OpenVINO IR model export by avbelova · Pull Request #1238 · roboflow/rf-detr

avbelova · 2026-07-27T11:28:41Z

What does this PR do?

This PR adds RF-DETR model export to OpenVINO IR (Intermediate Representation).

Type of Change

New feature (non-breaking change that adds functionality)
Documentation update

Testing

I have tested this change locally
I have added/updated tests for this change

Checklist

My code follows the style guidelines of this project
I have performed a self-review of my own code
I have commented my code where necessary, particularly in hard-to-understand areas
My changes generate no new warnings or errors
I have updated the documentation accordingly (if applicable)

ref: #1024

- Resolved conflicts in src/rfdetr/detr.py - Combined OpenVINO and ExecuTorch export features - Updated format documentation to include both openvino and executorch - Updated return type to include .xml (OpenVINO) and .pte (ExecuTorch) - Both formats now use direct conversion (no ONNX intermediate) - Merged upstream changes: ExecuTorch export, dataset improvements, CI updates

- Added [openvino] optional dependency to pyproject.toml - Updated all error messages to use 'pip install "rfdetr[openvino]"' - Updated OpenVINO export documentation with new installation instructions - Consistent with other optional dependencies (onnx, tflite, executorch, tensorrt)

- Update OpenVINO IR description to mention broader hardware support - Mention CPU (x86, ARM), GPU (Intel integrated & discrete), and AI accelerators (Intel NPU) - Make documentation more neutral and less Intel-specific - Simplify installation instructions - Remove redundant advantages section

CLAassistant · 2026-07-27T11:28:55Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ Borda
❌ avbelova
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

socket-security · 2026-07-27T11:29:42Z

Review the following changes in direct dependencies. Learn more about Socket for GitHub.

Diff	Package	Supply Chain Security	Vulnerability	Quality	Maintenance	License
	pypi/openvino@2026.3.1

View full report

codecov · 2026-07-27T14:22:53Z

Codecov Report

❌ Patch coverage is 85.49618% with 19 lines in your changes missing coverage. Please review.
✅ Project coverage is 87%. Comparing base (5a395c1) to head (644de42).
⚠️ Report is 1 commits behind head on develop.

Additional details and impacted files

@@           Coverage Diff            @@
##           develop   #1238    +/-   ##
========================================
- Coverage       87%     87%    -0%     
========================================
  Files          114     117     +3     
  Lines        15128   15251   +123     
========================================
+ Hits         13101   13205   +104     
- Misses        2027    2046    +19

🚀 New features to boost your workflow:

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

Borda · 2026-07-27T14:47:42Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

@avbelova could you ls check ^^ 🦝

Copilot

Pull request overview

Adds a new OpenVINO IR export path to RF-DETR, enabling direct PyTorch → OpenVINO conversion and documenting how to use the exported artifacts.

Changes:

Added an OpenVINO exporter and a small inference wrapper under src/rfdetr/export/_openvino/.
Integrated format="openvino" into RFDETR.export() and added an openvino optional dependency extra.
Updated export documentation to include OpenVINO IR export usage and examples.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
src/rfdetr/export/_openvino/README.md	New OpenVINO export usage guide and inference examples.
src/rfdetr/export/_openvino/inference.py	Adds a lightweight OpenVINO IR inference wrapper.
src/rfdetr/export/_openvino/exporter.py	Implements direct PyTorch → OpenVINO IR conversion and saving.
src/rfdetr/export/_openvino/init.py	Declares the OpenVINO export utilities package.
src/rfdetr/detr.py	Extends `RFDETR.export()` to support `format="openvino"`.
pyproject.toml	Adds `openvino` optional dependency extra.
docs/learn/export.md	Documents OpenVINO IR export workflow and examples.

Comments suppressed due to low confidence (2)

src/rfdetr/export/_openvino/README.md:88

The inference example loads output/inference_model.xml, but for RFDETRMedium the exporter will write output/rfdetr-medium.xml by default (because variant_name=self.size). This example should use the correct filename to avoid a copy/paste FileNotFoundError.

