### Basic

TensorFlow is an open source software library for machine learning developed by Google. This tutorial is designed to teach the basic concepts and how to use it.

#### First TensorFlow program

TensorFlow represents computations by linking op nodes into graphs. TensorFlow programs are structured into a construction phase and an execution phase. The following program:

1. Constructs a computation graph for a matrix multiplication.
2. Open a TensorFlow session and compute the matrix multiplication by executing the computation graph.
import tensorflow as tf

# Construct 2 op nodes (m1, m2) representing 2 matrix.
m1 = tf.constant([[3, 5]])     # (1, 2)
m2 = tf.constant([[2],[4]])    # (2, 1)

product = tf.matmul(m1, m2)    # A matrix multiplication op node

with tf.Session() as sess:     # Open a TensorFlow session to execute the graph.
result = sess.run(product) # Compute the result for “product”
print(result)              # 3*2+5*4: [[26]]


sess.run and _tensor.eval()_will return a NumPy array containing the result of the computation.

The above program hardwires the matrix as a constant. We will implement a new linear equation that feeds the graph with input data on execution.

import tensorflow as tf
import numpy as np

W = tf.constant([[3, 5]])

# Allow data to be supplied later during execution.
x = tf.placeholder(tf.int32, shape=(2, 1))
b = tf.placeholder(tf.int32)

# A linear model y = Wx + b
product = tf.matmul(W, x) + b

with tf.Session() as sess:
# Feed data into the place holder (x & b) before execution.
result = sess.run(product, feed_dict={x: np.array([[2],[4]]), b:1})
print(result)              # 3*2+5*4+1 = [[27]]


When we construct a graph (tf.constant, tf.get_variable, tf.matmul), we are just building a computation graph. No computation is actually perforned until we run it inside a session (sess.run).

Common tensor types in TensorFlow are:

• tf.Variable
• tf.Constant
• tf.Placeholder
• tf.SparseTensor

#### Train a linear model

Let’s do a simple linear regression with a linear model below.

We will supply the model with training data (x, y) and later compute the corresponding model parameter W & b.

import tensorflow as tf

### Define a model: a computational graph
# Parameters for a linear model y = Wx + b
W = tf.get_variable("W", initializer=tf.constant([0.1]))
b = tf.get_variable("b", initializer=tf.constant([0.0]))

# Placeholder for input and prediction
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)

# Define a linear model y = Wx + b
model = W * x + b

### Define a cost function, an optimizer and a trainer
# Define a cost function (Mean square error - MSE)
loss = tf.reduce_sum(tf.square(model - y))

# Optimizer with a 0.01 learning rate
train = optimizer.minimize(loss)

### Training (Fitting)
# Training data
x_train = [1.0, 2.0, 3.0, 4.0]
y_train = [1.5, 3.5, 5.5, 7.5]

with tf.Session() as sess:
# Retrieve the variable initializer op and initialize variable W & b.
sess.run(session.run(tf.global_variables_initializer()))
for i in range(1000):
sess.run(train, {x:x_train, y:y_train})
if i%100==0:
l_cost = sess.run(loss, {x:x_train, y:y_train})
print(f"i: {i} cost: {l_cost}")

# Evaluate training accuracy
l_W, l_b, l_cost  = sess.run([W, b, loss], {x:x_train, y:y_train})
print(f"W: {l_W} b: {l_b} cost: {l_cost}")
# W: [ 1.99999797] b: [-0.49999401] cost: 2.2751578399038408e-11


A typical TensorFlow program contains:

• Define a model
• Define a loss function and a trainer
• Training (fitting)

#### Model

Define the linear model y = Wx + b.

### Define a model: a computational graph
# Parameters for a linear model y = Wx + b
W = tf.get_variable("W", initializer=tf.constant([0.1]))
b = tf.get_variable("b", initializer=tf.constant([0.0]))

# Placeholder for input and prediction
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)

# Define a linear model y = Wx + b
model = W * x + b


We define both $W$ and $b$ as variables initialized as 0.1 and 0 respectively. Variables are trainable and can act as the parameters of a model.

