neural net toolbox: divideFcn = ''

Question

Fabio Muratore 2015-12-8

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/259381-neural-net-toolbox-dividefcn

评论： bear96 2020-1-1

Hello,

I read a lot of Q&A about neural nets on MATALB Answers lately. Someone (sadly I wasn't able to find the post again) mentioned to set the dividing funcion of the input data to ''. i.e.

NN = narxnet(0:5, 1:5, [10 10]); % creation of a narxnet
NN.divideFcn = ''; % usually I use 'divideblock' here

I get different results for NN.divideFcn = '' or NN.divideFcn = 'divideblock'.

My question is: How does this divideFcn behave. Or does this line of code lead to some default value for the divideFcn?

Thank you in advance for your answers.

0 个评论
显示 -2更早的评论隐藏 -2更早的评论

请先登录，再进行评论。

请先登录，再回答此问题。

Answer 1

Greg Heath 2015-12-9

0
链接

此回答的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/259381-neural-net-toolbox-dividefcn#answer_202654

在 MATLAB Online 中打开

In general you can specify

    1. The type of data division
    2. The trn/val/test ratios

If you don't specify anything, you will get the default of random data division with trn/val/test ratios 0.7/0.15/0.15.

If you specify ' ' or 'dividetrain', you will get no division i.e., 100/0/0.

If you only specify 'divideblock' the data will be divided into 3 solid blocks of trn/val/test with the default 0.7/0.15/0.15 ratios. However, you can also specify another set of block ratios. This is the datadivision I typically recommend for timeseries prediction.

The disappointing part of this thread is that you should have figured this out by

 1. Reading the help and doc documentations for the 
different datadivision options. 
 2. Demonstrating the options on a small data example. 
For example,
 input = 1:10;  target = input.^2;

Hope this helps.

Thank you for formally accepting my answer

4 个评论
显示 2更早的评论隐藏 2更早的评论

Greg Heath 2015-12-9

编辑：Greg Heath 2016-1-4

在 MATLAB Online 中打开

>> lookfor divide ...

divideblock - Partition indices into three sets using blocks of indices.

divideind - Partition indices into three sets using specified indices.

divideint - Partition indices into three sets using interleaved indices.

dividerand - Partition indices into three sets using random indices.

dividetrain - Partition indices into training set only.

% net.divideFcn = ' '; is the same as net.divideFcn = dividetrain';

>> help divideblock

 [trainInd,valInd,testInd] = 
           divideblock(Q,trainRatio,valRatio,testRatio)
takes a number of samples Q and divides up the sample indices 
1:Q between training, validation and test indices.
divideblock returns the first N1 indices as trainInd, the next 
N2 indices as valInd and the remaining N3 indices as testInd. 
The sum of N1, N2 and N3 is Q, and the proportions between them 
match the three ratios.
For example, here 250 samples are divided into 70% for training, 
15% for validation and 15% for testing.
    [trainInd,valInd,testInd] = divideblock(250,0.7,0.15,0.15)
Here is how to ensure a network will perform the same kind of 
data division when it is trained:
    net.divideFcn = 'divideblock';
    net.divideParam.trainRatio = 0.7;
    net.divideParam.valRatio = 0.15;
    net.divideParam.testRatio = 0.15.
See also divideind, divideint, dividerand, dividetrain.
    Reference page in Help browser
       doc divideblock

>> doc divideblock

....

% REPEAT FOR THE OTHER FOUR OPTIONS

Greg Heath 2015-12-10

编辑：Greg Heath 2016-1-4

在 MATLAB Online 中打开

If you look at the figures in the help and doc documentation, you will see that

 timedelaynet is trained with targets and external inputs; and 
deployed with external inputs that generate outputs.
 narnet is trained in the open-loop (OL) configuration with 
targets replacing external inputs. It is deployed in the 
closed-loop (CL) configuration with output feedback replacing 
the open-loop target inputs.
 narxnet is trained in the OL configuration with both external 
inputs and target inputs. It is deployed with external inputs 
and output feedback replacing OL target  inputs.