convolution-卷积神经网络

训练mnist数据集

结构组成：

input_image --> convolution1 --> pool1 --> convolution2 --> pool2 --> full_connecion1 --> full_connection2

# 卷积
import tensorflow as tf

import input_data

# 加载mnist数据集
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)


# 构建多层卷积网络
# 权重及偏置初始化, ReLU神经元 用一个较小的正数来初始化偏置项来打破对称性以及避免0梯度
def weight_variable(shape):
    """
    :param shape:二维tensor,第一个维度代表层中权重变量所连接（connect from）的单元数目，
    第二个维度代表层中权重变量所连接（connect to）到的单元数量
    :return: W
    """
    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)


# 卷积及池化
def conv2d(x, W):
    """
    卷积
    :param x:
    :param W:
    :return:
    """
    return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding="SAME")


def max_pool_2x2(x):
    """
    最大池化
    :param x:
    :return:
    """
    return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")


def compute_accuracy(v_xs, v_ys):
    """ 计算的准确率 """
    global prediction  # prediction value
    y_pre = sess.run(prediction, feed_dict={xs: v_xs})
    # 与期望的值比较 bool
    correct_pre = tf.equal(tf.argmax(y_pre, 1), tf.argmax(ys, 1))
    # 将bools转化为数字
    accuracy = tf.reduce_mean(tf.cast(correct_pre, tf.float32))
    result = sess.run(accuracy, feed_dict={xs: v_xs, ys: v_ys})
    return result


# 数据图片
xs = tf.placeholder("float", shape=[None, 784])  # size 28 * 28 =784
# 预期概率
ys = tf.placeholder("float", shape=[None, 10])  # 10: 矩阵维度(分类)
keep_prob = tf.placeholder(tf.float32)
x_image = tf.reshape(xs, [-1, 28, 28, 1])  # -1: 任意数量的图片; 28*28:图片的长宽; 1:灰色图片为1

# layer1
W_conv1 = weight_variable([5, 5, 1, 32])  # 5*5:patch过滤长宽, 1:起始输入一张图片, 32:out_size
b_conv1 = bias_variable([32])  # 32:上层输入的out_size
h_conv1 = tf.nn.relu(conv2d(x_image, W_conv1) + b_conv1)  # output_size=28*28*32
h_pool1 = max_pool_2x2(h_conv1)  # output_size=14*14*32 pool_strdes=2

# layer2
W_conv2 = weight_variable([5, 5, 32, 64])  # 64是训练中不断增加的高度，自定义
b_conv2 = bias_variable([64])
h_conv2 = tf.nn.relu(conv2d(h_pool1, W_conv2) + b_conv2)  # output_size=14*14*64
h_pool2 = max_pool_2x2(h_conv2)  # output_size=7*7*64 池化的步长为[2,2]

# func1 layer
W_fc1 = weight_variable([7*7*64, 1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2, [-1, 7*7*64])  # 将pool2铺平为7*7*64
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, W_fc1) + b_fc1)  # 矩阵相乘
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)  # 防止过拟合

# func2 layer
W_fc2 = weight_variable([1024, 10])  # 传入的1024， 判断0-9的数字one-hot，10来代表每个数字
b_fc2 = bias_variable([10])
prediction = tf.nn.softmax(tf.matmul(h_fc1_drop, W_fc2) + b_fc2)  # softmax分类

# the loss between prediction and really
cross_entropy = tf.reduce_mean(-tf.reduce_sum(ys * tf.log(prediction), reduction_indices=[1]))
tf.summary.scalar('loss', cross_entropy)  # 字符串类型的标量张量，包含一个Summaryprotobuf  1.1记录标量
# training
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)  # AdamOptimizer使用复杂模型

sess = tf.Session()
sess.run(tf.initialize_all_variables())


# start training
for i in range(1000):
    batch_x, batch_y = mnist.train.next_batch(100)
    sess.run(train_step, feed_dict={xs: batch_x, ys: batch_y, keep_prob: 0.1})
    if i % 50 == 0:
        print(compute_accuracy(mnist.test.images, mnist.test.labels))
print("Training Finished !!!")