How to apply Drop Out in Tensorflow to improve the accuracy of neural network?

Question

Drop-Out is regularization techniques. And I want to apply it to notMNIST data to reduce over-fitting to finish my Udacity Deep Learning Course Assignment.I have read the docs of tensorflow on how to call the tf.nn.dropout. And here is my code

# before proceeding further.
from __future__ import print_function
import numpy as np  
import tensorflow as tf
from six.moves import cPickle as pickle


pickle_file = 'notMNIST.pickle'

with open(pickle_file, 'rb') as f:
    save = pickle.load(f)
    train_dataset = save['train_dataset']
    train_labels = save['train_labels']
    valid_dataset = save['valid_dataset']
    valid_labels = save['valid_labels']
    test_dataset = save['test_dataset']
    test_labels = save['test_labels']
    del save  # hint to help gc free up memory
    print('Training set', train_dataset.shape, train_labels.shape)
    print('Validation set', valid_dataset.shape, valid_labels.shape)
    print('Test set', test_dataset.shape, test_labels.shape)


image_size = 28
num_labels = 10

def reformat(dataset, labels):
    dataset = dataset.reshape((-1, image_size * image_size)).astype(np.float32)
    # Map 1 to [0.0, 1.0, 0.0 ...], 2 to [0.0, 0.0, 1.0 ...]
    labels = (np.arange(num_labels) == labels[:,None]).astype(np.float32)
    return dataset, labels

    train_dataset, train_labels = reformat(train_dataset, train_labels)
    valid_dataset, valid_labels = reformat(valid_dataset, valid_labels)
    test_dataset, test_labels = reformat(test_dataset, test_labels)
    print('Training set', train_dataset.shape, train_labels.shape)
    print('Validation set', valid_dataset.shape, valid_labels.shape)
    print('Test set', test_dataset.shape, test_labels.shape)

    def accuracy(predictions, labels):
        return (100.0 * np.sum(np.argmax(predictions, 1) == np.argmax(labels, 1))  / predictions.shape[0])


# ReLU neuron
# param
training_epochs = 30
batch_size = 521
display_step = 1
n_input = 784 # img shape: 28*28
n_classes = 10 # MNIST total classes (0-9 digits)

# hyper-parameter
n_hidden_1 = 256 
learning_rate = 0.05
lambda_term = 0.01


graph = tf.Graph()
with graph.as_default():
    # init weights
    weights_hiden =  tf.Variable(tf.random_normal([n_input, n_hidden_1], stddev=np.sqrt(n_input)))
    weights_out = tf.Variable(tf.random_normal([n_hidden_1, n_classes], stddev=np.sqrt(n_hidden_1)))

    biases_hidden = tf.Variable(tf.random_normal([n_hidden_1]))
    biases_out = tf.Variable(tf.random_normal([n_classes]))

    x = tf.placeholder("float", [None, n_input])
    y = tf.placeholder("float", [None, n_classes])

    def model(x, weights_hiden, weights_out, biases_hidden, biases_out):
        # hidden layer with RELU activation
        layer_1 = tf.nn.relu(tf.add(tf.matmul(x, weights_hiden), biases_hidden))
        # apply DropOut to hidden layer
        keep_prob = tf.placeholder(tf.float32)  # DROP-OUT here
        drop_out = tf.nn.dropout(layer_1, keep_prob)  # DROP-OUT here
        # output layer with linear activation
        out_layer = tf.matmul(layer_1, weights_out) + biases_out
        return out_layer

    # Construct model
    pred = model(x, weights_hiden, weights_out, biases_hidden, biases_out)

    # Define loss and optimizer
    cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y) +
                          lambda_term * tf.nn.l2_loss(weights_hiden) + 
                          lambda_term * tf.nn.l2_loss(weights_out) +
                          lambda_term * tf.nn.l2_loss(biases_hidden) + 
                          lambda_term * tf.nn.l2_loss(biases_out))
    optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)


# run the graph
with tf.Session(graph=graph) as sess:
    tf.initialize_all_variables().run()
    print('Initialized')
    # Training cycle
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(train_dataset.shape[0]/batch_size)
        # Loop over all batches
        for i in range(total_batch):
            batch_x = train_dataset[(i*batch_size):((i*batch_size) + batch_size), :]
            batch_y = train_labels[(i*batch_size):((i*batch_size) + batch_size), :]
            # Run optimization op (backprop) and cost op (to get loss value)
            _, c = sess.run([optimizer, cost], feed_dict={x: batch_x, y: batch_y})
            # Compute average loss
            avg_cost += c / total_batch
        # Display logs per epoch step
        if epoch % display_step == 0:
            print("Epoch:", '%04d' % (epoch+1), "cost=", "{:.9f}".format(avg_cost))
    print("Optimization Finished!")

