Description
I am seeking clarification on the AverageMeter and metric reduction implementations in 3d_segmentation/swin_unetr_brats21_segmentation_3d.ipynb.
For instance, here is the AverageMeter implementation provided in the brats swinunetr tutorial:
class AverageMeter(object):
def __init__(self):
self.reset()
def reset(self):
self.val = 0
self.avg = 0
self.sum = 0
self.count = 0
def update(self, val, n=1):
self.val = val
self.sum += val * n
self.count += n
self.avg = np.where(self.count > 0, self.sum / self.count, self.sum)
It's similar to the ImageNet implementation apart from the last line.
ImageNet's last line: self.avg = self.sum / self.count
The reason why I think this could be problematic is because the tutorial does the following:
acc_func(y_pred=val_output_convert, y=val_labels_list)
acc, not_nans = acc_func.aggregate()
run_acc.update(acc.cpu().numpy(), n=not_nans.cpu().numpy())
This implies that anytime the model has a nan prediction and the ground truth is NOT NAN (or vice versa), the resulting dice coefficient is NOT penalized as though the prediction has a dice coefficient of zero.
For instance, let's assume the ground truth is has pixels in it, but the model predicts nothing for one class, resulting in these dice values: [0.0000, 0.8775, 0.1213]
The way that AverageMeter is currently used might simply ignore the 0 dice value in the first class if DiceMetric returned nan for the corresponding class.
However, NaNs are caused by the ground truth being NaN. What happens when the ground truth is NaN AND the predicted is not NaN, because that is also not desirable (i.e. false positive)?
Edit: please see my next comment for further clarification.
To Reproduce
Just run the notebook or use the AverageMeter implementation within your own code.
Expected behavior
Would it be possible to refactor the aggregation/average meter to yield zero if the ground truth is NaN and the predicted is NOT NaN?