tensorplay.onnx API

Export a TensorPlay model to ONNX.

Parameters:

• exported_program – an ExportedProgram, or a (model, *args, kwargs) sequence captured on the fly.

• f – file path or writable binary file object. When omitted the ModelProto is returned instead of being written.

• input_names – names for the graph inputs, in placeholder order.

• output_names – names for the graph outputs.

• opset_version – target ONNX opset (default 18, minimum 13).

• dynamic_axes – {value_name: {axis: axis_name}} (or a list of axis indices) marking dimensions that vary at runtime. Applies to both inputs and outputs.

• do_constant_folding – fold subgraphs whose inputs are all constants.

• verify – run the exported model under onnxruntime and compare against eager execution of exported_program.

• rtol/atol – tolerances used by verify.

• external_data – store initializers in a side-car file. None decides from the model size (models at or above the 2 GiB protobuf limit).

• external_data_location – side-car file name for external_data.

• check_model – run onnx.checker over the finished model.

Returns:

The onnx.ModelProto when f is None, else None.

Whether a captured call_function target has an ONNX lowering.

Capture a callable and return an executable graph program.

Parameters:

• model – an nn.Module or plain callable; child modules are inlined.

• args/kwargs – example inputs binding argument defaults.

• dynamic_shapes – dimension specification per argument (dict, sequence, ShapesCollection, or AdditionalInputs).

• strict – reserved for callers of the strict capture contract; capture validation is identical in both modes.

• preserve_module_call_signature – submodule paths whose call metadata is recorded in module_call_graph for module-level tooling.

Compare model under onnxruntime against the eager result.

Parameters:

• model – the built ModelProto.

• expected – what eager TensorPlay produced for example_inputs — a tensor, or an arbitrarily nested tuple/list of them.

• input_names – ONNX graph input names, in placeholder order.

• example_inputs – placeholder name -> example tensor.

• rtol/atol – tolerances forwarded to numpy.allclose().

• raise_on_mismatch – raise VerificationError instead of returning a failed result.