本章讲解循环神经网络的TensorFlow开发模式。其他TensorFlow学习请参考：TensorFlow 学习目录。
目录
一、TensorFlow中的cell类
如何理解cell？是熟练掌握RNN的基础。都知道MLP，多层感知器，除了输入层和输出层以外，其不同的隐藏层由不同个数的神经单元组成。每一个神经元就是一个权重加上一个偏置，RNN也可以用这样的神经元来构建网络，但是为了增加网络的记忆性，不断地对神经元进行复杂化，以此出现了GRU、LSTM，也就是说将原先的单一神经元变成了一个系列操作，并将这一系列操作运算封装起来，将其看成是一个复杂的神经元，也就是我们所说的cell，具体内容还是需要去学习一下。
有了Cell我们就可以像MLP那样构建RNN循环神经网络了，就完全把他当成一个神经元来理解。比如我们定义一个隐藏层的时候，需要定义隐藏层神经元个数，那么上述函数的定义就相当于是定义了一层隐藏层。
二、通过cell构建RNN
<1> 静态RNN构建
- cell：生成好的cell类对象
- inputs：一定是list
- initial_state：初始化cell状态
- dtype：期望输出和初始化state的类型
- sequenc_length：每一个输入的序列长度
- scope：命名空间
- 返回值：有两个一个是outputs一个是state，其outputs是一个list，输入多少个时序，list中就有多少时序的输出
<2> 动态RNN构建
- inputs：输入数据，一定是Tensor，[batch_size, max_time, ...]
- time_major：默认值为False，input的shape为[batch_size, max_time, ...]，如果将其改为True，input的shape为[max_time, batch_size, ...]
- 返回值：一个是结果cell，outputs是[batch_size, max_time, ...]的Tensor。
- 注意：在默认time_major的时候，输出的第一维度不是时间，所以要transpose一下，具体操作在代码中有体现
<3> 双向RNN构建
唯一的不同点是cell的输入有两个，一个是正向的cells，一个是反向的cells
三、RNN在MNIST数据集上的实现
其中有关MNIST加载使用问题，请参考TensorFlow 手写数字识别。
1、单层静态GRU
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 100
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
gru = tf.contrib.rnn.GRUCell(n_hidden)
outputs = tf.contrib.rnn.static_rnn(gru, inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(30):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9637
Epoch 2 acc= 0.9769
Epoch 3 acc= 0.9809
Epoch 4 acc= 0.9815
Epoch 5 acc= 0.9828
Epoch 6 acc= 0.9814
Epoch 7 acc= 0.9861
Epoch 8 acc= 0.9858
Epoch 9 acc= 0.9836
Epoch 10 acc= 0.9882
Epoch 11 acc= 0.9843
Epoch 12 acc= 0.987
Epoch 13 acc= 0.9872
Epoch 14 acc= 0.9877
Epoch 15 acc= 0.9885
Epoch 16 acc= 0.9894
Epoch 17 acc= 0.9881
Epoch 18 acc= 0.9868
Epoch 19 acc= 0.9864
Epoch 20 acc= 0.9891
Epoch 21 acc= 0.9867
Epoch 22 acc= 0.9873
Epoch 23 acc= 0.9873
Epoch 24 acc= 0.9885
Epoch 25 acc= 0.9894
Epoch 26 acc= 0.9898
Epoch 27 acc= 0.987
Epoch 28 acc= 0.9891
Epoch 29 acc= 0.9887
Epoch 30 acc= 0.98852、单层动态GRU
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 100
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
# neural network architicture
gru = tf.contrib.rnn.GRUCell(n_hidden)
outputs, _ = tf.nn.dynamic_rnn(gru, inputs_reshape, dtype=tf.float32)
outputs = tf.transpose(outputs, [1, 0, 2])
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(25):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9651
Epoch 2 acc= 0.9696
Epoch 3 acc= 0.9727
Epoch 4 acc= 0.9831
Epoch 5 acc= 0.9855
Epoch 6 acc= 0.9819
Epoch 7 acc= 0.9855
Epoch 8 acc= 0.9854
Epoch 9 acc= 0.9857
Epoch 10 acc= 0.9874
Epoch 11 acc= 0.9859
Epoch 12 acc= 0.986
Epoch 13 acc= 0.9858
Epoch 14 acc= 0.9883
Epoch 15 acc= 0.9876
Epoch 16 acc= 0.9863
Epoch 17 acc= 0.9892
Epoch 18 acc= 0.9895
Epoch 19 acc= 0.9886
Epoch 20 acc= 0.9868
Epoch 21 acc= 0.988
