Open zaptrem opened 1 year ago
Thanks @zaptrem.
After taking a look into this I see why it wasn't implemented in the first place. Handling operation for reduction = "none"
is non-trivial for any loss that has multiple sub-components. We have been thinking for a while to provide a different interface for computing losses that would involve a Transform
and a Distance
which would make this behavior more clear. We might have to wait for this interface in order to support this feature. However, this use-case provides more reasoning to work on this.
I will keep this issue open for now, but follow #16 if you are interested.
I'm currently trying to apply a different scalar to the loss of each element in a batch , but
SumAndDifferenceSTFTLoss
doesn't have the option to return element-wise losses (in most loss functions this isreduction="none"
). Is this feasible to add?