High error when evaluate mode with importNetworkFromPyTorch

Hi Matlab community,

I've built a simple regression model with pyTorch, and saved it to 'model.pt'. Then, I import it using importNetworkFromPyTorch and initialize with train data.

inputData = X_train';
dlX = dlarray(inputData, 'CB');
net = initialize(net, dlX); 

After having the net file, I added it Predict Block in simulink here: Predict

The result for regression is not good compare to result that I validate with python. I don't know if I did anything wrong here.

What happens when we initialize the network? do we need to use train data to set correct weights? Looking forward to hearing from you

Answer 1

[Sreeram]

Hi @Son Pham,

The “initialize” function automatically sets up the network using only the size and format information of the input data ‘dlX’. It does not consider the actual values of the input data, which means it does not set the learnable parameters based on the training data. This is explained in the “Note” section of the following documentation on "initialize":

https://www.mathworks.com/help/releases/R2024a/deeplearning/ref/dlnetwork.initialize.html#mw_3af1d784-472c-4f9b-872d-d8c26929d3e7

To ensure that the network is initialized with weights that reflect the training data, the documentation suggests to manually initialize the learnable parameters. Refer to the following documentation for further details:

https://www.mathworks.com/help/releases/R2024a/deeplearning/ug/initialize-learnable-parameters-for-custom-training-loop.html