W = tf.get_variable("W", initializer=tf.constant([0.1]))
b = tf.get_variable("b", initializer=tf.constant([0.0]))


The shape of a tensor is the dimension of a tensor. For example, a 5x5x3 matrix is a Rank 3 (3-dimensional) tensor with shape (5, 5, 3). By default, the data type (dtype) of a tensor is tf.float32. Here we initialize a tensor with shape (5, 5, 3) of int32 type with 0.

int_v = tf.get_variable("int_variable", [5, 5, 3], dtype=tf.int32,
initializer=tf.zeros_initializer)


#### Lost function and optimizer & trainer

Define the Mean Square Error (MSE) cost function:

loss = tf.reduce_sum(tf.square(model - y))


We define a gradient descent optimizer and trainer to find an optimal solution that can fit our training data with the minimum loss.

# Optimizer with a 0.01 learning rate
train = optimizer.minimize(loss)


#### Training (fitting)

Before any execution, we need to initialize all the parameters:

init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)


We train our data with 1000 iterations. For every 100 iterations, we compute the loss and print it out.

for i in range(1000):
sess.run(train, {x:x_train, y:y_train})
if i%100==0:
l_cost = sess.run(loss, {x:x_train, y:y_train})
print(f"i: {i} cost: {l_cost}")


Once 1000 iterations are done, we print out W, b and the loss:

# Evaluate training accuracy
l_W, l_b, l_cost  = sess.run([W, b, loss], {x:x_train, y:y_train})
print(f"W: {l_W} b: {l_b} cost: {l_cost}")
# W: [ 1.99999797] b: [-0.49999401] cost: 2.2751578399038408e-11


Here we model our training data as: $y = 2x - 0.5$

### Solving MNist

The MNIST dataset contains handwritten digits with examples shown as above. It has a training set of 60,000 examples and a test set of 10,000 examples. The following python file from TensorFlow mnist_softmax.py train a linear classifier for MNist digit recognition. The following model reaches an accuracy of 92%.

import argparse
import sys

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf

FLAGS = None

def main(_):
# Import data

# Create the model
x = tf.placeholder(tf.float32, [None, 784])
W = tf.Variable(tf.zeros([784, 10]))
b = tf.Variable(tf.zeros([10]))
y = tf.matmul(x, W) + b

# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, 10])

cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))

sess = tf.InteractiveSession()
tf.global_variables_initializer().run()
# Train
for _ in range(1000):
batch_xs, batch_ys = mnist.train.next_batch(100)
sess.run(train_step, feed_dict={x: batch_xs, y_: batch_ys})

# Test trained model
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
print(sess.run(accuracy, feed_dict={x: mnist.test.images,
y_: mnist.test.labels}))

if __name__ == '__main__':
parser = argparse.ArgumentParser()
help='Directory for storing input data')
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

# 0.9241


Read training, validation and testing dataset into “mnist”.

from tensorflow.examples.tutorials.mnist import input_data

def main(_):
# Import data


Each image is 28x28 = 784. We use a linear classifier to classify the handwritten image from either 0 to 9.

x = tf.placeholder(tf.float32, [None, 784])
W = tf.get_variable("W", [784, 10], initializer=tf.zeros_initializer)
b = tf.get_variable("b", [10], initializer=tf.zeros_initializer)
y = tf.matmul(x, W) + b


We use cross-entropy as the cost functions:

  cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y))


### Solving MNist with a fully connected networking

Now we replace the model using deep learning techniques. This example contains 2 hidden fully connected layers. The new model achieves an accuracy of 98%.

