MXNet on the iPhone

This is an example project running a binarized neural network on iOS. It can detect single handwritten digits.

Requirements

• OpenCV framework (e.g. from here)
• model trained on mnist (pre-trained binarized version included)

Usage

Add the OpenCV framework to the xcode project and then build the app and try it on your phone!

What it does

Under the hood, we need the amalgamated MXNet source, the c predict headers, a trained and saved model from mxnet and set some additional environment variables in xcode.

The mxnet script amalgamate_mxnet.sh will amalgamate mxnet and perform the changes necessary for ios as described in the amalgamation readme file.

The xcode project already contains the preprocessor settings required to build mxnet for ios:

• "MXNET_PREDICT_ONLY=1"
• "MXNET_USE_OPENCV=0"
• "MSHADOW_USE_CUDA=0"
• "MSHADOW_USE_SSE=0"
• "BINARY_WORD_32=1" (set to 32bit for ARM7 devices)
• "BINARY_WORD_64=0"

There are pre-trained models included in the projects.