Untraining the SalsaNext model Architecture

I used the following repository to evaluate my data (https://github.com/matlab-deep-learning/pretrained-salsanext), but it loads a pretrained SalsaNext model. I would like to know whether it is possible to untrain the model, or to train it from scratch without creating the layergraph (Since this model involves a large number of layers, and the work seems to be cumbersome, and error-prone).

Thanks

Answer 1

[Himanshu]

Hi,

I see that you want to import a pretrained SalsaNext model, reset its weights, and then export the model architecture without the pretrained weights.

Using the Deep Network Designer app in MATLAB provides a user-friendly interface to modify and export neural network architectures. You can import the model, reset the weights, and export the model without weights using Deep Network Designer. Please refer to the below steps:

1. Use the "deepNetworkDesigner" command to open the app.
2. Import the pretrained model into the app.
3. Manually reset the weights in the app or export the model and reset the weights programmatically.
4. Export the model without weights.

Please refer to the below documentations to learn more about Deep Network Designer App in MATLAB.

https://www.mathworks.com/help/deeplearning/gs/get-started-with-deep-network-designer.html

https://www.mathworks.com/help/deeplearning/ref/deepnetworkdesigner-app.html

I hope this helps.

[Serban]

Thanks a lot !