This project is our own implementation of a simple feedforward, fully connected neural network for number classification.
Note: This is a fork of an older project we originally created in 2015. In this fork, we've made some improvements but haven't modernized the codebase.
This is a Maven (Java) project currently developed using the IntelliJ IDE.
In addition to the source folder, there's a data folder containing two subfolders
- input - Training data (IDX files containing vectors of handwritten numbers)
- output - Created models (serialized Java objects)
Clone the repository and install the Maven dependencies.
This project contains two applications: one for creating and training models, and another for using the trained models.
In the nl.uitdehoogte.ann root package, you'll find the Main class containing the logic for both tasks. Since this is just a playground for us, there are no CLI options or configurations. If you want to experiment, simply modify the code and run the class.
We've created a simple UI for drawing and recognizing numbers. Its main Controller class can be found in the nl.uitdehoogte.ann.ui package.
As with the Main class, you'll need to manually set the desired model in the code before starting the app.
Once started, you can draw on the left using the primary mouse button. Then, click the Execute button to see which number the model predicts.
To clear the drawing, simply right-click on the canvas (i.e., use the secondary mouse button).
To test the model against the verification test set, you can use the (almost self-descriptive) Read Next button. This will randomly pick a sample from the test set, draw it on the canvas, and automatically click the Execute button.