Skip to content

Conversation

ahmedhshahin
Copy link
Contributor

Hello,

I think the current implementation of CoxLoss is not accurate, below I discuss the two issues I identified and their solutions:

  1. This line
    loss_cox = -torch.mean((theta - torch.log(torch.sum(exp_theta*R_mat, dim=1))) * censor)
    this part (theta - torch.log(torch.sum(exp_theta*R_mat, dim=1))) returns a 1d tensor, if censor is a 2d tensor (which is the case most of the time, as it is not squeezed in the function CoxLoss), then pytorch will have to broadcast, and this multiplication will return a 2d matrix.

Toy example:

theta = torch.tensor([-0.282,-0.1411,-0.1039,-0.0255])

exp_theta = torch.exp(theta) 
R_mat = np.array([[1,0,1,1],[1,1,1,1],[0,0,1,1],[0,0,0,1]])
R_mat = torch.FloatTensor(R_mat)
censor = torch.tensor([[0,1,0,0]]).T

print(R_mat)
tensor([[1., 0., 1., 1.],
        [1., 1., 1., 1.],
        [0., 0., 1., 1.],
        [0., 0., 0., 1.]])
print(censor.size())
torch.Size([4, 1])

print((theta - torch.log(torch.sum(exp_theta*R_mat, dim=1))))
tensor([-1.2491e+00, -1.3935e+00, -7.3312e-01, -2.2352e-08])

# cool, 1d vector, as expected. However, if censor is a 2d tensor, then it broadcasts
print((theta - torch.log(torch.sum(exp_theta*R_mat, dim=1))) * censor)
tensor([[-0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00],
        [-1.2491e+00, -1.3935e+00, -7.3312e-01, -2.2352e-08],
        [-0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00],
        [-0.0000e+00, -0.0000e+00, -0.0000e+00, -0.0000e+00]])
# this results in an inaccurate loss calculation

The fix to this is to ensure censor is a 1d tensor, as I propose here.

  1. Here you calculate the loss by averaging over the whole batch. I believe one should calculate the average over the subjects who experienced the event only.

Thanks

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

1 participant