Your ONNX export is more complicated than it looks

What it does

ONNX Simplifier takes a trained model in ONNX format and prunes the computational graph. It runs constant folding across the whole graph, replacing clusters of redundant operators—like the shape-gathering spaghetti that PyTorch sometimes emits for a simple reshape—with their actual computed values. The result is a smaller, cleaner model that still passes ONNX’s own validation check.

The interesting bit

The project ships both a Python package and a browser-based version that runs entirely offline via GitHub Pages, which is a neat trick for a graph-rewriting tool that handles potentially proprietary model files.

Key highlights

• Performs whole-graph inference to fold redundant operators into constants.
• Ships a fully offline web UI via GitHub Pages—no server upload required.
• Validates the rewritten graph against the original to ensure correctness.
• Adopted by deployment pipelines including YOLOv5, MMDetection, and ncnn.
• Exposes a minimal Python API returning both the cleaned model and a pass/fail check.

Verdict

Worth keeping in your toolkit if you regularly export from PyTorch to deployment runtimes like ncnn. Less useful if you already work with hand-tuned ONNX graphs that never picked up exporter cruft.