Machine Learning in Swift

My very first attempt to use coremltools

Since I was able to install coremltools as described in the previous post, it's time to try to convert an actual model and run it with MLKit. 

I have chosen "VGG Face CNN descriptor" from Model Zoo as the very first model to tray -- for no particular reason. I just needed a model to try. 

First, I downloaded the model from this link. It contains .prototxt, .affemodel, .txt files and a sample image. 

Second, I converted them into .mlmodel file using coremltools

 (p27)# python
 >>> import coremltools
 >>> caffe_model = ('VGG_FACE.caffemodel', 'VGG_FACE_deploy.prototxt')
 >>> labels = 'names.txt'
 >>> coreml_model = coremltools.converters.caffe.convert(caffe_model, class_labels=labels, image_input_names='data')
 >>> coreml_model.save('DeepFace.mlmodel')

Third, I created a new Xcode project (a single view iOS app), and imported 'DeepFace.mlmodel' file along with 'ak.png' file as a test image and wrote a following piece of code.

As a property of ViewController class,

  let model = DeepFace()

In viewDidLoad() function,

    let files = [
      "ak.png"
    ]
    for file in files {
      let image = UIImage(named: file)!
      let buffer = image.pixelBuffer()!
      if let output = try? model.prediction(data: buffer) {
        print(file, output.classLabel)
      } else {
        print(file, "failed")
      }
    }

It correctly detected the test image as "Aamir Khan".

I have also tried several other images from the internet (you need to resize and crop them into 224x244 image), and most of them were identified correctly (as long as the name is in the list). 

I have tried to upload the source code (along with .mlmodel file) to github, but the .mlmodel file is too big (580MB). You can easily create your own version by following steps above. 

Installing coremltools on MacOS

After watching ML Core in Depth (one of WWDC 2017 video), I have decided to try coremltools on my MacBook Pro (MacOS Sierra, Version 10.12.4). 

When I type,

 # pip install coremltools

it fails with the following error message:

Could not find a version that satisfies the requirement coremltools (from versions: )
No matching distribution found for coremltools

This is because Python on my MacOS is 3.5.2, while coremltools requires 2.7. 

To solve this problem, I have installed virtualenv by typing

 # pip install virtualenv

which will install vittualenv (globally). 

Then, I created a virtual environment "py27" with following command

 # virtualenv --python=/usr/bin/python2.7 py27

where py27 is the name of directory (to hold this environment). 

Once this environment is created, I can activate it by typing

 # source py27/bin/activate

You can verify it by typing

 (p27)# python --version
 Python 2.7.10

Now you are ready to install coremltools by typing

 (p27)# pip install -U coremltools

By the way, you can exit from this virtual environment by typing

 (p27)# deactivate