Your first TensorFlow programming with Jupyter

Google confidential | Do not distribute

Your first TensorFlow programming with Jupyter

Etsuji NakaiCloud Solutions Architect at Google2016/08/29 ver1.1

Etsuji Nakai

Cloud Solutions Architect at Google

The author of “Introduction to Machine Learning

Theory” (Japanese Book)

New book “ML programming with TensorFlow”

will be published soon!

$ who am i

Google's open source library for

machine intelligence

tensorflow.org launched in Nov 2015

Used by many production ML projects

What is TensorFlow?

Web based interactive data analysis

platform.

Can be used as a TensorFlow

runtime environment.

What is Jupyter?

How to use Jupyter on GCP? (Japanese Blog)http://enakai00.hatenablog.com/entry/2016/07/03/201117

● All calculations are done in a “Session”

● The session contains:○ Placeholders : where you put actual data

○ Variables : to be optimized by the algorithm

○ Functions : consisting of placeholders and variables

○ Training algorithm : to optimize the variables

Programming Paradigm of TensorFlow

Programming Paradigm of TensorFlow

● Three steps to write a program with TnesorFlow○ Define a model with placeholders, variables, functions.

○ Define a loss function and a training algorithm.

○ Run session to optimize the variables minimizing the loss

function.

Example: Least Squares Method

● Figure out a smooth curve which predicts

next year’s temperature.

● In matrix representation:Monthly average temperature in Tokyo.

VariablePlaceholder

Function

● Define a loss function

● In matrix representation:

Example: Least Squares Method

Placeholder

Function

Observed temperature

Prediction vs Observed Values

● The matrix representations can be directly

translated into TensorFlow codes.

Example: Least Squares Method

x = tf.placeholder(tf.float32, [None, 5])w = tf.Variable(tf.zeros([5, 1]))y = tf.matmul(x, w)t = tf.placeholder(tf.float32, [None, 1])loss = tf.reduce_sum(tf.square(y-t))

● Specify an optimization algorithm.

● Finally, prepare a session and run the optimization loop.

Example: Least Squares Method

sess = tf.Session()sess.run(tf.initialize_all_variables())i = 0for _ in range(100000): i += 1 sess.run(train_step, feed_dict={x:train_x, t:train_t}) if i % 10000 == 0: loss_val = sess.run(loss, feed_dict={x:train_x, t:train_t}) print ('Step: %d, Loss: %f' % (i, loss_val))

train_step = tf.train.AdamOptimizer().minimize(loss)

Putting actual data values in placeholders

● You can see the actual result at:○ http://goo.gl/Dojgp4

Example: Least Squares Method

Demo: More Interesting Examples!

Thank you!