Tensorflow.js: FullStack Machine Learning

TensorFlow ha conseguido acercar **Machine Learning** a profesionales y entusiastas de la tecnología en todo el mundo sin necesidad de ser un experto en Data Science.

Sin embargo, hasta hace poco, estábamos limitados a usarlo en el servidor y programar en Python, Java o C++. Pero ahora eso ha cambiado. Gracias a TensorFlow.js podemos emplear esta librería para entrenar o utilizar nuestros modelos tanto en cliente como en servidor y siempre en Javascript.

En esta charla realizaremos una introducción al mundo de Machine Learning y de TensorFlow y enseñaremos ejemplos de cómo podemos utilizarlo en nuestras aplicaciones para que al terminar estéis deseando llegar a casa para probarlo.

Transcript

1. Tensorflow.js Fullstack Machine learning
2. 2 I am Luis Talegón • Data Scientist at seedtag • Passionate about breakthrough ML technologies • Twitter: @luisetex
3. 3 I am Laura Morillo-Velarde • Tech Lead at seedtag • GDG Madrid Organizer / WTM Lead / Google Cloud GDE • Backend engineer • 11 years working with different technologies • Twitter: @Laura_Morillo
4. 1. Intro Knowing Tensorflow.js 4
5. What’s tensorflow? 5 TensorFlow is an open source software library for numerical computation using data-flow graphs. Use flexible and intuitive APIs to build and train models from scratch using the low-level library or the high-level API.
6. tensorflow.js 6
7. tensorflow.js 7 Layers API Ops API
8. Client side 8 × No drivers needed × Interactive × Use sensor data × Everything client side
9. Client side 9
10. node.js 10
11. 2. Core concepts Knowing the base concepts of Tensorflow.js 11
12. 12 What’s a tensor?
13. 5 Scalar 13 What’s a tensor?
14. 5 Vector Scalar 14 What’s a tensor? 5, 6, 0
15. 1 2 6 9 5 Matrix Vector Scalar 15 What’s a tensor? 5, 6, 0
16. 1 2 6 9 5 Matrix Vector Scalar 16 What’s a tensor? 5, 6, 0 Tensor 1, 2 5, 6 3, 4 9, 0
17. What’s A Variable? 17 Tensor 1, 2 5, 6 3, 4 9, 0
18. What’s A Variable? 18 Value2 3, 3 4, 4 1, 8 0, 0 Value1 1, 2 5, 6 3, 4 9, 0
19. WhAT’s AN operation? 19
20. WhAT’s AN operation? 20 Operation 1 Operation 2 Tensor 1 Tensor 2 Tensor 3
21. What’s a model? 21 f(x) X Y
22. What’s a model? 22 X Y Operation or Layer Operation or Layer
23. What’s AN Activation function? 23
24. What’s a loss function? 24
25. What’s an optimizer? 25
26. What are the epochs? 26
27. 3. server side example Let’s see how it works with Node.js 27
28. Super Hexagon 28
29. Hexagon.js 29 Find it at https://github.com/Schlipak/Hexagon.js
30. Super shit- 30
31. frustration 31
32. What to do 32 ?
33. solution 33 Just let your computer deal with it, lol
34. How do we learn? 34 By observation and recognizing patterns
35. how? 35
36. Steps 36 Let’s get started!
37. Steps 37 1. Get data
38. Steps 38 1. Get data 2. Preprocess data + augmentation
39. Steps 39 1. Get data 2. Preprocess data + augmentation 3. Build Deep Learning model
40. Steps 40 1. Get data 2. Preprocess data + augmentation 3. Build Deep Learning model 4. Train our model
41. Steps 41 1. Get data 2. Preprocess data + augmentation 3. Build Deep Learning model 4. Train our model 5. ???
42. Steps 42 1. Get data 2. Preprocess data + augmentation 3. Build Deep Learning model 4. Train our model 5. ??? 6. Profit!!
43. Data collection 43 Play the game (you might even get better!) and capture your screen as fast as possible
44. Image preprocessing 44
45. Image preprocessing 45
46. Image preprocessing 46
47. Network 47 Convolutional Layer Input Size: 200x200x1 32 filters 3x3 kernel Relu activated MaxPooling: 2x2 Flatten
48. Network 48 Fully connected: 128 Relu activated Dropout: 0.25 Fully connected: 128 Relu activated Dropout: 0.25 Fully connected: 3 Softmax activated
49. code 49
50. Training 50
51. All set! 51
52. Universal model 52
53. Give it the control 53 1. Capture 2. Preprocess 3. Throw it to the model!
54. Demo time! 54
55. 4. Client side example Let’s see how it works in the browser 55
56. Super Hexagon JEDI 56
57. Challenge accepted! 57
58. How? 58 Model ???
59. 59 Network MobileNet Fully connected: N Relu activated Fully connected: 3 Softmax activated Flatten
60. Demo time! 60
61. × https://js.tensorflow.org/ => Documentación oficial × https://github.com/Luisetex/HexAI => Demo server side × https://github.com/lauramorillo/tensorflow-hexagon-cam => Demo client side Links 61
62. Presentation template by SlidesCarnival Any questions? You can find us at: @luisetex & luis@seedtag.com @laura_morillo & laura.morillovelarde@gmail.com THANKS! 62