Customizing the Neural Network model and details in nnstart

Question

musa 2024-5-12

0
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2117861-customizing-the-neural-network-model-and-details-in-nnstart

回答： Pratyush Swain 2024-5-24

I am not so experienced in Matlab and it's tools. I am trying to learn essential apps with details.

I know nnstart (formerly nntool) and how to create a neural network, then train, and then get the results, following the instructions.

But I could not see the details screens (or code parts) that I can change/update the details of my neural network. The application screen implements everything and I cannot intervene in the details. Neither from app "Neural Networks Fitting" window, nor codes (I already cannot see each code).

Just for example, I want to build/update/change the weights, activation functions (selecting one or entering my own activation functions), neuron counts, input alingment etc. , also want to be able to break during the iterations and code breaks in some parts. Additionally draw a neural network visual scheme as well, if possible. Then I will be able to see the details during the implementation.

Is there a way in Neural network Fitting (at nnstart command) that I can get the control ?

Thank you

1 个评论
显示 -1更早的评论隐藏 -1更早的评论

musa 2024-5-12

I have found an another tool named Deep Network Designer. Looks like this tool allows to manage some parts of your new neural network. Not sure I can do everything or not.

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Pratyush Swain 2024-5-24

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/2117861-customizing-the-neural-network-model-and-details-in-nnstart#answer_1462916

在 MATLAB Online 中打开

Hi musa,

nnstart and its associated GUI tools (like the Neural Networks Fitting Tool) offer a straightforward path to creating and training neural networks, the depth of customization and interaction you can have with your model's internals is limited, such as adjusting weights, activation functions.

The Deep Network Designer allows for the drag-and-drop construction of neural network layers, enabling you to visually build and customize your network architecture. This includes selecting different types of layers, adjusting their parameters.

For more information, you can refer to following resources:

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

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

While it focuses on structural design (layer types, connections, etc.), for detailed adjustments like custom weights or activation functions, you may need to export the designed network and modify these properties programmatically.

Here's how you can perform customization of activation functions and manual initialization of weights of a simple feedforward Network in MATLAB:

% Create a feedforward network with 10 neurons in the hidden layer
net = feedforwardnet(10); 
 
% Use the logistic sigmoid activation function for the first hidden layer 
net.layers{1}.transferFcn = 'logsig'; 
% Use softmax activation function for the output layer
net.layers{2}.transferFcn = 'softmax'; 
 
% Randomly initialize input weights 
net.IW{1,1} = rand(size(net.IW{1,1}));
% Set all biases in the first hidden layer to 1
net.b{1} = ones(size(net.b{1})); 