# Load the exported model
model = OpenVINOInference("output/inference_model.xml")

src/rfdetr/export/_openvino/README.md:113

The benchmark_app example uses output/inference_model.xml, but the default filename for RFDETRMedium export is output/rfdetr-medium.xml (because variant_name=self.size). Update the command so it works as written.

```bash
benchmark_app -m output/inference_model.xml -data_shape [1,3,576,576]

</details>

- Add path traversal sanitization for variant_name parameter - Call model.export() before OpenVINO conversion to ensure proper export mode - Handle tuple outputs explicitly in ModelWrapper to prevent incorrect nesting - Use AUTO device selection instead of hardcoded CPU for better performance - Update documentation to reflect actual output filenames (rfdetr-medium.xml)

avbelova · 2026-08-31T15:41:26Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

@avbelova could you ls check ^^ 🦝

@Borda the corporate CLA is signed between Roboflow and Intel

Borda · 2026-08-31T18:42:19Z

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

@avbelova could you ls check ^^ 🦝

@Borda the corporate CLA is signed between Roboflow and Intel

Perfect! 🎉

OpenVINO package does not provide type stubs, causing mypy type check to fail. Add it to the list of external packages with ignore_missing_imports.

# Conflicts: # src/rfdetr/detr.py

… dispatch facade [resolve group] PR roboflow#1238 — items 9, 11, 12, 13, 14, 15, 18, 26, 29, 30, 31, 32 Adds "openvino" to _EXPORT_FORMATS (the feature was previously 100% unreachable), threads export_model/backbone_only through the new _export_openvino_format dispatch helper in _backend.py, rewrites ModelWrapper to raise on missing keys instead of silently dropping outputs, adds output_name/dynamic_batch/notes parity with sibling formats, narrows the mypy-failing model.export() call with a Protocol+cast, fixes the docformatter-corrupted export() docstring, adds openvino_precision control over save_model, reverts an unrelated ONNX import-guard restructure, and brings _openvino/__init__.py to sibling parity. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

…rop dead handle [resolve group] PR roboflow#1238 — items 19, 20, 21 Copies infer-request output tensors before returning instead of handing back views onto reused OpenVINO-managed memory (was silently corrupting any caller accumulating results across frames), adds device/cache_dir constructor parameters so kernel compilation caches across process starts, and drops the uncompiled read_model handle that kept a duplicate copy of the weights resident for no reason. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

[resolve group] PR roboflow#1238 — items 16, 22 Deletes the in-package _openvino/README.md (unprecedented pattern, never ships, already drifted from docs/learn/export.md) and ports its unique accurate content into the docs page instead. Fixes the OpenVINO inference example's mean/std arrays defaulting to float64 (silently doubling per-frame conversion cost and undoing an earlier .astype(np.float32)) and adds the missing np.ascontiguousarray call before inference. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

[resolve group] PR roboflow#1238 — item 7 Zero tests existed for the openvino export backend despite every sibling format (onnx/tflite/tensorrt/executorch/coreml) having a dedicated tests/export/test_*.py. Adds registry/dispatch/dependency-missing coverage that needs no openvino install, plus round-trip parity and inference tests gated behind pytest.importorskip("openvino"). Registers the e2e_openvino pytest marker alongside the existing e2e_executorch/e2e_coreml markers. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

The item-30 ModelWrapper rewrite (module-scope, tuple/list passthrough, raises instead of silently mapping) and item-29 output_name/precision changes landed in a separate worktree after this test file was written against the old dict-mapping/output_names contract. Rewrites TestExportOpenvinoModelWrapper to TestModelWrapper against the new module-scope class (tuple passthrough, list-to-tuple coercion, dict raises NotImplementedError, unsupported type raises TypeError); replaces the stale output_names-forwarding assertions with output_name/ dynamic_batch/notes coverage matching what detr.py actually forwards; fixes the save_model assertion to expect compress_to_fp16; and updates the missing-dependency test to assert on the exception message rather than a logger call the new _check_openvino_available() no longer makes. Caught by manually re-running the suite during the Step 9 QA gate after Phase 3 merge-back — the parallel worktree split meant items roboflow#7 and roboflow#30 had no visibility into each other's changes. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

