recipe(vitpose-base-simple): add CPU keypoint detection configs

[fuliucansheng]

Summary

Adds CPU recipe coverage for nielsr/vitpose-base-simple, a ViTPose keypoint-detection checkpoint. This is an L0 recipe-only contribution: current main resolves the model family but recipe-free winml config / winml build fail with empty exporter I/O, and this PR supplies the verified CPU fp32/fp16 recipe inputs and outputs.

Model metadata

What the model does: nielsr/vitpose-base-simple predicts human-pose keypoint heatmaps from an input image crop. WinML exports it as keypoint-detection with VitPoseForPoseEstimation.

Primary user stories:

Story	Confidence
A user supplies a pose-cropped RGB image to obtain 17 keypoint heatmaps for human pose estimation.	verified from checkpoint config, image processor metadata, and model architecture

Supported tasks:

Task	Surface	Confidence
keypoint-detection	Transformers / Optimum ONNX / WinML export surface	verified

Model architecture:

VitPoseForPoseEstimation
└── VitPoseBackbone
    ├── Patch embeddings (image 256x192, patch 16x16, channels 3, hidden 768)
    └── Encoder stack x12
        ├── self-attention (12 heads)
        └── MLP + GELU + residual/layernorm
└── Pose head -> heatmaps [B, 17, 64, 48]

Source/confidence: pinned checkpoint config resolves model_type: vitpose, architectures: [ViTPoseForPoseEstimation], image size [256, 192], patch size [16, 16], and backbone hidden size 768 with 12 layers and 12 attention heads.

Validation and support evidence

Baseline

Baseline was run from current origin/main at ace130f20e36df44ebf7de6cf0356235c8ee79fd with winml, version 0.2.0.

Probe	Result
winml inspect	PARTIAL: loader VitPoseForPoseEstimation, exporter VitPoseOnnxConfig, WinML class WinMLModelForGenericTask; task keypoint-detection; exporter input/output tensors are empty.
Optimum coverage	VENDOR-ONLY for vitpose: vendor ONNX tasks include keypoint-detection; WinML adds no extra task override.
Recipe-free config	FAIL: winml config -m nielsr/vitpose-base-simple -p fp32 --ep cpu --device cpu exits with Unexpected error: list index out of range.
Recipe-free build	FAIL: winml build -m nielsr/vitpose-base-simple ... exits with Build failed: list index out of range.
L3 task metric	HOST-BLOCKED: winml eval supports keypoint-detection, but the default local COCO keypoints dataset ~/.cache/winml/datasets/coco_keypoints_val2017 is absent on this host.

The checked-in recipe is seeded from existing same-family VitPose recipes and the checkpoint config: pixel_values [1,3,256,192] to heatmaps.

Goal

Effort: L0. Goal ceiling: L2. Outcome: L0 recipe-only.

Success means the checked-in CPU fp32/fp16 recipes repair the recipe-free build failure, produce structurally valid ONNX artifacts, run winml perf on CPU, and match the PyTorch reference heatmaps numerically with the same image tensor. L3 is blocked by missing local COCO keypoint data.

Outcome

Highest verified goal: L2 PASS on CPU for both fp32 and fp16. Coverage is full for the targeted EP set (cpu/cpu/fp32, cpu/cpu/fp16); no accelerator tuples are claimed.

Shipped recipes:

Recipe
examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp32_config.json
examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp16_config.json

Methodology friction observed: none. Model finding: this VitPose checkpoint needs explicit recipe I/O because current auto-config exposes empty tensors and fails with list index out of range.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Evidence
L0 build	CPUExecutionProvider / cpu	fp32	PASS	Recipe build complete in 33.2s; optimized model is loadable with input pixel_values [1,3,256,192] float32, output heatmaps [1,17,64,48], 419 nodes, FLOAT initializers, and 327.7 MB external data.
L0 build	CPUExecutionProvider / cpu	fp16	PASS	Recipe build complete in 29.9s with --precision fp16; fp16 stage complete in 4.1s; optimized model is loadable with 421 nodes, FLOAT16 initializers present, and 163.9 MB external data.
L1 perf	CPUExecutionProvider / cpu	fp32	PASS	Avg 34.97 ms, p50 34.07 ms, throughput 28.60 samples/sec, RAM total +389.6 MB.
L1 perf	CPUExecutionProvider / cpu	fp16	PASS	Avg 33.18 ms, p50 33.90 ms, throughput 30.14 samples/sec, RAM total +395.5 MB; perf reports model precision fp16.
L2 compare	CPUExecutionProvider / cpu	fp32	PASS	PyTorch vs ONNX heatmaps on the same deterministic image tensor: cosine 1.0, max absolute diff 2.952292561531067e-07.
L2 compare	CPUExecutionProvider / cpu	fp16	PASS	PyTorch vs ONNX heatmaps on the same deterministic image tensor: cosine 0.9999979734420776, max absolute diff 0.00021735578775405884.
L3 eval	CPUExecutionProvider / cpu	fp32	HOST-BLOCKED	winml eval -m ... --task keypoint-detection --samples 1 fails because default dataset ~/.cache/winml/datasets/coco_keypoints_val2017 is not present.
L3 eval	CPUExecutionProvider / cpu	fp16	HOST-BLOCKED	Same missing-dataset blocker as fp32; no task metric is reported.

