Digits: A neural net classifier trained in Python and deployed to iOS via Objective-C++

This project shows how to generate and train a series of handwritten digit classifiers in Python and deploy a final model (99% accuracy) in iOS. For more info on the training procedure and results see this post.

The project is divided into 2 parts:

Part 1: Training and Testing in Python (python folder):

1. Models were generated for a combination of hyper parameters and topologies and trained on the MNIST dataset.
2. The best 5 models were kept and tested both on a portion of the MNIST test set as well as custom images.
3. A final model was chosen out of the 5 to deploy to iOS.

Due to space considerations, the project only contains the final model, along with stats for the whole training session of all models.

Contents:

• digits.py: the entry point for training and testing.
• training.py: the training algorithm, including topology + hyperparameter search.
• testing.py: testing algorithms for both MNIST as well as custom images, including OpenCV code to preprocess custom images.
• statsplot.py: plotting of training statistics.
• data: the MNIST dataset plus custom images used in testing.
• training: a copy of the the final model and training stats from the session in which the final model as well as many others were generated.

Part 2: Deploying to iOS (ios folder):

An iOS app was implemented to take photographs, preprocess them and run the trained model on them to predict digits.

Contents:

• ViewController: Manages both the UI of the app and an instance of NNModel to do all preprocessing and prediction.
• NNModel: The iOS-side model, written in Objective-C++ to interoperate between Cocoa, OpenCV and the trained model.
• nn.hpp, arithmetic.hpp, activation.hpp: C++ implementation of subset of NNKit (originally written in Python), necessary to use the trained model in iOS.
• 3rdparty: C++ libraries used by the NNKit C++ subset.

Dependencies:

Python:

• Numpy
• NNKit
• OpenCV (if you want to preprocess custom images before testing).
• matplotlib (if you want to plot training stats).

iOS:

• Eigen (header only, included in project).
• JSON for Modern C++ (header only, included in project).
• OpenCV (iOS framework only).
• XCode

Installation:

After downloading or cloning the repo:

• Python: just pip install -r requirements.txt
• iOS / Xcode: download the OpenCV framework and link against it in the Xcode project.