Discretization error about pycox HOT 3 OPEN

A further question on discretization: is there a recommended value with which to set num_durations based on the variable time? I noticed an improvement with 20 bins compared to the 10 recommended in the tutorial (DeepHitSingle). The range of the variable goes from 0 to 1600.

Thanks again and I apologize for the previous question.

from pycox.

Thank you for the kind words, and I'm happy that you figured out your issue!
No need to apologize. If this is a reoccurring issue, we should probably check the event values and raise a proper error explaining what's going wrong with the inputs.

When it comes to the choice of num_durations, you can consider it a hyperparameter that needs to be tuned. Note that it affects the likelihood, so you cannot use the validation loss for tuning this parameter. Use something like the concordance index or brier score instead.
Alternatively, you can replace it with your own defined grid, if you have some knowledge what a good discretization grid would be. Or if you have discrete event times, you might just want to use those.

In the tutorial notebooks, I've just set some simple hyperparameters which are not tuned to be optimal. So you'll likely be able to to improve on the result of these examples, like you did with the 20 bins :)

from pycox.

In this paper we looked into some discretization schemes, so you can take a look there for more information

from pycox.