Implementation:PeterL1n BackgroundMattingV2 OnnxRuntime InferenceSession

Knowledge Sources	BackgroundMattingV2 ONNX Runtime
Domains	Model_Deployment, Testing
Last Updated	2026-02-09 00:00 GMT

Overview

Concrete external tool usage of onnxruntime.InferenceSession for validating exported ONNX matting models against PyTorch reference outputs.

Description

The ONNX validation in export_onnx.py uses onnxruntime.InferenceSession to load and run the exported ONNX model, then compares each output tensor element-wise against the original PyTorch model's output. Test inputs use different dimensions (1×3×720×1280) than the export dummy inputs (2×3×1080×1920) to verify dynamic axis support. Validation passes if the maximum absolute difference across all outputs is below 0.005.

Usage

Use after ONNX export to verify correctness. Enable via --validate flag. If validation fails, try different patch crop/replace methods or refine modes.

Code Reference

Source Location

• Repository: BackgroundMattingV2
• File: export_onnx.py
• Lines: 126-155

Signature

import onnxruntime

# Load ONNX model
sess = onnxruntime.InferenceSession(model_path: str)

# Run inference
outputs = sess.run(
    output_names=None,          # Return all outputs
    input_feed={
        'src': numpy.ndarray,   # (B, 3, H, W) float32
        'bgr': numpy.ndarray    # (B, 3, H, W) float32
    }
) -> List[numpy.ndarray]

Import

import onnxruntime
import numpy as np

I/O Contract

Inputs

Name	Type	Required	Description
model_path	str	Yes	Path to exported .onnx file
src	numpy.ndarray	Yes	Source image array (B, 3, H, W) float32
bgr	numpy.ndarray	Yes	Background image array (B, 3, H, W) float32

Outputs

Name	Type	Description
outputs	List[ndarray]	Model outputs as numpy arrays
validation result	str	"Validation passed." or raises exception

Usage Examples

Validate ONNX Export

import torch
import onnxruntime
import numpy as np
from model import MattingRefine

# PyTorch reference
model = MattingRefine(backbone='resnet50')
model.load_state_dict(torch.load('checkpoint.pth', map_location='cpu'), strict=False)
model.eval()

src = torch.randn(1, 3, 720, 1280)
bgr = torch.randn(1, 3, 720, 1280)

with torch.no_grad():
    out_torch = model(src, bgr)

# ONNX inference
sess = onnxruntime.InferenceSession('matting.onnx')
out_onnx = sess.run(None, {
    'src': src.numpy(),
    'bgr': bgr.numpy()
})

# Compare outputs
for a, b, name in zip(out_torch, out_onnx, ['pha', 'fgr', 'pha_sm', 'fgr_sm', 'err_sm', 'ref_sm']):
    e = torch.abs(a - torch.as_tensor(b)).max()
    print(f'"{name}" max diff: {e}')
    assert e < 0.005, f'Validation failed for {name}'

print('Validation passed.')

Related Pages

Implements Principle

• Principle:PeterL1n_BackgroundMattingV2_ONNX_validation