Delta

The recipe is not comparable to winml config output because winml config fails before writing a usable config. Relative to the failing auto path, the shipped recipes provide the missing I/O contract:

Field	Shipped value	Reason
export.input_tensors	pixel_values float32 [1,3,256,192], range [0,1]	Required by VitPoseForPoseEstimation.forward and the checkpoint image size.
export.output_tensors	heatmaps	Required output for keypoint detection; validated as [1,17,64,48].
loader	task keypoint-detection, class VitPoseForPoseEstimation, type vitpose	Matches current loader resolution and existing VitPose recipes.

The fp32 recipe uses quant: null; the fp16 recipe carries the fp16 quantization block and is built with --precision fp16. No source code or tests are changed, and examples/recipes/README.md is untouched.

Reducibility: this PR follows the existing VitPose recipe pattern and ships only exact CPU tuples verified here. The current auto-config failure is documented as a baseline gap; this model PR does not add a class-wide source fix.

Analyze summary

Static rule analysis completed with partial-success exit semantics: QNN/OpenVINO rows were complete, while VitisAI has no rule data and is all-unknown. This is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	patch embeddings; 12x encoder attention/MLP; pose heatmap head	536 exported nodes tagged during HTP export; 419 optimized ONNX ops analyzed	QNN NPU partial: GELU-related Div, Erf, Add, Mul; unknown Resize. OpenVINO NPU: unknown Resize. VitisAI: no rule data.
fp16	patch embeddings; 12x encoder attention/MLP; pose heatmap head	536 exported nodes tagged during HTP export; 421 optimized ONNX ops analyzed	QNN NPU partial: GELU-related Div, Erf, Add, Mul; unknown Resize. OpenVINO NPU: unknown Resize. VitisAI: no rule data.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	419 ops / 12 types	Add 110; MatMul 96; Reshape 50; Transpose 50; Mul 48; LayerNormalization 25; Softmax 12; Div 12	QNN: 406 supported, 12 partial, 0 unsupported, 1 unknown. OpenVINO: 418 supported, 0 partial, 0 unsupported, 1 unknown. VitisAI: 419 unknown.
fp16	421 ops / 13 types	Add 110; MatMul 96; Reshape 50; Transpose 50; Mul 48; LayerNormalization 25; Softmax 12; Div 12	QNN: 408 supported, 12 partial, 0 unsupported, 1 unknown. OpenVINO: 420 supported, 0 partial, 0 unsupported, 1 unknown. VitisAI: 421 unknown.

Reproduce commands

$MODEL = "nielsr/vitpose-base-simple"
$OUT = "temp/repro_nielsr_vitpose_base_simple"

uv run winml inspect -m $MODEL --format json
uv run winml config -m $MODEL -p fp32 --ep cpu --device cpu -o "$OUT/config_fp32.json"
uv run winml build -m $MODEL -o "$OUT/baseline" --ep cpu --device cpu --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp32_config.json -m $MODEL -o "$OUT/fp32" --ep cpu --device cpu --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/nielsr_vitpose-base-simple/cpu/cpu/keypoint-detection_fp16_config.json -m $MODEL -o "$OUT/fp16" --ep cpu --device cpu --precision fp16 --no-analyze --no-compile --rebuild
uv run winml perf -m "$OUT/fp32/model.onnx" --ep cpu --device cpu --iterations 5 --warmup 1
uv run winml perf -m "$OUT/fp16/model.onnx" --ep cpu --device cpu --iterations 5 --warmup 1
uv run winml eval --schema --task keypoint-detection
uv run winml eval -m "$OUT/fp32/model.onnx" --model-id $MODEL --task keypoint-detection --ep cpu --device cpu --samples 1
uv run winml analyze --model "$OUT/fp32/model.onnx" --ep all --output "$OUT/fp32/analyze_all.json"
uv run winml analyze --model "$OUT/fp16/model.onnx" --ep all --output "$OUT/fp16/analyze_all.json"