Example 4

This project is for deep MNIST for experts. The second implementation is the MLP, which contains construct several layers(conv, pooling etc...)


In [74]:
from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf

In [75]:
mnist = input_data.read_data_sets('MNIST_data', one_hot = True)


Extracting MNIST_data/train-images-idx3-ubyte.gz
Extracting MNIST_data/train-labels-idx1-ubyte.gz
Extracting MNIST_data/t10k-images-idx3-ubyte.gz
Extracting MNIST_data/t10k-labels-idx1-ubyte.gz

In [76]:
################## build a softmax regression model

# input data
x = tf.placeholder(tf.float32, shape = [None, 784])

# real label
y_ = tf.placeholder(tf.float32, shape=[None, 10])

Build a Multilayer Convolutional Network

This section will help to build more complex model thant the previous linear model(with softmax classifier).

Weight Initialization

To create this model, we're going to need to create a lot of weights and biases. One should generally initialize weights with a small amount of noise for symmetry breaking, and to prevent 0 gradients. Since we're using ReLU neurons, it is also good practice to initialize them with a slightly positive initial bias to avoid "dead neurons". Instead of doing this repeatedly while we build the model, let's create two handy functions to do it for us.


In [77]:
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

TensorFlow also gives us a lot of flexibility in convolution and pooling operations. How do we handle the boundaries? What is our stride size? In this example, we're always going to choose the vanilla version. Our convolutions uses a stride of one and are zero padded so that the output is the same size as the input. Our pooling is plain old max pooling over 2x2 blocks. To keep our code cleaner, let's also abstract those operations into functions.


In [78]:
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

We can now implement our first layer. It will consist of convolution, followed by max pooling. The convolution will compute 32 features for each 5x5 patch. Its weight tensor will have a shape of [5, 5, 1, 32]. The first two dimensions are the patch size, the next is the number of input channels, and the last is the number of output channels. We will also have a bias vector with a component for each output channel.


In [79]:
W_conv1 = weight_variable([5, 5, 1, 32])
b_conv1 = bias_variable([32])

To apply the layer, we first reshape x to a 4d tensor, with the second and third dimensions corresponding to image width and height, and the final dimension corresponding to the number of color channels.


In [80]:
x_image = tf.reshape(x, [-1, 28, 28, 1])

We then convolve x_image with the weight tensor, add the bias, apply the ReLU function, and finally max pool. The max_pool_2x2 method will reduce the image size to 14x14.


In [81]:
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)
h_pool1 = max_pool_2x2(h_conv1)

Second Convolutional Layer

In order to build a deep network, we stack several layers of this type. The second layer will have 64 features for each 5x5 patch.


In [82]:
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

Now that the image size has been reduced to 7x7, we add a fully-connected layer with 1024 neurons to allow processing on the entire image. We reshape the tensor from the pooling layer into a batch of vectors, multiply by a weight matrix, add a bias, and apply a ReLU.


In [83]:
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

To reduce overfitting, we will apply dropout before the readout layer. We create a placeholder for the probability that a neuron's output is kept during dropout. This allows us to turn dropout on during training, and turn it off during testing. TensorFlow's tf.nn.dropout op automatically handles scaling neuron outputs in addition to masking them, so dropout just works without any additional scaling.


In [84]:
keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

Readout Layer

Finally, we add a layer, just like for the one layer softmax regression above.


In [85]:
W_fc2 = weight_variable([1024, 10])
b_fc2 = bias_variable([10])

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

Train and Evaluate the Model

How well does this model do? To train and evaluate it we will use code that is nearly identical to that for the simple one layer SoftMax network above.

The differences are that:

  • We will replace the steepest gradient descent optimizer with the more sophisticated ADAM optimizer.
  • We will include the additional parameter keep_prob in feed_dict to control the dropout rate.
  • We will add logging to every 100th iteration in the training process.

Feel free to go ahead and run this code, but it does 20,000 training iterations and may take a while (possibly up to half an hour), depending on your processor.


