question: reader loss

[PaulLerner]

Hi,

I have trouble understanding this line in compute_loss in reader.py:

DPR/dpr/models/reader.py

Line 109 in a31212d

loss_tensor = loss_tensor.view(N, M, -1).max(dim=1)[0]

This keeps the maximum loss over all M passages, why? Why not summing or averaging?

Best regards,