CoxPHLoss does not handle batches where all samples are censored

[lenbrocki]

I'm training a custom model with the CoxPHLoss and have noticed that when using the Efron tie method the training will fail when a batch only contains censored events. The code giving the error is in lassonet/utils.py:

if hasattr(torch.Tensor, "scatter_reduce_"):
    # version >= 1.12
    def scatter_reduce(input, dim, index, reduce, *, output_size=None):
        src = input
        if output_size is None:
            output_size = index.max() + 1
        return torch.empty(output_size, device=input.device).scatter_reduce(
            dim=dim, index=index, src=src, reduce=reduce, include_self=False
        )

else:
    scatter_reduce = torch.scatter_reduce

When all samples are censored index will be an empty tensor and index.max() fails.
Also, if I understand correctly, the Cox likelihood would be zero in that case so that the log likelihood is not defined.
For now I have resorted to skipping these problematic batches, but I was thinking that it might be helpful to handle this edge case directly in CoxPHLoss. Not sure what's the best way of doing it though.

[louisabraham]

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?

[louisabraham]

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

[lenbrocki]

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.

[louisabraham]

I'll look at it tonight. But an empty product is 1, not 0, hence the log will be 0. Le lun. 8 janv. 2024, 09:57, lenbrocki ***@***.***> a écrit :

…

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. — Reply to this email directly, view it on GitHub <#52 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADEQQFJM3G6QQAZPKXGMPF3YNOYIRAVCNFSM6AAAAABA6M7RZ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBQGU4TQNJYGM> . You are receiving this because you commented.Message ID: ***@***.***>

[lenbrocki]

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.

[louisabraham]

I tried a fix, can you test it? Le lun. 8 janv. 2024, 10:37, lenbrocki ***@***.***> a écrit :

…

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. — Reply to this email directly, view it on GitHub <#52 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADEQQFIFJVKAUTX3FGRJWXTYNO44BAVCNFSM6AAAAABA6M7RZ6VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTQOBQGY2TIOBWHE> . You are receiving this because you commented.Message ID: ***@***.***>

[lenbrocki]

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

[louisabraham]

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

[lenbrocki]

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

[louisabraham]

Great!