Score SDE Score Matching

Use this skill when a recovery or implementation task needs the continuous denoising score-matching objective from the Score SDE paper. It is designed to consume marginal kernel outputs from a Score SDE implementation and can also run a deterministic reduced optimizer step for mechanism evidence.

Do not use this skill as a full neural-network trainer. It does not implement NCSN++, DDPM++, JAX, Flax, or dataset input pipelines.

Inputs

• Clean data batch x0.
• Marginal means and standard deviations from a forward SDE kernel.
• Gaussian noise values and sampled times.
• A score model callable or the bundled linear score fixture.
• Loss mode: unweighted or likelihood weighted by g(t)^2.

Outputs

• Perturbed examples.
• Conditional score targets for Gaussian perturbations.
• Per-example loss values and scalar average loss.
• Reduced optimizer trace with parameters before and after an update.

Workflow

1. Build a batch using build_score_matching_batch.
2. Evaluate score predictions with an injected model.
3. Compute the continuous denoising score-matching loss.
4. For reduced recovery, call optimizer_step on a LinearScoreModel.
5. Save loss_before, loss_after, params_before, and params_after when using the optimizer trace as experiment evidence.

Validation

Run:

python scripts/score_matching.py --self-test
python tests/test_score_matching.py

The tests check target construction, likelihood weighting, and a real parameter update.

Limitations

The reduced optimizer uses finite differences for portability. Full paper reproduction requires a large neural score model and many training steps, which are outside this skill's bounded test contract.