For each hidden layer:

with tf.name_scope('hidden1'):   # Create a name scope for hidden layer 1
weights = tf.Variable(       # Create a variable for weights initialized with truncated normal distribution
tf.truncated_normal([IMAGE_PIXELS, hidden1_units],
stddev=1.0 / math.sqrt(float(IMAGE_PIXELS))),
name='weights')
biases = tf.Variable(tf.zeros([hidden1_units]), name='biases') # Create avaraible for the biases
hidden1 = tf.nn.relu(tf.matmul(images, weights) + biases)      # Matrix multiplication followed by a RELU


Cost function using cross entropy.

def loss(logits, labels):
labels = tf.to_int64(labels)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels, logits=logits, name='xentropy')
return tf.reduce_mean(cross_entropy, name='xentropy_mean')


Training and evaluation:

def training(loss, learning_rate):
# Add a scalar summary for the snapshot loss.
tf.summary.scalar('loss', loss)

global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss, global_step=global_step)
return train_op

def evaluation(logits, labels):
correct = tf.nn.in_top_k(logits, labels, 1)
return tf.reduce_sum(tf.cast(correct, tf.int32))


The full code for defining the model:

import math
import tensorflow as tf

NUM_CLASSES = 10
IMAGE_SIZE = 28
IMAGE_PIXELS = IMAGE_SIZE * IMAGE_SIZE

def inference(images, hidden1_units, hidden2_units):
# Hidden 1
with tf.name_scope('hidden1'):
weights = tf.Variable(
tf.truncated_normal([IMAGE_PIXELS, hidden1_units],
stddev=1.0 / math.sqrt(float(IMAGE_PIXELS))),
name='weights')
biases = tf.Variable(tf.zeros([hidden1_units]),
name='biases')
hidden1 = tf.nn.relu(tf.matmul(images, weights) + biases)
# Hidden 2
with tf.name_scope('hidden2'):
weights = tf.Variable(
tf.truncated_normal([hidden1_units, hidden2_units],
stddev=1.0 / math.sqrt(float(hidden1_units))),
name='weights')
biases = tf.Variable(tf.zeros([hidden2_units]),
name='biases')
hidden2 = tf.nn.relu(tf.matmul(hidden1, weights) + biases)
# Linear
with tf.name_scope('softmax_linear'):
weights = tf.Variable(
tf.truncated_normal([hidden2_units, NUM_CLASSES],
stddev=1.0 / math.sqrt(float(hidden2_units))),
name='weights')
biases = tf.Variable(tf.zeros([NUM_CLASSES]),
name='biases')
logits = tf.matmul(hidden2, weights) + biases
return logits

def loss(logits, labels):
labels = tf.to_int64(labels)
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
labels=labels, logits=logits, name='xentropy')
return tf.reduce_mean(cross_entropy, name='xentropy_mean')

def training(loss, learning_rate):
# Add a scalar summary for the snapshot loss.
tf.summary.scalar('loss', loss)

global_step = tf.Variable(0, name='global_step', trainable=False)
train_op = optimizer.minimize(loss, global_step=global_step)
return train_op

def evaluation(logits, labels):
correct = tf.nn.in_top_k(logits, labels, 1)
return tf.reduce_sum(tf.cast(correct, tf.int32))


Here is the main program:

import argparse
import os
import sys
import time

import tensorflow as tf

from tensorflow.examples.tutorials.mnist import input_data
from tensorflow.examples.tutorials.mnist import mnist

# Basic model parameters as external flags.
FLAGS = None

def placeholder_inputs(batch_size):
images_placeholder = tf.placeholder(tf.float32, shape=(batch_size, mnist.IMAGE_PIXELS))
labels_placeholder = tf.placeholder(tf.int32, shape=(batch_size))
return images_placeholder, labels_placeholder

def fill_feed_dict(data_set, images_pl, labels_pl):
images_feed, labels_feed = data_set.next_batch(FLAGS.batch_size, FLAGS.fake_data)
feed_dict = {images_pl: images_feed, labels_pl: labels_feed,}
return feed_dict

