简单的卷积神经网络

一、建立一个简单的卷积神经网络

如上图所示，一个简单的卷积神经网络由卷积层，池化层，**层，全连接层组成。设计一个好的卷积神经网络不是简单的事情，包括卷积核大小的选择，比如1，3，5，7等几个常用的大小，还有卷积核的个数的选择，比如1，2，4，16，32等等，还有就是全连接层的**函数的选择等。下面我们就来简单看看卷积核个数和**函数的选择。

二、卷积核个数的选择

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import matplotlib.pyplot as plt
import csv

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)


def weight(shape):
    init = tf.random_normal(shape=shape, stddev=0.1)
    return tf.Variable(init)


def baise(shape):
    init = tf.random_normal(shape=shape, stddev=0.00001)
    return tf.Variable(init)


def conv(x, w, k, b):
    return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w, strides=[1, k, k, 1], padding="VALID"), b))


def max_pool(x, k):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding="VALID")


X = tf.placeholder(tf.float32, [None, 784])
Y = tf.placeholder(tf.float32, [None, 10])
X_input = tf.reshape(X, [-1, 28, 28, 1])
w1 = weight([5, 5, 1,16])
b1 = baise([16])
con1 = conv(X_input, w1, 1, b1)
pool1 = max_pool(con1, 2)
pool1 = tf.reshape(pool1, [-1, 12 * 12 * 16])
wc1 = weight([12 * 12 * 16, 10])
bc1 = baise([10])
fc1 = tf.matmul(pool1, wc1) + bc1
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=Y, logits=fc1))
correct_pred = tf.equal(tf.argmax(Y, 1), tf.argmax(fc1, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
step = tf.train.AdamOptimizer(0.01).minimize(loss)
init = tf.global_variables_initializer()
result_list = list()
result_list.append(["train_step", 'train_loss', "train_accuracy"])
with tf.Session() as sess:
    sess.run(init)
    costs = []
    for i in range(10):
        batch_x, batch_y = mnist.train.next_batch(64)
        cost, right, _ = sess.run([loss, accuracy, step], feed_dict={X: batch_x, Y: batch_y})
        costs.append(cost)
        print("第", i, "次迭代的损失:", cost, "accuracy:", '{:.2%}'.format(right))
        result_list.append([i, cost, right])
    result_file = open('result.csv', 'w', newline='')
    csv_write = csv.writer(result_file, dialect='excel')
    for row in result_list:
        csv_write.writerow(row)

    print("validation accuracy:",
          '{:.2%}'.format(accuracy.eval({X: mnist.validation.images, Y: mnist.validation.labels})))

利用MNIST数据集做了几个简单的实验，分别把卷积核的个数设置为1，2，4，8，16，32个，在选择mnist_batch为64，循环次数为10次的条件下得出如上数据，随着核数量的增加，训练集的准确率逐渐提高。随着核个数的增加，提取的特征也就越多，训练速度下降但准确率不断提高。

三、全连接层**函数的选择

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf
import matplotlib.pyplot as plt
import csv

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)


def weight(shape):
    init = tf.random_normal(shape=shape, stddev=0.1)
    return tf.Variable(init)


def baise(shape):
    init = tf.random_normal(shape=shape, stddev=0.00001)
    return tf.Variable(init)


def conv(x, w, k, b):
    return tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(x, w, strides=[1, k, k, 1], padding="VALID"), b))


def max_pool(x, k):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding="VALID")


X = tf.placeholder(tf.float32, [None, 784])
Y = tf.placeholder(tf.float32, [None, 10])
X_input = tf.reshape(X, [-1, 28, 28, 1])
w1 = weight([5, 5, 1,16])
b1 = baise([16])
con1 = conv(X_input, w1, 1, b1)
pool1 = max_pool(con1, 2)
pool1 = tf.reshape(pool1, [-1, 12 * 12 * 16])
wc1 = weight([12 * 12 * 16, 10])
bc1 = baise([10])
fc1 = tf.nn.elu(tf.matmul(pool1, wc1) + bc1)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=Y, logits=fc1))
correct_pred = tf.equal(tf.argmax(Y, 1), tf.argmax(fc1, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
step = tf.train.AdamOptimizer(0.01).minimize(loss)
init = tf.global_variables_initializer()
result_list = list()
result_list.append(["train_step", 'train_loss', "train_accuracy"])
with tf.Session() as sess:
    sess.run(init)
    costs = []
    for i in range(10):
        batch_x, batch_y = mnist.train.next_batch(64)
        cost, right, _ = sess.run([loss, accuracy, step], feed_dict={X: batch_x, Y: batch_y})
        costs.append(cost)
        print("第", i, "次迭代的损失:", cost, "accuracy:", '{:.2%}'.format(right))
        result_list.append([i, cost, right])
    result_file = open('result.csv', 'w', newline='')
    csv_write = csv.writer(result_file, dialect='excel')
    for row in result_list:
        csv_write.writerow(row)

    print("validation accuracy:",
          '{:.2%}'.format(accuracy.eval({X: mnist.validation.images, Y: mnist.validation.labels})))

fc1为全连接层，分别尝试了relu,sigmoid,sodtplus,linear,elu,tanh几个**函数，得出以下数据.

以上每个数据都可以在excel中画出图表，准确率最高的是linear，也就是fca=tf.matmul(pool1,wc1)+bc1。综合上述的两个实验可以得知，一般情况下，卷积核个数越多网络的准确率也会提高，但是是以计算速度为代价的。在卷积之后往往用的**函数是relu，全连接层用的是linear，输出层则用的softmax连接。具体到自己设计一个卷积神经网络时，可以先找找网上的网络构造，选取几个网络，在数据集上跑一次，综合比较不同的网络的结果。