In [86]:
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_, logits=y_conv))

#train_step = tf.train.GradientDescentOptimizer(0.5).minimize(cross_entropy)
train_step = tf.train.AdamOptimizer(1e-3).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))

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for i in range(20000):
        batch = mnist.train.next_batch(50)
        if i%100 == 0:
            train_accuracy = accuracy.eval(feed_dict={x:batch[0], y_: batch[1], keep_prob: 1.0})
            print("step %d, training accuracy %g"%(i, train_accuracy))
            train_step.run(feed_dict={x: batch[0], y_: batch[1], keep_prob: 0.5})

    print("test accuracy %g"% accuracy.eval(feed_dict={x: mnist.test.images, y_: mnist.test.labels, keep_prob: 1.0}))


step 0, training accuracy 0.02
step 100, training accuracy 0.06
step 200, training accuracy 0.2
step 300, training accuracy 0.22
step 400, training accuracy 0.3
step 500, training accuracy 0.26
step 600, training accuracy 0.56
step 700, training accuracy 0.34
step 800, training accuracy 0.48
step 900, training accuracy 0.54
step 1000, training accuracy 0.56
step 1100, training accuracy 0.54
step 1200, training accuracy 0.48
step 1300, training accuracy 0.46
step 1400, training accuracy 0.6
step 1500, training accuracy 0.56
step 1600, training accuracy 0.56
step 1700, training accuracy 0.7
step 1800, training accuracy 0.64
step 1900, training accuracy 0.72
step 2000, training accuracy 0.76
step 2100, training accuracy 0.84
step 2200, training accuracy 0.76
step 2300, training accuracy 0.76
step 2400, training accuracy 0.78
step 2500, training accuracy 0.78
step 2600, training accuracy 0.8
step 2700, training accuracy 0.86
step 2800, training accuracy 0.8
step 2900, training accuracy 0.82
step 3000, training accuracy 0.84
step 3100, training accuracy 0.8
step 3200, training accuracy 0.78
step 3300, training accuracy 0.88
step 3400, training accuracy 0.86
step 3500, training accuracy 0.88
step 3600, training accuracy 0.88
step 3700, training accuracy 0.88
step 3800, training accuracy 0.82
step 3900, training accuracy 0.78
step 4000, training accuracy 0.9
step 4100, training accuracy 0.94
step 4200, training accuracy 0.84
step 4300, training accuracy 0.92
step 4400, training accuracy 0.86
step 4500, training accuracy 0.9
step 4600, training accuracy 0.96
step 4700, training accuracy 0.88
step 4800, training accuracy 0.8
step 4900, training accuracy 0.8
step 5000, training accuracy 0.94
step 5100, training accuracy 0.94
step 5200, training accuracy 0.9
step 5300, training accuracy 0.94
step 5400, training accuracy 0.9
step 5500, training accuracy 0.94
step 5600, training accuracy 0.94
step 5700, training accuracy 0.9
step 5800, training accuracy 0.92
step 5900, training accuracy 0.9
step 6000, training accuracy 0.96
step 6100, training accuracy 0.88
step 6200, training accuracy 0.9
step 6300, training accuracy 0.98
step 6400, training accuracy 0.9
step 6500, training accuracy 0.92
step 6600, training accuracy 0.92
step 6700, training accuracy 0.92
step 6800, training accuracy 0.9
step 6900, training accuracy 0.96
step 7000, training accuracy 0.88
step 7100, training accuracy 0.96
step 7200, training accuracy 0.92
step 7300, training accuracy 0.96
step 7400, training accuracy 0.94
step 7500, training accuracy 0.84
step 7600, training accuracy 0.98
step 7700, training accuracy 0.94
step 7800, training accuracy 0.92
step 7900, training accuracy 0.96
step 8000, training accuracy 0.88
step 8100, training accuracy 0.94
step 8200, training accuracy 0.96
step 8300, training accuracy 0.86
step 8400, training accuracy 0.9
step 8500, training accuracy 0.94
step 8600, training accuracy 0.86
step 8700, training accuracy 0.98
step 8800, training accuracy 0.96
step 8900, training accuracy 0.92
step 9000, training accuracy 0.96
