An introduction to

TensorFlow!

1
 Agenda
Why TensorFlow

Graphs and Sessions

Linear Regression

Structuring your model

Managing experiments

3
 Why TensorFlow?
● Flexibility + Scalability
● Popularity

4
import tensorflow as tf

5
Graphs and Sessions

6
 Data Flow Graphs

TensorFlow separates definition of computations from their execution

Graph from TensorFlow for Machine Intelligence 7

 Data Flow Graphs

Phase 1: assemble a graph

Phase 2: use a session to execute operations in the graph.

Graph from TensorFlow for Machine Intelligence 8

What’s a tensor?

10
 What’s a tensor?

An n-dimensional array

0-d tensor: scalar (number)

1-d tensor: vector

2-d tensor: matrix

and so on

11
 Data Flow Graphs

import tensorflow as tf Visualized by TensorBoard

a = tf.add(3, 5)

12
 Data Flow Graphs

import tensorflow as tf Visualized by TensorBoard

a = tf.add(3, 5)

Why x, y?

TF automatically names the nodes when you don’t

explicitly name them.
x=3
y=5

13
 Data Flow Graphs

import tensorflow as tf Interpreted?

a = tf.add(3, 5)

3
Nodes: operators, variables, and constants
Edges: tensors 5 a

Tensors are data.

TensorFlow = tensor + flow = data + flow
(I know, mind=blown)

14
 Data Flow Graphs

import tensorflow as tf
a = tf.add(3, 5)
print(a)
3

5 a

>> Tensor("Add:0", shape=(), dtype=int32)

(Not 8)

15
 How to get the value of a?

Create a session, assign it to variable sess so we can call it later

Within the session, evaluate the graph to fetch the value of a

16
 How to get the value of a?

Create a session, assign it to variable sess so we can call it later

Within the session, evaluate the graph to fetch the value of a

import tensorflow as tf
a = tf.add(3, 5)
sess = tf.Session()
print(sess.run(a))
sess.close()

The session will look at the graph, trying to think: hmm, how can I get the value of a,
then it computes all the nodes that leads to a. 17
 How to get the value of a?

Create a session, assign it to variable sess so we can call it later

Within the session, evaluate the graph to fetch the value of a

import tensorflow as tf
a = tf.add(3, 5)
sess = tf.Session()
print(sess.run(a)) >> 8
sess.close() 8

The session will look at the graph, trying to think: hmm, how can I get the value of a,
then it computes all the nodes that leads to a. 18
 How to get the value of a?

Create a session, assign it to variable sess so we can call it later

Within the session, evaluate the graph to fetch the value of a

import tensorflow as tf
a = tf.add(3, 5)
sess = tf.Session()
with tf.Session() as sess:
print(sess.run(a)) 8
sess.close()

19
 tf.Session()

A Session object encapsulates the environment in which Operation objects are

executed, and Tensor objects are evaluated.

20
 tf.Session()

A Session object encapsulates the environment in which Operation objects are

executed, and Tensor objects are evaluated.

Session will also allocate memory to store the current values of variables.

21
 More graph
Visualized by TensorBoard
x = 2
y = 3
op1 = tf.add(x, y)
op2 = tf.multiply(x, y)
op3 = tf.pow(op2, op1)
with tf.Session() as sess:
op3 = sess.run(op3)

22
 Subgraphs
x = 2
y = 3 useless pow_op
add_op = tf.add(x, y)
mul_op = tf.multiply(x, y)
useless = tf.multiply(x, add_op)
pow_op = tf.pow(add_op, mul_op)
with tf.Session() as sess:
z = sess.run(pow_op)
add_op mul_op

Because we only want the value of pow_op and pow_op doesn’t

depend on useless, session won’t compute value of useless
→ save computation

23
 Subgraphs
Possible to break graphs into several
chunks and run them parallelly
across multiple CPUs, GPUs, TPUs,
or other devices

Example: AlexNet

Graph from Hands-On Machine Learning with Scikit-Learn and TensorFlow 24

 Distributed Computation
To put part of a graph on a specific CPU or GPU:

# Creates a graph.
with tf.device('/gpu:2'):
a = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], name='a')
b = tf.constant([1.0, 2.0, 3.0, 4.0, 5.0, 6.0], name='b')
c = tf.multiply(a, b)

# Creates a session with log_device_placement set to True.

sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))

# Runs the op.

print(sess.run(c))

25
 Why graphs

1. Save computation. Only run subgraphs that lead

to the values you want to fetch.

27
 Why graphs

1. Save computation. Only run subgraphs that lead

to the values you want to fetch.
2. Break computation into small, differential pieces
to facilitate auto-differentiation

28
 Why graphs

1. Save computation. Only run subgraphs that lead

to the values you want to fetch.
2. Break computation into small, differential pieces
to facilitate auto-differentiation
3. Facilitate distributed computation, spread the
work across multiple CPUs, GPUs, TPUs, or other
devices

29
 Why graphs

1. Save computation. Only run subgraphs that lead

to the values you want to fetch.
2. Break computation into small, differential pieces
to facilitate auto-differentiation
3. Facilitate distributed computation, spread the
work across multiple CPUs, GPUs, TPUs, or other
devices
4. Many common machine learning models are
taught and visualized as directed graphs
A neural net graph from Stanford’s
CS224N course
30
TensorBoard

31
 Your first TensorFlow program
import tensorflow as tf

a = tf.constant(2, name='a')
b = tf.constant(3, name='b')
x = tf.add(a, b, name='add')

with tf.Session() as sess:

print(sess.run(x))

32
 Visualize it with TensorBoard
import tensorflow as tf

a = tf.constant(2, name='a')
b = tf.constant(3, name='b') Create the summary writer after graph
definition and before running your session
x = tf.add(a, b, name='add')

writer = tf.summary.FileWriter('./graphs', tf.get_default_graph())

with tf.Session() as sess:
# writer = tf.summary.FileWriter('./graphs', sess.graph)
print(sess.run(x))
writer.close() # close the writer when you’re done using it
‘graphs’ or any location where you want to
keep your event files

33
 Run it
Go to terminal, run:

$ python [yourprogram].py
$ tensorboard --logdir="./graphs" --port 6006 6006 or any port you want

Then open your browser and go to: http://localhost:6006/

34
35
Constants, Sequences,
Variables, Ops

36
 Constants
import tensorflow as tf

a = tf.constant([2, 2], name='a')

b = tf.constant([[0, 1], [2, 3]], name= 'b')
x = tf.multiply(a, b, name='mul')
Broadcasting similar to NumPy
with tf.Session() as sess:
print(sess.run(x))

# >> [[0 2]
# [4 6]]

37
 Tensors filled with a specific value
tf.zeros([2, 3], tf.int32) ==> [[0, 0, 0], [0, 0, 0]]

# input_tensor is [[0, 1], [2, 3], [4, 5]] Similar to NumPy

tf.zeros_like(input_tensor) ==> [[0, 0], [0, 0], [0, 0]]

tf.fill([2, 3], 8) ==> [[8, 8, 8], [8, 8, 8]]

38
 Constants as sequences
tf.lin_space(start, stop, num, name=None)
tf.lin_space(10.0, 13.0, 4) ==> [10. 11. 12. 13.]

tf.range(start, limit=None, delta=1, dtype=None, name='range')

tf.range(3, 18, 3) ==> [3 6 9 12 15]
tf.range(5) ==> [0 1 2 3 4]
NOT THE SAME AS NUMPY SEQUENCES

Tensor objects are not iterable

for _ in tf.range(4): # TypeError

39
 Randomly Generated Constants

tf.random_normal
tf.truncated_normal
tf.random_uniform
tf.random_shuffle
tf.random_crop
tf.multinomial
tf.random_gamma

40
What’s wrong with constants?

Not trainable

45
 Constants are stored in graph definition
my_const = tf.constant([1.0, 2.0], name="my_const")

with tf.Session() as sess:

print(sess.graph.as_graph_def())

46
Constants are stored in graph definition

This makes loading graphs expensive when constants are big

47
Constants are stored in graph definition

This makes loading graphs expensive when constants are big

Only use constants for primitive types.

Use variables or readers for more data that requires more memory

48
 Variables
# create variables with tf.Variable
s = tf.Variable(2, name="scalar")
m = tf.Variable([[0, 1], [2, 3]], name="matrix")
W = tf.Variable(tf.zeros([784,10]))

# create variables with tf.get_variable

s = tf.get_variable("scalar", initializer=tf.constant(2))
m = tf.get_variable("matrix", initializer=tf.constant([[0, 1], [2, 3]]))
W = tf.get_variable("big_matrix", shape=(784, 10), initializer=tf.zeros_initializer())

49
 You have to initialize your variables
The easiest way is initializing all variables at once:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())

Initializer is an op. You need to execute it

within the context of a session

50
 You have to initialize your variables
The easiest way is initializing all variables at once:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())

Initialize only a subset of variables:

with tf.Session() as sess:
sess.run(tf.variables_initializer([a, b]))

51
 You have to initialize your variables
The easiest way is initializing all variables at once:
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())

Initialize only a subset of variables:

with tf.Session() as sess:
sess.run(tf.variables_initializer([a, b]))

Initialize a single variable

W = tf.Variable(tf.zeros([784,10]))
with tf.Session() as sess:
sess.run(W.initializer)

52
 Eval() a variable
# W is a random 700 x 100 variable object
W = tf.Variable(tf.truncated_normal([700, 10]))
with tf.Session() as sess:
sess.run(W.initializer)
print(W)

>> Tensor("Variable/read:0", shape=(700, 10), dtype=float32)

53
 tf.Variable.assign()
W = tf.Variable(10)
W.assign(100)
with tf.Session() as sess:
sess.run(W.initializer)
print(W.eval()) # >> ????

54
 tf.Variable.assign()
W = tf.Variable(10)
W.assign(100)
with tf.Session() as sess:
sess.run(W.initializer)
print(W.eval()) # >> 10

Ugh, why?

55
 tf.Variable.assign()
W = tf.Variable(10)
W.assign(100)
with tf.Session() as sess:
sess.run(W.initializer)
print(W.eval()) # >> 10

W.assign(100) creates an assign op.

That op needs to be executed in a session
to take effect.

56
 tf.Variable.assign()
W = tf.Variable(10)
W.assign(100)
with tf.Session() as sess:
sess.run(W.initializer)
print(W.eval()) # >> 10

--------

W = tf.Variable(10)
assign_op = W.assign(100)
with tf.Session() as sess:
sess.run(W.initializer)
sess.run(assign_op)
print(W.eval()) # >> 100
57
Placeholder

58
 A quick reminder

A TF program often has 2 phases:

1. Assemble a graph
2. Use a session to execute operations in the graph.

59
 Placeholders

A TF program often has 2 phases:

1. Assemble a graph
2. Use a session to execute operations in the graph.

⇒ Assemble the graph first without knowing the values needed for computation

60
 Placeholders

A TF program often has 2 phases:

1. Assemble a graph
2. Use a session to execute operations in the graph.

⇒ Assemble the graph first without knowing the values needed for computation

Analogy:
Define the function f(x, y) = 2 * x + y without knowing value of x or y.
x, y are placeholders for the actual values.

61
 Why placeholders?

We, or our clients, can later supply their own data when they
need to execute the computation.

62
 Placeholders

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder for a vector of 3 elements, type tf.float32
a = tf.placeholder(tf.float32, shape=[3])

b = tf.constant([5, 5, 5], tf.float32)

# use the placeholder as you would a constant or a variable

c = a + b # short for tf.add(a, b)

with tf.Session() as sess:

print(sess.run(c)) # >> ???

63
 Placeholders

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder for a vector of 3 elements, type tf.float32
a = tf.placeholder(tf.float32, shape=[3])

b = tf.constant([5, 5, 5], tf.float32)

# use the placeholder as you would a constant or a variable

c = a + b # short for tf.add(a, b)

with tf.Session() as sess:

print(sess.run(c)) # >> InvalidArgumentError: a doesn’t an actual value

64
Supplement the values to placeholders using
a dictionary

65
 Placeholders

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder for a vector of 3 elements, type tf.float32
a = tf.placeholder(tf.float32, shape=[3])

b = tf.constant([5, 5, 5], tf.float32)

# use the placeholder as you would a constant or a variable

c = a + b # short for tf.add(a, b)

with tf.Session() as sess:

print(sess.run(c, feed_dict={a: [1, 2, 3]})) # the tensor a is the key, not the string ‘a’

# >> [6, 7, 8]

66
 Placeholders

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder for a vector of 3 elements, type tf.float32
a = tf.placeholder(tf.float32, shape=[3])

b = tf.constant([5, 5, 5], tf.float32) Quirk:

shape=None means that tensor of any
# use the placeholder as you would a constant or a variable
shape will be accepted as value for
c = a + b # short for tf.add(a, b)
placeholder.
with tf.Session() as sess:
print(sess.run(c, feed_dict={a: [1, 2, 3]})) shape=None is easy to construct graphs
and great when you have different
# >> [6, 7, 8] batch sizes, but nightmarish for
debugging

67
 Placeholders

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder of type float 32-bit, shape is a vector of 3 elements
a = tf.placeholder(tf.float32, shape=[3])

# create a constant of type float 32-bit, shape is a vector of 3 elements

b = tf.constant([5, 5, 5], tf.float32)

# use the placeholder as you would a constant or a variable

Quirk:
c = a + b # Short for tf.add(a, b) shape=None also breaks all following
shape inference, which makes many
with tf.Session() as sess: ops not work because they expect
print(sess.run(c, {a: [1, 2, 3]})) certain rank.

# >> [6, 7, 8]

68
 Placeholders are valid ops

tf.placeholder(dtype, shape=None, name=None)

# create a placeholder of type float 32-bit, shape is a vector of 3 elements
a = tf.placeholder(tf.float32, shape=[3])

# create a constant of type float 32-bit, shape is a vector of 3 elements

b = tf.constant([5, 5, 5], tf.float32)

# use the placeholder as you would a constant or a variable

c = a + b # Short for tf.add(a, b)

with tf.Session() as sess:

print(sess.run(c, {a: [1, 2, 3]}))

# >> [6, 7, 8]

69
 What if want to feed multiple data points in?
You have to do it one at a time
with tf.Session() as sess:
for a_value in list_of_values_for_a:
print(sess.run(c, {a: a_value}))

70
Linear Regression
in TensorFlow

72
Model the linear relationship between:
● dependent variable Y
● explanatory variables X

73
 Want

Find a linear relationship between X and Y

to predict Y from X

76
 Model

Inference: Y_predicted = w * X + b

Mean squared error: E[(y - y_predicted) 2]

77
Phase 1: Assemble our graph

80
Step 2: Create placeholders for
inputs and labels

tf.placeholder(dtype, shape=None, name=None)

82
 Step 3: Create weight and bias

tf.get_variable(
No need to specify shape if
using constant initializer
name,

shape=None,

dtype=None,

initializer=None,

) 83
Step 4: Inference

Y_predicted = w * X + b

84
 Step 5: Specify loss function

loss = tf.square(Y - Y_predicted, name='loss')

85
 Step 6: Create optimizer

opt = tf.train.GradientDescentOptimizer(learning_rate=0.001)

optimizer = opt.minimize(loss)

86
 Phase 2: Train our model

Step 1: Initialize variables

Step 2: Run optimizer

(use a feed_dict to feed data into X and Y placeholders)

87
 Write log files using a FileWriter

writer = tf.summary.FileWriter('./graphs/linear_reg', sess.graph)

