The Scale Program is a utility intended to make it easier to prepare input data. The Scale Program enables you to take a raw data file (Training, Testing or Validation) and make it useable by Propagator for neural network training and testing. The Scale Program performs three basic functions:
1. Scales the input and target data and places it in an output file.
2. Sets the delimiters to tabs between items and a hard return at the end of each pattern.
3. Creates a report file detailing the parameter specifications of the input and output files.
Even if you do not change the scale of the data, you can use the Scale Program to change the delimiters to enable Propagator to read the data file. For example, the Scale Program changes data with comma delimiters to data with tab delimiters. This is useful because Propagator does not accept data with comma delimiters.
The Scale Program has a single data entry and display window with the following fields:
Number of Nodes (Input) — This is the number of input nodes in the network.
Scale Input Patterns — Select this check box to scale the input patterns. The New Range Low and High text boxes become active when Scale Input Patterns is selected. If the input values are not scaled, the delimiters are changed to a form acceptable to Propagator.
New Range Low & High (Input Patterns) — Used to select the range that you want the input data to be within after scaling. The scale range limits are -3.40e+38 for the low limit and 3.40e+38 for the high limit. The range to which the input data should be scaled depends largely on the transfer function used for training. If a linear transfer function is used, there is no need to scale the input data. If the sigmoid transfer function is used, the input data typically is scaled between 0 and 1.0. If the tanh transfer function is used, the input data typically is scaled between -1.0 and 1.0.
Number of Nodes (Target) — This is the number of output nodes in the network.
Scale Target Patterns — Select this check box to scale the target patterns. The New Range Low and High text boxes become active when Scale Target Patterns is selected. If the target values are not scaled, the delimiters are changed to a form acceptable to Propagator.
New Range Low & High (Target Patterns) — Used to select the range that you want the target data to be within after scaling. The scale range limits are -3.40e+38 for the low limit and 3.40e+38 for the high limit. The range to which the target data should be scaled depends largely on the transfer function used for training. If the linear transfer function is used, there is no need to scale the target data. If the sigmoid transfer function is used, the target data typically is scaled between 0.01 and 0.99. If the tanh transfer function is used, the target data typically is scaled between -0.99 and 0.99.
Select Input File... — Pressing this button presents an Open dialog box. The Input File containing the raw data that you want to scale is selected from the Open dialog. The name of the file is displayed in the Input File box.
Select Output File... — Pressing this button presents a Save As dialog box. You can select a name for the Output File containing the data that is scaled. The name of the file is displayed in the Output File box.
Progress — Displays the progress of the scaling when the Scale button has been pressed. The percent completion, the number of patterns completed and the total number of patterns are displayed numerically. A graphic bar actively displays the percent completion during scaling.
