ส่งออกเป็น ONNX

Sometimes your model must run somewhere PyTorch is not installed. ONNX is a shared format that lets one model run across many engines. 🌐

ONNX means Open Neural Network Exchange. It is a vendor-neutral file format that describes your model as a graph of standard operations.

With ONNX you train in PyTorch but deploy on ONNX Runtime, mobile, or the browser, without rewriting the model for each platform.

You export by giving PyTorch the model and a sample input. The exporter traces the run and writes a portable graph to disk.

That sample tensor must match your real input shape and dtype, because the exporter records the operations the data triggers.

Give clear input and output names so the serving code can bind data by name instead of guessing positions.

By default the batch size is fixed. Mark it as a dynamic axis so the exported model accepts any number of inputs at once.

Load the file with the onnx library and run the built-in checker to confirm the graph is valid before you ship it. ✅

ONNX Runtime is a fast engine that executes the exported model on CPU or GPU, often quicker than plain Python inference.

Run the same input through PyTorch and ONNX Runtime and compare outputs. They should match closely, proving the export is faithful.

Each export targets an opset version, the set of supported operations. Pick a version your target runtime understands to avoid errors.

You want a fixed batch to instead accept any size. What do you set?

You exported a PyTorch model to ONNX with a sample input, named axes, validated it, and ran it on ONNX Runtime anywhere. 🎉