h1loss output sum, fix for issue #103

[itskobold]

h1loss output is summed to produce a scalar, training would previously break if number of output channels > 1

#103

[kovachki]

For the loss to produce a scalar, the channel dimension must be added to the reduce_dims otherwise it will only reduce over the first dimension (this is the default behavior). Here is a 2d example:

loss = H1Loss(d=2, reduce_dims=[0,1])
x = torch.randn(5,10,128,128)
y = torch.randn(5,10,128,128)
loss(x,y)

This will be scalar.

It is done this way, so that the user has more control over what to do with each channel, for example, weighting each channel differently in your loss function.