Building ML Models with TensorFlow on GCP

For attendees of Cloud study Jam of GDG Ebolowa, in this presentation, we walk through the process of building a complete machine learning pipeline covering ingest, exploration, training, evaluation, deployment, and prediction.

Transcript

1. Building ML Models with TensorFlow on GCP @OBAMSerge AI/ML Evangelist
2. Yannick Serge Obam • Co- founder IsoMora • Teacher • AI/ML Evangelist • Organiser GDG Ebolowa
3. 1- Machine Learning with Tensorflow …………………………………………………………………………. 2- Machine Learning on Google Cloud ………………………………………………………………………... 3- What’s next ………………………………………………………………………. AGENDA
4. MACHINE LEARNING WITH TENSORFLOW
5. Machine Learning
7. AI Make machines solve problems commonly associated with human intelligence ML Make machines learn from experience instead of explicit programming DL Machine Learning using deep neural networks
8. Activity Detection if(speed<4){ status=WALKING; } if(speed<4){ status=WALKING; } else { status=RUNNING; } if(speed<4){ status=WALKING; } else if(speed<12){ status=RUNNING; } else { status=BIKING; } // Oh crap
9. Rules Data Traditional Programming Answers Answers Data Rules Machine Learning
10. Activity Detection 0101001010100101010 1001010101001011101 0100101010010101001 0101001010100101010 Label = WALKING 1010100101001010101 0101010010010010001 0010011111010101111 1010100100111101011 Label = RUNNING 1001010011111010101 1101010111010101110 1010101111010101011 1111110001111010101 Label = BIKING 1111111111010011101 0011111010111110101 0101110101010101110 1010101010100111110 Label = GOLFING (Sort of)
11. Training Phase Answers Data Rules Machine Learning Model Data Prediction s Inference Phase
12. There’s a revolution in machine learning
14. is making all of this possible
15. TensorFlow is an end-to-end open source platform for machine learning Tensorflow 2.0
17. What we see What the computer “sees”
19. Experiment with Transfer Learning Model A Model B
20. Implement one of the popular model architectures
21. Experiment with wide-n-deep Wide Models Deep Models Wide & Deep Models Sigmoid Rectified Linear Units Output Hidden Layers Dense Embeddings Sparse Features
22. 2.b Compute loss Weights Training data 1. Read, decompress, parse, perform data augmentation, batch 3. Update weights (2.a) Compute forward pass (2.c) Compute backward pass The ML training loop
23. Loss function parameter #1 loss parameter #2
24. Experiment with Keras Model import tensorflow as tf mnist = tf.keras.datasets.mnist (x_train, y_train),(x_test, y_test) = mnist.load_data() x_train, x_test = x_train / 255.0, x_test / 255.0 model = tf.keras.model.Sequential([ tf.keras.layers.Flatten(input_shape=(28, 28)), tf.keras.layers.Dense(512, activation=tf.nn.relu), tf.keras.layers.Dropout(0.2), tf.keras.layers.Dense(10,activation= tf.nn.softmax), ]) model.compile(loss='categorical_crossentropy', optimizer=‘adam', metrics=['accuracy']) model.fit(x_train, y_train, epochs=5) model.evaluate(x_test, y_test)
25. Hyperparameter Optimization Tuner Trainer learning_rate: 0.01, dropout_rate: 0.9 Train Validation Accuracy: 0.5 Existing systems/services: Spearmint, HyperOpt, Google HyperTune (May learn from feedback)
26. Raspberry Pi Android iOS GPU CPU Cloud TPU TensorFlow supports many platforms
27. TensorFlow runs in many languages... Go
28. MACHINE LEARNING ON GOOGLE CLOUD
29. Many Machine Learning Frameworks can handle toy problems!
30. To build effective ML, you need
31. As your data size increases, batching and distribution become important
32. To build effective ML, you need
33. Input necessary transformations
34. To build effective ML, you need :
36. Hyperparameter tuning might be nice
37. Need to autoscale predictio code
38. Who does the preprocessing
39. Cloud Machine Learning : Repeatable, scalable, tuned
40. In Datalab, start locally on sampled dataset
41. Then , scale it out to GCP using serverless technology
42. Data Collection and Storage - Data generation at scale - Storage based on datatypes Storage BigQuery Pub-Sub Functions
43. Train and serve ML Engine TPU GPU - Versioned APIs - Hypertuning
44. Supported features
45. Feature Reliability
51. KubeFlow - The complete stack! - Jupyter Notebooks - Training and Serving - Packaging TFX - tf.Transform - tf.DataValidation - TF Serving - TF Model Analysis No-Cloud? Multi-Cloud?
52. WHAT’S NEXT
53. 180 teraflops per device / 64 GB high-bandwidth memory Programmed via TensorFlow Now available in beta at cloud.google.com/tpu Cloud TPU
54. … at scale ResNet-50 on Cloud TPU v2 Pod Real data: 219,000+ images/sec Final accuracy: 93% excluding startup overhead Training time: (90 epochs) 8m 45s
55. Getting involved Openness Transparency Participation tensorflow.org/community
56. Ressources youtube.com/tensorflow twitter.com/tensorflow medium.com/tensorflow End-to-end Machine Learning with Tensorflow on GCP @Vikram_Tiwari Slide https://www.coursera.org/learn/end-to-end-ml-tensorflow-gcp
58. Thanks! Ask a question.