Epoch 22 acc= 0.9845
Epoch 23 acc= 0.9879
Epoch 24 acc= 0.9825
Epoch 25 acc= 0.98773、单层静态LSTM
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 100
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
lstm_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden, forget_bias=1.0) # forget_bias 添加到forget门的偏置
outputs, states = tf.contrib.rnn.static_rnn(lstm_cell, inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(30):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9396
Epoch 2 acc= 0.9623
Epoch 3 acc= 0.9637
Epoch 4 acc= 0.9769
Epoch 5 acc= 0.9808
Epoch 6 acc= 0.9832
Epoch 7 acc= 0.9825
Epoch 8 acc= 0.986
Epoch 9 acc= 0.9848
Epoch 10 acc= 0.9836
Epoch 11 acc= 0.9872
Epoch 12 acc= 0.9845
Epoch 13 acc= 0.9872
Epoch 14 acc= 0.9879
Epoch 15 acc= 0.9886
Epoch 16 acc= 0.9855
Epoch 17 acc= 0.9889
Epoch 18 acc= 0.9878
Epoch 19 acc= 0.9862
Epoch 20 acc= 0.9843
Epoch 21 acc= 0.9883
Epoch 22 acc= 0.9867
Epoch 23 acc= 0.9877
Epoch 24 acc= 0.9888
Epoch 25 acc= 0.9814
Epoch 26 acc= 0.9891
Epoch 27 acc= 0.9878
Epoch 28 acc= 0.9887
Epoch 29 acc= 0.9892
Epoch 30 acc= 0.98964、单层静态双向RNN
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden, forget_bias=1.0)
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden, forget_bias=1.0)
outputs, _, _ = tf.contrib.rnn.static_bidirectional_rnn(lstm_fw_cell, lstm_bw_cell, inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(10):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9279
Epoch 2 acc= 0.9477
Epoch 3 acc= 0.9543
Epoch 4 acc= 0.9663
Epoch 5 acc= 0.9725
Epoch 6 acc= 0.9765
Epoch 7 acc= 0.9765
Epoch 8 acc= 0.9743
Epoch 9 acc= 0.9808
Epoch 10 acc= 0.97945、单层动态双向RNN
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden, forget_bias=1.0)
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden, forget_bias=1.0)
outputs, states = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell, inputs_reshape, dtype=tf.float32)
outputs = tf.concat(outputs, 2)
outputs = tf.transpose(outputs, [1, 0, 2])
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(10):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9289
Epoch 2 acc= 0.9477
Epoch 3 acc= 0.9539
Epoch 4 acc= 0.9623
Epoch 5 acc= 0.9725
Epoch 6 acc= 0.969
Epoch 7 acc= 0.9764
Epoch 8 acc= 0.9801
Epoch 9 acc= 0.9793
Epoch 10 acc= 0.97846、多层静态LSTM
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
stacked_rnn = []
for i in range(3):
stacked_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
mcell = tf.contrib.rnn.MultiRNNCell(stacked_rnn)
outputs, states = tf.contrib.rnn.static_rnn(mcell, inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(30):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9467
Epoch 2 acc= 0.97
Epoch 3 acc= 0.9723
Epoch 4 acc= 0.9785
Epoch 5 acc= 0.9824
Epoch 6 acc= 0.9855
Epoch 7 acc= 0.9808
Epoch 8 acc= 0.9841
