CoreML - Models creations

A first look at last Apple great framework CoreML and the community around for models creations.

At the end we stop by models creation on the MNIST example.

Transcript

1. CoreML Models creation CocoHeads Nantes 14 September 2017 David Bonnet - @iGranDav
2. CoreML First Look
3. First Look • Bringing machine learning to everyone ! • Integrate your model and CoreML does all the job for you • Problem is « your model » • Apple helps by providing you some https:/ /developer.apple.com/machine-learning/
4. Demo
5. I want more!
6. Models from the ML community libsvm
7. Models from the ML community libsvm Neural networks Pipelines Tree Ensembles Linear models Support Vector Machines
8. CoreML Conversion Tool • Apple provide a CoreML conversion tool in python (standard for ML) https://developer.apple.com/documentation/coreml/converting_trained_models_to_core_ml import coremltools coreml_model = coremltools.converters.caffe.convert(‘my_caffe_model.caffemodel’) coreml_model.save(‘my_model.mlmodel') • All previous presented ML communities are supported • You can write your own converter if not yet supported https://pypi.python.org/pypi/coremltools
9. All together https://github.com/cocoa-ai/FacesVisionDemo • Using Vision framework to get face parts • 3 models from Caffee model zoo • Gender classification • Age classification • Emotion recognition https://github.com/caffe2/caffe2/wiki/Model-Zoo
10. Great! But I still want more!
11. Creating our own model • Let’s do an hello world: MNIST (number recognition) https://www.tensorflow.org/get_started/mnist/beginners https://colah.github.io/posts/2014-10-Visualizing-MNIST/
12. Creating our own model • Let’s do an hello world: MNIST (number recognition) https://www.tensorflow.org/get_started/mnist/beginners https://colah.github.io/posts/2014-10-Visualizing-MNIST/
13. Creating our own model • Let’s do an hello world: MNIST (number recognition) https://www.tensorflow.org/get_started/mnist/beginners https://colah.github.io/posts/2014-10-Visualizing-MNIST/
14. Convolutions • Big word to say « matrix operations » Simple box blur Edges detection http://www.aishack.in/tutorials/image-convolution-examples/
15. Max Pooling • Reduce the amount of data by downsampling it http://cs231n.github.io/convolutional-networks/
16. Rectifier Linear Unit - ReLU • An activation function defined as f(x) = max(0,x) https://en.wikipedia.org/wiki/Rectifier_(neural_networks)
17. Convolutional Neural Network http://cs231n.github.io/convolutional-networks/ • Example of a car picture recognizer
18. Demo https://sriraghu.com/2017/07/06/computer-vision-in-ios-coremlkerasmnist/ https://www.tensorflow.org/tutorials/layers
19. Models optimization • Prune them ( removing extra nodes ) • Reduce them ( combining nodes ) • Quantize them ( a Double can be a Int ) • Align them ( mmap for architecture optimization )
20. To play more • Awesome CoreML models • https:/ /github.com/likedan/Awesome-CoreML-Models • Dynamic CoreML • https:/ /github.com/zedge/DynamicCoreML/tree/master/GAN
21. -Steve Jobs « You can do anything! You just have to code it! »