Fix bug in loss for trees with one edge from leave to parent

[baumgach]

Hi there, thanks for making the code for this project public!

I think there may be a bug in the calculation of the hierarchical cross-entropy loss for the edge case of having only one edge between leaf and parent. Specifically, on this line https://github.com/fiveai/making-better-mistakes/blob/master/better_mistakes/model/losses.py#L77, j+1 is used to fill in the last column. However, when len(edges_from_leaf[i]) is 1, the loop for the denominator is never entered, so j=0, so onehot_den_np[i, :, 1] = 1.0. However, since there is only one leave, actually it should be onehot_den_np[i, :, 0] = 1.0.

The bug can be fixed by changing line 77 to this:

self.onehot_den[i, :, len(edges_from_leaf[i]) - 1] = 1.0

This should work in every case. Let me know if you agree with me.

[rhomu]

Thanks and good catch! Yes your bug fix makes sense... but it seems you have len(edges_from_leaf) instead of len(edges_from_leaf[i]) in the commit.

Happy to merge when it is fixed :-)