Epoch 9 acc= 0.9863
Epoch 10 acc= 0.9871
Epoch 11 acc= 0.9836
Epoch 12 acc= 0.9874
Epoch 13 acc= 0.9876
Epoch 14 acc= 0.9881
Epoch 15 acc= 0.9881
Epoch 16 acc= 0.9865
Epoch 17 acc= 0.9855
Epoch 18 acc= 0.9885
Epoch 19 acc= 0.9881
Epoch 20 acc= 0.9883
Epoch 21 acc= 0.9847
Epoch 22 acc= 0.9889
Epoch 23 acc= 0.9883
Epoch 24 acc= 0.9852
Epoch 25 acc= 0.9891
Epoch 26 acc= 0.9866
Epoch 27 acc= 0.9899
Epoch 28 acc= 0.9877
Epoch 29 acc= 0.9899
Epoch 30 acc= 0.98577、多层动态LSTM
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
lstm1 = tf.contrib.rnn.LSTMCell(n_hidden)
lstm2 = tf.contrib.rnn.LSTMCell(n_hidden*2)
mcell = tf.contrib.rnn.MultiRNNCell([lstm2, lstm1])
outputs, states = tf.nn.dynamic_rnn(mcell, inputs_reshape, dtype=tf.float32)
outputs = tf.transpose(outputs, [1, 0, 2])
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(20):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9621
Epoch 2 acc= 0.9589
Epoch 3 acc= 0.9782
Epoch 4 acc= 0.9825
Epoch 5 acc= 0.9845
Epoch 6 acc= 0.9874
Epoch 7 acc= 0.9873
Epoch 8 acc= 0.9866
Epoch 9 acc= 0.9872
Epoch 10 acc= 0.987
Epoch 11 acc= 0.9895
Epoch 12 acc= 0.989
Epoch 13 acc= 0.9892
Epoch 14 acc= 0.9886
Epoch 15 acc= 0.9904
Epoch 16 acc= 0.9902
Epoch 17 acc= 0.9894
Epoch 18 acc= 0.9881
Epoch 19 acc= 0.9889
Epoch 20 acc= 0.98698、多层混合静态LSTM+GRU
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
gru = tf.contrib.rnn.GRUCell(n_hidden*2)
lstm = tf.contrib.rnn.LSTMCell(n_hidden)
mcell = tf.contrib.rnn.MultiRNNCell([lstm, gru])
outputs, states = tf.contrib.rnn.static_rnn(mcell, inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(30):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9562
Epoch 2 acc= 0.974
Epoch 3 acc= 0.9817
Epoch 4 acc= 0.9848
Epoch 5 acc= 0.985
Epoch 6 acc= 0.9876
Epoch 7 acc= 0.9875
Epoch 8 acc= 0.9875
Epoch 9 acc= 0.9864
Epoch 10 acc= 0.986
Epoch 11 acc= 0.9877
Epoch 12 acc= 0.9896
Epoch 13 acc= 0.9892
Epoch 14 acc= 0.9891
Epoch 15 acc= 0.9861
Epoch 16 acc= 0.9866
Epoch 17 acc= 0.9875
Epoch 18 acc= 0.9875
Epoch 19 acc= 0.9872
Epoch 20 acc= 0.9878
Epoch 21 acc= 0.9885
Epoch 22 acc= 0.9889
Epoch 23 acc= 0.9883
Epoch 24 acc= 0.9883
Epoch 25 acc= 0.9888
Epoch 26 acc= 0.9844
Epoch 27 acc= 0.9892
Epoch 28 acc= 0.9884
Epoch 29 acc= 0.9895
Epoch 30 acc= 0.98729、多层双向RNN
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
# way 1: list多层RNN
# stacked_fw_rnn = []
# stacked_bw_rnn = []
# for i in range(2):
# stacked_fw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
# stacked_bw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
#
# outputs, _, _ = tf.contrib.rnn.stack_bidirectional_rnn(stacked_fw_rnn, stacked_bw_rnn, inputs_, dtype=tf.float32)
# way 2: multi双向RNN
stacked_fw_rnn = []
stacked_bw_rnn = []
for i in range(2):
stacked_fw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
