In this notebook, we convert our LeNet-5-inspired, MNIST-classifying, deep convolutional network from Keras to TensorFlow (compare them side by side) following Aymeric Damien's style.
In [1]:
import numpy as np
np.random.seed(42)
import tensorflow as tf
tf.set_random_seed(42)
In [2]:
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
In [3]:
epochs = 1
batch_size = 128
display_progress = 10 # after this many batches, output progress to screen
wt_init = tf.contrib.layers.xavier_initializer() # weight initializer
In [4]:
# input layer:
n_input = 784
# first convolutional layer:
n_conv_1 = 32
k_conv_1 = 3
# second convolutional layer:
n_conv_2 = 64
k_conv_2 = 3
# max pooling layer:
pool_size = 2
mp_layer_dropout = 0.25
# dense layer:
n_dense = 128
dense_layer_dropout = 0.5
# output layer:
n_classes = 10
In [5]:
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
In [7]:
# dense layer with ReLU activation:
def dense(x, W, b):
z = tf.add(tf.matmul(x, W), b)
a = tf.nn.relu(z)
return a
# convolutional layer with ReLU activation:
def conv2d(x, W, b, stride_length=1):
xW = tf.nn.conv2d(x, W, strides=[1, stride_length, stride_length, 1], padding='SAME')
z = tf.nn.bias_add(xW, b)
a = tf.nn.relu(z)
return a
# max-pooling layer:
def maxpooling2d(x, p_size):
return tf.nn.max_pool(x,
ksize=[1, p_size, p_size, 1],
strides=[1, p_size, p_size, 1],
padding='SAME'
)
In [8]:
bias_dict = {
'b_c1': tf.Variable(tf.zeros([n_conv_1])),
'b_c2': tf.Variable(tf.zeros([n_conv_2])),
'b_d1': tf.Variable(tf.zeros([n_dense])),
'b_out': tf.Variable(tf.zeros([n_classes]))
}
# calculate number of inputs to dense layer:
full_square_length = np.sqrt(n_input)
pooled_square_length = int(full_square_length / pool_size)
dense_inputs = pooled_square_length**2 * n_conv_2
weight_dict = {
'W_c1': tf.get_variable('W_c1',
[k_conv_1, k_conv_1, 1, n_conv_1], initializer=wt_init),
'W_c2': tf.get_variable('W_c2',
[k_conv_2, k_conv_2, n_conv_1, n_conv_2], initializer=wt_init),
'W_d1': tf.get_variable('W_d1',
[dense_inputs, n_dense], initializer=wt_init),
'W_out': tf.get_variable('W_out',
[n_dense, n_classes], initializer=wt_init)
}
In [9]:
def network(x, weights, biases, n_in, mp_psize, mp_dropout, dense_dropout):
# reshape linear MNIST pixel input into square image:
square_dimensions = int(np.sqrt(n_in))
square_x = tf.reshape(x, shape=[-1, square_dimensions, square_dimensions, 1])
# convolutional and max-pooling layers:
conv_1 = conv2d(square_x, weights['W_c1'], biases['b_c1'])
conv_2 = conv2d(conv_1, weights['W_c2'], biases['b_c2'])
pool_1 = maxpooling2d(conv_2, mp_psize)
pool_1 = tf.nn.dropout(pool_1, 1-mp_dropout)
# dense layer:
flat = tf.reshape(pool_1, [-1, weight_dict['W_d1'].get_shape().as_list()[0]])
dense_1 = dense(flat, weights['W_d1'], biases['b_d1'])
dense_1 = tf.nn.dropout(dense_1, 1-dense_dropout)
# output layer:
out_layer_z = tf.add(tf.matmul(dense_1, weights['W_out']), biases['b_out'])
return out_layer_z
In [10]:
predictions = network(x, weight_dict, bias_dict, n_input,
pool_size, mp_layer_dropout, dense_layer_dropout)
In [11]:
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predictions, labels=y))
optimizer = tf.train.AdamOptimizer().minimize(cost)
In [12]:
correct_prediction = tf.equal(tf.argmax(predictions, 1), tf.argmax(y, 1))
accuracy_pct = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) * 100
In [13]:
initializer_op = tf.global_variables_initializer()
In [ ]:
with tf.Session() as session:
session.run(initializer_op)
print("Training for", epochs, "epochs.")
# loop over epochs:
for epoch in range(epochs):
avg_cost = 0.0 # track cost to monitor performance during training
avg_accuracy_pct = 0.0
# loop over all batches of the epoch:
n_batches = int(mnist.train.num_examples / batch_size)
for i in range(n_batches):
# to reassure you something's happening!
if i % display_progress == 0:
print("Step ", i+1, " of ", n_batches, " in epoch ", epoch+1, ".", sep='')
batch_x, batch_y = mnist.train.next_batch(batch_size)
# feed batch data to run optimization and fetching cost and accuracy:
_, batch_cost, batch_acc = session.run([optimizer, cost, accuracy_pct],
feed_dict={x: batch_x, y: batch_y})
# accumulate mean loss and accuracy over epoch:
avg_cost += batch_cost / n_batches
avg_accuracy_pct += batch_acc / n_batches
# output logs at end of each epoch of training:
print("Epoch ", '%03d' % (epoch+1),
": cost = ", '{:.3f}'.format(avg_cost),
", accuracy = ", '{:.2f}'.format(avg_accuracy_pct), "%",
sep='')
print("Training Complete. Testing Model.\n")
test_cost = cost.eval({x: mnist.test.images, y: mnist.test.labels})
test_accuracy_pct = accuracy_pct.eval({x: mnist.test.images, y: mnist.test.labels})
print("Test Cost:", '{:.3f}'.format(test_cost))
print("Test Accuracy: ", '{:.2f}'.format(test_accuracy_pct), "%", sep='')
In [ ]:
In [ ]:
In [ ]: