Swift for TensorFlow - The Next Generation platform for Deep Learning.

Aprenderás a entrenar un modelo de Deep Learning utilizando Swift para TensorFlow, la nueva plataforma que Google está construyendo para Deep Learning; sus ventajas y cómo es que te permite crear algoritmos de una forma más sencilla y rápida.

Transcript

1. The Next Generation platform for Deep Learning. Swift for TensorFlow
2. Hello! I am Isabel Palomar 2
3. Table of contents • Getting started with AI ◦ Challenges • Why Swift? • Training a model with Swift • … Next Steps 3
4. 1. Getting started with ai How to get started with Artificial Intelligence
5. Curiosity: I want to create a custom model HOW? 5
6. “ 6 6 @jeremyphoward @math_rachel I started a Deep Learning Course last year at the University of San Francisco. MY STORY: TAKING MY FIRST DEEP LEARNING CLASS
7. 7 After the class….. The key outcome of this lesson is that we'll have trained an image classifier which can recognize pet breeds at state of the art accuracy. The key to this success is the use of transfer learning, which will be a key platform for much of this course. We also discuss how to set the most important hyper-parameter when training neural networks: the learning rate, using Leslie Smith's fantastic learning rate finder method. Finally, we'll look at the important but rarely discussed topic of labeling, and learn about some of the features that fastai provides for allowing you to easily add labels to your images. https://course.fast.ai/videos/?lesson=1
8. challenges…. ‐ Many courses, even basic, assume that you already know the subject. ‐ Reaching the final result without learning the basics is not good. 8
9. 2. Why swift ????????? Let’s understand the advantages of Swift
10. “ 10 10 FAST AI EMBRACING SWIFT FOR DEEP LEARNING
11. FAST AI AND SWIFT FOR TENSORFLOW 11 @clattner_llvm @jeremyphoward @math_rachel
12. SOMETHING IS HAPPENING….. 12 The AI industry is constantly changing and evolving
13. 13 Tools and infrastructure • Easy to learn and build • Fast • Quick implementations • Accuracy ARTIFICIAL INTELLIGENCE GOALS
14. PROGRAMMING LANGUAGES AND THE AI WORLD... Slow at runtime, poor support for parallel processing, but very easy to use. 14 Hard to use (and C++ is slow at compile time), but fast and (for C++) expressive. Unsafe (unless you use Typescript); somewhat slow (but easy to use and flexible) Poor support for general purpose programming, but fast and expressive. Verbose (but getting better, particularly if you use Kotlin), less flexible (due to JVM issues). Requires installation of a runtime, limited flexibility due to garbage collection, slow https://www.fast.ai/2019/01/10/swift-numerics/
15. 15 OHH… HELLO SWIFT! • Expressive, flexible, concise, safe, easy to learn and use, fast. Swift is developed to be open ◦ Community participation
16. 16 GETTING THE BEST FROM THE BEST Interoperability Python C, C++
17. 17 Swift for tensorflow First-class autodiff Differentiable programming gets first-class support in a general-purpose programming language. Next-generation APIs New APIs informed by the best practices of today, and the research directions of tomorrow, are both easier to use and more powerful. Builds on TensorFlow Building on TensorFlow, the Swift APIs give you transparent access to all low-level TensorFlow operators.
18. 18 High-quality tooling - COLAB Colaboratory is a free Jupyter notebook environment that requires no setup and runs entirely in the cloud. https://colab.research.google.com/
19. CREATING A MODEL WITH SWIFT
20. Let’s take a look of the implementation... 20 Classify handwritten digits MNIST Dataset
21. DEEP LEARNING CONCEPTS Data Task Model 21 Learning Algorithm Loss Function Evaluation
22. 1.- dATA Data is distinct pieces of information which acts as a fuel 22
23. Formats Data can come in different formats depending on what you want to solve 23
24. DATA FOR OUR EXAMPLE 24
25. How? Where do we get data from? Data curation is the organization and integration of data collected from various sources. 25 Techniques You can use techniques like Questionnaires and surveys, conducting interviews, using data scraping and data crawling techniques.
26. Public datasets • Google AI • UCI ML Repository • Data.gov.in • Kaggle Where do we get data from? Crowdsourcing Marketplaces • Amazon Mechanical Turk • Dataturks • Figure-eight 26
27. BACK TO OUR EXAMPLE... The MNIST database is a large database of handwritten digits that is commonly used for training and testing in the field of machine learning. https://en.wikipedia.org/wiki/MNIST_database 27
28. Load and explore the data set. STEP 1
29. LOAD THE DATASET 29 %include "/content/swift-datascience/MNIST.swift" // Load dataset let dataset = MNIST(batchSize: 128) // Get first 5 images let imgs = dataset.trainingImages.minibatch(at: 0, batchSize:5).makeNumpyArray()
30. EXPLORING THE MNIST DIGITS 30 // Display first 5 images for img in imgs{ plt.imshow(img.reshape(28,28)) plt.show() }
31. 2.- task The problem to resolve 31
32. Examples of tasks 32
33. Define the task STEP 2
34. TASK FOR OUR EXAMPLE 34 Classify handwritten digits
35. 3.- model The mathematical formulation of a task. 35
36. models 36 There are many models that are created over the years. Each model has its own advantages and disadvantages based on the type of data on which we are creating a model.
37. IMAGE CLASSIFICATION MODEL An image classification model is trained to recognize various classes of images. 37 When we subsequently provide a new image as input to the model, it will output the probabilities of the image representing each of the types it was trained on.
38. An example output might be as follows: Digit type Probability 0 0.00 1 0.02 ... 0.00 7 0.97 8 0.00 9 0.01 38 Based on the output, we can see that the classification model has predicted that the image has a high probability of representing a number 7
39. Defining the Structure of our Model STEP 3
40. In this example, we will create the LeNet-5 model using the Swift for TensorFlow Deep Learning Library Model for our example 40
41. Create a model using the Swift for TensorFlow API 41 import TensorFlow let epochCount = 100 let batchSize = 128 // The LeNet-5 model var classifier = Sequential { Conv2D<Float>(filterShape: (5, 5, 1, 6), padding: .same, activation: relu) AvgPool2D<Float>(poolSize: (2, 2), strides: (2, 2)) Conv2D<Float>(filterShape: (5, 5, 6, 16), activation: relu) AvgPool2D<Float>(poolSize: (2, 2), strides: (2, 2)) Flatten<Float>() Dense<Float>(inputSize: 400, outputSize: 120, activation: relu) Dense<Float>(inputSize: 120, outputSize: 84, activation: relu) Dense<Float>(inputSize: 84, outputSize: 10, activation: softmax) }
42. 4.- learning algorithm The Learning Algorithms also known as Optimization algorithms helps us to minimize Error 42
43. Is something you do everyday... You are optimizing variables and basing your personal decisions all day long, most of the time without even recognizing the process consciously https://mitsloan.mit.edu/ideas-made-to-matter/how-to-use -algorithms-to-solve-everyday-problems 43
44. First Order Optimization Algorithms  • Gradient Descent Types of learning algorithms Second Order Optimization Algorithms • Hessian https://towardsdatascience.com/types-of-optimization-algorithms-used-in-neural-networks-and- ways-to-optimize-gradient-95ae5d39529f 44
45. Create an optimizer STEP 4
46. Choosing Gradient Descent as the Optimizer Swift for TensorFlow has many optimization algorithms available for training. This model uses the SGD optimizer that implements the stochastic gradient descent (SGD) algorithm. 46 // Using Gradient Descent as the optimizer let optimizer = SGD(for: classifier, learningRate: 0.1)
47. OTHER OPTIMIZERS AVAILABLE ON THE SWIFT FOR TENSORFLOW API Swift4Tensorflow supports many additional optimizers. You can choose your pick based on your project: 47 ‐ AMSGrad ‐ AdaDelta ‐ AdaGrad ‐ AdaMax ‐ Adam ‐ Parameter ‐ RMSProp ‐ SGD
48. 5.- loss function How do we know which model is better? Loss function (also known as the error) answers this question. 48
