Implementing the TensorFlow MNIST example without feed_dict

cnn_mnist_queue.py

# coding: utf-8

# In[1]:

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


# In[9]:

mnist = input_data.read_data_sets('MNIST_data', one_hot=True)

BATCH_SIZE = 100


# In[10]:

#initialize weights
def weight_variable(name, shape):
    initial = tf.get_variable(name, shape=shape, initializer=tf.truncated_normal_initializer(stddev=0.1))
    return tf.Variable(initial)

#initialize neurons w/ small bias
def bias_variable(name, shape):
    initial = tf.get_variable(name, shape=shape, initializer=tf.constant_initializer(0.1))
    return tf.Variable(initial)


# In[11]:

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

#max pooling
def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1],
                          strides=[1, 2, 2, 1], padding='SAME')


# In[12]:

def input_pipeline(batch_size, test=False):
    
    if test:
        inputs, labels = mnist.test.images, mnist.test.labels
        #batch_size = len(mnist.test.images)
    else:
        inputs, labels = mnist.train.images, mnist.train.labels
    # min_after_dequeue defines how big a buffer we will randomly sample
    #   from -- bigger means better shuffling but slower start up and more
    #   memory used.
    # capacity must be larger than min_after_dequeue and the amount larger
    #   determines the maximum we will prefetch.  Recommendation:
    #   min_after_dequeue + (num_threads + a small safety margin) * batch_size
    min_after_dequeue = 1000
    capacity = min_after_dequeue + 3 * batch_size
    
    if test:
        example_batch, label_batch = tf.train.batch(
            [inputs, labels], batch_size=batch_size, capacity=capacity,
            enqueue_many=True)
    else:
        example_batch, label_batch = tf.train.shuffle_batch(
            [inputs, labels], batch_size=batch_size, num_threads=3, capacity=capacity,
            min_after_dequeue=min_after_dequeue, enqueue_many=True)
    return example_batch, label_batch
    


# In[13]:

with tf.variable_scope("queue"):
    #initialize batches (uses RandomShuffleQueue under the hood)
    batch, labels = input_pipeline(BATCH_SIZE)
    input_, y_true = batch, tf.cast(labels, tf.float32)


# In[14]:

with tf.variable_scope("layers"):
    
    #first convolutional layer
    #W_conv1 = weight_variable('W_conv1', [5, 5, 1, 32])
    W_conv1 = tf.get_variable('W_conv1', shape=[5, 5, 1, 32],
                              initializer=tf.truncated_normal_initializer(stddev=0.1))
    #b_conv1 = bias_variable('b_conv1', [32])
    b_conv1 = tf.get_variable('b_conv1', shape=[32],
                              initializer=tf.constant_initializer(0.1))
    
    x_image = tf.reshape(input_, [-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
    #W_conv2 = weight_variable('W_conv2', [5, 5, 32, 64])
    W_conv2 = tf.get_variable('W_conv2', shape=[5, 5, 32, 64],
                              initializer=tf.truncated_normal_initializer(stddev=0.1))
    
    #b_conv2 = bias_variable('b_conv2',[64])
    b_conv2 = tf.get_variable('b_conv2', shape=[64],
                              initializer=tf.constant_initializer(0.1))
    
    h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)
    h_pool2 = max_pool_2x2(h_conv2)
    
    #fully connected layer
    #W_fc1 = weight_variable('W_fc1', [7 * 7 * 64, 1024])
    W_fc1 = tf.get_variable('W_fc1', shape=[7 * 7 * 64, 1024], 
                            initializer=tf.truncated_normal_initializer(stddev=0.1))
    #b_fc1 = bias_variable('b_fc1',[1024])
    b_fc1 = tf.get_variable('b_fc1', shape=[1024],
                            initializer=tf.constant_initializer(0.1))
    
    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: this is where the magic happens
    
    this is the stock version
    """
    #holds probability that neuron's output is kept during dropout
    keep_prob = 0.5
    
    #TensorFlow's stock dropout function
    h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
    
    #readout layer
    #W_fc2 = weight_variable('W_fc2', [1024, 10])
    W_fc2 = tf.get_variable('W_fc2', shape=[1024,10],
                            initializer=tf.truncated_normal_initializer(stddev=0.1))
    #b_fc2 = bias_variable('b_fc2', [10])
    b_fc2 = tf.get_variable('b_fc2', shape=[10],
                            initializer=tf.constant_initializer(0.1))

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


# In[15]:

with tf.variable_scope("loss"):
    loss = tf.nn.softmax_cross_entropy_with_logits(logits=y_conv, labels=y_true)
    loss_op = tf.reduce_mean(loss)


# In[16]:

with tf.variable_scope("accuracy"):
    correct_pred = tf.cast(tf.equal(tf.argmax(y_conv,1), tf.argmax(y_true,1)), tf.float32)
    accuracy = tf.reduce_mean(correct_pred)
    #acc = tf.metrics.accuracy(y_true, y_conv)
    accuracy_ = tf.Print(accuracy, data=[accuracy], message="accuracy: ")
    #accuracy_curr = tf.Print(acc, data=[acc], message="accuracy: ")


# In[17]:

train_op = tf.train.AdamOptimizer(1e-3).minimize(loss_op, name="train_op")


# In[18]:

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    
    #print(labels.shape)
    
    #"""
    
    coord = tf.train.Coordinator()
    threads = tf.train.start_queue_runners(sess=sess, coord=coord)
    
    
    sess.run(accuracy_)
    
    batch_old, labels_old = batch, labels
    
    for i in range(20000):
        
        #_, loss = sess.run(train_op)
        _, loss, acc = sess.run([train_op, loss_op, accuracy])
        

        # We regularly check the loss
        if i % 100 == 0:
            print('iter:%d - loss:%f, acc:%f' % (i, loss, acc))
            sess.run(accuracy_)
            #train_accuracy = accuracy.eval([train_op, loss_op])
            #print("step %d: %.4g" % (i, train_accuracy))

    sess.run(accuracy_)
    
    
    #run test set
    input_, y_true = input_pipeline(len(mnist.test.images), test=True)
    
    acc = sess.run(accuracy)
    
    print("Test accuracy: " + str(acc))

    coord.request_stop()
    coord.join(threads)
    
    #"""