88
 See it on TensorBoard

Step 1: $ python linreg_starter.py

Step 2: $ tensorboard --logdir='./graphs'

89
90
91
Optimizers

111
 Optimizer

optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.01).minimize(loss)

_, l = sess.run([optimizer, loss], feed_dict={X: x, Y:y})

112
 Optimizer

optimizer = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss)

_, l = sess.run([optimizer, loss], feed_dict={X: x, Y:y})

Session looks at all trainable variables that loss depends on and update them

113
 Optimizer

Session looks at all trainable variables that optimizer depends on and update them

114
 Trainable variables

tf.Variable(initial_value=None, trainable=True,...)

Specify if a variable should be trained or not

By default, all variables are trainable

115
 List of optimizers in TF
tf.train.GradientDescentOptimizer

tf.train.AdagradOptimizer
Usually Adam works out-of-the-box better than
tf.train.MomentumOptimizer SGD

tf.train.AdamOptimizer

tf.train.FtrlOptimizer

tf.train.RMSPropOptimizer

...

116
 Name scope

TensorFlow doesn’t know what nodes should be

grouped together, unless you tell it to

127
 Name scope

Group nodes together with tf.name_scope(name)

with tf.name_scope(name_of_that_scope):

# declare op_1

# declare op_2

# ...

128
 Name scope
with tf.name_scope('data'):
iterator = dataset.make_initializable_iterator()
center_words, target_words = iterator.get_next()

with tf.name_scope('embed'):
embed_matrix = tf.get_variable('embed_matrix',
shape=[VOCAB_SIZE, EMBED_SIZE], ...)
embed = tf.nn.embedding_lookup(embed_matrix, center_words)

with tf.name_scope('loss'):
nce_weight = tf.get_variable('nce_weight', shape=[VOCAB_SIZE, EMBED_SIZE], ...)
nce_bias = tf.get_variable('nce_bias', initializer=tf.zeros([VOCAB_SIZE]))
loss = tf.reduce_mean(tf.nn.nce_loss(weights=nce_weight, biases=nce_bias, …)

with tf.name_scope('optimizer'):
optimizer = tf.train.GradientDescentOptimizer(LEARNING_RATE).minimize(loss)

129
TensorBoard

130
 Variable scope

Name scope vs variable scope

tf.name_scope() vs tf.variable_scope()

131
 Variable scope

Name scope vs variable scope

Variable scope facilitates variable sharing

132
 Variable sharing: The problem
def two_hidden_layers(x):
w1 = tf.Variable(tf.random_normal([100, 50]), name='h1_weights')
b1 = tf.Variable(tf.zeros([50]), name='h1_biases')
h1 = tf.matmul(x, w1) + b1

w2 = tf.Variable(tf.random_normal([50, 10]), name='h2_weights')

b2 = tf.Variable(tf.zeros([10]), name='2_biases')
logits = tf.matmul(h1, w2) + b2
return logits

133
 Variable sharing: The problem
def two_hidden_layers(x):
w1 = tf.Variable(tf.random_normal([100, 50]), name='h1_weights')
b1 = tf.Variable(tf.zeros([50]), name='h1_biases')
h1 = tf.matmul(x, w1) + b1

w2 = tf.Variable(tf.random_normal([50, 10]), name='h2_weights')

b2 = tf.Variable(tf.zeros([10]), name='2_biases')
logits = tf.matmul(h1, w2) + b2
return logits What will happen if we
make these two calls?
logits1 = two_hidden_layers(x1)
logits2 = two_hidden_layers(x2)

134
Sharing Variable: The problem

Two sets of
variables are
created.

You want all

your inputs to
use the same
weights and
biases!

