[parkjinman98]

Problem

YOLOv10*-face ONNX artifacts could not be postprocessed at all. Every one of
the four models raised:

ValueError: Wrong shape of input: torch.Size([1, 1, 300, 5])

build_postprocess dispatches these correctly to
YOLONMSFreeFaceDetectionPost, and extract_final_outputs already handles
qbcompiler's decode-enabled output triplet. But a single already-decoded
end2end tensor was only recognised at six columns, because
base._normalize_final_detection_tensor is called with a hardcoded
final_det_dim = 6 + n_extra (xyxy, confidence, class id).

A model with nc == 1 has a class column that is constant 0, and
mblt-model-ops strips it at export, so the shipped *-face ONNX emits
(1, N, 5). That never matched, fell through to raw split-head decoding, and
was rejected. The six-column object-detection form (YOLOv10m, YOLOv10s)
was unaffected and worked throughout.

Fix

YOLONMSFreeDetectionPost.extract_final_outputs now restores the implicit
class column when a single-class model presents the five-column form, then
delegates to the existing six-column path rather than duplicating any decode
logic.

At nc == 1 five columns are also the width of a converted xywh candidate
tensor (4 + nc), which is a different thing and belongs to non_e2e. The two
are separated by row count: an end2end export is top-k selected and never
exceeds max_det, while a converted tensor carries one row per anchor (8400 at
640x640). Only the bounded form is treated as decoded detections.

Validation

• All 24 *-face models in mblt-model-ops now agree with their model-local
  common/local_inference.py on identical inputs: 24/24 on ONNX and 24/24 on
  MXQ across all three decode families (9 anchor, 11 anchorless, 4 nmsfree),
  scores within 0.005.
• Before the fix the four nmsfree models failed 4/4 on ONNX; anchor and
  anchorless were already 20/20, so the change is scoped to the broken path.
• Six-column end2end is unchanged: YOLOv10m still returns 160 multi-class
  detections, YOLOv10s 183.
• Package suite: 766 passed. The 3 remaining failures
  (test_mxq_inference.py::test_mxq_classification_runs_with_shared_npu_options
  and two test_wrapper.py dflfree-pose cases) fail identically with this
  change reverted.
• mblt-model-ops can now drop face_detection from its parity harness
  EXCLUDED_TASKS once a release carrying this fix is pinned; with it applied,
  parity covers 350 models including all 24 face models.

One open question, not addressed here

While validating, a second divergence surfaced that is a product decision
rather than a bug fix, so it is deliberately left out of this PR.

mblt-vision caps NMS output at max_det=300 for every task. Every face model's
common/local_inference.py in mblt-model-ops is uncapped, as is
datasets/widerface/evaluator.py. On WiderFace this is reachable and material:
at the committed conf_thres: 0.001, 5 of 12 sampled 0--Parade images exceed
300 detections and one reaches 1277. Because the Hard subset supplying the
primary metric is exactly the dense, tiny-face regime, a 300-detection cap
discards real recall and will report a lower Hard AP than the repository does on
the same artifact.

The NMS-free path is unaffected — its selection is architecturally top-300, so
both sides already agree there.

Worth deciding whether YOLOFaceDetectionMixin should raise or remove the cap
for WiderFace, or whether the repository should adopt the 300 cap instead. The
parity adapters currently pin the difference to the cap alone: both sides must
agree on every detection mblt-vision returns, and the repository's surplus rows
must be strictly the lowest-scoring ones, so a genuine decode disagreement still
fails.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-03T01:03:35.690732Z	449cd7b	Manual request

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 7ddcd64064

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[parkjinman98]

@codex review

All three P2 findings from the 7ddcd64 review are fixed in 452d509, and the
threads are resolved. Please review 452d509 on its own terms rather than
re-deriving the earlier design — a summary of what changed and, more usefully,
which properties are now load-bearing, so the same three points are not
re-raised against code that no longer works that way.

What changed

1. Row count is gone as a discriminator. You were right that candidate
   count is not provenance. The tensor.shape[1] > self.max_det bound is
   deleted. The two five-column layouts are now separated by an exact
   property rather than a heuristic: a converted tensor carries precisely one
   row per anchor, and the configuration already knows its anchor count via
   anchors_as_tensor().shape[-1]. A tensor of exactly that height is returned
   as None and left to the conversion path.

   Your 64x64 three-head example is the direct regression case: anchors resolve
   to 84, and (B, 84, 5) is now left to filter_conversion/dual_topk, where
   the old bound accepted it because 84 <= 300.

2. Batch axis preserved. The old loop squeezed whichever leading axis was
   1, which dropped the batch when B > 1. It now squeezes axis 1 specifically.
   (2, 1, 300, 5) -> (2, 300, 6).

3. Non-E2E export contract honored. When self.e2e is false the restored
   tensor is passed through _stack_topk_outputs, matching the decoded-triplet
   branch above, so callers get (B, max_det, 6) rather than a per-image list.