def do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_set):
true_count = 0
steps_per_epoch = data_set.num_examples // FLAGS.batch_size
num_examples = steps_per_epoch * FLAGS.batch_size
for step in range(steps_per_epoch):
feed_dict = fill_feed_dict(data_set, images_placeholder, labels_placeholder)
true_count += sess.run(eval_correct, feed_dict=feed_dict)
precision = float(true_count) / num_examples
print('  Num examples: %d  Num correct: %d  Precision @ 1: %0.04f' % (num_examples, true_count, precision))

def run_training():

with tf.Graph().as_default():
images_placeholder, labels_placeholder = placeholder_inputs(FLAGS.batch_size)

logits = mnist.inference(images_placeholder, FLAGS.hidden1, FLAGS.hidden2)
loss = mnist.loss(logits, labels_placeholder)

train_op = mnist.training(loss, FLAGS.learning_rate)
eval_correct = mnist.evaluation(logits, labels_placeholder)

saver = tf.train.Saver()
init = tf.global_variables_initializer()

summary = tf.summary.merge_all()

sess = tf.Session()

summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)

sess.run(init)

for step in range(FLAGS.max_steps):
start_time = time.time()

feed_dict = fill_feed_dict(data_sets.train, images_placeholder, labels_placeholder)

_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)

duration = time.time() - start_time

if step % 100 == 0:
print('Step %d: loss = %.2f (%.3f sec)' % (step, loss_value, duration))
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.flush()

if (step + 1) % 1000 == 0 or (step + 1) == FLAGS.max_steps:
checkpoint_file = os.path.join(FLAGS.log_dir, 'model.ckpt')
saver.save(sess, checkpoint_file, global_step=step)

print('Training Data Eval:')
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.train)

print('Validation Data Eval:')
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.validation)

print('Test Data Eval:')
do_eval(sess, eval_correct, images_placeholder, labels_placeholder, data_sets.test)

def main(_):
if tf.gfile.Exists(FLAGS.log_dir):
tf.gfile.DeleteRecursively(FLAGS.log_dir)
tf.gfile.MakeDirs(FLAGS.log_dir)
run_training()

if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--learning_rate', type=float, default=0.01, help='Initial learning rate.')
parser.add_argument('--max_steps', type=int, default=2000, help='Number of steps to run trainer.')
parser.add_argument('--hidden1', type=int, default=128, help='Number of units in hidden layer 1.')
parser.add_argument('--hidden2', type=int, default=32, help='Number of units in hidden layer 2.')
parser.add_argument('--batch_size', type=int, default=100, help='Batch size.  Must divide evenly into the dataset sizes.')
parser.add_argument('--input_data_dir', type=str, default=os.path.join(os.getenv('TEST_TMPDIR', '/tmp'),  'tensorflow/mnist/input_data'), help='Directory to put the input data.')
parser.add_argument('--log_dir', type=str, default=os.path.join(os.getenv('TEST_TMPDIR', '/tmp'), 'tensorflow/mnist/logs/fully_connected_feed'), help='Directory to put the log data.')
parser.add_argument('--fake_data', default=False, help='If true, uses fake data for unit testing.', action='store_true')

FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)


Collecting data for TensorBoard In mnist.py:

def training(loss, learning_rate):
...
tf.summary.scalar('loss', loss)
...

def run_training():
...
loss = mnist.loss(logits, labels_placeholder)
train_op = mnist.training(loss, FLAGS.learning_rate)

summary = tf.summary.merge_all()
sess = tf.Session()
summary_writer = tf.summary.FileWriter(FLAGS.log_dir, sess.graph)

...
for step in range(FLAGS.max_steps):
...
_, loss_value = sess.run([train_op, loss], feed_dict=feed_dict)
...

if step % 100 == 0:
summary_str = sess.run(summary, feed_dict=feed_dict)
summary_writer.flush()


Further accuracy improvement can be achieved by:

• Increase the number of iterations.
• Change to a CNN architect.
• Replace the regularization with more advanced methods like batch normalization or dropout.
• Fine tuning of the learning rate in the Adam optimizer and the lambda in the L2 regularization.

