What's the behavior of BCEWithLogitsLoss for multi-label data

1

Hi,

I have a noob question about the behavior of BCEWithLogitsLoss. I tested its loss with some data

image
image1696×1400 253 KB
.

My understanding is that the loss between 0.999(x) and 1(y) should be 0.3135 and loss between 0.999(x) and 0(y) 1.3125. Thats the case when target and output are of size(1, #sample), which is binary classification. However, when I reshape the data to size(#size, #class) to simulate multi-label, the loss is different (0.3143, 0.6931).

So how will BCEWithLogitsLoss calculate loss for multilabel. My understanding for the doc is they should be the same. Thank you!

2

You are passing the output and target in the wrong order in your second example.
Pass the output first, then the target, and you’ll get the same result. :wink:

1 Like
3

Thank you! I feel bad for myself lol