…s_from_dataset() [resolve No.9] Review by copilot-pull-request-reviewer (PR roboflow#1238): "commit 18229b9's docformatter reflow corrupted the whole export() doc..." Challenge: evidence=VALID suggestion=VALID resolution=as-suggested (residual hunk missed by the prior arch-worktree commit) --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

…enames [resolve No.18] Review by foundry:sw-engineer (PR roboflow#1238): "save_model(ov_model, output_xml) passes no precision argument, so stored..." Challenge: evidence=VALID suggestion=VALID resolution=as-suggested (doc-only gap: docstring already had it, export.md did not) --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

…openvino dep [resolve No.25] Review by foundry:doc-scribe (PR roboflow#1238): "The OpenVINOInference docstring's Example: block carries `# doctest: +SK..." Challenge: evidence=VALID suggestion=VALID resolution=as-suggested (also applied to the analogous _backend.py block from item roboflow#30) [resolve No.28] Review by foundry:linting-expert (PR roboflow#1238): "openvino>=2024.0 has no upper bound. Every sibling caps deliberately..." Challenge: evidence=VALID suggestion=partial (capped at <2027.0, next calendar major; no python_version marker added -- openvino's own requires_python is >=3.10 with no upper bound, so a marker isn't justified) --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

…into openvino_export # Conflicts: # src/rfdetr/detr.py

- [resolve No.H-1] /review finding by foundry:sw-engineer (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): guard OpenVINOInference.infer() with a per-instance lock; shared infer_request is not thread-safe per OpenVINO upstream (openvinotoolkit/openvino#2349, #24509) - [resolve No.H-2] /review finding by foundry:sw-engineer (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): hoist the openvino+dynamic_batch rejection into detr.py's existing fail-fast block, matching the executorch/coreml precedent; add a test asserting make_infer_image is not called - [resolve No.M-8] /review finding by foundry:sw-engineer (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): validate dtype/contiguity at the infer() boundary instead of silently doubling buffer size on a float64/non-contiguous input --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

[resolve No.H-3] /review finding by oss:cicd-steward (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): the e2e_openvino pytest marker was registered but wired to no CI job, so openvino.convert_model has never run against a real RF-DETR graph in any automated run. Add an openvino-parity job mirroring coreml-parity/executorch-parity (portable wheel, runs on ubuntu-latest); add it to export-priority-guardian's required jobs; add the currently-missing `and not e2e_openvino` exclusion to ci-tests-cpu.yml's default filter (latent until an env installs openvino, becomes live the moment one does). --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

- [resolve No.H-4] /review finding by foundry:sw-engineer (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): add a passthrough tier (ImportError/NotImplementedError/TypeError/ValueError) before export_openvino's broad `except Exception`, mirroring export_coreml's precedent, so ModelWrapper's documented NotImplementedError/TypeError reach the caller instead of arriving relabeled as RuntimeError - [resolve No.H-5] /review finding by foundry:qa-specialist (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): extract the shared `_check_openvino_available()` choke point into inference.py's __init__ (previously an inline try/except with its own message) so both export_openvino and OpenVINOInference share one patchable check; rewrite test_raises_import_error_before_file_check to monkeypatch it instead of relying on openvino actually being absent from the environment - test-coverage follow-up for M-8 (previous commit 1ad6f4c): add TestOpenVINOInferenceInputValidation covering infer()'s float64 and non-contiguous rejection paths --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

