Multi Layer Perceptron MNIST

Load tensorflow library and MNIST data

import tensorflow as tf

# Import MNIST data
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)

print('Test shape:',mnist.test.images.shape)
print('Train shape:',mnist.train.images.shape)

Extracting /tmp/data/train-images-idx3-ubyte.gz
Extracting /tmp/data/train-labels-idx1-ubyte.gz
Extracting /tmp/data/t10k-images-idx3-ubyte.gz
Extracting /tmp/data/t10k-labels-idx1-ubyte.gz
Test shape: (10000, 784)
Train shape: (55000, 784)

Neural network parameters

# Parameters
learning_rate = 0.001
training_epochs = 15
batch_size = 100
display_step = 1

# Network Parameters
n_hidden_1 = 256 # 1st layer number of features
n_hidden_2 = 256 # 2nd layer number of features
n_input = 784 # MNIST data input (img shape: 28*28)
n_classes = 10 # MNIST total classes (0-9 digits)

Build graph

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

# Create model
def multilayer_perceptron(x, weights, biases):
    # Use tf.matmul instead of "*" because tf.matmul can change it's dimensions on the fly (broadcast)
    print( 'x:', x.get_shape(), 'W1:', weights['h1'].get_shape(), 'b1:', biases['b1'].get_shape())        
    # Hidden layer with RELU activation
    layer_1 = tf.add(tf.matmul(x, weights['h1']), biases['b1']) #(x*weights['h1']) + biases['b1']
    layer_1 = tf.nn.relu(layer_1)

    # Hidden layer with RELU activation
    print( 'layer_1:', layer_1.get_shape(), 'W2:', weights['h2'].get_shape(), 'b2:', biases['b2'].get_shape())        
    layer_2 = tf.add(tf.matmul(layer_1, weights['h2']), biases['b2']) # (layer_1 * weights['h2']) + biases['b2'] 
    layer_2 = tf.nn.relu(layer_2)

    # Output layer with linear activation
    print( 'layer_2:', layer_2.get_shape(), 'W3:', weights['out'].get_shape(), 'b3:', biases['out'].get_shape())        
    out_layer = tf.matmul(layer_2, weights['out']) + biases['out'] # (layer_2 * weights['out']) + biases['out']    
    print('out_layer:',out_layer.get_shape())

    return out_layer

Initialize weights and construct the model

# Store layers weight & bias
weights = {
    'h1': tf.Variable(tf.random_normal([n_input, n_hidden_1])),    #784x256
    'h2': tf.Variable(tf.random_normal([n_hidden_1, n_hidden_2])), #256x256
    'out': tf.Variable(tf.random_normal([n_hidden_2, n_classes]))  #256x10
}
biases = {
    'b1': tf.Variable(tf.random_normal([n_hidden_1])),             #256x1
    'b2': tf.Variable(tf.random_normal([n_hidden_2])),             #256x1
    'out': tf.Variable(tf.random_normal([n_classes]))              #10x1
}

# Construct model
pred = multilayer_perceptron(x, weights, biases)

x: (?, 784) W1: (784, 256) b1: (256,)
layer_1: (?, 256) W2: (256, 256) b2: (256,)
layer_2: (?, 256) W3: (256, 10) b3: (10,)
out_layer: (?, 10)

Define Loss function, and Optimizer

# Cross entropy loss function
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(pred, y))

# On this case we choose the AdamOptimizer
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)

Launch graph

# Initializing the variables
init = tf.initialize_all_variables()

# Launch the graph
with tf.Session() as sess:
    sess.run(init)

    # Training cycle
    for epoch in range(training_epochs):
        avg_cost = 0.
        total_batch = int(mnist.train.num_examples/batch_size)
        # Loop over all batches
        for i in range(total_batch):
            batch_x, batch_y = mnist.train.next_batch(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"))
    # To keep sizes compatible with model
    print ("Accuracy:", accuracy.eval({x: mnist.test.images, y: mnist.test.labels}))

Epoch: 0001 cost= 152.289635962
Epoch: 0002 cost= 39.134648348    
...
Epoch: 0015 cost= 0.850344581
Optimization Finished!
Accuracy: 0.9464