In next section, we will cover the CNN and dropout implementation.

### MNist with a Convolution network (CNN)

To push the accuracy higher, we will create a model with 2 CNN layers followed by 2 hidden fully connected (FC) layers and the final linear classifier. We also apply:

• a 5x5 filter for both CNN layers.
• a 2x2 max pooling max(z11, z12, z21, z22) for both CNN layers.
• Use RELU max(0, z) for both CNN and FC layer.
• Use dropout for regularization.
• Use cross entropy cost function with Adam optimizer.

The following code reaches an accuracy of 99.4% with little parameter tuning.

import argparse
import sys

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf
import numpy as np

FLAGS = None

def main(_):
# Import data

### Building a model with 2 Convolution layers
### followed by 2 fully connected hidden layers and a linear classification layer.
x = tf.placeholder(tf.float32, [None, 784])

# Parameters for the 2 convolution layer
with tf.variable_scope("CNN"):
cnn_W1 = tf.get_variable("W1", [5, 5, 1, 32], initializer=tf.truncated_normal_initializer(stddev=0.1))
cnn_b1 = tf.get_variable("b1", [32], initializer=tf.constant_initializer(0.1))
cnn_W2 = tf.get_variable("W2", [5, 5, 32, 64], initializer=tf.truncated_normal_initializer(stddev=0.1))
cnn_b2 = tf.get_variable("b2", [64], initializer=tf.constant_initializer(0.1))

# Parameters for 2 hidden layers with dropout and the linear classification layer.
# 3136 = 7 * 7 * 64
W1 = tf.get_variable("W1", [3136, 1000], initializer=tf.truncated_normal_initializer(stddev=np.sqrt(2.0 / 3136)))
b1 = tf.get_variable("b1", [1000], initializer=tf.zeros_initializer)
W2 = tf.get_variable("W2", [1000, 100], initializer=tf.truncated_normal_initializer(stddev=np.sqrt(2.0 / 1000)))
b2 = tf.get_variable("b2", [100], initializer=tf.zeros_initializer)
W3 = tf.get_variable("W3", [100, 10], initializer=tf.truncated_normal_initializer(stddev=np.sqrt(2.0 / 100)))
b3 = tf.get_variable("b3", [10], initializer=tf.zeros_initializer)

keep_prob = tf.placeholder(tf.float32)

# First CNN with RELU and max pooling.
x_image = tf.reshape(x, [-1, 28, 28, 1])
cnn1 = tf.nn.conv2d(x_image, cnn_W1, strides=[1, 1, 1, 1], padding='SAME')
z1 = tf.nn.relu(cnn1 + cnn_b1)
h1 = tf.nn.max_pool(z1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# Second CNN
cnn2 = tf.nn.conv2d(h1, cnn_W2, strides=[1, 1, 1, 1], padding='SAME')
z2 = tf.nn.relu(cnn2 + cnn_b2)
h2 = tf.nn.max_pool(z2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')

# First FC layer with dropout.
h2_flat = tf.reshape(h2, [-1, 3136])
h_fc1 = tf.nn.relu(tf.matmul(h2_flat, W1) + b1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

# Second FC
h_fc2 = tf.nn.relu(tf.matmul(h_fc1_drop, W2) + b2)
h_fc2_drop = tf.nn.dropout(h_fc2, keep_prob)

# Linear classification.
y = tf.matmul(h_fc2_drop, W3) + b3

# True label
labels = tf.placeholder(tf.float32, [None, 10])

# Cost function & optimizer
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=y) )

init = tf.global_variables_initializer()
with tf.Session() as sess:
sess.run(init)
# Train
for i in range(10001):
batch_xs, batch_ys = mnist.train.next_batch(100)
sess.run(train_step, feed_dict={x: batch_xs, labels: batch_ys, keep_prob:0.5})
if i%50==0:
# Test trained model
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
result = sess.run(accuracy, feed_dict={x: mnist.test.images,
labels: mnist.test.labels,
keep_prob:1.0})
print(f"Iteration {i}: accuracy = {result}")

if __name__ == '__main__':
parser = argparse.ArgumentParser()
help='Directory for storing input data')
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)