… outputs - [resolve No.M-4] /review finding by foundry:qa-specialist (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): precision, device, and cache_dir were only ever exercised at their defaults. Add TestExportOpenvinoPrecision (float32/float16 compress_to_fp16 forwarding, invalid-precision ValueError) and TestOpenVINOInferenceDeviceAndCache (device forwarded to compile_model, CACHE_DIR set before compile, cache_dir=None skips set_property) — all stubbed, no real openvino needed - [resolve No.M-5] /review finding by foundry:qa-specialist, foundry:challenger (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): ModelWrapper's 3-tuple (masks/keypoints) path was docstring-named but untested, and the gated e2e suite covered detection + backbone-only only. Add test_three_tuple_output_passes_through_unchanged (unit) plus openvino_segmentation_export/openvino_keypoint_export fixtures and their e2e parity tests (gated behind -m e2e_openvino, now wired into CI by the previous commit) --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

… signature - [resolve No.M-1] /review finding by foundry:doc-scribe (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): the OpenVINO inference doc example resized with PIL's default (bicubic) resampling, diverging from predict()'s explicit antialias=False bilinear resize. Rewrite the example to use torchvision.transforms.functional (to_tensor/resize/normalize), matching every sibling exporter's doc example - [resolve No.M-2] /review finding by foundry:doc-scribe: "OpenVINO Model Outputs" documented boxes as "(x, y, w, h)" when the actual convention is normalized cxcywh (center_x, center_y, width, height) — fix the doc table and annotate the inference example's unpacking - [resolve No.M-3] /review finding by foundry:doc-scribe: RFDETR.export()'s Raises: block was missing the ValueError path for an invalid coreml_precision/openvino_precision value - [resolve No.M-6] /review finding by foundry:sw-engineer: export_openvino's signature diverged from every sibling converter (reversed output_dir-first arg order, no keyword-only marker, output_dir: str narrower than str | os.PathLike[str]). Realign to (model, input_tensors, output_dir, *, ...) matching export_coreml/export_executorch; update the one _backend.py call site (already all-kwargs, no behavior change) and all positional test call sites - [resolve No.M-7] /review finding by foundry:sw-engineer: the export-mode switch (hasattr/getattr + callable + .export()) was duplicated three times across _backend.py in two different idioms. Extract a single _switch_to_export_mode() helper (built on the existing _ExportableModule Protocol) and use it in all three dispatch functions - [resolve No.M-10] /review finding by codex (report: .reports/review/2026-09-07T19-00-24Z/review-report.md): "tight numeric parity" overclaimed for openvino_precision="float32" — it controls IR *storage* precision only; execution precision still depends on the compiled device. Soften the claim in detr.py, exporter.py, and docs/learn/export.md Confirmed no regression: pre-commit run --all-files clean; tests/export/test_openvino_export.py 27 passed (6 e2e skipped, no openvino installed locally — wired into CI by a prior commit); ruff/mypy clean on all touched files. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

Positional max-abs-diff over all ~300 raw two-stage query-selection outputs was comparing mostly low-confidence background candidates whose near-tied objectness scores flip order across backends -- not evidence of a broken export. Detection/segmentation/keypoint fixtures now use a real photo (matching the CoreML suite's own pattern) and compare only the top-10 highest-confidence queries; masks compare in sigmoid space since mask logits span a much wider range than boxes/labels. All parity assertions now pin OpenVINO's execution precision to float32 via raw ov.Core (OpenVINO's ARM CPU plugin otherwise executes in fp16 regardless of IR storage precision -- measured 0.11 vs 0.0059 on the backbone). The backbone-only test keeps its structured input (no topk involved) with a loosened, documented tolerance covering a real, currently unexplained ~60x OpenVINO-vs-CoreML backbone precision gap that is tracked separately from this fix. --- Co-authored-by: claude[bot] <209825114+claude[bot]@users.noreply.github.com>

