nll_loss requires log-probabilities as input, which is not given in our code

[moritzschaefer]

@MarcoMorik

https://pytorch.org/docs/stable/generated/torch.nn.NLLLoss.html states that

The input given through a forward call is expected to contain log-probabilities of each class. [...] Obtaining log-probabilities in a neural network is easily achieved by adding a LogSoftmax layer in the last layer of your network. You may use CrossEntropyLoss instead, if you prefer not to add an extra layer.

    "relative_ce": partial(torch.nn.functional.nll_loss, reduction='sum'),
    "relative_cdf": lambda output, label: torch.nn.functional.nll_loss((output+1e-10).log(), label, reduction="sum")

The way we use these losses does not seem to guarantee that our inputs correspond to log-probabilities. Let me know if you have thoughts about this and if not, let's maybe normalize the outputs of the comparison to adhere to the requirements