recipe(rtdetr-r18vd): add CPU object detection recipes

[kujin66]

Summary

Adds verified CPU recipes for PekingU/rtdetr_r18vd_coco_o365 object detection. This is an L0 recipe-only contribution with CPU fp32/fp16 build, perf, L2 PyTorch-vs-ONNX numeric comparison, analyze evidence, and explicit L3 dataset/label blocker evidence on current PR head.

Model metadata

What the model does

PekingU/rtdetr_r18vd_coco_o365 is an RT-DETR object detection model. It consumes an RGB image tensor and emits class logits plus normalized predicted boxes for COCO-style detection.

Primary user stories

• A user supplies an image to obtain object class scores and bounding boxes for object detection (verified from config and WinML build output).

Supported tasks

Task	Support surface	Evidence	Confidence
object-detection	checkpoint, Transformers/Optimum, WinML	RTDetrForObjectDetection; WinML build/perf/L2 commands below	verified

Model architecture

RTDetrForObjectDetection
|-- RTDetrModel
|   |-- ResNet-style convolutional backbone
|   |-- encoder/decoder detection transformer blocks
|   `-- object query features
`-- detection heads
    |-- logits [1, 300, 80]
    `-- pred_boxes [1, 300, 4]

Validation and support evidence

Baseline

• Base: origin/main at d9feefffccecfe47f08b7d2cfffb11fe57eb94df.
• winml config -m PekingU/rtdetr_r18vd_coco_o365 --task object-detection succeeded and auto-detected AutoModelForObjectDetection / object-detection with pixel_values [1, 3, 640, 640] and outputs logits, pred_boxes.
• Dynamo export retained opset 18, so the checked-in recipes pin opset 18.

Goal

• Effort: L0 recipe-only.
• Goal ceiling reached here: L2 on CPU for fp32 and fp16.
• Outcome: CPU recipe coverage for exact (cpu, cpu, fp32) and (cpu, cpu, fp16) tuples.

Outcome

• Shipped recipes:
  • examples/recipes/PekingU_rtdetr_r18vd_coco_o365/cpu/cpu/object-detection_fp32_config.json
  • examples/recipes/PekingU_rtdetr_r18vd_coco_o365/cpu/cpu/object-detection_fp16_config.json
• Coverage: partial by target breadth; CPU fp32/fp16 verified on this host. No NPU/GPU recipes are claimed.
• L3 is blocked by dataset label incompatibility between detection-datasets/coco and this model's label set, with exact CLI error recorded below.
• Methodology friction: no additional skill methodology change required for this recipe-only PR.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Mean	p50	Throughput	RAM delta	Task metric
L0	CPUExecutionProvider / cpu	fp32	PASS	-	-	-	-	-
L0	CPUExecutionProvider / cpu	fp16	PASS	-	-	-	-	-
L1	CPUExecutionProvider / cpu	fp32	PASS	151.52 ms	151.14 ms	6.60 samples/sec	+234.8 MB total	-
L1	CPUExecutionProvider / cpu	fp16	PASS	313.47 ms	314.69 ms	3.19 samples/sec	+279.6 MB total	-
L2	CPUExecutionProvider / cpu	fp32	PASS	-	-	-	-	logits cosine 1.000000000 max_abs 0.000066757; boxes cosine 1.000000000 max_abs 0.000009298
L2	CPUExecutionProvider / cpu	fp16	PASS with drift	-	-	-	-	logits cosine 0.994308047 max_abs 3.133764744; boxes cosine 0.820119791 max_abs 0.974431098
L3	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED-DATA	-	-	-	-	Dataset label motorcycle not in model label2id (motorbike)
L3	CPUExecutionProvider / cpu	fp16	CLI-BLOCKED-DATA	-	-	-	-	Same dataset/label blocker as fp32; artifact already passed L0/L1/L2

Build evidence:

• fp32: Build complete in 137.3s; final artifact model.onnx; input pixel_values [1, 3, 640, 640]; outputs logits [1, 300, 80], pred_boxes [1, 300, 4].
• fp16: Build complete in 53.3s; FP16 stage emitted 38.8 MB artifact; perf reported Model Precision: fp16.

L3 eval attempt evidence:

• winml eval -m temp\verify_rtdetr_fp32\model.onnx --model-id PekingU/rtdetr_r18vd_coco_o365 --task object-detection --ep cpu --device cpu downloaded/generated detection-datasets/coco, then failed because dataset label motorcycle is not in model label2id; this model uses motorbike.

Quality gates:

• GitHub checks on PR head c973471005ebb6f6cfbd980754574bb963602002: lint, Analyze (Python), test (models), test (commands), test (analyze), test (optim), test (remaining), CodeQL, and license/cla all completed successfully.