stacked_bw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
mcell_fw = tf.contrib.rnn.MultiRNNCell(stacked_fw_rnn)
mcell_bw = tf.contrib.rnn.MultiRNNCell(stacked_bw_rnn)
outputs, _, _ = tf.contrib.rnn.stack_bidirectional_rnn([mcell_fw], [mcell_bw], inputs_, dtype=tf.float32)
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(10):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
输出1：
Epoch 1 acc= 0.9471
Epoch 2 acc= 0.9619
Epoch 3 acc= 0.9756
Epoch 4 acc= 0.9766
Epoch 5 acc= 0.9768
Epoch 6 acc= 0.9837
Epoch 7 acc= 0.9808
Epoch 8 acc= 0.9851
Epoch 9 acc= 0.9816
Epoch 10 acc= 0.9841
输出2：
Epoch 1 acc= 0.9455
Epoch 2 acc= 0.9697
Epoch 3 acc= 0.9654
Epoch 4 acc= 0.9769
Epoch 5 acc= 0.9805
Epoch 6 acc= 0.9818
Epoch 7 acc= 0.9851
Epoch 8 acc= 0.9874
Epoch 9 acc= 0.9859
Epoch 10 acc= 0.986610、多层动态双向RNN
import tensorflow as tf
import numpy as np
import get_Dataset
x_train, y_train, x_test, y_test = get_Dataset.get_Dataset(name='mnist')
n_input = 28
n_steps = 28
n_hidden = 50
n_classes = 10
batch_size = 32
inputs = tf.placeholder(dtype=tf.float32, shape=[None, n_steps*n_input])
labels = tf.placeholder(dtype=tf.float32, shape=[None, n_classes])
test_labels = tf.placeholder(dtype=tf.int64, shape=[None])
inputs_reshape = tf.reshape(inputs, [-1, n_steps, n_input])
inputs_ = tf.unstack(inputs_reshape, n_steps, 1) # return a list
# neural network architicture
stacked_fw_rnn = []
stacked_bw_rnn = []
for i in range(2):
stacked_fw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
stacked_bw_rnn.append(tf.contrib.rnn.LSTMCell(n_hidden))
mcell_fw = tf.contrib.rnn.MultiRNNCell(stacked_fw_rnn)
mcell_bw = tf.contrib.rnn.MultiRNNCell(stacked_bw_rnn)
outputs, _, _ = tf.contrib.rnn.stack_bidirectional_dynamic_rnn([mcell_fw], [mcell_bw], inputs_reshape, dtype=tf.float32)
outputs = tf.transpose(outputs, [1, 0, 2])
logits = tf.contrib.layers.fully_connected(outputs[-1], n_classes, activation_fn=None)
# build_loss
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=labels))
optimizer = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss)
# evaluate
correct_pred = tf.equal(tf.argmax(logits, 1), test_labels)
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(10):
for batch in range(int(x_train.shape[0]/batch_size)):
batchx = x_train[batch*batch_size: (batch+1)*batch_size]
batchy = y_train[batch*batch_size: (batch+1)*batch_size]
feed_dict = {
inputs: batchx,
labels: batchy
}
_ = sess.run(optimizer, feed_dict=feed_dict)
acc = sess.run(accuracy, feed_dict={inputs: x_test, test_labels: y_test})
print ("Epoch ", epoch+1, "acc=", acc)输出
Epoch 1 acc= 0.9406
Epoch 2 acc= 0.9627
Epoch 3 acc= 0.9675
Epoch 4 acc= 0.9706
Epoch 5 acc= 0.9792
Epoch 6 acc= 0.9823
Epoch 7 acc= 0.984
Epoch 8 acc= 0.9866
Epoch 9 acc= 0.9853
Epoch 10 acc= 0.9844