Accepted dimensionality / shapes

[e-pet]

Hi Maarten!

I am a little confused by the tensor shapes that are or are not being accepted.

In my training loop, I actually have 1-D shapes for preds / sensitive features / labels.

But that is not accepted:

import torch
import fairret.statistic

preds = torch.tensor([0.1, 0.2, 0.3, 0.4])
sens = torch.tensor([0, 0, 1, 1])
labels = torch.tensor([0, 1, 0, 1])

fairret.statistic.TruePositiveRate()(preds, sens, labels)

This raises an error message because of this assumption.

If I artificially add another dimension this runs without error, but I still cannot compute a loss:

import torch
import fairret.statistic
import fairret.loss

# len(shape) == 2 now
preds = torch.tensor([[0.1, 0.2, 0.3, 0.4]])
sens = torch.tensor([[0, 0, 1, 1]])
labels = torch.tensor([[0, 1, 0, 1]])

# "works"
fairret.statistic.TruePositiveRate()(preds, sens, labels)
# fails with "ValueError: Invalid dimensions (1, 8) for Parameter value." in some low-level cvxpy function
fairret.loss.KLProjectionLoss(fairret.statistic.TruePositiveRate())(preds, sens, labels)

Could you provide any hints as to how this is supposed to be used?

I believe it might also be helpful to catch some of these issues early and provide descriptive error messages / suggestions how to fix them. :-)

Cheers
-- Eike