    # Test model
    correct_prediction = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
    # Calculate accuracy
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
    print("Test data accuracy:", accuracy.eval({x: test_dataset, y: test_labels}))
    print("Valid data accuracy:", accuracy.eval({x: valid_dataset, y: valid_labels}))

The tf.nn.dropout is called in function model(), but after I applied the DropOut technique to the neural network, the accuracy did seem any change, here is the result:

Epoch: 0001 cost= 579980.086977807
Epoch: 0002 cost= 238859.802382506
Epoch: 0003 cost= 90672.733752856
Epoch: 0004 cost= 32649.040985028
Epoch: 0005 cost= 11325.878361874
Epoch: 0006 cost= 3866.805511076
Epoch: 0007 cost= 1357.785540469
Epoch: 0008 cost= 519.381747333
Epoch: 0009 cost= 225.359804119
Epoch: 0010 cost= 110.099476707
Epoch: 0011 cost= 55.212384386
Epoch: 0012 cost= 28.469241683
Epoch: 0013 cost= 14.511494627
Epoch: 0014 cost= 6.567228943
Epoch: 0015 cost= 3.186372240
Epoch: 0016 cost= 1.701917576
Epoch: 0017 cost= 1.041632473
Epoch: 0018 cost= 0.843376874
Epoch: 0019 cost= 0.786183911
Epoch: 0020 cost= 0.775412846
Epoch: 0021 cost= 0.782965020
Epoch: 0022 cost= 0.796788171
Epoch: 0023 cost= 0.814522117
Epoch: 0024 cost= 0.832090579
Epoch: 0025 cost= 0.849197715
Epoch: 0026 cost= 0.867473578
Epoch: 0027 cost= 0.889561496
Epoch: 0028 cost= 0.921837020
Epoch: 0029 cost= 16.655304543
Epoch: 0030 cost= 1.421570476
Optimization Finished!
Test data accuracy: 0.8775
Valid data accuracy: 0.8069

How can I apply DropOut by Tensorflow to improve the accuracy of the network? Thank you!

Answer 1

In the graph, I'd suggest to move keep_prob = tf.placeholder(tf.float32) outside of the model function to make it global.

with graph.as_default():
    ...
    x = tf.placeholder("float", [None, n_input])
    y = tf.placeholder("float", [None, n_classes])
    keep_prob = tf.placeholder(tf.float32)

    def model(x, weights_hiden, weights_out, biases_hidden, biases_out, keep_prob):
        # hidden layer with RELU activation
        layer_1 = tf.nn.relu(tf.add(tf.matmul(x, weights_hiden), biases_hidden))
        # apply DropOut to hidden layer
        drop_out = tf.nn.dropout(layer_1, keep_prob)  # DROP-OUT here
        # output layer with linear activation
        out_layer = tf.matmul(drop_out, weights_out) + biases_out
        return out_layer
    ...

When running session, feed a desired keep_prob value during training time, and feed 1.0 to keep_prob during reference (validation and/or testing) time.

# run the graph
with tf.Session(graph=graph) as sess:
    tf.initialize_all_variables().run()
    ...
    for epoch in range(training_epochs):
        ...
        for i in range(total_batch):
            batch_x = ...
            batch_y = ...
            # Run optimization op (backprop) and cost op (to get loss value)
            # Feed a value < 1.0 for keep prob during training
            _, c = sess.run([optimizer, cost], feed_dict={x: batch_x, y: batch_y, keep_prob : 0.5})
    ...
    # Feed 1.0 for keep prob during testing
    print("Test data accuracy:", accuracy.eval({x: test_dataset, y: test_labels, keep_prob : 1.0}))
    print("Valid data accuracy:", accuracy.eval({x: valid_dataset, y: valid_labels, keep_prob : 1.0}))