Release v0.24.1#81
Merged
Merged
Conversation
Signed-off-by: Sébastien Taylor <[email protected]>
Contributor
There was a problem hiding this comment.
Pull request overview
Patch release v0.24.1 intended to fix a contract mismatch where Decoder::normalized_boxes() (and C/Python equivalents) reported the schema annotation rather than the post-decode coordinate space actually returned to callers.
Changes:
- Update
Decoder::normalized_boxes()semantics/docs to report the post-decode coordinate space. - Update Rust/C/Python documentation to describe the corrected contract.
- Update/extend tests and bump workspace + Python package versions to
0.24.1.
Reviewed changes
Copilot reviewed 12 out of 13 changed files in this pull request and generated 3 comments.
Show a summary per file
| File | Description |
|---|---|
| NOTICE | Update crate version attributions to 0.24.1. |
| crates/python/src/decoder.rs | Update Python API docs for normalized_boxes / input_dims. |
| crates/python/pyproject.toml | Bump Python package version to 0.24.1. |
| crates/decoder/tests/per_scale_parity.rs | Update per-scale parity assertion to new normalized_boxes() contract. |
| crates/decoder/tests/decoder_normalized_flag.rs | Update assertions for new contract and add regression test for input_dims == None. |
| crates/decoder/src/decoder/tests.rs | Update internal test expectation for Hailo YOLOv8-seg schema. |
| crates/decoder/src/decoder/mod.rs | Implement new normalized_boxes() logic + updated rustdoc. |
| crates/decoder/src/decoder/builder.rs | Adjust per-scale normalization commentary to match new contract. |
| crates/capi/src/decoder.rs | Update C API docs describing hal_decoder_normalized_boxes. |
| crates/capi/include/edgefirst/hal.h | Update public header docs for hal_decoder_normalized_boxes. |
| CHANGELOG.md | Add v0.24.1 changelog entry describing the contract fix. |
| Cargo.toml | Bump workspace + internal crate versions to 0.24.1. |
| Cargo.lock | Lockfile updates reflecting 0.24.1 versions. |
Comment on lines
302
to
+339
| @@ -316,21 +325,35 @@ impl Decoder { | |||
| /// # } | |||
| /// ``` | |||
| pub fn normalized_boxes(&self) -> Option<bool> { | |||
| self.normalized | |||
| // Report the effective post-decode coordinate space. The | |||
| // internal `self.normalized` is the *input* policy fed to | |||
| // `yolo::maybe_normalize_boxes_in_place`; that helper divides | |||
| // by `input_dims` exactly when `normalized == Some(false)` and | |||
| // `input_dims` is a non-zero `(w, h)`. The output the caller | |||
| // sees is therefore normalized in that case — surface that. | |||
| match (self.normalized, self.input_dims) { | |||
| (Some(true), _) => Some(true), | |||
| (Some(false), Some((w, h))) if w != 0 && h != 0 => Some(true), | |||
| (Some(false), _) => Some(false), | |||
| (None, _) => None, | |||
| } | |||
Comment on lines
+961
to
+966
| /// Reports the **post-decode** state, not the raw schema annotation: | ||
| /// when the schema declares pixel-space outputs but | ||
| /// :attr:`input_dims` is known, the decoder internally divides bbox | ||
| /// channels by ``(W, H)`` before returning, so the boxes the caller | ||
| /// receives are in ``[0, 1]`` and this getter returns ``True``. | ||
| /// Callers must not re-normalize in that case. |
Comment on lines
+1337
to
+1346
| * Reports the **post-decode** state, not the raw schema annotation: when | ||
| * the schema declares pixel-space outputs but `hal_decoder_input_dims()` | ||
| * is known, the decoder internally divides bbox channels by `(W, H)` | ||
| * before returning, so the boxes the caller receives are in `[0, 1]` | ||
| * and this function returns `1`. Only when no input dimensions are | ||
| * available does pixel-space leak out and this function return `0`. | ||
| * | ||
| * Callers MUST NOT re-normalize when this function returns `1`; | ||
| * dividing already-normalized coordinates by `(W, H)` again collapses | ||
| * detections to ~0. |
Signed-off-by: Sébastien Taylor <[email protected]>
