˹A7̳ >> Tensorflowѧϰʼ_̳ >> 5.

5.4TensorflowʵCNNMNIST

 

 tensorflow һ CNN Ĵ룬
Ҫ4layerֱ:

convolutional layer1 + max pooling;

һҪpooling

convolutional layer2 + max pooling;

fully connected layer1 + dropout;

fully connected layer2 to prediction.

import tensorflow as tf

from tensorflow.examples.tutorials.mnist import input_data# number 1 to 10 data

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

def compute_accuracy(v_xs, v_ys):

    global prediction

    y_pre = sess.run(prediction, feed_dict={xs: v_xs, keep_prob: 1})

    correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(v_ys,1))

    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))

    result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys, keep_prob: 1})

    return result

#  normal ֲ#  shape ͿԷweightbiası

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)

# 2ά convolutional ͼ

def conv2d(x, W):

    # stride [1, x_movement, y_movement, 1]

    # Must have strides[0] = strides[3] = 1

    # strides ǿ󲽳ȡϢ

    return tf.nn.conv2d(x, W, strides=[1111], padding='SAME')        

#  pooling ͼ

def max_pool_2x2(x):

    # stride [1, x_movement, y_movement, 1]

    # poolingԸ粽ʧϢ

    return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME')        

# define placeholder for inputs to network

xs = tf.placeholder(tf.float32, [None784])         # 78428x28

ys = tf.placeholder(tf.float32, [None10])

keep_prob = tf.placeholder(tf.float32)

x_image = tf.reshape(xs, [-128281])            # һ1ʾǺڰ׵# print(x_image.shape)  # [n_samples, 28,28,1]

## 1. conv1 layer ###  x_imageĺ1Ӻ32

W_conv1 = weight_variable([55132])                 # patch 5x5, in size 1, out size 32

b_conv1 = bias_variable([32])# һconvolutional㣬ټһԻĴrelu

h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)             # output size 28x28x32# pooling󣬳СΪ14x14

h_pool1 = max_pool_2x2(h_conv1)                                     # output size 14x14x32

## 2. conv2 layer ### Ѻ32Ӻ64

W_conv2 = weight_variable([5,53264])                 # patch 5x5, in size 32, out size 64

b_conv2 = bias_variable([64])# ڶconvolutional

h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)             # output size 14x14x64# pooling󣬳СΪ7x7

h_pool2 = max_pool_2x2(h_conv2)                                     # output size 7x7x64

## 3. func1 layer ### ɵĸ߱1024

W_fc1 = weight_variable([7*7*641024])

b_fc1 = bias_variable([1024])# [n_samples, 7, 7, 64] ->> [n_samples, 7*7*64]# poolingĽƽ

h_pool2_flat = tf.reshape(h_pool2, [-17*7*64])

h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)

h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

## 4. func2 layer ### һ㣬1024size 10 softmax ʽзĴ

W_fc2 = weight_variable([102410])

b_fc2 = bias_variable([10])

prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)

 

# the error between prediction and real data

cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction),

                                              reduction_indices=[1]))       # loss

train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

 

sess = tf.Session()# important step

sess.run(tf.initialize_all_variables())

for i in range(1000):

    batch_xs, batch_ys = mnist.train.next_batch(100)

    sess.run(train_step, feed_dict={xs: batch_xs, ys: batch_ys, keep_prob: 0.5})

    if i % 50 == 0:

        print(compute_accuracy(

            mnist.test.images, mnist.test.labels))