static_rnn和dynamic_rnn的区别主要在于实现不同。
-
static_rnn会把RNN展平,用空间换时间。 gpu会吃不消(个人测试结果)
-
dynamic_rnn则是使用for或者while循环。
调用static_rnn实际上是生成了rnn按时间序列展开之后的图。打开tensorboard你会看到sequence_length个rnn_cell
stack在一起,只不过这些cell是share
weight的。因此,sequence_length就和图的拓扑结构绑定在了一起,因此也就限制了每个batch的sequence_length必须是一致。调用dynamic_rnn不会将rnn展开,而是利用tf.while_loop这个api,通过Enter, Switch, Merge,
LoopCondition, NextIteration等这些control
flow的节点,生成一个可以执行循环的图(这个图应该还是静态图,因为图的拓扑结构在执行时是不会变化的)。在tensorboard上,你只会看到一个rnn_cell,
外面被一群control
flow节点包围着。对于dynamic_rnn来说,sequence_length仅仅代表着循环的次数,而和图本身的拓扑没有关系,所以每个batch可以有不同sequence_length。
static_rnn
导包、加载数据、定义变量
import tensorflow as tf
tf.reset_default_graph() #流式计算图形graph 循环神经网络 将名字相同重置了图
import datetime #打印时间
import os #保存文件
from tensorflow.examples.tutorials.mnist import input_data
# minst测试集
mnist = input_data.read_data_sets('../', one_hot=True)
# 每次使用100条数据进行训练
batch_size = 100
# 图像向量
width = 28
height = 28
# LSTM隐藏神经元数量
rnn_size = 256
# 输出层one-hot向量长度的
out_size = 10
声明变量
def weight_variable(shape, w_alpha=0.01):
initial = w_alpha * tf.random_normal(shape)
return tf.Variable(initial)
def bias_variable(shape, b_alpha=0.1):
initial = b_alpha * tf.random_normal(shape)
return tf.Variable(initial)
# 权重及偏置
w = weight_variable([rnn_size, out_size])
b = bias_variable([out_size])
将数据转化成RNN所要求的数据
# 按照图片大小申请占位符
X = tf.placeholder(tf.float32, [None, height, width])
# 原排列[0,1,2]transpose为[1,0,2]代表前两维装置,如shape=(1,2,3)转为shape=(2,1,3)
# 这里的实际意义是把所有图像向量的相同行号向量转到一起,如x1的第一行与x2的第一行
x = tf.transpose(X, [1, 0, 2])
# reshape -1 代表自适应,这里按照图像每一列的长度为reshape后的列长度
x = tf.reshape(x, [-1, width])
# split默任在第一维即0 dimension进行分割,分割成height份,这里实际指把所有图片向量按对应行号进行重组
x = tf.split(x, height)
构建静态的循环神经网络
# LSTM
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(rnn_size)
# 这里RNN会有与输入层相同数量的输出层,我们只需要最后一个输出
outputs, status = tf.nn.static_rnn(lstm_cell, x, dtype=tf.float32)
#取最后一个进行矩阵乘法
y_conv = tf.add(tf.matmul(outputs[-1], w), b)
# 最小化损失优化
Y = tf.placeholder(dtype=tf.float32,shape = [None,10])
#损失使用的交叉熵
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=y_conv, labels=Y))
optimizer = tf.train.AdamOptimizer(0.01).minimize(loss)
# 计算准确率
correct = tf.equal(tf.argmax(y_conv, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))
模型的训练
# 启动会话.开始训练
saver = tf.train.Saver()
session = tf.Session()
session.run(tf.global_variables_initializer())
step = 0
acc_rate = 0.90
while 1:
batch_x, batch_y = mnist.train.next_batch(batch_size)
batch_x = batch_x.reshape((batch_size, height, width))
session.run(optimizer, feed_dict={X:batch_x,Y:batch_y})
# 每训练10次测试一次
if step % 10 == 0:
batch_x_test = mnist.test.images
batch_y_test = mnist.test.labels
batch_x_test = batch_x_test.reshape([-1, height, width])
acc = session.run(accuracy, feed_dict={X: batch_x_test, Y: batch_y_test})
print(datetime.datetime.now().strftime('%c'), ' step:', step, ' accuracy:', acc)
# 偏差满足要求,保存模型
if acc >= acc_rate:
# os.sep = ‘/’
model_path = os.getcwd() + os.sep + str(acc_rate) + "mnist.model"
saver.save(session, model_path, global_step=step)
break
step += 1
session.close()
Wed Dec 18 10:08:45 2019 step: 0 accuracy: 0.1006
Wed Dec 18 10:08:46 2019 step: 10 accuracy: 0.1009
Wed Dec 18 10:08:46 2019 step: 20 accuracy: 0.1028
…
Wed Dec 18 10:08:57 2019 step: 190 accuracy: 0.9164
dynamic_rnn
加载数据,声明变量
import tensorflow as tf
tf.reset_default_graph()
from tensorflow.examples.tutorials.mnist import input_data
# 载入数据
mnist = input_data.read_data_sets("../", one_hot=True)
# 输入图片是28
n_input = 28
max_time = 28
lstm_size = 100 # 隐藏单元 可调
n_class = 10 # 10个分类
batch_size = 100 # 每次50个样本 可调
n_batch_size = mnist.train.num_examples // batch_size # 计算一共有多少批次
Extracting …/train-images-idx3-ubyte.gz
Extracting …/train-labels-idx1-ubyte.gz
Extracting …/t10k-images-idx3-ubyte.gz
Extracting …/t10k-labels-idx1-ubyte.gz
占位符、权重
# 这里None表示第一个维度可以是任意长度
# 创建占位符
x = tf.placeholder(tf.float32,[None, 28*28])
# 正确的标签
y = tf.placeholder(tf.float32,[None, 10])
# 初始化权重 ,stddev为标准差
weight = tf.Variable(tf.truncated_normal([lstm_size, n_class], stddev=0.1))
# 初始化偏置层
biases = tf.Variable(tf.constant(0.1, shape=[n_class]))
构建动态RNN、损失函数、准确率
# 定义RNN网络
def RNN(X, weights, biases):
# 原始数据为[batch_size,28*28]
# input = [batch_size, max_time, n_input]
input_ = tf.reshape(X,[-1, max_time, n_input])
# 定义LSTM的基本单元
# lstm_cell = tf.contrib.rnn.BasicLSTMCell(lstm_size)
lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(lstm_size)
# final_state[0] 是cell state
# final_state[1] 是hidden stat
outputs, final_state = tf.nn.dynamic_rnn(lstm_cell, input_, dtype=tf.float32)
display(final_state)
results = tf.nn.softmax(tf.matmul(final_state[1],weights)+biases)
return results
# 计算RNN的返回结果
prediction = RNN(x, weight, biases)
# 损失函数
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=prediction, labels=y))
# 使用AdamOptimizer进行优化
train_step = tf.train.AdamOptimizer(1e-4).minimize(loss)
# 将结果存下来
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
# 计算正确率
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))
LSTMStateTuple(c=<tf.Tensor ‘rnn/while/Exit_3:0’ shape=(?, 100) dtype=float32>, h=<tf.Tensor ‘rnn/while/Exit_4:0’ shape=(?, 100) dtype=float32>)
训练数据
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_size):
# 取出下一批次数据
batch_xs,batch_ys = mnist.train.next_batch(batch_size)
sess.run(train_step, feed_dict={x: batch_xs,y: batch_ys})
if(batch%100==0):
print(str(batch)+"/" + str(n_batch_size))
acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
print("Iter" + str(epoch) + " ,Testing Accuracy = " + str(acc))
if acc >0.9:
saver.save(sess,'./rnn_dynamic')
break
0/550
100/550
200/550
300/550
400/550
500/550
Iter0 ,Testing Accuracy = 0.5903
…
Iter5 ,Testing Accuracy = 0.9103