How can I use fitnet coefficients of my older neural network in training a new one?

Hi everyone,

I am using neural network for function fitting using fitnet command. First, I trained it using some inputs my input is a matrix 231x5000 and target is 8x5000. Now I want to add new inputs to my input matrix and to train a new neural network. The new input would be (231 + y)x5000 and the target is the same as on old problem 8x5000. How can I use the coefficients that I obtained from the first model in order to increase the computation time?

Answer 1

[Prasanna]

Hi Vahagn,

To use the weights and biases from your first trained neural network to initialize a new network with additional inputs, you can refer these steps:

• Extract the weights and biases from the first model using the ‘getwb’ function.
• Create the new network with the updated input size using the ‘fitnet’ method. The ‘fitnet’ method returns a function fitting neural network with a new hidden layer size.
• Initialize the new network with the extracted weights using the ‘setwb’ function. You may need to adjust the weights to match the new input size accordingly.
• Train the new network with the updated input data.

For more information regarding the functions used, refer the following documentations:

• ‘getwb’: https://www.mathworks.com/help/deeplearning/ref/getwb.html
• ‘fitnet’: https://www.mathworks.com/help/deeplearning/ref/fitnet.html
• ‘setwb’: https://www.mathworks.com/help/deeplearning/ref/setwb.html
• Initialize neural network to set of weights: https://www.mathworks.com/matlabcentral/answers/50961-how-to-initialize-the-neural-network-to-a-set-of-weights
• Set a specific weight for a connection in neural networks: https://www.mathworks.com/matlabcentral/answers/499607-set-a-specific-weight-for-a-connection-in-neural-networks