Importing pytorch models in matlab using importNetworkFromPyTorch

Hello,

I am trying to import the pre-trained pytorch model in matlab using the importNetworkFromPyTorch command supported by deep learning toolbox. However I am getting an error as below

Error using pytorchmex

Traced model failed to load. Trace the model in the fully supported version of PyTorch as described in Deep

Learning Toolbox Converter for PyTorch Models.

Error in nnet.internal.cnn.pytorch_importer.architecture.ModuleManager/loadModule (line 28)

PropertyCell = nnet.internal.cnn.pytorch_importer.architecture.pytorchmex(this.Filename);

Error in nnet.internal.cnn.pytorch_importer.architecture.ModuleManager (line 16)

PropertyCell = loadModule(this);

Error in nnet.internal.cnn.pytorch_importer.architecture.util.translatePyTorchFile (line 9)

nnet.internal.cnn.pytorch_importer.architecture.ModuleManager(filename);

Error in nnet.internal.cnn.pytorch_importer.architecture.importNetworkFromPyTorch (line 18)

importManager = nnet.internal.cnn.pytorch_importer.architecture.util.translatePyTorchFile(filename,

customLayerPath);

Error in importNetworkFromPyTorch (line 36)

Network = nnet.internal.cnn.pytorch_importer.architecture.importNetworkFromPyTorch(modelfile, varargin{:});

Error in mnist2mat (line 1)

net = importNetworkFromPyTorch("mnist_cnn.pt");

Answer 1

[MathWorks Deep Learning Toolbox Team]

The model must be traced in PyTorch first before importing into MATLAB. Please see PyTorch documentation fo some details on how it's done. https://pytorch.org/docs/stable/generated/torch.jit.trace.html

You can also read this blog post for additional information: https://blogs.mathworks.com/deep-learning/2022/10/04/whats-new-in-interoperability-with-tensorflow-and-pytorch/

As a simple example, try something similar to the following in PyTorch:

# This example loads a pretrained PyTorch model from torchvision,

# traces it with example inputs, and saves the trace as a .pt file.

import torch
from torchvision import models

# Load the model with pretrained weights

model = models.mobilenet_v2(pretrained=True)

# Call "eval" to ensure that layers like batch norm and dropout are set to

# inference mode

model.eval()

# Move the model to the CPU

model.to("cpu")

# Create example inputs

X = torch.rand(1, 3, 224, 224)

# Trace model with the example input

traced_model = torch.jit.trace(model.forward, X)

# Save the traced model to a .pt file

traced_model.save('traced_mnasnet.pt')