Questions for paper
dungscout96 opened this issue · comments
Young Truong commented
Are we normalizing the locations to [0 1]?https://github.com/sccn/sound2meg/blob/d490f91f957eb95c764e6a85ea1a75eef87de8f7/code/utils.py#L104C53-L104C53
The spatial drop out is not being implemented correctly. We are not choosing channel from input channels (currently random instead) and we are including them in the softmax (zeroed out) instead of excluding them entirely:
Line 97 in d490f91
Arnaud Delorme commented
There was another problem, which was that we should remove the dropped channels (instead of setting their weight to 0) from the softmax calculation.