PR 4 of #508 — InferenceLayer ABC + SingleInstanceLayer (#512)

[gitttt-1234]

Summary

Closes #512. Lock down the InferenceLayer abstraction with the proof-of-pattern single-instance subclass — and prove parity end-to-end against the PR 0 golden. Every subsequent layer (PR 6) and the Predictor orchestrator (PR 8) follows this template.

What lands

• sleap_nn/inference/layers/configs.py — attrs.frozen PreprocessConfig + PostprocessConfig
• sleap_nn/inference/layers/base.py — InferenceLayer ABC: abstract preprocess / postprocess, concrete predict / __call__ / warmup. Includes _to_4d_float_tensor helper that accepts (H,W), (H,W,C), (C,H,W), (B,H,W,C), (B,C,H,W) numpy or torch — gives every subclass a uniform input contract for the new direct-numpy API
• sleap_nn/inference/layers/single_instance.py — SingleInstanceLayer. Decodes confmaps via ops.peaks.find_global_peaks, applies the full coord ladder via ops.coord, returns Outputs

Headline result

End-to-end parity proven by test_single_instance_layer_parity_vs_pr0_golden:

new_outputs = SingleInstanceLayer(...).predict(image_4d)
np.testing.assert_allclose(
    new_outputs.pred_keypoints.squeeze(1).numpy(),
    golden[\"pred_instance_peaks\"],
    atol=1e-5, rtol=1e-5,
)  # PASSES

The new layer stack — np.ndarray in, Outputs dataclass out — matches the captured-from-old-code golden within 1e-5 atol/rtol on the same fixed input. Linchpin of the refactor: as long as this passes, every subsequent PR (5–14) is verifiably parity-preserving on the single-instance path.

Test plan

tests/inference/layers/ — 22 cases:

• InferenceLayer ABC enforcement (cannot instantiate directly; rejects non-ModelBackend)
• _to_4d_float_tensor: 5 input shapes parametrized + numpy/torch + rejection of unsupported types/ranks
• Parity vs PR 0 golden within 1e-5
• Direct numpy + torch APIs
• 2D grayscale + 4D batched inputs
• return_confmaps off-by-default + opt-in populates Outputs.pred_confmaps
• Synthetic-input coord ladder verification

SingleInstanceLayer.from_checkpoint(...) is deferred to PR 8 (#516) where checkpoint-load logic consolidates. For now the constructor takes an already-built ModelBackend; tests build via the existing Predictor.from_model_paths and pull modules out of inference_model.

Full inference suite: 240 passed, 8 skipped, 2 xfailed.
Parity regen: 30/30 specs match.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 75.07297% with 427 lines in your changes missing coverage. Please review.
✅ Project coverage is 61.66%. Comparing base (00f2d64) to head (33b0c52).

Files with missing lines	Patch %	Lines
sleap_nn/inference/layers/exported.py	0.00%	65 Missing ⚠️
...p_nn/inference/layers/backends/tensorrt_backend.py	25.88%	63 Missing ⚠️
sleap_nn/inference/factory.py	59.15%	58 Missing ⚠️
sleap_nn/inference/layers/centroid.py	62.38%	41 Missing ⚠️
sleap_nn/inference/layers/backends/onnx_backend.py	37.50%	40 Missing ⚠️
sleap_nn/inference/predictor.py	73.77%	32 Missing ⚠️
sleap_nn/inference/layers/topdown_multiclass.py	47.27%	29 Missing ⚠️
sleap_nn/cli.py	85.71%	23 Missing ⚠️
sleap_nn/inference/streaming.py	88.57%	12 Missing ⚠️
sleap_nn/inference/filters.py	92.30%	11 Missing ⚠️
... and 8 more

Additional details and impacted files

@@                           Coverage Diff                           @@
##           divya/inf-refactor-03-torch-backend     #534      +/-   ##
=======================================================================
- Coverage                                63.65%   61.66%   -1.99%     
=======================================================================
  Files                                      109      129      +20     
  Lines                                    17918    19607    +1689     
=======================================================================
+ Hits                                     11406    12091     +685     
- Misses                                    6512     7516    +1004     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.