On-device ML with TensorFlow Lite DevFest Vancouver Margaret Maynard-Reid, 9/7/2019 @margaretmz

@margaretmz | #MachineLearning #GDE Slides Slides for this talk are posted on speakerdeck: bit.ly/on-device-ml-tflite-devfest-vancouver Click on download PDF to access the links 2

@margaretmz | #MachineLearning #GDE Topics ● Intro to TF 2.0 & tf.Keras ● On-device ML options ● E2E tf.Keras to TFLite to Android ○ train a model from scratch ○ convert to TFLite ○ deploy to mobile and IoT ● TFLite models on microcontroller & Coral Edge TPU 3

Intro AI, ML, Deep Learning Computer Vision TensorFlow, Keras 4

@margaretmz | #MachineLearning #GDE AI vs. ML vs. Deep Learning Artificial Intelligence Machine Learning Deep Learning: - Computer Vision - NLP ….

@margaretmz | #MachineLearning #GDE Examples of computer vision 6 Generative Adversarial Networks (GANs) Generating new images Image classification Is this a cat? Object detection Drawing bounding boxes around the objects Dance Like @I/O Segmentation, pose, GPU on-device Other examples: - Photos enhancement - Style transfer - OCR - Face keypoints

@margaretmz | #MachineLearning #GDE Deep Learning - getting started ● Deep learning Frameworks: ○ TensorFlow (>129k stars on Github) ← most popular! ○ PyTorch ○ Caffe (1 & 2) ○ Theano… ● Languages: Python, Swift, Javascript etc. ● IDE - Colab ● Popular neural networks: ○ CNN (Convolutional Neural Networks) ○ RNN (Recurrent Neural Networks) ○ Generative Models (Auto encoder, GANs) ○ ... 7

@margaretmz | #MachineLearning #GDE TensorFlow model building APIs TensorFlow is a deep learning framework for both research & production Write TensorFlow code in C++, Python, Java, R, Go, SWIFT, JavaScript Deploy to CPU, GPU, TPU, Mobile, Android Things, Raspberry Pi tf.* tf.layers tf.keras Custom Estimator Premade Estimator ← Low level ← Mid level (moving to tf.keras in TF 2.0) ← High level ← Model in a box ← Distributed execution, tf serving 8 TensorFlow 2.0 Beta just got announced! | My Notes on TensorFlow 2.0

@margaretmz | #MachineLearning #GDE tf.Keras vs Keras No 1:1 mapping between tf.Keras and Keras 9 tf.keras - part of the TensorFlow core APIs import tensorflow as tf # import TensorFlow from tensorflow import keras # import Keras Keras remains an independent open-source project, with backend: ● TensorFlow (Protip: use tf.keras, instead of Keras + TF as backend) ● Theano ● CNTK...

@margaretmz | #MachineLearning #GDE tf.Keras model building APIs ● Sequential - the easiest way ● Functional - more flexibility ● Model subclassing - extend a Model class Learn more in Josh Gordon’s blog: What are Symbolic and Imperative APIs in TensorFlow 2.0? 10

@margaretmz | #MachineLearning #GDE Blog.tensorflow.org TensorFlow and ML learning resources Tensorflow.org Deep learning with Python by Francois Chollet TensorFlow on Youtube TensorFlow on Twitter #AskTensorFlow #TensorFlowMeets Collection of interactive ML examples (blogpost | website) 11 Interested in learning about TensorFlow 2.0 and try it out? Read My Notes on TensorFlow 2.0 TensorFlow Dev Summit 2019 By Aurélien Géron

@margaretmz | #MachineLearning #GDE Anaconda, TensorFlow & Keras Why use a virtual environment? Ease of upgrade/downgrade of tensorflow ● Download anaconda here ● Create a new virtual environment $ conda create -n [my-env-name] ● Activate the virtual environment you created $ conda activate [my-env-name] ● Install TensorFlow beta $ pip install tensorflow==2.0.0-beta1 My blog post Anaconda, Jupyter Notebook, TensorFlow, Keras 12

