91 lines
4.5 KiB
Python
91 lines
4.5 KiB
Python
import numpy as np
|
||
# 在文件开头添加兼容性设置
|
||
import tensorflow.compat.v1 as tf
|
||
tf.disable_v2_behavior()
|
||
|
||
# 加载MNIST数据集，通过设置 one_hot=True 来使用独热编码标签
|
||
# 独热编码：对于每个图片的标签 y，10 位中仅有一位的值为 1，其余的为 0。
|
||
mnist = tf.keras.datasets.mnist
|
||
(x_train, y_train), (x_test, y_test) = mnist.load_data()
|
||
|
||
|
||
# 权重正态分布初始化函数
|
||
def weight_variable(shape):
|
||
# 生成截断正态分布随机数,shape表示生成张量的维度，mean是均值(默认=0.0)，stddev是标准差。
|
||
# 取值范围为 [ mean - 2 * stddev, mean + 2 * stddev ]，这里为[-0.2, 0.2]
|
||
initial = tf.truncated_normal(shape, stddev=0.1)
|
||
return tf.Variable(initial)
|
||
|
||
|
||
# 偏置量初始化函数
|
||
def bias_variable(shape):
|
||
initial = tf.constant(0.1, shape=shape) # value=0.1, shape是张量的维度
|
||
return tf.Variable(initial)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
# 预处理数据
|
||
x_train = x_train.reshape(-1, 784).astype('float32') / 255.0
|
||
x_test = x_test.reshape(-1, 784).astype('float32') / 255.0
|
||
y_train = tf.keras.utils.to_categorical(y_train, 10)
|
||
y_test = tf.keras.utils.to_categorical(y_test, 10)
|
||
|
||
# 替换原来的mnist变量使用方式
|
||
mnist_dataset = (x_train, y_train), (x_test, y_test)
|
||
|
||
# 后续代码中所有mnist.train.xxx需要改为mnist_dataset[0][0]和mnist_dataset[0][1]
|
||
# 所有mnist.test.xxx需要改为mnist_dataset[1][0]和mnist_dataset[1][1]
|
||
print(x_train.shape[0]) # 输出训练集样本数 60000
|
||
print(x_test.shape[0]) # 输出测试集样本数 10000
|
||
|
||
# 为训练数据集的输入 x 和标签 y 创建占位符
|
||
x = tf.placeholder(tf.float32, [None, 784]) # 保持原样（已通过兼容性导入）
|
||
y = tf.placeholder(tf.float32, [None, 10])
|
||
keep_prob = tf.placeholder(tf.float32)
|
||
|
||
# 创建神经网络第1层，输入层，激活函数为relu
|
||
W_layer1 = weight_variable([784, 500])
|
||
b_layer1 = bias_variable([500])
|
||
h1 = tf.add(tf.matmul(x, W_layer1), b_layer1) # W * x + b
|
||
h1 = tf.nn.relu(h1)
|
||
# 创建神经网络第2层，隐藏层，激活函数为relu
|
||
W_layer2 = weight_variable([500, 1000])
|
||
b_layer2 = bias_variable([1000])
|
||
h2 = tf.add(tf.matmul(h1, W_layer2), b_layer2) # W * h1 + b，h1为第1层的输出
|
||
h2 = tf.nn.relu(h2)
|
||
# 创建神经网络第3层，隐藏层，激活函数为relu
|
||
W_layer3 = weight_variable([1000, 300])
|
||
b_layer3 = bias_variable([300])
|
||
h3 = tf.add(tf.matmul(h2, W_layer3), b_layer3) # W * h2 + b，h2为第2层的输出
|
||
h3 = tf.nn.relu(h3)
|
||
# 创建神经网络第4层，输出层，激活函数为softmax
|
||
W_layer4 = weight_variable([300, 10])
|
||
b_layer4 = bias_variable([10])
|
||
predict = tf.add(tf.matmul(h3, W_layer4), b_layer4) # W * h3 + b，h3为第3层的输出
|
||
y_conv = tf.nn.softmax(tf.matmul(h3, W_layer4) + b_layer4)
|
||
# 计算交叉熵代价函数
|
||
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=predict, labels=y))
|
||
# 使用Adam下降算法优化交叉熵代价函数
|
||
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
|
||
# 预测是否准确的结果存放在一个布尔型的列表中
|
||
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y, 1)) # argmax返回的矩阵行中的最大值的索引号
|
||
# 求预测准确率
|
||
accuracy = tf.reduce_mean(tf.cast(correct_prediction, 'float')) # cast将布尔型的数据转换成float型的数据；reduce_mean求平均值
|
||
|
||
# 初始化
|
||
init_op = tf.global_variables_initializer()
|
||
|
||
with tf.Session() as sess:
|
||
sess.run(init_op)
|
||
for i in range(550): # 训练样本为55000，分成550批，每批为100个样本
|
||
# 手动实现next_batch
|
||
batch_index = np.random.randint(0, x_train.shape[0], 100)
|
||
batch = (x_train[batch_index], y_train[batch_index])
|
||
train_accuracy = accuracy.eval(feed_dict={x: batch[0], y: batch[1], keep_prob: 1.0})
|
||
test_accuracy = accuracy.eval(feed_dict={x: x_test, y: y_test})
|
||
print('step %d, training accuracy %g, test accuracy %g' % (i, train_accuracy, test_accuracy))
|
||
# 每一步迭代，都会加载100个训练样本，然后执行一次train_step，并通过feed_dict，用训练数据替代x和y张量占位符。
|
||
sess.run(train_step, feed_dict={x: batch[0], y: batch[1], keep_prob: 0.5})
|
||
# 显示最终在测试集上的准确率
|
||
print(
|
||
'test accuracy %g' % accuracy.eval(feed_dict={x: x_test, y: y_test, keep_prob: 1.0})) |