Control the epochs while training a neural network

12 次查看（过去 30 天）
Dingguo Lu
Dingguo Lu 2012-9-5
I am trying to train a BP neural network with the following codes. I mean to run the iterations for 1000 epochs. However, when the net.trainParam.goal = 0 is achieved, the training process will stop, which is much less than 1000.
How can I set some parameters so that I can train the neural network for 1000 times. I can to plot MSE VS. epoch.
Thanks!
%%%%%%%%%%%%%%%%%%%% clear all; close all; clc;
% Number of Inputs(n), Outputs(r) and neurons in hidden layer(m) n = 1; r = 1; m = 12; % Number of training values (epochs) epochs = 1000; % Input value range x_min = -1; x_max = 1; for k = 1:n x_train(k,:) = x_min + (x_max-x_min)* rand(epochs,1); end % Desired values for random vector x_train y_des = 2*x_train.^2 + 1;
net = feedforwardnet(m);
net.trainParam.epochs = 1000; % epoch net.trainParam.show = 10; % show frequency net.trainParam.goal = 0; % objective MSE
net = train(net, x_train, y_des);

请先登录，再回答此问题。

采纳的回答

Greg Heath
Greg Heath 2012-9-5
The purpose of training is to reduce mse to a reasonably low value in as few epochs as possible. When training is sufficiently long, the plot of mse will asymptotically decrease to a horizontal straight line at mse = 0.
Therefore, your request makes no sense to me.
In fact, if the training target is standardized (zero-mean/unit-variance rows) via the functions zscore or mapstd, there is no practical reason to reduce mse below
~mean(var(transpose(target)))/100.
Hope this helps.
Greg

更多回答（0 个）

类别

在 Help Center 和 File Exchange 中查找有关 Sequence and Numeric Feature Data Workflows 的更多信息

标签

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by