@margaretmz | #MachineLearning #GDE Google Colab What is Google Colab? ● Jupyter Notebook ○ stored on Google Drive ○ running on Google’s VM in the cloud ● Free GPU and TPU! ● TensorFlow is already installed ● Save and share from your Drive ● Save directly to GitHub 13 Check out these learning resources ● My blog on Colab ● TF team’s blog on Colab ● Laurence’ Video Build a deep neural network in 4 mins with TensorFlow in Colab ● Paige’s video How to take advantage of GPUs & TPUs for your ML project ● Sam’s blog Keras on TPUs in Colab Launch Colab from colab.research.google.com/

@margaretmz | #MachineLearning #GDE TensorBoard in Colab TensorBoard now integrated in Colab! ● Debug ● Monitor ● Visualize Lab - https://www.tensorflow.org/tensorboard/r2/tensorboard_in_notebooks 14

ML Pipeline 15

On-device ML What are your options? 16

@margaretmz | #MachineLearning #GDE TensorFlow for edge devices 17 2015 TF open sourced 2016 TF mobile 2017 TF Lite developer preview 2018 ML Kit 2019 TF Mobile deprecated ML Kit improves TF Lite exits dev preview More than just mobile apps: ● Microcontrollers ● Edge TPUs

@margaretmz | #MachineLearning #GDE TensorFlow Lite ● For deploying to edge devices ● Works with Inception & MobileNet ● May not support all operations ● Supports ○ Mobile: Android & IOS ○ Android Things ○ Raspberry Pi ○ Microcontroller ○ Edge TPU 18

@margaretmz | #MachineLearning #GDE Optimization TFLite model optimization toolkit ● Quantization - convert 32 bit floating point to fixed point (e.g. 8-bit int) ○ Post-training quantization ○ Quantization-aware training ● Pruning - eliminating unnecessary values in the weight tensor Android: ● GPU delegate ● Android NNAPI 19

@margaretmz | #MachineLearning #GDE Image labelling OCR Face detection Barcode scanning Landmark detection Smart reply (coming soon) Object detection & Tracking Translation (56 languages) AutoML ML Kit 20 Brings Google’s ML expertise to mobile developers in a powerful and easy-to-use package. Powered by TF Lite and hosted on Firebase Base APIs: Custom models ● Dynamic model downloads ● A/B testing (via Firebase remote Configuration) ● Model compression & conversion (from TensorFlow to TF Lite)

@margaretmz | #MachineLearning #GDE Android ML with TensorFlow Your options: ● With ML Kit ○ (Out of the box) Base APIs ○ Custom model ● Direct deploy to Android ○ Custom model 21 Custom Models ● Download pre trained models ● Retrain model ● Train your own from scratch ○ data ○ train ○ convert ○ inference Note: you can use AutoML to train but no easy implementation on mobile until recently

@margaretmz | #MachineLearning #GDE End to end: model training to inference 22 Model ● tf.Keras (TensorFlow) ● Python libraries: Numpy, Matplotlib etc SavedModel or Keras model Serving ● Cloud ● Web ● Mobile ● IoT ● Micro controllers ● Edge TPU Training Inference Data

@margaretmz | #MachineLearning #GDE Data ● Existing datasets ○ Part of the deep learning framework: ■ MNIST, CIFAR10, FASHION_MNIST, IMDB movie reviews etc ○ Open datasets: ■ MNIST, MS-COCO, IMAGENet, CelebA etc ○ Kaggle datasets: https://www.kaggle.com/datasets ○ Google Dataset search tool: https://toolbox.google.com/datasetsearch ○ TF 2.0: TFDS ● Collect your own data 23

@margaretmz | #MachineLearning #GDE Models Options of getting a model: ● Download a pre-trained model (here): Inception-v3, mobilenet etc. ● Transfer learning with a pre-trained model ○ Feature extraction or fine tuning on pre-trained model ○ TensorFlow hub (https://www.tensorflow.org/hub/) ● Train your own model from scratch (example in this talk) 24

@margaretmz | #MachineLearning #GDE Model saving, conversion, deployment ● Model saving - SavedModel or Keras model ● Model conversion ○ Convert the model to tflite format ○ Validate the converted model before deploy ● Deploy TFLite for inference 25

Datasets Train model Convert to TFLite Deploy for inference End to End tf.Keras to TFLite to Android Train a model from scratch 26

