From the documentation here, it seems that the loss is mainly the element-wise squared differences between two vectors.
reduction argument should be able to handle special cases where
sum or no-reduction is required.
Does this argument make sense?
Given the default
reduction argument value, this is actually doing an MSE.
Is you proposition to introduce a new
nn.SELoss that is the same but with the default
I believe that would make more sense semantically. Writing MSELoss and then not doing mean-squared-error looks counter-intuitive.
In any case, MSELoss is a single metric and should not be overloaded with arguments that make it flexible and thus confusing. Perhaps deprecating it in favour of SELoss, or deprecating the
reduction param so that MSELoss doesn’t require any param would be something that could be considered.