5. Network training
Before the network can be used it must be trained using sample patterns. Some example patterns are supplied with the software in the ‘/training_patterns’ folder. In this folder there is a series of sub folders each representing an example of a computer generated character. The user can make their own patterns by using the format described in section 8. Training is activated via the ‘operations’ menu. All networks that are eligible for training appear in a list beside the ‘train’ menu item. Any network that has already been trained is removed from this list. By clicking this menu item opens the train dialog shown in figure two.
Figure 2: The train dialog
To train the network the user must select the locations of the folders containing each training set. These are individually added to the list. The ‘browse’ button in used to select the required directory. After this has been complete the dialog will be updated showing the number of patterns found. Ideally this should be between 5 and 20 to avoid network saturation. Before a class can be added to the training list it must be given a name using the text box above the ‘add training set’ button. The set is added once this button is pressed, the list will update showing the newly added training set. Figure 2 shows 4 training sets that have been added by selecting 4 individual folders.
The ‘train’ button will only start the training process when one or more training sets have been added to the list. If this has been achieved then a progress bar will appear showing the training progress. After this has completed the trained network will be added to the list beside the ‘recognition’ and ‘batch recog’ functions. It is recommended that you save your network now.