How to customize performance functions(Neural Network)
17 views (last 30 days)
I would like to train a network by writing a new custom performance function. Here's how to customize the performance function. As explained above, the best way to create a custom performance function by using "mse.m:" and the +MSE function package as a template, that is, consider using the template provided by "+mse" to create a new package "+newfcn".
This is the contents of the `+mse` folder. What makes up the cost function seems to be structured in these atom functions. However, their exact roles are unclear.
apply.m backprop.m forwardprop.m perfwb.m
name.m type.m normalize.m parameterInfo.m
Since the code is simple and uncommented, it's hard for me to understand how they implement the functionality. Also, for reference, `apply`, `backprop`, and `forwardprop` functions are concatenated below:
function perfs = apply(t,y,e,param)
perfs = e .* e;
function dperf = forwardprop(dy,t,y,e,param)
dperf = bsxfun(@times,dy,-2*e);
function dy = backprop(t,y,e,param)
dy = -2*e;
I'm having a hard time trying to understand how these work together to compose a Mean Squared Error function. It seems like the classical implementation of MSE was decomposed into steps to achieve more flexibility, but combined with no documentation it's become very hard to understand.
What are the roles of the cost function's subfunctions, and how they should be implemented? Where do I need to modify the code if I want to implement a custom performance function？
Thanks in advance.
David Willingham on 6 Sep 2022
The code you refer to above is referencing the older classic neural networks framework, which is primarily used for shallow neural networks.
We have a newer, updated framework for deep learning that allows you to fully customize how yo can train a model. More information on this can be found here.
Can you provide more information on the network you are looking to customize?