step 9100, training accuracy 0.98
step 9200, training accuracy 0.94
step 9300, training accuracy 0.9
step 9400, training accuracy 0.92
step 9500, training accuracy 0.88
step 9600, training accuracy 0.94
step 9700, training accuracy 0.94
step 9800, training accuracy 0.96
step 9900, training accuracy 0.9
step 10000, training accuracy 0.96
step 10100, training accuracy 0.88
step 10200, training accuracy 0.92
step 10300, training accuracy 0.94
step 10400, training accuracy 0.98
step 10500, training accuracy 0.94
step 10600, training accuracy 0.92
step 10700, training accuracy 0.94
step 10800, training accuracy 0.86
step 10900, training accuracy 0.96
step 11000, training accuracy 0.88
step 11100, training accuracy 0.9
step 11200, training accuracy 0.92
step 11300, training accuracy 0.98
step 11400, training accuracy 0.98
step 11500, training accuracy 0.9
step 11600, training accuracy 0.98
step 11700, training accuracy 0.94
step 11800, training accuracy 0.98
step 11900, training accuracy 1
step 12000, training accuracy 0.9
step 12100, training accuracy 0.96
step 12200, training accuracy 0.96
step 12300, training accuracy 0.96
step 12400, training accuracy 0.94
step 12500, training accuracy 0.94
step 12600, training accuracy 0.98
step 12700, training accuracy 1
step 12800, training accuracy 0.88
step 12900, training accuracy 0.98
step 13000, training accuracy 0.9
step 13100, training accuracy 0.92
step 13200, training accuracy 0.98
step 13300, training accuracy 0.94
step 13400, training accuracy 0.98
step 13500, training accuracy 0.96
step 13600, training accuracy 0.98
step 13700, training accuracy 0.94
step 13800, training accuracy 0.94
step 13900, training accuracy 0.96
step 14000, training accuracy 0.9
step 14100, training accuracy 0.98
step 14200, training accuracy 0.96
step 14300, training accuracy 0.98
step 14400, training accuracy 0.92
step 14500, training accuracy 0.94
step 14600, training accuracy 0.98
step 14700, training accuracy 0.88
step 14800, training accuracy 0.94
step 14900, training accuracy 0.94
step 15000, training accuracy 0.94
step 15100, training accuracy 0.94
step 15200, training accuracy 0.94
step 15300, training accuracy 0.94
step 15400, training accuracy 0.92
step 15500, training accuracy 0.96
step 15600, training accuracy 0.98
step 15700, training accuracy 0.94
step 15800, training accuracy 0.94
step 15900, training accuracy 0.94
step 16000, training accuracy 0.94
step 16100, training accuracy 0.9
step 16200, training accuracy 0.96
step 16300, training accuracy 0.9
step 16400, training accuracy 0.96
step 16500, training accuracy 0.9
step 16600, training accuracy 0.96
step 16700, training accuracy 1
step 16800, training accuracy 0.96
step 16900, training accuracy 0.94
step 17000, training accuracy 0.96
step 17100, training accuracy 0.96
step 17200, training accuracy 0.96
step 17300, training accuracy 0.98
step 17400, training accuracy 0.94
step 17500, training accuracy 0.86
step 17600, training accuracy 0.94
step 17700, training accuracy 0.98
step 17800, training accuracy 0.98
step 17900, training accuracy 0.92
step 18000, training accuracy 0.94
step 18100, training accuracy 0.94
step 18200, training accuracy 1
step 18300, training accuracy 0.98
step 18400, training accuracy 0.96
step 18500, training accuracy 0.98
step 18600, training accuracy 0.98
step 18700, training accuracy 0.94
step 18800, training accuracy 0.96
step 18900, training accuracy 0.92
step 19000, training accuracy 0.94
step 19100, training accuracy 0.98
step 19200, training accuracy 0.94
step 19300, training accuracy 0.96
step 19400, training accuracy 0.9
step 19500, training accuracy 0.98
step 19600, training accuracy 0.94
step 19700, training accuracy 0.96
step 19800, training accuracy 0.92
step 19900, training accuracy 0.98
test accuracy 0.9609

In [ ]:


In [ ]: