ONNX export

The model exports to a single ONNX graph with its three outputs, so it can run on CPU-only machines or in any ONNX Runtime host without PyTorch.

python -m infer.export_onnx --ckpt outputs/v5/best.pt --out outputs/v5/depthwizard.onnx

Signature

	Name	Type	Shape
Input	`image`	float32	(batch, 3, 512, 512): encoder-normalised RGB at 0.5 m
Output	`height_m`	float32	(batch, 1, 512, 512): nDSM in metres
Output	`seg`	int32	(batch, 1, 512, 512): class id
Output	`height_std_m`	float32	(batch, 1, 512, 512): Head B spread (uncertainty)

Only the batch dimension is dynamic. The graph is traced at batch 2 and checked at batch 1. Tiling, stretching, blending and TTA stay in Python (predict_scene), exactly as for PyTorch.

Files

File	Size	Notes
`depthwizard.onnx`	3.96 MB	graph
`depthwizard.onnx.data`	about 1.3 GB	external weights; must sit next to the graph
`depthwizard.onnx.json`	—	preprocessing recipe, I/O description, opset, source checkpoint
`parity.json`	—	PyTorch vs ONNX comparison

The requested opset was 17; the file is opset 18 because the opset-17 conversion failed and the exporter kept 18.

Parity

Max |Δ height|

0.00024m

sample RGB; tolerance 0.05 m

Mean |Δ height|

1.6e-5m

sample RGB

Class agreement

100%

sample RGB and noise inputs

{
  "tol_m": 0.05,
  "ok": true,
  "checks": {
    "sample rgb": { "height_max_abs_m": 0.000236, "std_max_abs_m": 0.00035, "seg_agree": 1.0 },
    "noise":      { "height_max_abs_m": 0.000039, "std_max_abs_m": 0.000078, "seg_agree": 1.0 }
  }
}