• Train a CNN with only images
• Convert pre-trained keras models to Core ML models
• Train and Convert a CNN into a Core ML Model
• Edit attributes of existing Core ML Models

Swifty-ML allows developers with no experience in ML to create classifiers and improve their apps. It's also good for playing around!

• Keras
• coremltools
• Python 3.6

To train a model using Swifty-ML, run the following command in terminal.

python modeler.py train path/to/data inceptionv3

This usually takes a very long time, even with a good laptop, but once finished, there will be two files. A Core ML Model and a Keras .h5 model. Saved in exported_models/coreml and exported_models/keras respectively. The system also saves the state of the model during training after each epoch.

To edit a coreml model...

python modeler.py edit path/to/coreml/model.mlmodel

To convert a Keras model (.h5) to CoreML format

python modeler.py convert-pre-trained path/to/keras/model.h5

The format for training and validation images must be the following

.
├── training
│   ├── class1
│   ├── class2
│   └── class3
└── validation
    ├── class1
    ├── class2
    └── class3

If I was training a network on images of the different flags, I would have a structure like so...

.
├── training
│   ├── belgium
│   ├── canada
│   └── unites_states
└── validation
    ├── belgium
    ├── canada
    └── unites_states

Note: training and validation cannot change, but the sub folders' names can, and will be the class labels for the images inside them.

As of now, Core ML Tools does not provide support for Python 3.6 via pip. Instead, you will need to download the latest version from their GitHub and install it manually.

https://github.com/apple/coremltools

After downloading the project, cd into the project and run

python setup.py install

As of now, only a few model architectures are supported. In the future, all models supported by both coremltools and keras will be supported.

Future models include

• Xception
• VGG16
• VGG19
• ResNet50
• InceptionV3
• InceptionResNetV2
• MobileNet