This Notebook implements the TensorFlow advanced tutorial which uses a Multilayer Convolutional Network on the MNIST dataset
In [1]:
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import matplotlib.pyplot as plt
import time
Import MNIST Data
In [2]:
mnist = input_data.read_data_sets("../datasets/MNIST/", one_hot=True)
Look at sizes of training, validation and test sets Each image is 28 X 28 pixels Labels are in one hot encoding for use with softmax
In [42]:
print(mnist.train.num_examples)
print(mnist.validation.num_examples)
print(mnist.test.num_examples)
plt.imshow(mnist.train.images[10004].reshape(28,28),cmap="Greys")
plt.show()
print (mnist.train.labels[10004])
Declare Variables
In [4]:
x = tf.placeholder(tf.float32, [None, 784])
y_ = tf.placeholder(tf.float32, [None, 10])
Weight Initialization for ReLU nodes
In [5]:
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)
Convolution and Pooling operations
In [6]:
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],
strides=[1, 2, 2, 1], padding='SAME')
First Convolutional Layer The convolution will compute 32 features for each 5x5 patch. The max_pool_2x2 method will reduce the image size to 14x14.
In [7]:
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])
x_image = tf.reshape(x, [-1, 28, 28, 1])
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)
Second Convolutional Layer The second layer will have 64 features for each 5x5 patch.
In [8]:
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)
h_pool2 = max_pool_2x2(h_conv2)
Densely Connected Layer The image size has been reduced to 7x7 and we add a fully-connected layer with 1024 neurons to allow processing on the entire image.
In [9]:
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)
Dropout
In [10]:
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
Readout Layer
In [11]:
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])
y_conv = tf.matmul(h_fc1_drop, W_fc2) + b_fc2
Implement Model
In [12]:
cross_entropy = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
Define function that runs the model for given number of batches and returns the training time and accuracy on the validations and test data sets.
In [36]:
def train_and_test_model(batches,batches_per_epoch,verbose=False):
start = time.time()
epoch = 1
results = []
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(batches):
batch = mnist.train.next_batch(50)
if i % 100 == 0 and verbose:
train_accuracy = accuracy.eval(feed_dict={
x: batch[0], y_: batch[1], keep_prob: 1.0})
print('step %d, training accuracy %g, elapsedtime %g' % (i, train_accuracy, time.time() - start))
train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})
if (i+1) % batches_per_epoch == 0:
test_accuracy = accuracy.eval(feed_dict={
x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0})
validation_accuracy = accuracy.eval(feed_dict={
x: mnist.validation.images, y_: mnist.validation.labels, keep_prob: 1.0})
if verbose:
print('Done with test/val accuracy elapsed time %g' % (time.time() - start))
train_accuracy = accuracy.eval(feed_dict={
x: mnist.train.images[0:10000], y_: mnist.train.labels[0:10000], keep_prob: 1.0})
if verbose:
print('Done with train accuracy elapsed time %g' % (time.time() - start))
time_elapsed = time.time() - start
if verbose:
print(epoch,i+1, time_elapsed, train_accuracy, validation_accuracy, test_accuracy)
results.append((epoch,time_elapsed, train_accuracy, validation_accuracy, test_accuracy))
epoch += 1
return results
Try different number of epochs
In [38]:
results=train_and_test_model(7700,1100,verbose=True)
for r in results:
print(r)
| Num Epochs | Train Time | Training Accuracy | Validation Accuracy | Test Accuracy |
|---|---|---|---|---|
| 1 | 236.2375 | 0.9658 | 0.9690 | 0.9679 |
| 2 | 524.6372 | 0.9800 | 0.9790 | 0.9756 |
| 3 | 782.5051 | 0.9869 | 0.9836 | 0.9820 |
| 4 | 1073.9373 | 0.9907 | 0.9860 | 0.9849 |
| 5 | 1317.5896 | 0.9933 | 0.9888 | 0.9878 |
| 6 | 1553.6760 | 0.9956 | 0.9904 | 0.9880 |
| 7 | 1781.1032 | 0.9965 | 0.9912 | 0.9896 |
The tutorial achieved 99.2% accuracy after 20000 batches, >18 epoch, in about 30 minutes. This notebook on my laptop, ran 7 epochs in about 30 minutes and achieved around 99% accuracy. We will use this as the benchmark to aim for the numpy based models.
In [ ]: