设计思路
实现代码
# -*- coding:utf-8 -*-
import tensorflow as tf
import numpy as np
from tensorflow.contrib import rnn
from tensorflow.examples.tutorials.mnist import input_data
#根据电脑情况设置 GPU
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
# 1、定义数据集
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
print(mnist.train.images.shape)
#2、定义模型超参数
lr = 1e-3
# batch_size = 128
batch_size = tf.placeholder(tf.int32) #采用占位符的方式，因为在训练和测试的时候要用不同的batch_size。注意类型必须为 tf.int32
input_size = 28 # 每个时刻的输入特征是28维的，就是每个时刻输入一行，一行有 28 个像素
timestep_size = 28 # 时序持续长度为28，即每做一次预测，需要先输入28行
hidden_size = 256 # 每个隐含层的节点数
layer_num = 2 # LSTM layer 的层数
class_num = 10 # 最后输出分类类别数量，如果是回归预测的话应该是 1
_X = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, class_num])
keep_prob = tf.placeholder(tf.float32)
#3、LSTM模型的搭建、训练、测试
#3.1、LSTM模型的搭建
X = tf.reshape(_X, [-1, 28, 28]) #RNN 的输入shape = (batch_size, timestep_size, input_size)，把784个点的字符信息还原成 28 * 28 的图片
lstm_cell = rnn.BasicLSTMCell(num_units=hidden_size, forget_bias=1.0, state_is_tuple=True) #定义一层 LSTM_cell，只需要说明 hidden_size, 它会自动匹配输入的 X 的维度
lstm_cell = rnn.DropoutWrapper(cell=lstm_cell, input_keep_prob=1.0, output_keep_prob=keep_prob) #添加 dropout layer, 一般只设置 output_keep_prob
mlstm_cell = rnn.MultiRNNCell([lstm_cell] * layer_num, state_is_tuple=True) #调用 MultiRNNCell来实现多层 LSTM
init_state = mlstm_cell.zero_state(batch_size, dtype=tf.float32) #用全零来初始化state
#3.2、LSTM模型的运行：构建好的网络运行起来
#T1、调用 dynamic_rnn()法
# ** 当 time_major==False 时， outputs.shape = [batch_size, timestep_size, hidden_size]，所以，可以取 h_state = outputs[:, -1, :] 作为最后输出
# ** state.shape = [layer_num, 2, batch_size, hidden_size],或者，可以取 h_state = state[-1][1] 作为最后输出，最后输出维度是 [batch_size, hidden_size]
# outputs, state = tf.nn.dynamic_rnn(mlstm_cell, inputs=X, initial_state=init_state, time_major=False)
# h_state = outputs[:, -1, :] # 或者 h_state = state[-1][1]
#T2、自定义LSTM迭代按时间步展开计算：为了更好的理解 LSTM 工作原理把T1的函数自己来实现
#(1)、可以采用RNNCell的 __call__()函数，来实现LSTM按时间步迭代。
outputs = list()
state = init_state
with tf.variable_scope('RNN'):
for timestep in range(timestep_size):
if timestep > 0:
tf.get_variable_scope().reuse_variables()
(cell_output, state) = mlstm_cell(X[:, timestep, :], state) # 这里的state保存了每一层 LSTM 的状态
outputs.append(cell_output)
h_state = outputs[-1]
#3.3、LSTM模型的训练
# 定义 softmax 的连接权重矩阵和偏置：上面 LSTM 部分的输出会是一个 [hidden_size] 的tensor，我们要分类的话，还需要接一个 softmax 层
# out_W = tf.placeholder(tf.float32, [hidden_size, class_num], name='out_Weights')
# out_bias = tf.placeholder(tf.float32, [class_num], name='out_bias')
W = tf.Variable(tf.truncated_normal([hidden_size, class_num], stddev=0.1), dtype=tf.float32)
bias = tf.Variable(tf.constant(0.1,shape=[class_num]), dtype=tf.float32)
y_pre = tf.nn.softmax(tf.matmul(h_state, W) + bias)
#定义损失和评估函数
cross_entropy = -tf.reduce_mean(y * tf.log(y_pre))
train_op = tf.train.AdamOptimizer(lr).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(y,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
sess.run(tf.global_variables_initializer())
for i in range(2000):
_batch_size = 128
batch = mnist.train.next_batch(_batch_size)
if (i+1)%200 == 0:
train_accuracy = sess.run(accuracy, feed_dict={
_X:batch[0], y: batch[1], keep_prob: 1.0, batch_size: _batch_size})
# 已经迭代完成的 epoch 数: mnist.train.epochs_completed
print("Iter%d, step %d, training accuracy %g" % ( mnist.train.epochs_completed, (i+1), train_accuracy))
sess.run(train_op, feed_dict={_X: batch[0], y: batch[1], keep_prob: 0.5, batch_size: _batch_size})
# 计算测试数据的准确率
print("test accuracy %g"% sess.run(accuracy, feed_dict={_X: mnist.test.images, y: mnist.test.labels,
keep_prob: 1.0, batch_size:mnist.test.images.shape[0]}))
