Digit Recognizer

The goal of this project is to recognize handwritten digits. It uses deep learning to train a handwritten digit recognizer model by using MNIST data as the training data. The model is built with Keras and Tensorflow. The accompanying Java app can be used to manually test the model. It uses Deeplearning4J to read the Keras model and enables the user to draw a digit that is then used as an input for the model.

Running the Java app

Requirements

• Java 1.8 or newer
• Maven 3.2.2 or newer

Building and running

Clone the repo

git clone https://github.com/Miksu82/DigitRecognizer

Build and run the Java app

cd app
maven clean package
java -jar target/digitdetector-1.0-SNAPSHOT.jar

Modifying the model

Requirements

This is my setup that I have tested with. Older versions may also work.

• Java 1.8
• Python 2.7
• Keras 1.2.2
• Tensorflow 1.0.1
• Numpy 1.12.0

Retraining the model

The repository contains prebuilt model saved to

trainer/mnist.h5f

The current accuracy is around 97 %. If you want to you tweak the parameters and retrain the model you can do that by modifying

trainer/mnist-trainer.py

and while in trainer directory run

python mnist-trainer.py

The model will always be saved to your working directory so if you run the script from somewhere else the Java app can not find the new model.

Note

This is not meant to be state-of-art (or even close) digit recognizer. The project goal was to learn a bit about deep learning and see how the most used deep learning frameworks are used.

For more information read the related blog post