135
 tf.get_variable()
tf.get_variable(<name>, <shape>, <initializer>)

If a variable with <name> already exists, reuse it

If not, initialize it with <shape> using <initializer>

136
 tf.get_variable()
def two_hidden_layers(x):
assert x.shape.as_list() == [200, 100]
w1 = tf.get_variable("h1_weights", [100, 50], initializer=tf.random_normal_initializer())
b1 = tf.get_variable("h1_biases", [50], initializer=tf.constant_initializer(0.0))
h1 = tf.matmul(x, w1) + b1
assert h1.shape.as_list() == [200, 50]
w2 = tf.get_variable("h2_weights", [50, 10], initializer=tf.random_normal_initializer())
b2 = tf.get_variable("h2_biases", [10], initializer=tf.constant_initializer(0.0))
logits = tf.matmul(h1, w2) + b2
return logits
logits1 = two_hidden_layers(x1)
logits2 = two_hidden_layers(x2)

137
 tf.get_variable()
def two_hidden_layers(x):
assert x.shape.as_list() == [200, 100]
w1 = tf.get_variable("h1_weights", [100, 50], initializer=tf.random_normal_initializer())
b1 = tf.get_variable("h1_biases", [50], initializer=tf.constant_initializer(0.0))
h1 = tf.matmul(x, w1) + b1
assert h1.shape.as_list() == [200, 50]
w2 = tf.get_variable("h2_weights", [50, 10], initializer=tf.random_normal_initializer())
b2 = tf.get_variable("h2_biases", [10], initializer=tf.constant_initializer(0.0))
logits = tf.matmul(h1, w2) + b2
return logits
ValueError: Variable h1_weights already exists,
logits1 = two_hidden_layers(x1)
disallowed. Did you mean to set reuse=True in
logits2 = two_hidden_layers(x2) VarScope?

138
 tf.variable_scope()
def two_hidden_layers(x):
assert x.shape.as_list() == [200, 100]
w1 = tf.get_variable("h1_weights", [100, 50], initializer=tf.random_normal_initializer())
b1 = tf.get_variable("h1_biases", [50], initializer=tf.constant_initializer(0.0))
h1 = tf.matmul(x, w1) + b1
assert h1.shape.as_list() == [200, 50]
w2 = tf.get_variable("h2_weights", [50, 10], initializer=tf.random_normal_initializer())
b2 = tf.get_variable("h2_biases", [10], initializer=tf.constant_initializer(0.0))
logits = tf.matmul(h1, w2) + b2
return logits
Put your variables within a scope and reuse all
with tf.variable_scope('two_layers') as scope: variables within that scope
logits1 = two_hidden_layers(x1)
scope.reuse_variables()
139
logits2 = two_hidden_layers(x2)
tf.variable_scope()
Only one set of
variables, all within
the variable scope
“two_layers”

They take in two

different inputs

140
tf.variable_scope()
tf.variable_scope
implicitly creates a
name scope

141
 Reusable code?
def two_hidden_layers(x):
assert x.shape.as_list() == [200, 100]
w1 = tf.get_variable("h1_weights", [100, 50], initializer=tf.random_normal_initializer())
b1 = tf.get_variable("h1_biases", [50], initializer=tf.constant_initializer(0.0))
h1 = tf.matmul(x, w1) + b1
assert h1.shape.as_list() == [200, 50]
w2 = tf.get_variable("h2_weights", [50, 10], initializer=tf.random_normal_initializer())
b2 = tf.get_variable("h2_biases", [10], initializer=tf.constant_initializer(0.0))
logits = tf.matmul(h1, w2) + b2
return logits
with tf.variable_scope('two_layers') as scope:
logits1 = two_hidden_layers(x1)
scope.reuse_variables()
142
logits2 = two_hidden_layers(x2)
 Layer ‘em up
def fully_connected(x, output_dim, scope):
with tf.variable_scope(scope, reuse=tf.AUTO_REUSE) as scope:
w = tf.get_variable("weights", [x.shape[1], output_dim], initializer=tf.random_normal_initializer())
b = tf.get_variable("biases", [output_dim], initializer=tf.constant_initializer(0.0))
return tf.matmul(x, w) + b

def two_hidden_layers(x):
Fetch variables if they
already exist
h1 = fully_connected(x, 50, 'h1')
h2 = fully_connected(h1, 10, 'h2') Else, create them

with tf.variable_scope('two_layers') as scope:

logits1 = two_hidden_layers(x1)
logits2 = two_hidden_layers(x2)

