Pseudo Huber Loss

[hao-pt]

Could you share some insights into why you use a fixed coefficient c like this: c = 0.00054 * math.sqrt(math.prod(input.shape[1:])) rather than a specific user's defined value like 0.001 in the paper.

[Kinyugo]

I obtained that specific scaling from table 1.

[hao-pt]

Thank you, I just found that in the paper.