看 RNN 的 paper 大多数集中在 RNNcell 内部构建，少数涉及 units 之间交互，
Tensorflow 提供了几种最流行的 RNN 变种类，但没有 CNN 编写方便，这里分享一段使用 tf.scan 构建 GRUcell 代码，可以作为自定义 RNNcell 的参考。
import numpy as np
import pandas as pd
import tensorflow as tf
import pylab as pl
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("/tmp/data/", one_hot=True)
%matplotlib inline
class GRUcell(object):
def __init__(self):
self.in_length= 28
self.in_width= 28
self.hidden_layer_size = 2000
self.out_classes = 10
self.Wr = tf.Variable(tf.zeros([self.in_width, self.hidden_layer_size]))
self.Wz = tf.Variable(tf.zeros([self.in_width, self.hidden_layer_size]))
self.W_ = tf.Variable(tf.zeros([self.in_width, self.hidden_layer_size]))
self.Ur = tf.Variable(tf.truncated_normal([self.hidden_layer_size, self.hidden_layer_size]))
self.Uz = tf.Variable(tf.truncated_normal([self.hidden_layer_size, self.hidden_layer_size]))
self.U_ = tf.Variable(tf.truncated_normal([self.hidden_layer_size, self.hidden_layer_size]))
self.Wout = tf.Variable(tf.truncated_normal([self.hidden_layer_size, self.out_classes], mean=0., stddev=.1))
self.bout = tf.Variable(tf.truncated_normal([self.out_classes], mean=0., stddev=.1))
self.inX = tf.placeholder(shape=[None, self.in_length, self.in_width], dtype=tf.float32)
self.initial_hidden = tf.matmul(self.inX[:,0,:], tf.zeros([self.in_width, self.hidden_layer_size]))
self.X = tf.transpose(self.inX, perm=[1,0,2])
def GRU(self, hidden_states_previous, current_input_X):
"""
GRU topology unit
Note that the input order above is for the fn function
The two tensors are entered for the fn function,
the first tensor is the output calculated in the previous step,
and the second tensor is the input value at this time
"""
hp = hidden_states_previous
x = current_input_X
r = tf.sigmoid(tf.matmul(x, self.Wr) + tf.matmul(hp, self.Ur))
z = tf.sigmoid(tf.matmul(x, self.Wz) + tf.matmul(hp, self.Uz))
h_ = tf.tanh(tf.matmul(x, self.W_) + tf.matmul(r*hp ,self.U_))
h = tf.multiply(hp,z) + tf.multiply((1-z),h_)
return h
def PRO_TS(self):
"""
Perform recursive operations in time series
Iterates through time/ sequence to get all hidden state
Input format : [in_length, batch_size, in_width]
Output format : [in_length, batch_size, hidden_layer_size]
"""
return tf.scan(fn= self.GRU, elems=self.X, initializer=self.initial_hidden)
def Full_Connection_Layer(self, batch_hidden_layer_states):
"""
The hidden layer state input is converted to
output through the full connection layer
Input format : [batch_size, hidden_layer_size]
Output format : [batch_size, out_classes]
"""
return tf.nn.relu(tf.nn.bias_add(tf.matmul(batch_hidden_layer_states, self.Wout), self.bout))
def deal_hidden_layer(self):
"""
Handle all state output of hidden layer
Input format : [in_length, batch_size, hidden_layer_size]
Output format : [in_length, batch_size, out_classes]
"""
#all_hidden_states = self.PRO_TS()
#return tf.map_fn(self.Full_Connection_Layer, all_hidden_states)
return tf.map_fn(self.Full_Connection_Layer, self.PRO_TS())
def last_output(self):
tp = tf.reverse(self.deal_hidden_layer(), axis=[0])[0,:,:]
return tf.nn.softmax(tp)
y = tf.placeholder(tf.float32, shape=[None, 10],name='inputs')
rnn = GRUcell()
output = rnn.last_output()
cross_entropy = -tf.reduce_sum(y * tf.log(output))
train_step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cross_entropy)
correct_prediction = tf.equal(tf.argmax(y,1), tf.argmax(output,1))
accuracy = (tf.reduce_mean(tf.cast(correct_prediction, tf.float32)))
sess=tf.InteractiveSession()
sess.run(tf.global_variables_initializer())
batch_size = 32
ss = []
for i in range(5000):
batch_x, batch_y = mnist.train.next_batch(batch_size)
batch_x = batch_x.reshape((batch_size, 28, 28))
sess.run(train_step, feed_dict={rnn.inX:batch_x, y:batch_y})
t = sess.run(accuracy, feed_dict={rnn.inX:batch_x, y:batch_y})
ss.append(t)
ttt = pd.Series(ss)
ttt.plot()
使用 Tensorflow version 1.0 python 3.6
源代码地址： https://uqer.io/community/share/58a9332bf1973300597ae209
Select Language