recipe(vitpose-base): add CPU keypoint detection recipes

[kujin66]

Summary

Adds verified CPU recipes for usyd-community/vitpose-base keypoint 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 blocker evidence on current PR head.

Model metadata

What the model does

usyd-community/vitpose-base is a ViTPose top-down human pose estimation model. It consumes an RGB image tensor and emits 17 COCO-style keypoint heatmaps.

Primary user stories

• A user supplies a person crop/image tensor to obtain body keypoint heatmaps for pose estimation (verified from config and WinML export/build output).

Supported tasks

Task	Support surface	Evidence	Confidence
keypoint-detection	checkpoint, Transformers, WinML	VitPoseForPoseEstimation; WinML build/perf/L2 commands below	verified

Model architecture

VitPoseForPoseEstimation
|-- VitPoseBackbone
|   |-- patch embeddings (pixel_values -> patch tokens)
|   `-- encoder stack (ViT self-attention + MLP blocks)
`-- pose estimation head (heatmaps: [1, 17, 64, 48])

Validation and support evidence

Baseline

• Base: origin/main at d9feefffccecfe47f08b7d2cfffb11fe57eb94df.
• winml config -m usyd-community/vitpose-base --task keypoint-detection succeeded and auto-detected VitPoseForPoseEstimation / keypoint-detection with pixel_values [1, 3, 256, 192] and heatmaps output. Auto-config resolved device/EP to DML/GPU; this PR adds explicit CPU coverage recipes.

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/usyd-community_vitpose-base/cpu/cpu/keypoint-detection_fp32_config.json
  • examples/recipes/usyd-community_vitpose-base/cpu/cpu/keypoint-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 missing local COCO keypoint dataset files, 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	84.93 ms	84.05 ms	11.77 samples/sec	+381.8 MB total	-
L1	CPUExecutionProvider / cpu	fp16	PASS	104.74 ms	104.27 ms	9.55 samples/sec	+392.8 MB total	-
L2	CPUExecutionProvider / cpu	fp32	PASS	-	-	-	-	cosine 1.000000000, max_abs 0.000000276
L2	CPUExecutionProvider / cpu	fp16	PASS	-	-	-	-	cosine 0.999951883, max_abs 0.000411466
L3	CPUExecutionProvider / cpu	fp32	CLI-BLOCKED-DATA	-	-	-	-	Dataset ~/.cache/winml/datasets/coco_keypoints_val2017 not found
L3	CPUExecutionProvider / cpu	fp16	CLI-BLOCKED-DATA	-	-	-	-	Same dataset blocker as fp32; artifact already passed L0/L1/L2

Build evidence:

• fp32: Build complete in 83.1s; final artifact model.onnx; input pixel_values [1, 3, 256, 192]; output heatmaps [1, 17, 64, 48].
• fp16: Build complete in 64.6s; FP16 stage emitted 171.8 MB artifact; perf reported Model Precision: fp16.

L3 eval attempt evidence:

• winml eval -m temp\verify_usyd_vitpose_base_fp32\model.onnx --model-id usyd-community/vitpose-base --task keypoint-detection --ep cpu --device cpu failed after schema-compatible setup because the default local dataset path had no data files: Couldn't find any data file at C:\Users\jinkun\.cache\winml\datasets\coco_keypoints_val2017.

Quality gates:

• GitHub checks on PR head 3a2af091e8b831cc9d225c846c4681c4402cf981: lint, Analyze (Python), test (models), test (commands), test (analyze), test (optim), test (remaining), CodeQL, and license/cla all completed successfully.

Delta

• Recipe-only change, copied from the established ViTPose recipe pattern and pinned to usyd-community/vitpose-base CPU coverage.
• The production recipe README is untouched.

Analyze summary

Static rule analysis completed on both fp32 and fp16 artifacts; this is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	ViTPose patch embeddings; encoder attention/MLP blocks; heatmap head	539/539 ONNX nodes tagged during export	QNN partial: Add, Div, Erf, Mul; ConvTranspose unknown
fp16	ViTPose patch embeddings; encoder attention/MLP blocks; heatmap head	539/539 ONNX nodes tagged during export	QNN partial: Add, Div, Erf, Mul; ConvTranspose unknown

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	423 ops / 13 types	Add 110; MatMul 96; Reshape 50; Transpose 50; Mul 48	TensorRT/OpenVINO 421 supported / 2 unknown; QNN 409 supported / 12 partial / 2 unknown
fp16	425 ops / 14 types	Add 110; MatMul 96; Reshape 50; Transpose 50; Mul 48	TensorRT/OpenVINO 423 supported / 2 unknown; QNN 411 supported / 12 partial / 2 unknown

Rule-less EPs (CUDA, MIGraphX, DML): all operators unknown in static rules.

Reproduce commands

$OUT='temp/verify_usyd_vitpose_base_fp32'
uv run winml build -c examples\recipes\usyd-community_vitpose-base\cpu\cpu\keypoint-detection_fp32_config.json -m usyd-community/vitpose-base -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_usyd_vitpose_base_fp16'
uv run winml build -c examples\recipes\usyd-community_vitpose-base\cpu\cpu\keypoint-detection_fp16_config.json -m usyd-community/vitpose-base -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_usyd_l2.py
uv run winml eval -m temp\verify_usyd_vitpose_base_fp32\model.onnx --model-id usyd-community/vitpose-base --task keypoint-detection --ep cpu --device cpu

$env:WINMLCLI_RULES_DIR='<rules-dir>'
uv run winml analyze --model temp\verify_usyd_vitpose_base_fp32\model.onnx --ep all --output temp\verify_usyd_vitpose_base_fp32\analyze_all.json
uv run winml analyze --model temp\verify_usyd_vitpose_base_fp16\model.onnx --ep all --output temp\verify_usyd_vitpose_base_fp16\analyze_all.json
Remove-Item Env:WINMLCLI_RULES_DIR

[kujin66]

APPROVE (skill reviewer verdict)

Reviewed PR #1189 against the adding-model-support reviewer checklist.

• Shipment state: draft PR, model-scale-by-skill label present, head 3a2af091e8b831cc9d225c846c4681c4402cf981.
• Diff scope: recipe-only Lane B, exactly examples/recipes/usyd-community_vitpose-base/cpu/cpu/keypoint-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; L3 has an exact data blocker for missing local COCO keypoint dataset files.
• Analyze evidence: fp32 and fp16 static analyze summaries are recorded with component-level and op-level rows.
• 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.