How to transfer a trained model from Pytorch to Matlab?

1

Hi, Everyone. I have a trained CNN model in Pytorch and I want to transfer it to the Matlab. is there any solution for this purpose?
Thank you for advising me on this topic.
Reza.

2

You can extract the state dictionary of a model using .state_dict() and load it into a similar model designed in MATLAB.

1 Like
3

Can you explain more about this? I’m New to the PyTorch and actually, I’m Confuse about this issue

4

It seems they use different format. I saved my .state_dict() as a .mat file as:

import numpy, scipy.io
scipy.io.savemat(params[‘restore_folder’] +
‘encoder’+params[‘restore_model’]+’.mat’, mdict=encoder.state_dict())

However, when I load the .mat file to matlab it is able to read the names of the encoder layers but they are not proper structures and matlab gives an error.

Any suggestions?

5

@Adam_Stevens: A better way is probably to use the ONNX as a model exchange tool:

6

Unfortunately ONNX import in Matlab seems to be underdeveloped and works only for smaller set of architectures. I don’t think it works for other than CNNs at the moment.

Does anyone know about other options?

Thanks,
Jan

7

Maybe error is due to the name of value contaion ‘.’, which is not supported in matlab. You can replace ‘.’ with ‘_’.