[M-Colley]

Summary

CORNLoss now also accepts probabilistic targets: a (N, J) tensor with a distribution over the classes for every example, for instance a soft label, the uniform distribution over a set of acceptable classes, or the empirical distribution of several annotators. Until now its docstring said "CORN loss does not support probabilistic targets".

With such targets it is the soft-CORN loss: CORN's hard subset membership 1[y >= k] and hard indicator 1[y > k] are replaced by their expectations P(y >= k) and P(y > k) under the target distribution,

L = - sum_i sum_{k=0}^{J-2} [ P_i(y > k) log sigmoid(z_ik) + p_ik log(1 - sigmoid(z_ik)) ]
    / sum_i sum_{k=0}^{J-2} P_i(y >= k)

• For one-hot targets it equals the existing loss exactly.
• Up to the normalisation it is the cross-entropy between the target distribution and the class probabilities that the CORN logits define, P(y=j) = prod_{k<j} sigmoid(z_k) (1 - sigmoid(z_j)), so it is a proper scoring rule for the target distribution.
• The hard-label path is not touched: y_true of shape (N,) runs the same code as before. A 2-D y_true takes a new vectorised path, _soft_forward.
• It uses the suffix sums P(y >= k) (not 1 - cumsum) and logsigmoid, so it stays finite for extreme logits.

This is not a new loss (it is the continuation-ratio likelihood that CORN already optimises, applied to a soft target); the PR only adds support for probabilistic targets, which fits the soft labelling methods already in the library.

Benchmark

The docstring and docs/references.bib cite a benchmark of this loss against eleven others for learning from set-valued ordinal labels: https://github.com/M-Colley/soft-corn-benchmark (12 losses including soft cross-entropy, soft-CORAL / soft-OR, OLDL, partial-label likelihood and the hard-label CORN / CORAL / CE; five real datasets, synthetic families, fine-tuned backbones, and an in-vehicle study with natively set-valued answers). In short, on the frozen-feature benchmark soft-CORN is clearly better on likelihood than the hard-label losses, OLDL and partial-label likelihood, essentially tied with soft cross-entropy, and practically tied with soft-CORAL on likelihood while better on set-MAE and at larger training sizes; the report (REPORT.md) lists where it loses as well.

Tests

Added to dlordinal/losses/tests/test_corn_loss.py:

• one-hot probabilistic targets give exactly the loss of the class indices;
• the loss equals the CORN-pmf cross-entropy divided by the expected size of the conditional sets (independent reference);
• a hand computation for a uniform-over-set target {1, 2};
• zero loss for a point mass with saturated logits;
• finite loss and gradients for extreme logits;
• gradient descent on one example recovers the target distribution;
• ValueError for wrong target or logit shapes.

I also checked the new path against coral_pytorch.losses.corn_loss on one-hot targets and against an independent implementation of soft-CORN on 60 random cases (maximum difference 5e-7 in float32).

pytest dlordinal/losses: 237 passed (the CUDA parametrisations are skipped on the CPU-only environment I used). ruff, isort and black as configured in .pre-commit-config.yaml are clean, and the Sphinx docs build renders the new math and the citation (the only docs warning is in GaussianUncertaintyLossWrapper.forward, which this PR does not touch).

Notes

• The README asks contributors to contact the maintainers before doing any work; I went straight to a PR, so please say if you would rather see a different shape for this (a separate class or a flag instead of dispatching on y_true.ndim) and I will change it.
• Probabilistic targets are not validated to be non-negative or to sum to one; the docstring states the requirement.
• I did not touch CHANGELOG.md / changelog_versions, which look like they are written at release time.

🤖 Generated with Claude Code

[victormvy]

Thanks for the solid PR!

To answer your notes:

• I think that dispatching on y_true.ndim is super clean and keeps the API straightforward.
• Documenting the requirement in the docstring instead of adding runtime checks makes total sense for performance.
• Regarding the changelog, no problem at all, we usually update that at release time anyway.

One quick thing: it looks like the benchmark repository linked in the docstring is currently set to private. Could you please make it public so users can access it?

I'll do a quick code review in the meantime, but overall this looks ready to go. Thanks again for the great contribution!

[M-Colley]

Ah thanks! I have made the benchmark public :)

https://github.com/M-Colley/soft-corn-benchmark/

[victormvy]

Thanks!

[M-Colley]

Just for completeness, this work was done together with @DaniPulidoE :)