TensorFlow学习（9）循环神经网络、保存模型和载入模型
RNN
RNN 和 BP 都有梯度消失的问题，信号会逐渐减弱
LSTM
主要用于文本和语音等序列化问题
LSTM 隐藏层的神经元不像BP的那样的神经元，而是如下比较复杂的block
第一项的输入影响了第4和第6项的输出结果
手写数字识别使用卷积神经网络案列
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# 输入图片是 28*28
n_inputs = 28 # 输入一行，一行有28个数据，每一次输入图片的一行
max_time = 28 # 一共28行，一共输入28次
lstm_size = 100 # 隐藏层单元个数
n_classes = 10 # 10 个分类
batch_size = 50 # 每批次50个样本
n_batch = mnist.train.num_examples // batch_size # 计算有多少个批次
x = tf.placeholder(tf.float32, [None, 784]) #[50,784]
y = tf.placeholder(tf.float32, [None, 10])
# 初始化权值
weights = tf.Variable(tf.truncated_normal([lstm_size, n_classes], stddev=0.1)) # [100,10]
biases = tf.Variable(tf.constant(0.1, shape=[n_classes]))
# 定义RNN网络
def RNN(X, weights, biases):
inputs = tf.reshape(X, [-1, max_time, n_inputs]) # [-1, 28, 28]
# 定义 LSTM 基本 CELL，即隐藏层的block
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
# inputs 的格式固定 [batch_size, max_time, n_inputs]
# final_state [state, batch_size, cell.state_size]，final_state 是记录时间序列最后一次输出的结果
# final_state[0] 是 cell_state
# final_state[1] 是 hidden_state
# outputs 记录了每一次时间序列输出的结果，如果max_time是1，则是记录第一次，最后一次的输出结果与final_state 相同
# 'outputs' is a tensor of shape [batch_size, max_time, cell_state_size]
outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, inputs, dtype=tf.float32)
results = tf.nn.softmax(tf.matmul(final_state[1], weights) + biases)
return results
# 计算RNN返回结果
prediction = RNN(x, weights, biases)
# 交叉熵代价函数
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction))
# 优化器
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
# 结果存放于布尔列表
correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
# 计算准确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# 用于保存模型
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(6):
for batch in range(n_batch):
batch_xs, batch_ys = mnist.train.next_batch(batch_size)
sess.run(train_step, feed_dict={x: batch_xs, y: batch_ys})
acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
print('Iter: ' + str(epoch) + ', Testing Accuracy= ' + str(acc))
# 保存模型
saver.save(sess, 'net/rnn_mnist.ckpt')hidden_state 就是整个 block 的输出，即下面的 ht，cell_state 就是下面的 ct
上面保存了模型，这里载入模型进行测试
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
# 输入图片是 28*28
n_inputs = 28 # 输入一行，一行有28个数据，每一次输入图片的一行
max_time = 28 # 一共28行，一共输入28次
lstm_size = 100 # 隐藏层单元个数
n_classes = 10 # 10 个分类
batch_size = 50 # 每批次50个样本
n_batch = mnist.train.num_examples // batch_size # 计算有多少个批次
x = tf.placeholder(tf.float32, [None, 784]) #[50,784]
y = tf.placeholder(tf.float32, [None, 10])
# 初始化权值
weights = tf.Variable(tf.truncated_normal([lstm_size, n_classes], stddev=0.1)) # [100,10]
biases = tf.Variable(tf.constant(0.1, shape=[n_classes]))
# 定义RNN网络
def RNN(X, weights, biases):
inputs = tf.reshape(X, [-1, max_time, n_inputs]) # [-1, 28, 28]
# 定义 LSTM 基本 CELL，即隐藏层的block
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
# inputs 的格式固定 [batch_size, max_time, n_inputs]
# final_state [state, batch_size, cell.state_size]，final_state 是记录时间序列最后一次输出的结果
# final_state[0] 是 cell_state
# final_state[1] 是 hidden_state
# outputs 记录了每一次时间序列输出的结果，如果max_time是1，则是记录第一次，最后一次的输出结果与final_state 相同
# 'outputs' is a tensor of shape [batch_size, max_time, cell_state_size]
outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, inputs, dtype=tf.float32)
results = tf.nn.softmax(tf.matmul(final_state[1], weights) + biases)
return results
# 计算RNN返回结果
prediction = RNN(x, weights, biases)
# 交叉熵代价函数
cross_entropy = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction))
# 优化器
train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)
# 结果存放于布尔列表
correct_prediction = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
# 计算准确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
print(sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels}))
# 载入模型
saver.restore(sess, 'net/rnn_mnist.ckpt')
print(sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels}))载入模型后准确率明显提高，说明用的是已经训练好的参数