# Iteration 10000: accuracy = 0.9943000078201294


Define the convolution layer with a 5x5 filter using RELU activation following by a 2x2 max pool:

with tf.variable_scope("CNN"):
cnn_W1 = tf.get_variable("W1", [5, 5, 1, 32], initializer=tf.truncated_normal_initializer(stddev=0.1))
cnn_b1 = tf.get_variable("b1", [32], initializer=tf.constant_initializer(0.1))
cnn_W2 = tf.get_variable("W2", [5, 5, 32, 64], initializer=tf.truncated_normal_initializer(stddev=0.1))
cnn_b2 = tf.get_variable("b2", [64], initializer=tf.constant_initializer(0.1))


We can add scope to a variable by _ tf.variable_scope_. Here, cnn_W1 will have the name ‘CNN/W1:0’.

# First CNN with RELU and max pooling.
x_image = tf.reshape(x, [-1, 28, 28, 1])
cnn1 = tf.nn.conv2d(x_image, cnn_W1, strides=[1, 1, 1, 1], padding='SAME')
z1 = tf.nn.relu(cnn1 + cnn_b1)
h1 = tf.nn.max_pool(z1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')


We flatten the 2D features into a 1D array for the fully connected layer. We apply dropout for the regularization.

# First FC layer with dropout.
h2_flat = tf.reshape(h2, [-1, 3136])
h_fc1 = tf.nn.relu(tf.matmul(h2_flat, W1) + b1)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)


Further possible accuracy improvement:

• Apply ensemble learning.
• Use a smaller filter like 3x3.
• Whitening of the input image.
• Further tuning of the learning rate and dropout parameter.

Here is another CNN implementation example from the TensorFlow distribution:

import argparse
import sys
import tempfile

from tensorflow.examples.tutorials.mnist import input_data

import tensorflow as tf

FLAGS = None

def deepnn(x):
with tf.name_scope('reshape'):
x_image = tf.reshape(x, [-1, 28, 28, 1])

with tf.name_scope('conv1'):
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)

with tf.name_scope('pool1'):
h_pool1 = max_pool_2x2(h_conv1)

with tf.name_scope('conv2'):
W_conv2 = weight_variable([5, 5, 32, 64])
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)

with tf.name_scope('pool2'):
h_pool2 = max_pool_2x2(h_conv2)

with tf.name_scope('fc1'):
W_fc1 = weight_variable([7 * 7 * 64, 1024])
b_fc1 = bias_variable([1024])

h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

with tf.name_scope('dropout'):
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

with tf.name_scope('fc2'):
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])

y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
return y_conv, keep_prob

def conv2d(x, W):
return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME')

def max_pool_2x2(x):
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],

def weight_variable(shape):
initial = tf.truncated_normal(shape, stddev=0.1)
return tf.Variable(initial)

def bias_variable(shape):
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial)

def main(_):

x = tf.placeholder(tf.float32, [None, 784])

y_ = tf.placeholder(tf.float32, [None, 10])

y_conv, keep_prob = deepnn(x)

with tf.name_scope('loss'):
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_,
logits=y_conv)
cross_entropy = tf.reduce_mean(cross_entropy)

with tf.name_scope('accuracy'):
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
correct_prediction = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(correct_prediction)

graph_location = tempfile.mkdtemp()
print('Saving graph to: %s' % graph_location)
train_writer = tf.summary.FileWriter(graph_location)

with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(20000):
batch = mnist.train.next_batch(50)
if i % 100 == 0:
train_accuracy = accuracy.eval(feed_dict={
x: batch[0], y_: batch[1], keep_prob: 1.0})
print('step %d, training accuracy %g' % (i, train_accuracy))
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

print('test accuracy %g' % accuracy.eval(feed_dict={
x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))

if __name__ == '__main__':
parser = argparse.ArgumentParser()
default='/tmp/tensorflow/mnist/input_data',
help='Directory for storing input data')
FLAGS, unparsed = parser.parse_known_args()
tf.app.run(main=main, argv=[sys.argv[0]] + unparsed)


