Tensorflow is a open source library which is especially useful for building neural networks. In this notebook, we:

Build a neural network to recognize neural networks.

Visualize the architecture of the neural network using tensorboard, a feature of tensorflow.

First, though, we need to import the library itself and import our dataset:

In [27]:

```
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
```

In [31]:

```
def weight_variable(shape, name):
"""
Returns:
initial: a tensor of weights. This tensor is instantiated with shape 'shape' and its entries consist of
independent and identically distributed (iid) draws from a truncated normal distribution with mean 0 and standard
deviation 0.1
"""
# initial = tf.truncated_normal(shape, stddev=0.1)
initial = tf.random_uniform(shape,-0.1, 0.1)
return tf.Variable(initial, name = name)
def bias_variable(shape, name):
"""
Returns:
initial: a tensor of biases. This tensor is instantiated with shape 'shape' and its entries consist of
independent and identically distributed (iid) draws from a truncated normal distribution with mean 0 and standard
deviation 0.1
"""
initial = tf.constant(0.1, shape=shape)
return tf.Variable(initial, name = name)
def conv2d(x, W, name):
"""
Arguments:
x: a four-dimensional input tensor
W: a four-dimensional filter tensor to be used in computing convolution.
Returns:
convolution: a four-dimensional tensor, x convolved by W according to strides and padding.
(For details on this special convolution, see https://www.tensorflow.org/api_docs/python/tf/nn/conv2d)
"""
convolution = tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME', name = name)
return convolution
def max_pool_2x2(x, name):
"""
Arguments:
x: a four-dimensional input tensor
Returns:
maxpool: a four-dimensional tensor, each entry is the max element of the subtensor dictated by ksize,
strides, and padding.
(For details on this operation, see https://www.tensorflow.org/api_docs/python/tf/nn/max_pool)
"""
return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name = name)
```

Now, we will decide the topology of our neural net. We will have two convolutional layers, a fully connected layer, and an output layer:

In [32]:

```
#Clear away any old graphs
tf.reset_default_graph()
#Use an interactive session, so that we don't have to specify a session when we evaluate tensors from the computation
#graph.
sess = tf.InteractiveSession()
with tf.name_scope('Neural_Net') as scope:
#Placholders for input to neural net
x = tf.placeholder(tf.float32, shape=[None, 784])
y_ = tf.placeholder(tf.float32, shape=[None, 10])
with tf.name_scope('Layer_1'):
W_conv1 = weight_variable([5, 5, 1, 32], 'Weights')
b_conv1 = bias_variable([32], 'Biases')
x_image = tf.reshape(x, [-1,28,28,1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1, 'Convolution') + b_conv1)
h_pool1 = max_pool_2x2(h_conv1, 'Pool')
with tf.name_scope('Layer_2'):
W_conv2 = weight_variable([5, 5, 32, 64], 'Weights')
b_conv2 = bias_variable([64], 'Biases')
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2, 'Convolution') + b_conv2)
h_pool2 = max_pool_2x2(h_conv2, 'Pool')
with tf.name_scope('Fully_Connected_Layer'):
W_fc1 = weight_variable([7 * 7 * 64, 1024], 'Weights')
b_fc1 = bias_variable([1024], 'Biases')
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)
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
with tf.name_scope('Output_Layer'):
W_fc2 = weight_variable([1024, 10], 'Weights')
b_fc2 = bias_variable([10], 'Biases')
y_conv=tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2, name = 'Softmax')
with tf.name_scope('Cost') as scope:
cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_ * tf.log(y_conv), reduction_indices=[1]))
with tf.name_scope('Optimizer') as scope:
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
with tf.name_scope('Accuracy') as scope:
correct_prediction = tf.equal(tf.argmax(y_conv,1), tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
w1_hist = tf.summary.histogram('W_conv1', W_conv1)
w2_hist = tf.summary.histogram('W_conv2', W_conv2)
accuracy_sum = tf.summary.scalar('accuracy', accuracy)
cost_sum = tf.summary.scalar('cost', cross_entropy)
merged = tf.summary.merge_all()
writer = tf.summary.FileWriter('tf_logs', sess.graph)
```

In [33]:

```
sess.run(tf.global_variables_initializer())
for i in range(1500):
batch = mnist.train.next_batch(50)
if i%10 == 0:
summary, _, train_accuracy = sess.run([merged, train_step, accuracy], feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
writer.add_summary(summary, i)
else:
_ = sess.run([train_step], feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
if i%100 == 0:
print("step %d, training accuracy %g"%(i, train_accuracy))
```

In [34]:

```
writer.close()
```

Tensorboard yields the following visualizations of the graph, accuracy, loss/cost, and the weights:

(Main graph, a bit unravelled)