tensorflow 1.x practical tutorial (9) - convolutional neural network

Target

This article aims to introduce the introductory knowledge points and practical examples of tensorflow. I hope that all novice students can be proficient in basic operations related to tensorflow after learning.

Simple Convolutional Neural Network Implementation

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

mnist = input_data.read_data_sets("MNIST", one_hot=True)
batch_size = 64
n_batches = mnist.train.num_examples // batch_size

def weight_variable(shape):
    return tf.Variable(tf.random_normal(shape, stddev=0.1))
def biases_variable(shape):
    return tf.Variable(tf.constant(0.1, shape=shape))
def conv2d(x, w):
    return tf.nn.conv2d(x, w, strides=[1,1,1,1], padding='SAME') # strides 表示步长，padding 表示填充方式
def max_pool_2x2(x):
    return tf.nn.max_pool(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding='SAME') # ksize 表示卷积核大小

x = tf.placeholder(shape=[None, 784], dtype=tf.float32)
y = tf.placeholder(shape=[None, 10], dtype=tf.float32)
x_ = tf.reshape(x, [-1, 28, 28, 1]) # 因为图片本身就是 28*28*1 的大小，28 是图片像素点的长和宽的个数，1 是每个像素点像素维度

w1 = weight_variable([5,5,1,32]) # 这个权重是靠公式可以算出来的
b1 = biases_variable([32]) # 保证和 w 的最后一个维度相同即可

h1 = tf.nn.relu(conv2d(x_, w1) + b1) # 使用 relu 激活函数
p1 = max_pool_2x2(h1) # 最大池化操作

w2 = weight_variable([5,5,32,64])
b2 = biases_variable([64])

h2 = tf.nn.relu(conv2d(p1, w2) + b2) # 使用 relu 激活函数
p2 = max_pool_2x2(h2) # 最大池化操作

w_fc1 = weight_variable([7*7*64, 1024]) # 全联接层操作
b_fc1 = biases_variable([1024]) # 保证和 w_fc1 最后一个维度相同

p2_flat = tf.reshape(p2, [-1, w_fc1.shape[0]])
h_fc1 = tf.nn.tanh(tf.matmul(p2_flat, w_fc1) + b_fc1)

keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)

w_fc2 = weight_variable([1024, 10]) # 输出层操作
b_fc2 = biases_variable([10]) # 保证和 w_fc2 最后一个维度相同
prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, w_fc2) + b_fc2)

loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))
opt = tf.train.AdamOptimizer(0.001).minimize(loss)
correct = tf.equal(tf.argmax(y,1), tf.argmax(prediction, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))

with tf.Session() as sess:
    sess.run(tf.global_variables_initializer())
    for epoch in range(10):
        for _ in range(n_batches):
            xx, yy = mnist.train.next_batch(batch_size)
            sess.run(opt, feed_dict={x:xx, y:yy, keep_prob:0.5})
        acc, l = sess.run([accuracy, loss], feed_dict={x:mnist.test.images, y:mnist.test.labels, keep_prob:1.0})
        print(epoch, l, acc)
        
复制代码

result output

0 1.5783486 0.8818
1 1.4810778 0.9803
2 1.4758472 0.9855
3 1.472993 0.9884
4 1.4741219 0.9866
5 1.4728734 0.9882
6 1.4742823 0.9869
7 1.4712367 0.9898
8 1.4690293 0.9922
9 1.473154 0.988
10 1.4709185 0.9904
复制代码

point one

For the basic knowledge of convolutional neural network, you can find it online. There are many learning materials, including convolution kernel, step size, padding, etc.

point two

Compared with the simple multi-hidden layer network + Dropout introduced in the previous article, the accuracy rate is only 97.8%. The accuracy rate of using the convolutional neural network can be easily increased to 99%, indicating that the convolutional neural network is naturally suitable for processing image data. However, as the complexity of the model increases, the resources and time spent on training also increase.

Reference in this article

Reference for this article: blog.csdn.net/qq_19672707…