AI for Art & Design with TensorFlow Lite Margaret Maynard-Reid, Nov 17 2020 #dcAmericas @margaretmz

#dcAmericas | @margaretmz | #TFLite Topics ● Computer vision tasks ● Generative art & design ● Datasets & resources AI for Art & Design TensorFlow Lite Android Samples & tutorials ● Sample apps ● Awesome TFLite ● E2E TFLite Tutorials ● What is TFLite? ● TFLite on Android ● Tools & resources 2

#dcAmericas | @margaretmz | #TFLite Art & Design ● Photography ● Sculpture ● Drawing ● Painting ● Fashion design ● Product design ● UI/UX design 3

Computer Vision Tasks Complexity Image classification (single-label) Classify an image to a class Examples Painting style or artist Van Gogh Image classification (multi-label) Classify an image to multiple classes Examples Movie poster genre action, sci-fi Feature extraction Extracting latent features of an image with CNN models Examples Visual search find similar fashion Object detection Identify one or multiple objects within an image and their locations with bounding boxes. detect UI elements Segmentation Classify whether each pixel of the image belongs to a certain class segment UI elements Generative models (GANs) Two or more models trained simultaneously Examples - Generate new images - Super res - Image-to-image 4

Generative Art & Design AI for Art & Design with TFLite 5

#dcAmericas | @margaretmz | #TFLite Deep Dream 2015 6

Use CNN neural networks Content image + style image = styled image Content Style Generated Neural Style Transfer: 50 Shades of Miaw (link) A few more samples of generated images… Style Transfer 2015 7

#dcAmericas | @margaretmz | #TFLite Generative Adversarial Networks 2014 What are GANs? Generative Adversarial Networks (GANs) have at least two network models which compete against each other ... 8

#dcAmericas | @margaretmz | #TFLite GANs variations GANs are fun but difficult to train! ● 2014 Ian Goodfellow’s original GAN’s paper ● 2015 DCGAN (Deep Convolutional GAN) (paper) ● 2014 cGAN (Conditional GAN) - Generate images conditioned on a class (paper) ● 2016 Pix2Pix – paired image to image translation (paper) ● 2017 CycleGAN – unpaired Image to image translation ● 2018 StyleGAN – Generate images with some control on image characteristics (paper) ● Many other GAN variations… 9

#dcAmericas | @margaretmz | #TFLite ● Training datasets must be paired ● A type of conditional GAN Link to TensorFlow tutorial Pix2Pix 2016 10

#dcAmericas | @margaretmz | #TFLite CycleGAN 2017 ● Training datasets are unpaired ● Horses <--> zebras ● Summer <--> winter ● Photos to paintings (Monet, Van Gogh etc.) Link to TensorFlow tutorial 11

#dcAmericas | @margaretmz | #TFLite BigGAN 2018 Large scale GAN that generates high fidelity class-conditional images. Link to paper | TF Hub biggan512, biggan-256, biggan-128 12

#dcAmericas | @margaretmz | #TFLite Control the styles of the generated images ● Pose ● Face shape ● Age ● Gender ● Hair color 4x4 --> 1024x1024 8GPUs trained for 1 week Current examples: faces, bedrooms and cats Link to video StyleGAN 2018 13

#dcAmericas | @margaretmz | #TFLite Sketch2Code Microsoft: Turn your whiteboard sketches to working code in seconds with Sketch2Code 14

#dcAmericas | @margaretmz | #TFLite Fashion++ Facebook AI: “Fashion++: Minimal Edits for Outfit Improvement” (news | paper) 15

Datasets Art ● The Art Institute of Chicago - 5000 high res images ● Best Artworks of All Time ● Art Images: Drawing/Painting/Sculptures/Engravings ● Overwatch Heroes Recognition ● Chinese Fine Art ● Museum of Modern Art Collection UI/UX Design ● Rico mobile app datasets ● Kaggle ○ Common Mobile/Web App Icons ○ Icons-50 16