Per-scale decoders: accessor reports post-decode coordinate space (Some(false) + valid input_dims upgrades to Some(true)). All other decoders: raw schema annotation returned verbatim. C header regenerated via cbindgen from updated Rust doc. CHANGELOG entry tightened to match actual fix scope. Signed-off-by: Sébastien Taylor <[email protected]>
- Flip accessor assertions in pixel_space_input_with_normalized_false_decodes and split_schema_pixel_space_with_normalized_false_decodes from Some(true) to Some(false): non-per-scale paths surface the raw schema flag. - Rewrite file-level doc-block to describe the per-path contract and point readers at Decoder::normalized_boxes rustdoc; note which sub-claims each test section pins. - Update normalized_false_without_input_dims_reports_false rationale from "input_dims absent so helper cannot run" to "non-per-scale always returns raw flag". - Add normalized_none_reports_none (non-per-scale + None -> None). - Add normalized_true_reports_true (non-per-scale + Some(true) -> Some(true)). - Add non_per_scale_normalized_false_with_input_dims_reports_false_but_normalizes: accessor returns Some(false) AND decoded boxes land in [0,1], pinning both the under-claiming flag and the internal normalizer firing. Signed-off-by: Sébastien Taylor <[email protected]>
Signed-off-by: Sébastien Taylor <[email protected]>
Signed-off-by: Sébastien Taylor <[email protected]>
Signed-off-by: Sébastien Taylor <[email protected]>
Contributor
Author
PR feedback addressed (6 follow-up commits,
|
| Suite | Passed | Failed | Ignored |
|---|---|---|---|
--lib --features tracker |
390 | 0 | 4 |
decoder_normalized_flag |
8 | 0 | 0 |
per_scale_parity |
29 | 0 | 2 |
| doc-tests | 61 | 0 | 0 |
edgefirst-hal-capi |
1 | 0 | 0 |
All gates green locally: cargo check --workspace, make format lint, verify_version, make sbom.
Known remaining gaps (potential follow-up tickets)
ModelType::YoloSplitSegDet: helper fires ondecode(yolo.rs:1282) but NOT on the tracker macroprocess_tracked_yolo_segmentation_split!(~postprocess.rs:1773) or on_protovariants. Asymmetric — accessor correctly under-claimsSome(false). Same-pattern fix would mirror what landed forYoloSegDet.ModelType::YoloSegDet2Way: helper not called anywhere in either tracker or non-tracker paths. Likely emits pixel-space whennormalized: false. Architect's investigation surfaced this as a likely-pre-existing bug independent of the original contract leak. Worth a JIRA ticket.
Both can be deferred — they don't regress against the pre-0.24.1 behavior; the accessor simply doesn't claim a contract it can't honor.
Test plan
- CI: full Rust + Python test matrix
- CI: SBOM + license policy gate
- On-target: gst bridge smoke test on the platform that triggered the original
coords ≈ 0regression, exercising both per-scale andYoloSegDetdecoders (including tracked variants)
…ry points Signed-off-by: Sébastien Taylor <[email protected]>
Four paths now invoke maybe_normalize_boxes_in_place uniformly: per-scale, YoloSegDet, YoloSplitSegDet, YoloSegDet2Way. Update Doxygen, Python docstrings, and CHANGELOG to name all four and drop the stale two-path claim. Regenerate C header via cbindgen. Signed-off-by: Sébastien Taylor <[email protected]>
Update decoder_normalized_flag.rs for fa8e919: flip split_schema_pixel_space_with_normalized_false_decodes to assert Some(true) now that YoloSplitSegDet normalizes uniformly; replace the stale non_per_scale_non_yolo_segdet test (premise invalidated) with detection_only_normalized_false_with_input_dims_reports_false_raw using a YoloDet schema (still returns raw flag); add yolo_split_segdet_tracker_path_normalizes and yolo_segdet_2way_decode_normalizes to pin the new uniform-normalization contract for the tracker and decode entry points on those types; tighten the file-level doc-block to reflect the complete coverage. Signed-off-by: Sébastien Taylor <[email protected]>
Contributor
Author
Round 4 — full uniform-normalization closure (commits
|
| ModelType | decode | decode_proto | decode_tracked | decode_tracked_proto |
|---|---|---|---|---|
| Per-scale | Y | Y | Y | Y |
YoloSegDet |
Y | Y | Y | Y |
YoloSplitSegDet |
Y | Y (new) | Y (new) | Y (new) |
YoloSegDet2Way |
Y (new) | Y (new) | Y (new) | Y (new) |
Total helper callsites: 19 (was 11 at the start of this round).
