5.1.4 NetPlot

The command NetPlot can be used to illustrate the derived FF network or the evolution of the training. Depending on how the option DataFormat is set, the command can be used in very different ways.

Illustrate models and training of models.

If the input dimension is one or two, the default is to plot the estimated function in the range of the supplied data. In the one-dimensional case, the data is also shown.

When NetPlot is applied to an FF network, it takes the following options.

Options of NetPlot.

In addition to these, you can submit options to modify the graphical output. Depending on the chosen option for DataFormat, the graphic is created by BarChart, BarChart3D, MultipleListPlot, ListPlot, Plot3D, or Histogram.

If the input dimension is higher than two, then the default is to plot the numerical values of the hidden neurons versus the data. This can be obtained also in one- and two-dimensional problems by choosing DataFormatHiddenNeurons. Notice, however, that plotting the neurons versus data only makes sense if the input signal vectors are placed in some kind of order.

The option DataFormat takes any of the following values:

FunctionPlot: plots the mapping using the range of the supplied data. It can only be used if the model has one or two inputs.

NetOutput: plots the network output versus the given output. A perfect fit corresponds to a straight line with slope 1 through the origin.

ErrorDistribution: gives a histogram of the errors when the model is applied to submitted data. You can modify the presentation of the result using any of the options applicable to Histogram.

HiddenNeurons: gives the values of the hidden neurons when the model is evaluated on the data. This function makes most sense when it is applied to dynamic models.

ParameterValues: plots the parameters versus the training iterations.This is only possible for the training record.

LinearParameters: plots the parameters of the linearization versus data. This function makes most sense when it is applied to dynamic models.

The following three possibilities are primarily intended for classification models:

Classifier: shows the borders between different classes. It can only be used with two-input models.

BarChart: illustrates the classification result with bar charts.

ClassPerformance: plots the improvement of the classification for each class versus the number of training iterations. Correctly classified samples are marked with diamonds and solid lines, while incorrectly classified samples are indicated with stars and dashed lines. This is only possible for the training record.

If you submit a training record instead of an FF network, then depending on which option is given, you obtain a graphic array of the corresponding results as a function of the number of training iterations. For a large number of iterations, it is advisable to set the option Intervals to a larger integer, thus controlling the size of the length of the graphic array.

Examples where NetPlot is used to evaluate FF networks are given in Section 5.3, Classification with Feedforward Networks and Section 5.2, Examples.