@margaretmz | #MachineLearning #GDE MNIST dataset ● 60,000 train set and 10,000 test set ● 28x28x1 grayscale images ● 10 classes: 0, 1, 2, 3, 4, 5, 6, 7, 8, 9 ● Popular for computer vision ○ “hello world” tutorial or ○ benchmarking ML algorithms 27

@margaretmz | #MachineLearning #GDE Training the model in Colab Launch sample code on Colab → mnist_tfkeras_to_tflite.ipynb 1. Import data 2. Define model architecture 3. Train the model 4. Model saving & conversion ○ Save a Keras model ○ convert to tflite format 28

@margaretmz | #MachineLearning #GDE A typical CNN model architecture MNIST example: ● Convolutional layer (definition) ● Pooling layer (definition) ● Dense (fully-connected layer) definition 29 input conv pool conv pool conv pool Dense 0 1 2 3 4 5 6 7 8 9

@margaretmz | #MachineLearning #GDE Inspect the model - in python code In python code, after defining the model architecture, use model.summary() to show the model architecture 30

@margaretmz | #MachineLearning #GDE Virtualize model Use a visualization tool: ● TensorBoard ● Netron (https://github.com/lutzroeder/Netron) Drop the .tflite model into Netron and see the model visually 31

@margaretmz | #MachineLearning #GDE TensorFlow Lite Converter Convert Keras model → a tflite model with the tflite converter There are two options: 1. Command line 2. Python API Note: ● you can convert from SavedModel as well, ● GraphDef and tf.Session are no longer supported in 2.0 for TFLite conversion. Read details on tflite converter on TF documentation here 32

@margaretmz | #MachineLearning #GDE Tflite convert through command line To convert a tf.keras model to a tflite model: $ tflite_convert \ $--output_file=mymodel.tflite \ $ --keras_model_file=mymodel.h5 33

@margaretmz | #MachineLearning #GDE Tflite convert through Python code Note: converter API is different between TF 1.13, 1.14, 2.0 Alpha & nightly # Create a converter converter = tf.contrib.lite.TFLiteConverter.from_keras_model_file(keras_model) # Set quantize to true converter.post_training_quantize=True # Convert the model tflite_model = converter.convert() # Create the tflite model file tflite_model_name = "mymodel.tflite" open(tflite_model_name, "wb").write(tflite_model) 34

@margaretmz | #MachineLearning #GDE Validate the tflite model Protip: validate the tflite model in python after conversion - # Load TFLite model and allocate tensors. interpreter = tf.lite.Interpreter(model_path="converted_model.tflite") interpreter.allocate_tensors() # Get input and output tensors. input_details = interpreter.get_input_details() output_details = interpreter.get_output_details() # Test model on random input data. input_shape = input_details[0]['shape'] input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32) interpreter.set_tensor(input_details[0]['index'], input_data) interpreter.invoke() Tflite_results = interpreter.get_tensor(output_details[0]['index']) # Test the TensorFlow model on random input data. tf_results = model(tf.constant(input_data)) # Compare the result. for tf_result, tflite_result in zip(tf_results, tflite_results): 35

@margaretmz | #MachineLearning #GDE Validate the tflite model Protip: validate the tflite model in python after conversion - 36 TensorFlow result TFLite result Compare results # Test the TensorFlow model on random Input data. tf_result = model(tf.constant(input_data)) # Load TFLite model and allocate tensors. interpreter = tf.lite.Interpreter(model_path="converted_model.tflite") interpreter.allocate_tensors() # Get input and output tensors. input_details = interpreter.get_input_details() output_details = interpreter.get_output_details() # Test model on random input data. input_shape = input_details[0]['shape'] input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32) interpreter.set_tensor(input_details[0]['index'], input_data) interpreter.invoke() tflite_result = interpreter.get_tensor(output_details[0]['index']) # Compare the result. for tf_result, tflite_result in zip(tf_result, tflite_result): np.testing.assert_almost_equal(tf_result, tflite_result, decimal=5)

@margaretmz | #MachineLearning #GDE Tflite on Android Android sample code DigitRecognizer, step by step: ● Place tf.lite model under assets folder ● Update build.gradle dependencies ● Input image - custom view, gallery or camera ● Data preprocessing ● Classify with the model ● Post processing ● Display result in UI 37

