Huber Loss in Torch?

[Eren Gölge]

Do you know any Huber Loss implementation in Torch7? I like to do a outlier robust regression with a CNN. I guess it is a good alternative to MSEcriterion. Or do you know any other way to fight against outliers?

[Lev E Givon]

On Fri, Dec 2, 2016 at 9:33 AM, Eren Gölge via torch7
<torch7+APn2wQfdWeEqhQail37iMKFIz...@googlegroups.com>
wrote:

> Do you know any Huber Loss implementation in Torch7? I like to do a outlier
> robust regression with a CNN. I guess it is a good alternative to
> MSEcriterion. Or do you know any other way to fight against outliers?

You might want to try the SmoothL1Criterion:

https://github.com/torch/nn/blob/master/doc/criterion.md#nn.SmoothL1Criterion
--
Lev E. Givon, PhD
http://lebedov.github.io