TensorFlow Neural Networks on iOS

Slide 1

Slide 1 text

TensorFlow on iOS Taylan Pince

Slide 2

Slide 2 text

Using TensorFlow, CoreML, Metal Performance Shaders, Accelerate BNNs, Keras and What the Heck is a Neural Network Anyway? Taylan Pince

Slide 3

Slide 3 text

No content

Slide 4

Slide 4 text

No content

Slide 5

Slide 5 text

No content

Slide 6

Slide 6 text

No content

Slide 7

Slide 7 text

No content

Slide 8

Slide 8 text

No content

Slide 9

Slide 9 text

No content

Slide 10

Slide 10 text

No content

Slide 11

Slide 11 text

No content

Slide 12

Slide 12 text

No content

Slide 13

Slide 13 text

No content

Slide 14

Slide 14 text

No content

Slide 15

Slide 15 text

No content

Slide 16

Slide 16 text

No content

Slide 17

Slide 17 text

No content

Slide 18

Slide 18 text

What is a Neural Network Anyway?

Slide 19

Slide 19 text

No content

Slide 20

Slide 20 text

if x + y > b { return blue } else { return orange }

Slide 21

Slide 21 text

if (xWeight * x) + (yWeight * y) > b { return blue } else { return orange }

Slide 22

Slide 22 text

No content

Slide 23

Slide 23 text

No content

Slide 24

Slide 24 text

No content

Slide 25

Slide 25 text

No content

Slide 26

Slide 26 text

No content

Slide 27

Slide 27 text

No content

Slide 28

Slide 28 text

No content

Slide 29

Slide 29 text

Network Training Basics

Slide 30

Slide 30 text

1. Data Gathering & Balancing 2. Preprocessing 3. Training 4. Testing Results

Slide 31

Slide 31 text

No content

Slide 32

Slide 32 text

No content

Slide 33

Slide 33 text

No content

Slide 34

Slide 34 text

No content

Slide 35

Slide 35 text

No content

Slide 36

Slide 36 text

No content

Slide 37

Slide 37 text

No content

Slide 38

Slide 38 text

No content

Slide 39

Slide 39 text

import tensorflow as tf matrix_size = 224 * 224 category_size = 150 with tf.name_scope("data"): d1 = tf.placeholder(tf.float32, [None, matrix_size], name="image_data") d2 = tf.placeholder(tf.float32, [None, category_size], name="category_data") with tf.name_scope("model"): weights = tf.Variable(tf.zeros([matrix_size, category_size]), name="weights") bias = tf.Variable(tf.zeros([category_size]), name="bias")

Slide 40

Slide 40 text

No content

Slide 41

Slide 41 text

Slide 42

Slide 42 text

Neural Networks on iOS

Slide 43

Slide 43 text

TensorFlow CoreML Metal Performance Shaders Accelerate

Slide 44

Slide 44 text

C++ library Adds around 40MB to final app size Cannot use Bitcode Cannot use GPU

Slide 45

Slide 45 text

Use freeze_graph & optimize_for_inference Import final pb file into Xcode project

Slide 46

Slide 46 text

tensorflow::GraphDef graph; tensorflow::Session *session; ReadBinaryProto(tensorflow::Env::Default(), path, &graph); tensorflow::NewSession(options, &session); session->Create(graph); tensorflow::Tensor x( tensorflow::DT_FLOAT, tensorflow::TensorShape({ 1, 224 * 224 }) ); std::vector outputs; session->Run(inputs, nodes, {}, &outputs);

Slide 47

Slide 47 text

Limited support for training engines and layer types Custom models need conversion Picks CPU or GPU automatically

Slide 48

Slide 48 text

Pretrained Models Inception v3 VGG16 MobileNet SqueezeNet

Slide 49

Slide 49 text

Custom Models Convert Caffe or Keras models with coremltools Import mlmodel file into Xcode project

Slide 50

Slide 50 text

let model = VNCoreMLModel(for: graph().model) let request = VNCoreMLRequest(model: model) { [unowned self] request, error in results.forEach({ (result) in print("\(result.identifier)") }) } } let handler = VNImageRequestHandler(ciImage: image) DispatchQueue.global(qos: .userInitiated).async { do { try handler.perform([request]) } catch { print(error) } }

Slide 51

Slide 51 text

Low-level API behind CoreML Always runs on GPU Got tons of love with iOS11 updates

Slide 52

Slide 52 text

Convert pb file into a binary Metal can read: A list of floating point numbers

Slide 53

Slide 53 text

Delivering Updates

Slide 54

Slide 54 text

Recap Train with TensorFlow + Keras Use CoreML if you can Use TF if you need multi-platform

Slide 55

Slide 55 text

Further Reading Matthijs Hollemans (MachineThink.net) Reza Shirazian (reza.codes)