143
Manage Experiments
 tf.train.Saver
saves graph’s variables in binary files

146
 Saves sessions, not graphs!

tf.train.Saver.save(sess, save_path, global_step=None...)

tf.train.Saver.restore(sess, save_path)

147
 Save parameters after 1000 steps
# define model
model = SkipGramModel(params)

# create a saver object

saver = tf.train.Saver()

with tf.Session() as sess:

for step in range(training_steps):
sess.run([optimizer])

# save model every 1000 steps

if (step + 1) % 1000 == 0:
saver.save(sess,
'checkpoint_directory/model_name',
global_step=step)

148
 Specify the step at which the model is saved
# define model
model = SkipGramModel(params)

# create a saver object

saver = tf.train.Saver()

with tf.Session() as sess:

for step in range(training_steps):
sess.run([optimizer])

# save model every 1000 steps

if (step + 1) % 1000 == 0:
saver.save(sess,
'checkpoint_directory/model_name',
global_step=step)

149
 Global step

global_step = tf.Variable(0, dtype=tf.int32, trainable=False, name='global_step')

Very common in
TensorFlow program

150
 Global step

global_step = tf.Variable(0,
dtype=tf.int32,
trainable=False,
name='global_step')

optimizer = tf.train.AdamOptimizer(lr).minimize(loss, global_step=global_step)

Need to tell optimizer to increment global step

This can also help your optimizer know when

to decay learning rate

151
 Restore variables

saver.restore(sess, 'checkpoints/name_of_the_checkpoint')

e.g. saver.restore(sess, 'checkpoints/skip-gram-99999')

Still need to first build

graph

155
 Restore the latest checkpoint
# check if there is checkpoint
ckpt = tf.train.get_checkpoint_state(os.path.dirname('checkpoints/checkpoint'))

# check if there is a valid checkpoint path

if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)

1. checkpoint file keeps track of the latest

checkpoint
2. restore checkpoints only when there is a valid
checkpoint path

156
 tf.summary
Why matplotlib when you can summarize?

157
 tf.summary

Visualize our summary statistics during our training

tf.summary.scalar
tf.summary.histogram
tf.summary.image

158
 Step 1: create summaries

with tf.name_scope("summaries"):
tf.summary.scalar("loss", self.loss)
tf.summary.scalar("accuracy", self.accuracy)
tf.summary.histogram("histogram loss", self.loss)
summary_op = tf.summary.merge_all()

merge them all into one summary op to

make managing them easier

159
 Step 2: run them

loss_batch, _, summary = sess.run([loss,

optimizer,
summary_op])

Like everything else in TF, summaries are ops.

For the summaries to be built, you have to run
it in a session

160
 Step 3: write summaries to file

writer.add_summary(summary, global_step=step)

Need global step here so the model knows

what summary corresponds to what step

161
 Putting it together
tf.summary.scalar("loss", self.loss)
tf.summary.histogram("histogram loss", self.loss)
summary_op = tf.summary.merge_all()

saver = tf.train.Saver() # defaults to saving all variables

with tf.Session() as sess:

sess.run(tf.global_variables_initializer())
ckpt = tf.train.get_checkpoint_state(os.path.dirname('checkpoints/checkpoint'))
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)

writer = tf.summary.FileWriter('./graphs', sess.graph)

for index in range(10000):
...
loss_batch, _, summary = sess.run([loss, optimizer, summary_op])
writer.add_summary(summary, global_step=index)

if (index + 1) % 1000 == 0:
saver.save(sess, 'checkpoints/skip-gram', index)
162
See summaries on TensorBoard

163
Scalar loss

164
Histogram loss

165
Toggle run to compare experiments

166