I am trying to create a model that outputs the PMF only for the first m time steps in the future, but can handle samples with survival > m. As far as I understand the paper, phi_m+1(x) is not a real output of the network but just set to 0, which makes sense at it can be inferred (by the softmax) from the rest of the networks output due to the sum to 1 constraint. I assume that's also the reason for pad_col (might be the answer to #152),
I am trying to create a model that outputs the PMF only for the first m time steps in the future, but can handle samples with survival > m. As far as I understand the paper, phi_m+1(x) is not a real output of the network but just set to 0, which makes sense at it can be inferred (by the softmax) from the rest of the networks output due to the sum to 1 constraint. I assume that's also the reason for pad_col (might be the answer to #152),
https://github.com/havakv/pycox/blob/0e9d6f9a1eff88a355ead11f0aa68bfb94647bf8/pycox/models/loss.py#L83
which adds the static output of 0 for phi_m+1. However, I do not understand the purpose of the exception
https://github.com/havakv/pycox/blob/0e9d6f9a1eff88a355ead11f0aa68bfb94647bf8/pycox/models/loss.py#L75-L78
as the function apparently can handle survival beyond m.
Could you maybe explain why this constraint would be still needed or if it could be safely removed?