49. Classification losses: • Mean Square Error/L2 Loss • Mean Absolute Error/L1 Loss Regression losses: • Hinge Loss/Multi-class SVM Loss • Cross Entropy • Loss/Negative Log Likelihood LOSS FUNCTIONS To know which model is good for our data, we compute the loss function by comparing the predicted outputs to actual output. 49
50. Choose a loss function STEP 5
51. Choose a loss function Our model will calculate its loss using the softmaxCrossEntropy(logits:labels:) function which takes the model's class probability predictions and the desired label, and returns the average loss across the examples. 51 let loss = softmaxCrossEntropy(logits: ŷ, labels: y)
52. 6.- evaluation To compute a score for our ML model, we need to evaluate it. 52
53. evaluation There are many performance metrics to achieve this. One simple such metric is Accuracy. 53
54. Evaluation/performance metrics 54 • Confusion matrix • Area Under the ROC curve (AUC — ROC) • Root Mean Squared Error (RMSE) (used in regression)
55. 5.- Model Training and Evaluation STEP 6
56. Model training 56 print("Beginning training...") struct Statistics { var correctGuessCount: Int = 0 var totalGuessCount: Int = 0 var totalLoss: Float = 0 } // The training loop - DATA for epoch in 1...epochCount { var trainStats = Statistics() } let trainAccuracy = Float(trainStats.correctGuessCount) / Float(trainStats.totalGuessCount) let testAccuracy = Float(testStats.correctGuessCount) / Float(testStats.totalGuessCount) print(""" [Epoch \(epoch)] \ Training Loss: \(trainStats.totalLoss), \ Training Accuracy: \(trainStats.correctGuessCount)/\(trainStats.totalGuessCount) \ (\(trainAccuracy)), \ Test Loss: \(testStats.totalLoss), \ Test Accuracy: \(testStats.correctGuessCount)/\(testStats.totalGuessCount) \ (\(testAccuracy)) """) }
57. TRAINING THE MODEL TO CLASSIFY DIGITS - STATS AND EVALUATION 57
58. DEEP LEARNING CONCEPTS Data Task Model 58 Learning Algorithm Loss Function Evaluation
59. LET’S EXPLORE COLABORATORY Colaboratory is a free Jupyter notebook environment that requires no setup and runs entirely in the cloud. https://colab.research.google.com
60. 4 Next steps Getting involved in the Swift for Tensorflow world
61. We did it!!! Now you are ready to start building your first custom Deep Learning model. 61
62. Using the model - Tensorflow platform 62 TensorFlow is an end-to-end open source platform for machine learning. • Web Apps • Back end - APIs https://www.tensorflow.org/
63. Tensorflow lite options - Mobile & iot TensorFlow Lite provides all the tools you need to convert and run TensorFlow models on mobile, embedded, and IoT devices. 63
64. SWIFT FOR TENSORFLOW It is an exciting time 64 TOOLS PLATFORMS APIS FRAMEWORKS
65. “ Believe me…. You don’t need a PHD degree! 65
66. There are so many resources where you can start learning…. Fast AI • Practical Deep Learning for Coders: ◦ https://course.fast.ai/ ◦ https://course.fast.ai/part2. Swift and tensorflow - Colaboration with Chris Lattner • Embracing Swift for Deep Learning ◦ https://www.fast.ai/2019/03/06/fastai-swift/ 66 IF YOU DON’T FEEL READY….
67. Udacity • Intro to Artificial Intelligence: ◦ https://www.udacity.com/course/intro-to-artificial-intelli gence--cs271 • Intro to Machine Learning: ◦ https://www.udacity.com/course/intro-to-machine-learn ing--ud120 • TensorFlow for Deep Learning ◦ https://www.udacity.com/course/intro-to-tensorflow-for -deep-learning--ud187 ◦ 67 Learning about ai, ml and deep learning
68. • Official documentation and resources ◦ https://www.tensorflow.org/swift • Model training walkthrough ◦ https://www.tensorflow.org/swift/tutorials/model_traini ng_walkthrough • Guide to learn swift from scratch for Data Science ◦ https://www.analyticsvidhya.com/blog/2019/10/compre hensive-guide-learn-swift-from-scratch-data-science/ • Swift for Machine learning (podcast episode) ◦ https://www.swiftbysundell.com/podcast/58/ ◦ 68 Swift for tensorflow
69. 69 YOU CAN BUILD GREAT THINGS! https://ai.google/social-good/ There a lot of problems to solve! 69
70. 70 SOME EXTRA MOTIVATION http://blog.indeed.com/2019/06/28/top-10-ai-jobs-salaries-cities/
71. 71 THANKS! You can find me at @isabelpalomar