan, we ignore nans in metrics as some labels may be missing.

Thanks.