Deliberate decisions — please do not re-raise these as defects

• Ties are read as converted. When a configuration's anchor count equals an
  export's row count the tensor is treated as converted. This is the
  conservative direction on purpose: the full decode path runs, and a genuine
  export of that shape fails loudly on raw-head decoding rather than silently
  skipping candidate selection. Preferring the opposite would trade a loud
  failure for a silent behavior change.

• The restore is gated on self.nc == 1. Multi-class models are untouched;
  six-column end2end handling is entirely unchanged. YOLOv10m still returns
  160 multi-class detections and YOLOv10s 183.

• The max_det=300 NMS cap is out of scope for this PR. It is described in
  the PR body as an open question, not an oversight. mblt-vision caps every
  task at 300 while mblt-model-ops' face inference and its WiderFace evaluator
  are uncapped, which is material on crowd images. Deciding that changes
  published accuracy numbers, so it is deliberately not bundled into a crash
  fix. The NMS-free path is unaffected because its selection is already
  top-300.

Validation already performed, if useful as context

• Package suite: 766 passed. The 3 failures
  (test_mxq_inference.py::test_mxq_classification_runs_with_shared_npu_options
  and two test_wrapper.py dflfree-pose cases) reproduce identically with this
  branch's changes reverted, so they are pre-existing and unrelated.
• All 24 *-face models in mblt-model-ops agree with their model-local
  inference on identical inputs: 24/24 on ONNX and 24/24 on MXQ, across all
  three decode families (9 anchor, 11 anchorless, 4 nmsfree).
• mblt-model-ops parity now covers 350 models including all 24 face models,
  green against this branch.

The most valuable thing to scrutinize is point 1: whether anchor count is a
sound invariant for identifying a converted tensor in every configuration this
class supports, including any where anchors_as_tensor() is unavailable (that
path returns None and the restore is skipped).

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 452d5099fd

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[parkjinman98]

@codex review

Fixed in 449cd7b; thread resolved. Please review 449cd7b.

You were right, and the finding also caught me asserting something false: my
previous comment claimed the unresolved-anchor path "returns None and the
restore is skipped". It did not — tensor.shape[1] == None is false for every
integer row count, so the guard disappeared instead of skipping. I reproduced it
before fixing: with anchors forced unresolvable, a converted (1, 8400, 5)
tensor (an exact anchor-count match, the clearest possible converted input) was
restored to (1, 8400, 6) and treated as final.

_anchor_grid_size() is now resolved into a local first, and restoration is
skipped when it is None, so the guard is fail-closed and consistent with the
tie-breaking rationale already in the docstring.

Verified behavior after the fix

Input	anchors resolvable	Result
converted (1, 8400, 5)	no	None — skipped
export (1, 300, 5)	no	None — skipped
export (1, 300, 5)	yes	restored (1, 300, 6)
converted (1, 8400, 5)	yes	None — left to conversion
converted (1, 84, 5) @ 64x64	yes	None — left to conversion

The shipped path is unaffected: anchors resolve at construction, before any
inference, so this is defensive only.

Still deliberate, please do not re-raise as defects

Restating these because they are settled decisions, not oversights, and the
previous round confirmed the resolved threads are staying resolved:

• Ties and unresolved anchors both resolve toward "converted." In both
  cases the full decode path runs and a genuine export of that shape fails
  loudly on raw-head decoding. This is chosen over the opposite direction,
  which would trade a loud failure for a silent behavior change.
• The restore is gated on self.nc == 1. Multi-class is untouched and
  six-column end2end handling is unchanged (YOLOv10m 160 detections,
  YOLOv10s 183).
• The max_det=300 NMS cap is out of scope, documented as an open question
  in the PR body. mblt-vision caps every task at 300 while mblt-model-ops' face
  inference and its WiderFace evaluator are uncapped, which is material on
  crowd images. Deciding it changes published accuracy numbers, so it is not
  bundled into a crash fix. The NMS-free path is unaffected — its selection is
  already top-300.
• The 3 failing tests in this repo are pre-existing.
  test_mxq_inference.py::test_mxq_classification_runs_with_shared_npu_options
  and two test_wrapper.py dflfree-pose cases reproduce identically with this
  branch's changes reverted. 766 pass.

Re-validated at 449cd7b

• Package suite: 766 passed, 3 pre-existing failures.
• All 24 *-face models in mblt-model-ops agree with their model-local
  inference: 24/24 on ONNX across the 9 anchor / 11 anchorless / 4 nmsfree
  families.
• mblt-model-ops parity green across 350 models, including all 24 face models.

If a further finding is a style or structural preference rather than a
behavioral defect, please say so explicitly so it can be weighed against the
above rather than treated as a blocker.

[chatgpt-codex-connector]

Codex Review: Didn't find any major issues. You're on a roll.

Reviewed commit: 449cd7b53b

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".