Verify Numerical Consistency of a LiteRT or PyTorch ExportedProgram Model in Python

Since R2026a

This page describes how to verify numerical consistency for a Python^® model when you run it natively in Python environment and when you run it by calling a MATLAB^® entry-point function. The mInvoke_torch entry-point function loads a PyTorch ExportedProgram file by using the loadPyTorchExportedProgram function and performs prediction by calling the invoke function.

Prerequisite

To call MATLAB and execute MATLAB commands from within a Python environment as a computational engine, you must install the MATLAB Engine API for Python. For more information, see Install MATLAB Engine API for Python.

You also need to install the MATLAB Coder™ Support Package for PyTorch^® and LiteRT Models. For more information, see Install MATLAB Coder Support Package for PyTorch and LiteRT Models.

Create the MATLAB Function

Create a mInvoke_torch function. The function loads a PyTorch ExportedProgram model file by using the loadPyTorchExportedProgram function. It then performs prediction by passing the PyTorchExportedProgram object to the invoke function.

function out = mInvoke_torch(network, input)
persistent net;

if isempty(net)
    net = loadPyTorchExportedProgram(network);
end

out = invoke(net,input);
end

Execute the mInvoke_torch Function in Python

Start the MATLAB Engine in Python.

import matlab.engine
eng = matlab.engine.start_matlab()

Use this code to add the MATLAB file on path and then specify the path to the PyTorch ExportedProgram file. Replace pathtoMfiles with the path to mInvoke_torch function, and pathtoModel with the path to the PyTorch ExportedProgram file.

eng.addpath(r"/pathtoMfiles", nargout=0)
pt2_path = r"/pathtoModel/Model.pt2"

Create test input x_torch in Python. Then cast the input data to single data type.

import torch
import numpy as np

x_torch = torch.randn(1, 3, 224, 224, dtype=torch.float32)
x_np = x_torch.numpy()

x_mat_single = eng.single(x_np)

Call the MATLAB function by passing the input data x_mat_single.

y_matlab = eng.mInvoke_torch(pt2_path, x_mat_single)

Convert the MATLAB output to NumPy.

y_matlab_np = np.array(y_matlab).astype(np.float32)

Execute the PyTorch ExportedProgram in Python

Load the PyTorch ExportedProgram file in Python. Create a callable nn.Module by using .module().

from torch.export import load

exported_program = torch.export.load(pt2_path)     # returns an ExportedProgram
model_ep = exported_program.module()  # get a callable nn.Module

Invoke the nn.Module by passing the same input tensor previously used to call the MATLAB function.

y_python = model_ep(x_torch).detach().numpy().astype(np.float32)

Compare Numerical Outputs

MATLAB might return a row vector of size 1-by-N while the Python model returns a 1‑D vector. For element‑wise comparison, reshape the MATLAB output to remove the singleton dimension so it matches the shape of the Python output.

if y_python.shape != y_matlab_np.shape:
    y_matlab_np = y_matlab_np.reshape(y_python.shape)

Calculate the maximum absolute and relative error between MATLAB and Python outputs.

max_abs_err = float(np.max(np.abs(y_python - y_matlab_np)))
max_rel_err = float(np.max(np.abs(y_python - y_matlab_np) / (np.abs(y_python) + 1e-12)))

print(
    "Comparison summary:\n"
    f"  Max absolute error : {max_abs_err:.3e}\n"
    f"  Max relative error : {max_rel_err:.3e}"
)