[El1as42]

Summary

• Adds tests/onnx_graph_parity.py: checks that a self-contained graph from src/inference/onnx_export.py (the one the browser demo runs) still reproduces detect_birds.py, the path the CLI and Streamlit use. tests/model_format_parity.py only covers the plain network exports, which run inside detect_birds.py.
• The reference is configured from the graph's ONNX metadata: source model, species mapping, merge method, NMS IoU, conf and gap. That way the two sides differ only in implementation. Audio is resampled with librosa for both.
• Same matching as the existing test: merged boxes are matched one-to-one at IoU ≥ 0.5, using the matcher and PASS/WARN thresholds of model_format_parity.py.
• The verify_onnx docstring in onnx_export.py now points to the new test.

python tests/onnx_graph_parity.py --onnx docs/models/Just-Bird_fp16.onnx

Results

Run in birdbox-onnx (both sides on the CPU), on tests/test.wav plus the 12 parity soundscapes (13 recordings in total):

Graph	Matched	Extra	Mean IoU	Verdict
Just-Bird fp32 (fresh export)	1228/1228	0	1.0000	PASS
Just-Bird fp16 (docs/models)	1226/1228	3	0.9955	PASS

The fp32 graph matches the Python pipeline exactly. The remaining fp16 differences come from fp16 alone.

Correction to #25: the "ONNX fp32 99.7 %, mean IoU 0.999" row there came from a run where the .pt reference was on the GPU. Those 0.3 % were GPU-vs-CPU rounding, not the graph.

Test plan

• python tests/onnx_graph_parity.py --onnx docs/models/Just-Bird_fp16.onnx <fp32 export> in birdbox-onnx: both PASS, exit code 0
• Graphs of the other mappings (song merge, e.g. Western-US) were not run

🤖 Generated with Claude Code