recipe(BiRefNet_lite-ONNX): image segmentation recipes

[fuliucansheng]

Summary

Adds CPU recipe coverage for onnx-community/BiRefNet_lite-ONNX as a direct Hub ONNX image-segmentation/background-removal model. This is an L0 recipe-only contribution: no WinML exporter code is changed, and the checked-in recipes document verified CPU fp32/fp16 direct-ONNX builds for the published onnx/model.onnx and onnx/model_fp16.onnx artifacts.

The highest completed goal is L0 PASS. CPU perf was attempted but did not complete within the host time budget for either precision, and task eval is blocked on dataset/mask provenance rather than reported as a fabricated metric.

Model metadata

What the model does: onnx-community/BiRefNet_lite-ONNX produces a single-channel foreground/background mask from an RGB image. The checkpoint is published as a direct ONNX artifact for background removal, mask generation, and image segmentation.

Primary user stories:

Story	Confidence
A user supplies an RGB image and receives a foreground mask suitable for background removal or cutout workflows.	verified from the model card tags, ONNX graph I/O, and preprocessor metadata

Supported tasks:

Task	Surface	Confidence
image-segmentation	Direct ONNX model refs onnx/model.onnx and onnx/model_fp16.onnx	verified for build/graph structure

Model architecture:

BiRefNet_lite direct ONNX image segmentation graph
|-- Image preprocessing contract (external)
|   |-- RGB image resized to 1024 x 1024
|   `-- Rescale and normalize with ImageNet mean/std
|-- Swin-style hierarchical vision backbone / feature encoder
|-- Multi-scale decoder with attention-like ASPP/deformable-convolution regions
`-- Binary mask head producing output_image [1, 1, 1024, 1024]

Source/confidence: graph I/O, task, dtype, and preprocessor contract are verified from the pinned direct ONNX artifacts and repository metadata. The high-level architecture tree is inferred from the published model_type: swin, base-model tag, and ONNX node scopes; no source-code architecture change is made in this PR.

Validation and support evidence

Baseline

Baseline was refreshed on current origin/main at e405dc6638dff6cac96b1539ee0cea678cc6d183 with winml, version 0.2.0. The PR branch merge-base is the same commit.

Probe	Result
Root HF model inspect	winml inspect -m onnx-community/BiRefNet_lite-ONNX --format json resolves the repository root as model_type: swin, task feature-extraction, loader AutoModel, exporter SwinOnnxConfig, and input pixel_values [1,3,224,224]. That is not the published BiRefNet segmentation ONNX graph shipped by this PR.
Root HF recipe-free build	FAIL: uv run winml build -m onnx-community/BiRefNet_lite-ONNX ... attempts HuggingFace export and fails because the repo has no PyTorch/safetensors/TF/Flax weights.
Optimum probe	swin has vendor ONNX tasks feature-extraction, image-classification, and masked-im; WinML adds no override tasks.
Direct ONNX source refs	PASS: onnx/model.onnx has 16400 nodes, opset 17, input input_image [1,3,1024,1024] FLOAT, output output_image [1,1,1024,1024] FLOAT, and 437 FLOAT initializers. onnx/model_fp16.onnx has 16446 nodes with the same I/O and 437 FLOAT16 initializers.
Direct ONNX auto-config	PASS: winml config -m .../onnx/model.onnx and .../model_fp16.onnx emit export: null, optim: {}, quant: null, compile: null; they do not infer loader.task.
Direct ONNX recipe-free build	PASS: fp32 direct build completes in 25.7s. fp16 direct build also completes, but without skip_optimize it produces a mixed/fp32-sized optimized artifact: 4654 nodes with 450 FLOAT, 256 INT64, and only 40 FLOAT16 initializers plus a 243 MB external-data file.
L3 task metric	DATA-PROVENANCE-BLOCKED: winml eval --schema --task image-segmentation is available, but no authoritative dataset revision, split, binary mask semantics, and label mapping were established for this background-removal artifact.

Goal

Effort: L0. Goal ceiling: L1 attempted on CPU. Outcome: L0 recipe-only.

