ph1ps / mnist-coreml

Predict handwritten digits with CoreML

MNIST for CoreML (CNN)

Description

This is the MNIST dataset implemented in Apple's new framework CoreML. The MNIST dataset can predict handwritten (drawn) digits from an image and outputs a prediction from 0-9. The model was built with Keras 1.2.2.

To test this model you can open the MNISTPrediction.xcodeproj and run it on your device (iOS 11 and Xcode 9 is required). To test further images just add them to the project and replace my testing with yours.

An example of a handdrawn digit would look like this:

Be aware that your images have to have a black background and white line color!

Furthermore your images resolution has to be 28x28px. If yours is bigger just use my UIImage rescaling extension I wrote. The line width has to be thick enough to be recognized as a digit.

Information about the model

This CNN model achieves up to 99.5% of accuracy and the structure is as follows:

mnist-coreml's Issues

'Conv2D' object has no attribute 'inbound_nodes'

When I try to run the python script to get a coreML model , I get an error saying 'Conv2D' object has no attribute 'inbound_nodes'. How do I resolve this?