Shapes not documented properly in MarginRankingLoss Β #42884
Description
π Documentation
This issue refers to #42864, but the original issue was closed by the author while I was writing my response to it.
These are the docs at the moment:
Based on the source the correct input/target shape is:
- Input 1: (N, β) where β means, any number of additional dimensions
- Input 2: (N, β) same shape as Input 1, or it has to be broadcastable to that shape
- Target: (N, β), same shape as Input 1 or Input 2, or it has to be broadcastable to that shape
So I think this part in the docs should be fixed and an example of how to use the loss would be helpful.