Accessor (fa8e919)
fn legacy_path_normalizes_uniformly(&self) -> bool {
matches!(
self.model_type,
ModelType::YoloSegDet { .. }
| ModelType::YoloSplitSegDet { .. }
| ModelType::YoloSegDet2Way { .. }
)
}Tests (c252df5)
- Flipped two now-stale assertions (
split_schema_pixel_space_with_normalized_false_decodes, the YoloSplitSegDet under-claim regression). - Refactored
non_per_scale_non_yolo_segdet_normalized_false_with_input_dims_reports_falseintodetection_only_normalized_false_with_input_dims_reports_false_rawusing aYoloDetschema (which intentionally stays at the raw-flag contract). - Added two new tracker-path regressions:
yolo_split_segdet_tracker_path_normalizes— pinsYoloSplitSegDetdecode_trackedupgrade + boxes in[0,1].yolo_segdet_2way_decode_normalizes— pinsYoloSegDet2Waydecodeupgrade + boxes in[0,1].
- Updated file-level doc-block to describe the now-complete four-path uniform contract.
Docs (73d04bc)
- C Doxygen / regenerated C header: rewritten from "two normalizing paths" to "four normalizing paths"; explicitly names the still-raw-flag families (
YoloDet,YoloSplitDet,YoloEndToEnd*,ModelPack*). - Python getter docstrings: same widening, using
:attr:cross-refs. - CHANGELOG rewritten in place to describe the complete fix: implementation + accessor halves, and the intentional raw-flag exceptions.
Test summary
| Suite | Passed | Failed | Ignored |
|---|---|---|---|
--lib --features tracker (unit) |
390 | 0 | 4 |
decoder_normalized_flag --features tracker |
10 | 0 | 0 |
per_scale_parity |
29 | 0 | 2 |
Integration (decoder_from_edgefirst_json, decoder_capacity, decode_vs_proto_parity) |
7 | 0 | 0 |
| doc-tests | 61 | 0 | 0 |
CAPI test_decoder_normalized_boxes |
1 | 0 | 0 |
All gates green: cargo check --workspace, make format lint, verify_version (after Round 1 bump), SBOM + license policy clean from round 1.
Confirmed remaining gaps (intentional; suitable for follow-up tickets, NOT release blockers)
YoloEndToEnd*: schemas declaringnormalized: falsepaired with these models would still receive pixel-space boxes — the embedded-NMS coordinate contract makes the right answer unclear. Worth a JIRA ticket to clarify whetherYoloEndToEnd*schemas can legitimately declarenormalized: falseand, if so, whether the helper should fire there too.ModelPack*: separate model family with its own conventions. Same situation — needs a clarification ticket.YoloSegDet2Waydecode_trackedtest: the implementation is in place (macro atpostprocess.rs:1974now calls the helper), but no synthetic test exercises this entry point because constructing a 2-way schema + ByteTrack stack programmatically is non-trivial. Thedecodepath IS pinned by the newyolo_segdet_2way_decode_normalizestest. Low priority — covered implicitly by source-level audit.
Test plan
- CI: full Rust + Python test matrix
- CI: SBOM + license policy gate
- On-target: gst bridge smoke test exercising all four ModelType families (per-scale,
YoloSegDet,YoloSplitSegDet,YoloSegDet2Way) via bothdecodeanddecode_trackedto confirm coordinates land in[0, 1]for all of them.
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Patch release fixing the
Decoder::normalized_boxes()accessor contract leak: the public flag now describes the post-decode coordinate space the caller actually receives, not the raw schema annotation.When a schema declared
normalized: falseand the decoder also knewinput_dims, every decode path internally divided bbox channels by(W, H)viayolo::maybe_normalize_boxes_in_place(EDGEAI-1303) — butnormalized_boxes()still returnedSome(false)("pixel-space"). Downstream callers that trusted the flag (notably the GStreamer bridge) re-normalized themselves and collapsed detections to ~0.The accessor now returns
Some(true)whenever the internal helper runs, andSome(false)only wheninput_dimsis unknown and pixel-space genuinely leaks out. No control-flow changes — the in-place helper still triggers on the same input policy. Only the public accessor and surrounding docs move; the C / Python equivalents inherit the fix automatically.Three existing assertions that enshrined the buggy state flipped to expect
Some(true):crates/decoder/tests/decoder_normalized_flag.rs(combined-Detection and SplitSegDet)crates/decoder/src/decoder/tests.rs(Hailo YOLOv8-seg schema)crates/decoder/tests/per_scale_parity.rs(per-scale parity helper)New regression
normalized_false_without_input_dims_reports_falsepins theSome(false)branch via a programmatic builder that leavesinput_dimsunset.Test plan
coords ≈ 0regression and confirm segmentation boxes land in[0, 1]