### Reshape Numpy

Find the shape of a Numpy array and reshape it.

import tensorflow as tf
import numpy as np

### ndarray shape
x = np.array([[2, 3], [4, 5], [6, 7]])
print(x.shape)          # (3, 2)

x = x.reshape((2, 3))
print(x.shape)          # (2, 3)

x = x.reshape((-1))
print(x.shape)          # (6,)

x = x.reshape((6, -1))
print(x.shape)          # (6, 1)

x = x.reshape((-1, 6))
print(x.shape)          # (1, 6)


### Reshape TensorFlow

Find the shape of a tensor and reshape it

import tensorflow as tf
import numpy as np

### Tensor
W = tf.get_variable("W", [4, 5], initializer=tf.random_uniform_initializer(-1, 1))

print(W.get_shape())    # Get the shape of W (4, 5)

W = tf.reshape(W, [10, 2])
print(W.get_shape())    # (10, 2)

W = tf.reshape(W, [-1])
print(W.get_shape())    # (20,)

W = tf.reshape(W, [5, -1])
print(W.get_shape())    # (5, 4)


tf.unique(x) returns a 1D tensor contains all unique elements. The shape is dynamic which depends on “x” and need to evaluate at runtime:

import tensorflow as tf
import numpy as np

c = tf.constant([1, 2, 3, 1])
y, _ = tf.unique(c)     # y only contains the unique elements.

print(y.get_shape())    # (?,) This is a dynamic shape. Only know in runtime

y_shape = tf.shape(y)   # Define an op to get the dynamic shape.

init = tf.global_variables_initializer()

with tf.Session() as sess:
sess.run(init)
print(sess.run(y_shape))   # [3] contains 3 unique elements


### Initialize variables

Initialize variables with constant:

import tensorflow as tf
import numpy as np

v1 =  tf.get_variable("v1", [5, 5, 3])   # A tensor with shape (5, 5, 3) filled with random values

v2 = tf.get_variable("v2", shape=(), initializer=tf.zeros_initializer())

v3 = tf.get_variable("v3", initializer=tf.constant(2))    # 2, float32 scalar
v4 = tf.get_variable("v4", initializer=tf.constant([2]))  # [2]
v5 = tf.get_variable("v5", initializer=tf.constant([[2, 3], [4, 5]]))  # [[2, 3], [4, 5]]

v6 = tf.get_variable("v6", initializer=tf.constant(2.0), dtype=tf.float64, trainable=True)


Note: when we use tf.constant in tf.get_variable, we do not need to specify the tensor shape.

Fill with 0, 1 or specific values.

v1 = tf.get_variable("v1", [3, 2], initializer=tf.zeros_initializer)
v2 = tf.get_variable("v2", [3, 2], initializer=tf.ones_initializer)

# [[ 1.  2.], [ 3.  4.], [ 5.  6.]]
v3 = tf.get_variable("v3", [3, 2], initializer=tf.constant_initializer([1, 2, 3, 4, 5, 6]))

# [[ 1.  2.], [ 2.  2.], [ 2.  2.]]
v4 = tf.get_variable("v4", [3, 2], initializer=tf.constant_initializer([1, 2]))


Randomized the value of variables:

import tensorflow as tf
import numpy as np

W = tf.get_variable("W", [784, 256], initializer=tf.truncated_normal_initializer(stddev=np.sqrt(2.0 / 784)))
Z = tf.get_variable("z", [4, 5], initializer=tf.random_uniform_initializer(-1, 1))


### Slicing

subdata = data[:, 3]
subdata = data[:, 0:10]


