ONNX Export Workflow

SpatialHub provides standalone export scripts under tools/export/ to convert supported model architectures into ONNX format.

---

Standalone Execution via uv

Export scripts declare isolated dependencies using PEP 723 inline script metadata, enabling execution via uv run without modifying your primary runtime environment:

uv run tools/export/export_<model>.py [options]

---

Standard Export Pipeline

1. Architecture Initialization: Instantiates the model architecture from local source definitions or standard model repositories.
2. Weight Restoration: Loads checkpoint weights and sets layers to evaluation mode (model.eval()).
3. Graph Tracing: Executes torch.onnx.export with configured dynamic batch and spatial dimension axes.
4. Graph Validation: Verifies structural graph integrity using onnx.checker.check_model.
5. Serialization: Writes the .onnx model file to disk and unifies external tensor data when applicable.

---

Supported Export Scripts

Model	Source Architecture	Export Script	Required Checkpoint / Source
EfficientLoFTR	upstream/efficient_loftr	tools/export/export_efficient_loftr.py	eloftr_outdoor.ckpt
Depth Anything 3	upstream/depth_anything_3	tools/export/export_depth_anything_3.py	Hugging Face Model ID / Weights
DINOv2	PyTorch Hub (torchvision)	tools/export/export_dinov2.py	PyTorch Hub model variant
FastSAM	Ultralytics	tools/export/export_fastsam.py	FastSAM-x.pt
SAM	Segment Anything	tools/export/export_sam.py	sam_vit_h_4b8939.pth
FoundationPose	upstream/foundationpose	tools/export/export_foundationpose.py	weights/ directory

---

Common CLI Options

Most export utilities share the following standard command-line parameters:

Parameter	Type	Default	Description
--checkpoint	str	Contextual	Path to source weights file or model repository ID.
--output-folder	str	Contextual	Destination directory for exported .onnx files.
--opset	int	17	ONNX Operator Set version.