Success means the checked-in CPU fp32/fp16 recipes build structurally valid direct ONNX artifacts with the expected image-segmentation I/O. L1 CPU perf was attempted and recorded as timeout on this host; L2 numeric comparison is not claimed because this PR consumes published ONNX artifacts directly rather than exporting a matched PyTorch checkpoint; L3 is blocked by dataset and semantic provenance.

Outcome

Highest verified goal: L0 PASS on CPU for both fp32 and fp16. Coverage is full for the targeted direct-ONNX build tuples (cpu/cpu/fp32, cpu/cpu/fp16); no accelerator tuples are claimed.

Shipped recipes:

Recipe
examples/recipes/onnx-community_BiRefNet_lite-ONNX/cpu/cpu/image-segmentation_fp32_config.json
examples/recipes/onnx-community_BiRefNet_lite-ONNX/cpu/cpu/image-segmentation_fp16_config.json

No methodology friction observed. The fp16 skip_optimize: true field is a model-specific direct-ONNX recipe delta, not a skill or CLI methodology change. No source, skill, or examples/recipes/README.md files are included in this model PR.

Per-EP/device/precision results

Tier	EP / Device	Precision	Verdict	Evidence
L0 build	CPUExecutionProvider / cpu	fp32	PASS	Recipe build complete in 23.0s from onnx/model.onnx; optimized final graph has input input_image [1,3,1024,1024] FLOAT, output output_image [1,1,1024,1024] FLOAT, 4099 nodes, FLOAT/INT64 initializers, and colocated external data.
L0 build	CPUExecutionProvider / cpu	fp16	PASS	Recipe build complete in 0.9s from onnx/model_fp16.onnx with --precision fp16 and skip_optimize: true; final graph keeps 16446 nodes and 437 FLOAT16 initializers, with the same FLOAT input/output mask contract.
L1 perf	CPUExecutionProvider / cpu	fp32	TIMEOUT	winml perf with 3 iterations and 1 warmup did not complete within about 904s on this host; no latency/throughput/memory JSON was produced.
L1 perf	CPUExecutionProvider / cpu	fp16	TIMEOUT	winml perf with 1 iteration and 0 warmup did not complete within about 304s on this host; no latency/throughput/memory JSON was produced.
L3 eval	CPUExecutionProvider / cpu	fp32	DATA-PROVENANCE-BLOCKED	CLI schema exists for image-segmentation, but no authoritative dataset/revision/split and foreground-mask semantic mapping were established for this checkpoint.
L3 eval	CPUExecutionProvider / cpu	fp16	DATA-PROVENANCE-BLOCKED	Same dataset and mask-semantics provenance blocker as fp32.

Available runtime providers on this host: ['AzureExecutionProvider', 'CPUExecutionProvider'].

Delta

The fp32 recipe adds the missing loader task to the direct-ONNX CPU auto-config shape. The fp16 recipe also preserves the published fp16 ONNX graph by disabling the optimization path that otherwise expands the artifact back toward fp32.

Recipe	JSON pointer	Baseline value	Shipped value	Reason
fp32	/loader/task	absent	image-segmentation	Direct ONNX auto-config cannot infer task intent from the artifact alone; this recipe records the checkpoint-specific segmentation task.
fp16	/loader/task	absent	image-segmentation	Same task declaration as fp32.
fp16	/skip_optimize	absent	true	Preserve the published onnx/model_fp16.onnx graph and its 437 FLOAT16 initializers. Recipe-free direct build re-optimizes into a mixed/fp32-sized artifact.

No source code or tests are changed, and examples/recipes/README.md is untouched.

Reducibility: recipe-only is appropriate because the deltas are direct-artifact task intent and checkpoint-specific fp16 preservation. A class-wide Swin exporter or task-resolution change would not correctly describe the repository-root swin feature-extraction config versus the shipped segmentation ONNX graph.

Analyze summary

Static rule analysis produced parseable JSON but exited nonzero because rule derivation lacks required shape information for some GatherND/Conv checks and reports Sum unsupported during QNN/OpenVINO rule evaluation. This is compatibility analysis, not runtime execution.