avbelova added 6 commits July 23, 2026 15:23

Add export to OpenVINO IR

af5ae4b

docs: Add OpenVINO IR export documentation to export.md

fe56e28

Merge upstream/develop

1a30a98

avbelova requested review from Borda, SkalskiP, isaacrob and probicheaux as code owners July 27, 2026 11:28

fix(pre-commit): 🎨 auto format pre-commit hooks

25158c4

Borda requested a review from Copilot July 27, 2026 14:47

Copilot started reviewing on behalf of Borda July 27, 2026 14:48 View session

Copilot AI reviewed Jul 27, 2026

View reviewed changes

Merge branch 'develop' into openvino_export

6ba596a

github-actions Bot added the has conflicts label Jul 28, 2026

Borda mentioned this pull request Jul 30, 2026

Expand model export format support #1024

Open

Merge branch 'develop' into openvino_export

cbfaa64

github-actions Bot removed the has conflicts label Aug 31, 2026

pre-commit-ci Bot and others added 3 commits August 31, 2026 14:08

fix(pre-commit): 🎨 auto format pre-commit hooks

a73301a

fix(pre-commit): 🎨 auto format pre-commit hooks

1a218ae

fix: Add openvino to mypy ignore_missing_imports list

35cf237

OpenVINO package does not provide type stubs, causing mypy type check to fail. Add it to the list of external packages with ignore_missing_imports.

Borda added the enhancement New feature or request label Sep 2, 2026

avbelova added 2 commits September 2, 2026 16:46

Merge branch 'develop' into openvino_export

15730f2

fix: wrap export documentation lines

18229b9

github-actions Bot added the has conflicts label Sep 7, 2026

Borda reviewed Sep 7, 2026

View reviewed changes

Comment thread src/rfdetr/detr.py Outdated

Borda reviewed Sep 7, 2026

View reviewed changes

Comment thread src/rfdetr/detr.py Outdated

Borda and others added 3 commits September 7, 2026 16:38

linting

4515fd3

Merge remote-tracking branch 'origin/develop' into openvino_export

5441774

# Conflicts: # src/rfdetr/detr.py

Merge branch 'develop' into openvino_export

4e21206

github-actions Bot removed the has conflicts label Sep 7, 2026

pre-commit-ci Bot and others added 17 commits September 7, 2026 16:55

fix(pre-commit): 🎨 auto format pre-commit hooks

4ff835a

Merge branch 'openvino_export' of https://github.com/avbelova/rf-detr …

e1298ab

…into openvino_export # Conflicts: # src/rfdetr/detr.py

Merge remote-tracking branch 'origin/develop' into openvino_export

3bf0ddc

Borda reviewed Sep 8, 2026

View reviewed changes

Comment thread .github/workflows/ci-integrations.yml Outdated

Apply suggestion from @Borda

644de42

Borda approved these changes Sep 8, 2026

View reviewed changes

Borda merged commit 39c2d3a into roboflow:develop Sep 8, 2026
118 of 119 checks passed

Navigation Menu

Uh oh!

OpenVINO IR model export - #1238

OpenVINO IR model export#1238
Borda merged 37 commits into
roboflow:developfrom
avbelova:openvino_export

avbelova commented Jul 27, 2026 •

edited by Borda

Loading

Uh oh!

CLAassistant commented Jul 27, 2026 •

edited

Loading

Uh oh!

socket-security Bot commented Jul 27, 2026 •

edited

Loading

Uh oh!

codecov Bot commented Jul 27, 2026 •

edited

Loading

Uh oh!

Borda commented Jul 27, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

avbelova commented Aug 31, 2026

Uh oh!

Borda commented Aug 31, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

avbelova commented Jul 27, 2026 • edited by Borda Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Type of Change

Testing

Checklist

Uh oh!

CLAassistant commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

socket-security Bot commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

codecov Bot commented Jul 27, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codecov Report

Uh oh!

Borda commented Jul 27, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

avbelova commented Aug 31, 2026

Uh oh!

Borda commented Aug 31, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

avbelova commented Jul 27, 2026 •

edited by Borda

Loading

CLAassistant commented Jul 27, 2026 •

edited

Loading

socket-security Bot commented Jul 27, 2026 •

edited

Loading

codecov Bot commented Jul 27, 2026 •

edited

Loading