### Utilities function

Concat and split

import tensorflow as tf

t1 = [[1, 2], [3, 4]]
t2 = [[5, 6], [7, 8]]
tf.concat([t1, t2], 0) # [[1, 2], [3, 4], [5, 6], [7, 8]]
tf.concat([t1, t2], 1) # [[1, 2, 5, 6], [3, 4, 7, 8]]

value = tf.get_variable("value", [4, 10], initializer=tf.zeros_initializer)

s1, s2, s3 = tf.split(value, [2, 3, 5], 1)
# s1 shape(4, 2)
# s2 shape(4, 3)
# s3 shape(4, 5)

# Split 'value' into 2 tensors along dimension 1
s0, s1= tf.split(value, num_or_size_splits=2, axis=1)  # s0 shape(4, 5)



Generate a one-hot vector

import tensorflow as tf

# Generate a one hot array using indexes
indexes = tf.get_variable("indexes", initializer=tf.constant([2, 0, -1, 0]))

target = tf.one_hot(indexes, 3, 2, 0)

init = tf.global_variables_initializer()

with tf.Session() as sess:
sess.run(init)
print(sess.run(target))
# [[0 0 2]
# [2 0 0]
# [0 0 0]
# [2 0 0]]


### Casting

s0 = tf.cast(s0, tf.int32)
s0 = tf.to_int64(s0)


During training, we may interest in the gradients for each variable. For example, from the gradients, we may tell how well the gradient descent is working for the deep network. To expose the gradient, replace the following code:

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


With:

global_step = tf.Variable(0)



import tempfile

import tensorflow as tf
import urllib.request
import numpy as np

FLAGS = None

tf.logging.set_verbosity(tf.logging.INFO)

if train_data:
train_file_name = train_data
else:
train_file = tempfile.NamedTemporaryFile(delete=False)
urllib.request.urlretrieve(
train_file.name)
train_file_name = train_file.name
train_file.close()
return train_file_name

training_local_file = ""

filename=training_local_file, target_dtype=np.int, features_dtype=np.float64)

print(f"data shape = {training_set.data.shape}")      # (3320, 7)
print(f"label shape = {training_set.target.shape}")   # (3320,)


### Evaluate & print a tensor

A quick way to evaluate a Tensor in particular for debugging.

m1 = tf.constant([[3, 5]])
m2 = tf.constant([[2],[4]])
product = tf.matmul(m1, m2)

with tf.Session() as sess:
v = product.eval()
t = tf.Print(v, [v])  # tf.Print return the first parameter
result = t + 1  # v will be printed only if t is accessed
result.eval()


### InteractiveSession

TensorFlow provides another way to execute a computational graph using tf.InteractiveSession which is more convenient for an ipython environment.

import tensorflow as tf
import numpy as np

sess = tf.InteractiveSession()

m1 = tf.get_variable("m1", initializer=tf.constant([[3, 5]]))
m2 = tf.placeholder(tf.int32, shape=(2, 1))
product = tf.matmul(m1, m2)

m1.initializer.run()   # Run the initialization op (and what it depends)

v1 = m1.eval()    # Evaluate a tensor
p = product.eval(feed_dict={m2: np.array([[1], [2]])}) # with feed

print(f"{v1}, {p}")

# Close the Session when we're done.
sess.close()


### Layers Module (Pre-built DNN layers)

Convolution

conv1 = tf.layers.conv2d(inputs=input_layer, filters=32, kernel_size=[5, 5], padding="same", activation=tf.nn.relu)


Maximum pool

pool1 = tf.layers.max_pooling2d(inputs=conv1, pool_size=[2, 2], strides=2)


Dense layer with ReLU

pool2_flat = tf.reshape(pool2, [-1, 7 * 7 * 64])
dense = tf.layers.dense(inputs=pool2_flat, units=1024, activation=tf.nn.relu)


Dropout

dropout = tf.layers.dropout(inputs=dense, rate=0.4, training=mode == tf.estimator.ModeKeys.TRAIN)


One hot

onehot_labels = tf.one_hot(indices=tf.cast(labels, tf.int32), depth=10)