Component-level summary

Artifact	Architecture coverage	Mapping	Actionable EP findings
fp32	external preprocessing contract; Swin-style feature encoder; multi-scale decoder; binary mask head	4099 optimized ONNX ops analyzed; architecture-region mapping is partial because this is a direct ONNX artifact without HTP export module tags	QNN/OpenVINO have partial outcomes for high-rank reshape/transpose and shape-dependent arithmetic. Unknown rule outcomes include GatherND, Sum, Conv, and related decoder ops. VitisAI has no useful op-level support rules for this graph on this host.

Op-level summary

Artifact	Graph	Dominant ops	EP roll-up
fp32	4099 ops / 27 types	Unsqueeze 1361; Reshape 568; Add 287; Mul 275; Slice 260; Transpose 240; Concat 159; MatMul 150; Conv 102; Clip 80	QNN: 19 supported, 11 partial, 0 unsupported, 9 unknown op entries. OpenVINO: 21 supported, 9 partial, 0 unsupported, 9 unknown. VitisAI: 27 unknown.

Reproduce commands

$OUT = "temp/birefnet-repro"

uv run winml --version
git rev-parse origin/main
uv run winml inspect -m onnx-community/BiRefNet_lite-ONNX --format json
uv run winml config -m onnx-community/BiRefNet_lite-ONNX/onnx/model.onnx --ep cpu --device cpu -o "$OUT/config_cpu_fp32.json" --overwrite
uv run winml config -m onnx-community/BiRefNet_lite-ONNX/onnx/model_fp16.onnx --ep cpu --device cpu -o "$OUT/config_cpu_fp16.json" --overwrite

uv run winml build -m onnx-community/BiRefNet_lite-ONNX -o "$OUT/root_baseline" --device cpu --ep cpu --no-analyze --no-optimize --no-compile --rebuild
uv run winml build -m onnx-community/BiRefNet_lite-ONNX/onnx/model.onnx -o "$OUT/baseline_fp32" --device cpu --ep cpu --no-analyze --no-compile --rebuild
uv run winml build -m onnx-community/BiRefNet_lite-ONNX/onnx/model_fp16.onnx -o "$OUT/baseline_fp16" --device cpu --ep cpu --no-analyze --no-compile --rebuild

uv run winml build -c examples/recipes/onnx-community_BiRefNet_lite-ONNX/cpu/cpu/image-segmentation_fp32_config.json -m onnx-community/BiRefNet_lite-ONNX/onnx/model.onnx -o "$OUT/fp32" --device cpu --ep cpu --precision fp32 --no-analyze --no-compile --rebuild
uv run winml build -c examples/recipes/onnx-community_BiRefNet_lite-ONNX/cpu/cpu/image-segmentation_fp16_config.json -m onnx-community/BiRefNet_lite-ONNX/onnx/model_fp16.onnx -o "$OUT/fp16" --device cpu --ep cpu --precision fp16 --no-analyze --no-compile --rebuild

uv run python -c "import onnx, collections; from onnx import TensorProto; paths=['$OUT/fp32/model.onnx','$OUT/fp16/model.onnx'];\nfor p in paths:\n    m=onnx.load(p, load_external_data=False)\n    dims=lambda v:[d.dim_value or d.dim_param for d in v.type.tensor_type.shape.dim]\n    init=collections.Counter(TensorProto.DataType.Name(i.data_type) for i in m.graph.initializer)\n    ops=collections.Counter(n.op_type for n in m.graph.node)\n    print(p, len(m.graph.node), [(i.name, dims(i), TensorProto.DataType.Name(i.type.tensor_type.elem_type)) for i in m.graph.input], [(o.name, dims(o), TensorProto.DataType.Name(o.type.tensor_type.elem_type)) for o in m.graph.output], init.most_common(), ops.most_common(10))"
uv run winml perf -m "$OUT/fp32/model.onnx" --device cpu --ep cpu --iterations 3 --warmup 1 --memory --format json -o "$OUT/fp32/perf_cpu.json" --overwrite
uv run winml perf -m "$OUT/fp16/model.onnx" --device cpu --ep cpu --iterations 1 --warmup 0 --memory --format json -o "$OUT/fp16/perf_cpu.json" --overwrite
uv run winml eval --schema --task image-segmentation
uv run winml analyze --model "$OUT/fp32/model.onnx" --ep all --output "$OUT/fp32/analyze_all.json"

[fuliucansheng]

Closing per request.