sskpeng / RoboND-DeepLearning-Note

TensorFlow for Deep Learning

This document for note in the course material.

Tensor

tf.constant() returns a constant tensor

Session

A "TensorFlow Session" is an environment for running a graph.

with tf.Session() as sess:
    output = sess.run(hello_constant)
    print(output)

Input

tf.placeholder() returns a tensor that gets its value from data passed to the tf.session.run() function, allowing you to set the input right before the session runs.

x = tf.placeholder(tf.string)

with tf.Session() as sess:
    output = sess.run(x, feed_dict={x: 'Hello World'})

Math

Addition, subtraction and Multiplication

tf.add(a, b) returns a + b

tf.subtract(a, b) returns a - b

tf.multiply(a, b) returns a * b

tf.divide(a, b) returns a / b

Converting types

tf.subtract(tf.cast(tf.constant(2.0), tf.int32), tf.constant(1))   # 1

TensorFlow Softmax

x = tf.nn.softmax([2.0, 1.0, 0.2])

TensorFlow Cross-Entropy

softmax = tf.placeholder(tf.float32)
one_hot = tf.placeholder(tf.float32)

cross_entropy = -tf.reduce_sum(tf.multiply(one_hot, tf.log(softmax)))

with tf.Session() as sess:
    print(sess.run(cross_entropy, feed_dict={softmax: [0.7, 0.2, 0.1], one_hot: [1.0, 0.0, 0.0]}))