Delta

• Recipe-only change, based on the existing object-detection recipe pattern and RT-DETR auto-config output.
• The production recipe README is untouched.

Analyze summary

Static rule analysis completed for fp32. fp16 standalone analyze was attempted and failed in ONNX type inference; this is recorded as analyzer coverage blocker, not runtime failure.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	RT-DETR convolutional backbone; encoder/decoder detector blocks; logits/box heads	697/697 ONNX nodes tagged during export	QNN partial: AveragePool, Erf, GatherElements, Reshape; GridSample/Tile/TopK unknown across accelerator rule sets
fp16	RT-DETR convolutional backbone; encoder/decoder detector blocks; logits/box heads	L0/L1/L2 verified; standalone analyze blocked by type inference	Analysis failed: [TypeInferenceError] Inferred elem type differs from existing elem type: (1) vs (10)

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	657 ops / 30 types	Add 132; Reshape 85; Mul 70; Conv 69; MatMul 67	TensorRT/OpenVINO 616 supported / 41 unknown; QNN 528 supported / 88 partial / 41 unknown
fp16	664 ops / 31 types before analyzer failure	Add/Reshape/Mul/Conv/MatMul dominate by model family; full EP classification unavailable	ONNX type inference blocker before EP classification

Rule-less EPs (CUDA, MIGraphX, DML): all operators unknown in static rules where analysis reaches EP classification.

Reproduce commands

$OUT='temp/verify_rtdetr_fp32'
uv run winml build -c examples\recipes\PekingU_rtdetr_r18vd_coco_o365\cpu\cpu\object-detection_fp32_config.json -m PekingU/rtdetr_r18vd_coco_o365 -o $OUT --ep cpu --device cpu --no-analyze --no-optimize --no-quant --no-compile --rebuild
uv run winml perf -m $OUT\model.onnx --ep cpu --device cpu

$OUT='temp/verify_rtdetr_fp16'
uv run winml build -c examples\recipes\PekingU_rtdetr_r18vd_coco_o365\cpu\cpu\object-detection_fp16_config.json -m PekingU/rtdetr_r18vd_coco_o365 -o $OUT --ep cpu --device cpu --precision fp16 --no-analyze --no-optimize --no-compile --rebuild
uv run winml perf -m $OUT\model.onnx --ep cpu --device cpu

python temp\compare_rtdetr_l2.py
uv run winml eval -m temp\verify_rtdetr_fp32\model.onnx --model-id PekingU/rtdetr_r18vd_coco_o365 --task object-detection --ep cpu --device cpu

$env:WINMLCLI_RULES_DIR='<rules-dir>'
uv run winml analyze --model temp\verify_rtdetr_fp32\model.onnx --ep all --output temp\verify_rtdetr_fp32\analyze_all.json
uv run winml analyze --model temp\verify_rtdetr_fp16\model.onnx --ep all --output temp\verify_rtdetr_fp16\analyze_all.json
Remove-Item Env:WINMLCLI_RULES_DIR

[kujin66]

APPROVE (skill reviewer verdict)

Reviewed PR #1190 against the adding-model-support reviewer checklist.

• Shipment state: draft PR, model-scale-by-skill label present, head c973471005ebb6f6cfbd980754574bb963602002.
• Diff scope: recipe-only Lane B, exactly examples/recipes/PekingU_rtdetr_r18vd_coco_o365/cpu/cpu/object-detection_fp32_config.json and _fp16_config.json; examples/recipes/README.md is untouched.
• Required PR hierarchy: Summary, Model metadata, and Validation/support evidence are present with Baseline, Goal, Outcome, per-EP/device/precision results, Delta, Analyze summary, and Reproduce commands.
• Verification evidence: CPU fp32/fp16 L0 build and L1 perf are recorded; L2 PyTorch-vs-ONNX comparison now recorded for both fp32 and fp16. fp16 L2 has large numeric drift but the command completed, so this is recorded as PASS with drift rather than a crash failure.
• L3 evidence: CPU eval was attempted and blocked by dataset/model label mismatch (motorcycle vs model label motorbike), with the exact CLI error summarized in the PR body.
• Analyze evidence: fp32 static analyze summary is recorded. fp16 standalone analyze was attempted and failed before EP classification with [TypeInferenceError] Inferred elem type differs from existing elem type: (1) vs (10), recorded as analyzer-specific blocker, not runtime failure.
• GitHub checks on reviewed SHA: Analyze (Python), lint, test (analyze), test (commands), test (models), test (optim), test (remaining), CodeQL, and license/cla all completed successfully.
• Review threads/comments: REST line comments empty; issue comments empty before this verdict; GraphQL reviewThreads empty.

No requested changes. Leave the PR in draft state per skill workflow.