#dcAmericas | @margaretmz | #TFLite Magenta https://magenta.tensorflow.org/ | https://github.com/magenta 17

https://ami.withgoogle.com/ 18

#dcAmericas | @margaretmz | #TFLite AI & the Future of UX https://uxdesign.cc/ai/home 19

#dcAmericas | @margaretmz | #TFLite AI for Art & Design AI will assist and inspire artists and designers Everyone can now make art and design! 20

TensorFlow Lite AI for Art & Design with TFLite 21

#dcAmericas | @margaretmz | #TFLite What is an ML model? • A complex mathematical function • A representation of learnings from training data 22

#dcAmericas | @margaretmz | #TFLite Model training, conversion & inference 23 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

#dcAmericas | @margaretmz | #TFLite Why on device? It’s all about that camera! 24

#dcAmericas | @margaretmz | #TFLite What is TensorFlow Lite? A framework with a set of tools for deploying ML models to mobile & embedded: ● A converter ● An interpreter ● Ops / kernels ● Interface to hardware accelerations: ○ NN AIP ○ Edge TPu ● A bunch of other tools... 25

#dcAmericas | @margaretmz | #TFLite tf.Keras to TFLite to Android 26

#dcAmericas | @margaretmz | #TFLite Android app with TensorFlow Lite ● ML model: ○ Train / find a model ○ Convert to .tflite format ○ Run inference in Python ○ Benchmark (optional) ○ Add model metadata ● Build Android app ○ UI navigation ○ Input image: canvas, gallery or camera ○ Load model & run inference ○ Display results in UI 27

#dcAmericas | @margaretmz | #TFLite TensorFlow Lite tools 1. Train a model ● Teacheable machine ● Model Maker ( w/ transfer learning) ● ML Kit AutoML Vision ● TensorFlow Hub 28 2. Convert ● Converter ● Optimization toolkit ● Benchmarking tool 3. Deploy ● ML Model Binding ● Task API

https://pair.withgoogle.com/ 29

Android Apps and Tutorials AI for Art & Design with TFLite 30

#dcAmericas | @margaretmz | #TFLite ML Kit Digital Ink Recognition Recognize handwritten text and drawings ● Text models: 300+ languages ● Non-text models ○ Autodraw ○ Emoji ○ Shapes Blog: Digital Ink Recognition in ML Kit Sample app: Overview | GitHub 31 Example: Draw & recognize emojis

Part 2: ML Model Binding from Android Studio TensorFlow Lite Model Maker Part 1: TensorFlow Lite Model Maker ● Image/Text classification 32

#dcAmericas | @margaretmz | #TFLite Style transfer Google Arts & Culture app: https://play.google.com/store/apps/details?id=com.google.android.apps.cultural Blog post: 4/2/2020 - Transform your photo in the style of an iconic artist 33

Selfie2Anime ● U-GAT-IT paper ● Image to image translation ● Written in TF1 code ● Unpaired selfie and anime images ● 50 & 100 epochs checkpoints ● Collaboration with Sayak & Khanh ● Links: project repo | blog 34

#dcAmericas | @margaretmz | #TFLite Cartoonizer ● Model: Whitebox CartoonGAN ○ Faster & smaller ○ Not just for selfies ● Link to project repo ● Link to TF Hub models ● Collaboration with Sayak & Khanh ● Links: project repo | blog 35

#dcAmericas | @margaretmz | #TFLite Segmentation + Style Transfer Demo by Khanh LeViet TFLite inference in Python, Already implemented in Android (and iOS)... (project repo) 36

#dcAmericas | @margaretmz | #TFLite 2020 E2E TFLite Tutorials GitHub repo 37

Awesome TensorFlow Lite A curated list of models, apps, tutorials and learning resources… Thanks to the ML GDE and Android community! Support of TensorFlow Lite team https://github.com/margaretmz/awesome-tflite 38

#dcAmericas | @margaretmz | #TFLite Follow me on Twitter, Medium or GitHub to learn more about deep learning, TensorFlow and on-device ML. @margaretmz @margaretmz margaretmz Thank you! 39