@margaretmz | #MachineLearning #GDE Dependencies Update build.gradle to include tensorflow lite android { // Make sure model doesn't get compressed when app is compiled aaptOptions { noCompress "tflite" } } dependencies { …. // Add dependency for TensorFlow Lite compile 'org.tensorflow:tensorflow-lite:[version-number]’ } Place the mnist.tflite model file under /assets folder 38

@margaretmz | #MachineLearning #GDE Input - image data Input to the classifier is an image, your options: ● Draw on canvas from custom View ● Get image from Gallery or a 3rd party camera ● Live frames from Camera2 API Make sure the image dimensions (shape) matches what your classifier expects ● 28x28x1- MNIST or FASHION_MNIST gray scale image ● 299x299x3 - Inception V3 ● 256x256x3 - MobileNet 39

@margaretmz | #MachineLearning #GDE Image preprocessing ● Convert Bitmap to ByteBuffer ● Normalize pixel values to be a certain range ● Convert from color to grayscale, if needed 40

@margaretmz | #MachineLearning #GDE Run inference Load the model file located under the assets folder Use the TensorFlow Lite interpreter to run inference on the input image 41

@margaretmz | #MachineLearning #GDE Post processing The output is an array of probabilities, each correspond to a category Find the category with the highest probability and output result to UI 42

@margaretmz | #MachineLearning #GDE Summary ● Training with tf.Keras is easy ● Model conversion to TFLite is easier ● Android implementation is still challenging & error-prone: (Hopefully this gets improved in the future!) ○ Validate tflite model before deploy to Android ○ Image pre-processing ○ Input tensor shape? ○ Color or grayscale? ○ Post processing My blog post: E2E tf.Keras to TFLite to Android 43

@margaretmz | #MachineLearning #GDE TFLite demo app Check out the Demo app in TensorFlow repo Clone tensorflow project from github git clone https://www.github.com/tensorflow/tensorflow Then open the tflite Android demo from Android Studio /tensorflow/lite/java/demo 44

@margaretmz | #MachineLearning #GDE More TFLite examples More TensorFlow examples → 45

@margaretmz | #MachineLearning #GDE Inference with GPU ● Face contour detection ● Link to blog post: TensorFlow Lite Now Faster with Mobile GPUs (Developer Preview) 46

@margaretmz | #MachineLearning #GDE Posenet example ● PoseNet model on Android ● Camera live frames ● Display key body parts in real time ● Link to blog post: Track human poses in real-time on Android with TensorFlow Lite 47

@margaretmz | #MachineLearning #GDE TFLite on microcontroller ● Tiny models on tiny computers ● Consumes much less power than CPUs - days on a coin battery ● Tiny RAM and Flash available ● Opens up voice interface to devs More info here - ● Doc - https://www.tensorflow.org/lite/guide/microcontroller ● Code lab - https://g.co/codelabs/sparkfunTF ● Purchase - https://www.sparkfun.com/products/15170 48

@margaretmz | #MachineLearning #GDE Coral edge TPU (beta) - hardware for on-device ML acceleration Link to codelab: https://codelabs.developers.google.com/codelabs/edgetpu-classifier/index.html#0 ● Dev board (+ camera module) ● USB Accelerator (+ camera module + Raspberry Pi) Coral Edge TPU 49

@margaretmz | #MachineLearning #GDE Coral Edge TPU MobileNet SSD model running on TPU Inference time: < ~20 ms > ~60 fps 50

@margaretmz | #MachineLearning #GDE Coral Edge TPU demo MobileNet SSD model running on CPU Inference time > ~390ms ~ 3fps 51

@margaretmz | #MachineLearning #GDE Upcoming ● Why the future of machine learning is tiny? - Pete Warden ● Deploying to mobile and IoT will get much easier ● TFLite will have many more features ● Federated learning ● On device training 52

@margaretmz | #MachineLearning #GDE Thank you! 53 Follow me on Twitter, Medium or GitHub to learn more about Deep learning, TensorFlow and on-device ML @margaretmz @margaretmz margaretmz