Confusing documentation for mx.sym.MakeLoss

[fhieber]

The documentation of mx.sym.MakeLoss is highly confusing. To my understanding the only thing MakeLoss does is that it wraps an existing symbol and defines that it does not require a head gradient when used in optimization. Further, it seems that the output of the forward() call to a MakeLoss-symbol is the output of the forward() call of the wrapped symbol. That is, MakeLoss just passes through its input data in forward().

However, the documentation states the following:
"The output of this function is the gradient of loss with respect to the input data."

What does this mean? I read it as if the output of forward() is the same as the output of backward(), namely the gradient of the symbol, MakeLoss wraps. But this does not seem to be true.

[andrewfayres]

@mxnet-label-bot add [documentation, question]

[andrewfayres]

@mxnet-label-bot add [doc]