[ Rustem Arzymbetov] Intro to TensorFlow

Presentation from GDG DevFest Ukraine 2016.
Learn more at: https://devfest.gdg.org.ua

Transcript

1. Intro to TensorFlow Rustem Arzymbetov rustema@google.com
2. • Open source Machine Learning library • Especially useful for Deep Learning • For research and production • Apache 2.0 license
3. Agenda: 1. Deep Learning 2. TensorFlow framework 3. Best ways to do Deep Learning with TensorFlow
4. Agenda: 1. Deep Learning 2. TensorFlow framework 3. Best ways to do Deep Learning with TensorFlow
5. Images: Wikimedia Source: https://commons.wikimedia.org/wiki/File:Red_Apple.jpg https://en.wikipedia.org/wiki/Orange_(fruit)#/media/File:Orange-Whole-%26-Split.jpg
6. spotted salamander fire salamander retriever Recognizing images with Inception https://research.googleblog.com/2016/03/train-your-own-image-classifier-with.html
7. SPAM!!!
8. [glacier] Google Photos
12. What is Deep Learning?
13. • The modern reincarnation of Artificial Neural Networks from the 1980s and 90s. • A collection of simple trainable mathematical units, which collaborate to compute a complicated function. • Compatible with supervised, unsupervised, and reinforcement learning. What is Deep Learning?
14. Loosely inspired by what (little) we know about the biological brain. Higher layers form higher levels of abstraction. What is Deep Learning?
15. Commonalities with real brains: • Each neuron is connected to a small subset of other neurons. • Based on what it sees, it decides what it wants to say. • Neurons learn to cooperate to accomplish a task. What is Deep Learning?
16. Each neuron implements a relatively simple mathematical function: Y = T( W * x + b) But a composition of 106-109 such functions is surprisingly powerful. What is Deep Learning?
17. ConvNets
18. Agenda: 1. Deep Learning 2. TensorFlow framework 3. Best ways to do Deep Learning with TensorFlow
19. A multidimensional array. A graph of operations.
20. Data flow graphs Computation is defined as a directed acyclic graph (DAG) to optimize an objective function • Graph is defined in high-level language (Python) • Graph is compiled and optimized • Graph is executed (in parts or fully) on available low level devices (CPU, GPU) • Data (tensors) flow through the graph • TensorFlow can compute gradients automatically
21. Architecture • Core in C++ Core TensorFlow Execution System CPU GPU Android iOS ...
22. Architecture • Core in C++ • Different front ends ◦ Python and C++ today, community may add more Core TensorFlow Execution System CPU GPU Android iOS ... C++ front end Python front end ...
23. Toy TensorFlow program
24. Basic TensorFlow usage example Want to predict real estate price from square footage • Collect data (square footage, price) • Come up with a model (linear dependence: y = W * x + b) • Describe the model in TensorFlow • Train the model • Predict prices
25. Linear regression y = W x + b learned X (footage) Y (price)
26. Defining the graph import tensorflow as tf W = tf.Variable(tf.ones([1, 2])) b = tf.Variable(tf.ones([1])) x = tf.placeholder(tf.float32, shape=(2,1)) y = tf.add(tf.matmul(W, x), b) * + X W b y
27. Defining the graph: variables import tensorflow as tf W = tf.Variable(tf.ones([1, 2])) b = tf.Variable(tf.ones([1])) x = tf.placeholder(tf.float32, shape=(2,1)) y = tf.add(tf.matmul(W, x), b) * + X W b y
28. Defining the graph: placeholders import tensorflow as tf W = tf.Variable(tf.ones([1, 2])) b = tf.Variable(tf.ones([1])) x = tf.placeholder(tf.float32, shape=(2,1)) y = tf.add(tf.matmul(W, x), b) * + X W b y
29. Defining the graph: formula import tensorflow as tf W = tf.Variable(tf.ones([1, 2])) b = tf.Variable(tf.ones([1])) x = tf.placeholder(tf.float32, shape=(2,1)) y = tf.add(tf.matmul(W, x), b) * + X W b y
30. Executing the graph output = [[1 , 1]] * [[3] ,[4]] + 1 output = 8 with tf.Session() as sess: tf.initialize_all_variables().run() output = sess.run(y,feed_dict={x:[[3],[4]]}) x = tf.placeholder(tf.float32, shape=(2,1)) W x b
31. Given (x, y) pairs learn W, b to minimize the distance from predicted y to known y. Updating variables: learning W and b W x + b X (footage) Y (price)
32. Loss functions distance = y - labels sq_distance = distance * distance cost = tf.reduce_sum(sq_distance) X (footage) Y (price)
33. Minimize the loss: build the graph distance = y - labels sq_distance = distance * distance cost = tf.reduce_sum(sq_distance) # Change W, b to minimize cost update_step = tf.GradientDescentOptimizer(learning_rate).minimize(cost)
34. Minimize the loss: run the ops # Define x, W, b, y, cost as before update_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost) with tf.Session() as sess: tf.initialize_all_variables().run() for i in xrange(100): sess.run(update_step, feed_dict={x: ..., labels: ...}) Your data
35. TensorBoard: visualizing learning
36. Setting up TensorBoard In your python code: summary_writer = tf.train.SummaryWriter('/path/to/logs', sess.graph) # Tell the summary_writer what you want to log summary_writer.add_summary(measure, time_step) In the command line: % tensorboard --logdir=/path/to/logs
37. Agenda: 1. Deep Learning 2. TensorFlow framework 3. Best ways to do Deep Learning with TensorFlow
38. playground.tensorflow.org
39. TensorFlow Playground
40. Great docs & tutorials
41. http://tensorflow.org/whitepaper2015.pdf
42. (1) Use a Cloud-based API (2) Run your own pretrained model (3) Use an existing model architecture ◦ retrain or fine tune on your dataset (4) Develop your own machine learning models ◦ solve new problems More flexible, but more effort required Your options
43. cloud.google.com/translate cloud.google.com/speech cloud.google.com/vision cloud.google.com/natural-language (1) Use Cloud-based APIs
44. cloud.google.com/translate cloud.google.com/speech cloud.google.com/vision (1) Use Cloud-based APIs cloud.google.com/natural-language
45. Google Cloud Vision API
46. (2) Using a Pretrained Model yourself with TensorFlow www.tensorflow.org/tutorials/image_recognition/index.html 1. Load graph of trained model 2. Start doing inference!
47. (2) Use pretrained models with tf.gfile.FastGFile(model_path, 'rb') as f: graph_def = tf.GraphDef() graph_def.ParseFromString(f.read()) _ = tf.import_graph_def(graph_def, name='') with tf.Session() as sess: softmax_tensor = sess.graph.get_tensor_by_name('softmax:0') predictions = sess.run(softmax_tensor, {'DecodeJpeg/contents:0': image_data}) load model inference
48. (3) Training your own models cloud.google.com/ml OR Run open-source release on • your own physical machines • virtual machines in a cloud hosting environment
49. (4) Develop your own learning models https://www.tensorflow.org/versions/master/get_started/basic_usage.html • Define a computational graph for your model • Start training!
50. Thanks! Q&A time http://tensorflow.org http://github.com/tensorflow/tensorflow