export

R2026b

Save underlying PyTorch model in different PyTorch file format

Since R2026b

Syntax

export(model,format,fileName)

export(___,X1,...,XM)

pythonObject = export(___)

Description

export(model,format,fileName) saves the underlying PyTorch^® model to the specified file in the specified format. When using this syntax, format is "full", "weights(zipped)", "weights(safetensors)", or "scripted". The ModelType of the model must be "torch.nn.Module".

export(___,X1,...,XM) saves the PyTorch model as an Exported Program or Traced model, passing example input data to trace the model. When using this syntax, format is "exportedProgram" or "traced".

pythonObject = export(___) also returns the Python^® object saved to a file.

Examples

Export Model as Full, Weights-Only, or Scripted Format

Export a PyTorch model containing a torch.nn.Module as a full model, weights-only, or scripted model file.

Load a multi-input, multi-output model, specifying data-transfer settings.

model = pyTorchModel("mimo32_Full.pt",...
    NumInputs=3, ...
    InputDimensionOrder={[3 4 1 2],[2 3 1],[2 3 1]},...
    InputNumDimensions={4 3 3},...
    InputDataType={"float32","float32","float32"},...
    OutputDimensionOrder={[1 2 3 4],[1 2 3]});

Save as a full model.

export(model,"full","mimo32_full_2.pt")

Save as a scripted model.

export(model,"scripted","mimo32_scripted_2.pt")

Save as a weights-only model using torch.save (zipped format).

export(model,"weights(zipped)","mimo32_weights_2.pth")

Save as a weights-only model using safetensors.

export(model, "weights(safetensors)","mimo32_weights_2.safetensors")

Export Model as Exported Program or Traced Format

Export a PyTorch model containing a torch.nn.Module as an exported program or traced model file, providing example inputs.

Load a multi-input, multi-output model, specifying data-transfer settings.

model = pyTorchModel("mimo32_Full.pt",...
    NumInputs=3, ...
    InputDimensionOrder={[3 4 1 2],[2 3 1],[2 3 1]},...
    InputNumDimensions={4 3 3},...
    InputDataType={"float32","float32","float32"},...
    OutputDimensionOrder={[1 2 3 4],[1 2 3]});

Create example input to trace the model.

X1 = rand([3 4 1 2]);
X2 = rand([6 1 5]);
X3 = rand([8 1 7]);

Save as an exported program.

export(model,"exportedProgram","mimo32_exported_2.pt2",X1,X2,X3)

Save as a traced model.

export(model,"traced","mimo32_traced_2.pt",X1,X2,X3)

Capture Python Object Output from Export

Export a model as an exported program and capture the returned Python ExportedProgram object.

Load a model and export it as an exported program, capturing the returned Python object.

model = pyTorchModel("fooBarModel.pt",InputDataType="float32");
X = rand(17,2);
ep = export(model,"exportedProgram","fooBarModel.pt2",X);

Input Arguments

model — Reference to a PyTorch model

PyTorchModel object

Reference to a PyTorch model, specified as a PyTorchModel object. ModelType of the model must be "torch.nn.Module".

format — Target file format

"full" | "weights(zipped)" | "weights(safetensors)" | "exportedProgram" | "traced" | "scripted"

Target file format, specified as one of these values:

• "full" — Full model (architecture and weights saved with torch.save(model,file)).

• "weights(zipped)" — Weights-only in zipped format (saved with torch.save(model.state_dict(),file)).

• "weights(safetensors)" — Weights-only in safetensors format.

• "scripted" — Scripted model (saved with torch.jit.save).

• "exportedProgram" — Exported Program (saved with torch.export.save). Requires example input X1,...,XM.

• "traced" — Traced model (saved with torch.jit.save). Requires example input X1,...,XM.

Data Types: string | char

fileName — Name of output file

string | character vector

Name of the file for saving the model, specified as a string or character vector.

Data Types: char | string

X1,...,XM — Example input data for tracing

numeric arrays | logical arrays | gpuArray objects | py.* objects | ...

Example input data on which to trace the model, specified as MATLAB^® numeric arrays, logical arrays, gpuArray objects, py.* objects, or other MATLAB data types compatible with the MATLAB Python interface. This argument is required when format is "exportedProgram" or "traced".

Numeric arrays are converted to torch.Tensor objects using the data-transfer settings stored in the model.

Output Arguments

pythonObject — Python object saved to file

py.* object

Python object that was saved to file, returned as a py.* object. Depending on the format, this can be a subclass of torch.nn.Module, a dict, an ExportedProgram, or a RecursiveScriptModule.

Algorithms

CPU and GPU Behavior

For export formats that require the model to be traced on example inputs ("exportedProgram" and "traced"):

• If any input is a gpuArray or a py.torch.Tensor whose device attribute is "cuda*", then the GPU of the first such input is selected as the execution device. Otherwise, CPU is used. Using GPU with PyTorchModel requires Parallel Computing Toolbox™.

• All MATLAB numeric array inputs and py.torch.Tensor inputs in X1,...,XM are copied to the selected device if not already there.

• The model parameters are moved to the selected device if not already there.

Version History

Introduced in R2026b