CoxPHLoss does not handle batches where all samples are censored about lassonet HOT 10 CLOSED

Hello, can you produce a minimal reproducible example?

Also, if I understand correctly, the Cox likelihood would be zero in that case so that the log likelihood is not defined.

wouldn't it be one? Can you test with the Breslow approximation?

from lassonet.

I think that if the sets are empty, the log-likelihood is just zero.

from lassonet.

Sorry for the late response. A minimal example would be:

import torch
from lassonet.cox import CoxPHLoss

loss = CoxPHLoss("breslow")
labels = torch.tensor([[5.0, 0], [2.0, 0]])
hazards = torch.tensor([5.0, 2.0])
print(loss(hazards, labels)) 
#prints nan 

loss = CoxPHLoss("efron")
labels = torch.tensor([[5.0, 0], [2.0, 0]])
hazards = torch.tensor([5.0, 2.0])
print(loss(hazards, labels)) 
#fails with RuntimeError: max(): Expected reduction dim to be specified for input.numel() == 0. Specify the reduction dim with the 'dim' argument.

The error for the Efron method happens because of what I've described above. I think the nan in the breslow case happens because the likelihood is zero. This can for example be seen from your paper https://arxiv.org/pdf/2208.09793.pdf in equation 1: if all $\delta_i$ are zero, the product is zero and then the log of this is not defined.

from lassonet.

from lassonet.

Oh I wasn't aware of that, but yes you're right of course. Then the problem becomes why nan is returned and not 0.

from lassonet.

from lassonet.

The CoxPHLoss now correctly returns 0. But when I'm trying to use the fixed loss in training I'm getting for batches where all samples have $\delta_i = 0$ this error:

RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

from lassonet.

I think I managed to find a better fix :) Can you test again?

from lassonet.

Sorry again for the delay. Yes, it's working now!

from lassonet.

Great!

from lassonet.