Tresholding the prediction image to binary before sending to loss function

[KFrank]

Hi Andrei (and Einrone and Arul)!

@Andrei_Cristea

Please do check out InnovArul’s reply and link above as well where is suggesting you try something called “straight through estimator”

For completeness, let me link to some further comments on this issue
that I posted in the aforementioned thread:

The parameters of the model with custom loss function doesn't upgraded thorough its learning over epochs

Hi Arul! I am aware that this “straight-through estimator” trick is recommended from time to time as a work-around for zero gradients (or a broken computation graph) in the presence of thresholding. Let me expand a little on what I see as its problems: First, it is easy to implement and use “soft” thresholds for which you get useful (and correct) gradients. So, what is the benefit of using something that has demonstrable problems? (The fact that you might want to look at a performance…

Best.

K. Frank