def rnn_net(x):
# 将x 从二维转成3维 目的使用rnn（输入需是3维数据）
inputs=slim.embed_sequence(tf.cast(x,tf.int32),vocab_size=len(lex),embed_dim=n_layer_1) # [none,len(train_dataset[0][0]),n_layer_1] 3维
cell = tf.contrib.rnn.BasicLSTMCell(n_layer_1, state_is_tuple=True)
cell = tf.nn.rnn_cell.MultiRNNCell([cell] * 4, state_is_tuple=True) # 多层核
initial_state = cell.zero_state(batch_size, tf.float32)
outputs, last_state = tf.nn.dynamic_rnn(cell, inputs, initial_state=initial_state, scope='rnnlm')
# outputs shape [batch_size,len(train_dataset[0][0]),n_layer_1]
results=slim.fully_connected(outputs[:,-1,:],n_output_layer,activation_fn=None) # [batch,2]
return results
def rnn_net_1(x):
cell = tf.nn.rnn_cell.BasicLSTMCell(n_layer_1, state_is_tuple=True)
cell = tf.nn.rnn_cell.MultiRNNCell([cell] * 2, state_is_tuple=True)
initial_state = cell.zero_state(batch_size, tf.float32)
inputs = slim.embed_sequence(
tf.cast(x, tf.int32), vocab_size=len(lex),
embed_dim=n_layer_1) # # [batch_size,-1,n_layer_1]
outputs, last_state = tf.nn.dynamic_rnn(cell,
inputs,
initial_state=initial_state,
scope='rnnlm')
# outputs shape [batch_size,-1,n_layer_1]
# outputs = tf.reshape(outputs, [-1, n_layer_1])
logits = slim.fully_connected(outputs[:, -1, :],
n_output_layer,
activation_fn=None) # [-1,2]
return logits
def cnn_net(x):
inputs = slim.embed_sequence(tf.cast(x, tf.int32), vocab_size=len(lex),
embed_dim=n_layer_1) # [batch_size,len(train_dataset[0][0]),n_layer_1] 3维
branch1=tf.layers.conv1d(inputs,n_layer_1,3,padding='valid',activation=tf.nn.relu,kernel_regularizer=tf.nn.l2_loss)
branch2 = tf.layers.conv1d(inputs, n_layer_1, 4, padding='valid', activation=tf.nn.relu,
kernel_regularizer=tf.nn.l2_loss)
branch3 = tf.layers.conv1d(inputs, n_layer_1, 5, padding='valid', activation=tf.nn.relu,
kernel_regularizer=tf.nn.l2_loss)
network = tf.concat([branch1, branch2, branch3], 1) # # [batch_size,-1,n_layer_1]
network = tf.expand_dims(network, 2) # [batch_size,-1,1,n_layer_1]
network = tf.reduce_max(network, [1, 2]) # [batch_size,n_layer_1]
results = slim.fully_connected(network, n_output_layer, activation_fn=None) # [batch,2]
return results
def neural_network(vocabulary_size, embedding_size=128, num_filters=128):
'''
使用RNN， 动态Rnn，静态rnn，多层rnn
:param vocabulary_size: 词汇量的大小
:param embedding_size: 嵌套大小
:param num_filters: 隐藏层节点数
:return:
'''
# embedding layer
with tf.device('/cpu:0'), tf.name_scope("embedding"):
# W = tf.Variable(tf.random_uniform([vocabulary_size, embedding_size], -1.0, 1.0))
# embedded_chars = tf.nn.embedding_lookup(W, X) # [-1,8,128]
# embedded_chars=tflearn.embedding(X, vocabulary_size, embedding_size)
embedded_chars = slim.embed_sequence(
X, vocabulary_size, embedding_size) # X 必须是二维的 [-1,8,128]
# embedded_chars_expanded = tf.expand_dims(embedded_chars, -1) # [-1,8,128,1] 变成4维方便使用卷积
# 方法一 多层rnn+动态rnn
'''
cell = rnn.BasicLSTMCell(num_units=num_filters)
cell = rnn.DropoutWrapper(cell=cell, input_keep_prob=1.0, output_keep_prob=dropout_keep_prob)
cells = rnn.MultiRNNCell([cell] * 4, state_is_tuple=True) # 4个cell
init_state = cells.zero_state(batch_size, dtype=tf.float32)
outputs, final_state = tf.nn.dynamic_rnn(cells, embedded_chars, initial_state=init_state, time_major=False)
# 相当于 以下4层
outputs, final_state = tf.nn.dynamic_rnn(cell, embedded_chars, initial_state=init_state, time_major=False)
outputs, final_state = tf.nn.dynamic_rnn(cell, outputs, initial_state=init_state, time_major=False)
outputs, final_state = tf.nn.dynamic_rnn(cell, outputs, initial_state=init_state, time_major=False)
outputs, final_state = tf.nn.dynamic_rnn(cell, outputs, initial_state=init_state, time_major=False)
# # outputs shape [batch,8,128] outputs[:, -1, :] shape [batch,128]
results=tf.layers.dense(outputs[:, -1, :],num_classes,activation=None)
'''
# 方法二：多层rnn+静态rnn
'''
cell = rnn.BasicLSTMCell(num_units=num_filters)
cell = rnn.DropoutWrapper(cell=cell, input_keep_prob=1.0, output_keep_prob=dropout_keep_prob)
cells = rnn.MultiRNNCell([cell] * 4, state_is_tuple=True) # 4个cell
init_state = cells.zero_state(batch_size, dtype=tf.float32)
inputs = tf.unstack(embedded_chars, 8, 1) # 拆成8个[-1,128]序列
outputs, final_state = tf.nn.static_rnn(cells, inputs, initial_state=init_state)
# outputs 8个[batch,128]
results = tf.layers.dense(outputs[-1], num_classes, activation=None)
'''
# 方法三：动态rnn
'''
cell = rnn.BasicLSTMCell(num_units=num_filters * 2)
init_state = cell.zero_state(batch_size, dtype=tf.float32)
outputs, final_state = tf.nn.dynamic_rnn(cell, embedded_chars, initial_state=init_state)
# # outputs shape [batch,8,256] outputs[:, -1, :] shape [batch,256]
results=tf.layers.dense(outputs[:, -1, :],num_classes,activation=None)
'''
# 方法四：静态rnn
cell = rnn.BasicLSTMCell(num_units=num_filters * 2)
init_state = cell.zero_state(batch_size, dtype=tf.float32)
inputs = tf.unstack(embedded_chars, 8, 1) # 拆成8个[-1,128]序列
outputs, final_state = tf.nn.static_rnn(cell,
inputs,
initial_state=init_state)
# outputs 8个[batch,256]
results = tf.layers.dense(outputs[-1], num_classes, activation=None)
# '''
return results
return text
# 我们可以通过将标记列表转换回单词来重新创建除标点符号和其他符号之外的文本
# tokens_to_string(x_train_tokens[1])
batch_size = 64
# ---------改成layer API--------------#
from tensorflow.contrib import slim, rnn
x = tf.placeholder(tf.float32, [None, 544])
Y = tf.placeholder(tf.float32, [
None,
])
inputs = slim.embed_sequence(ids=tf.cast(x, tf.int32),
vocab_size=num_words,
embed_dim=8,
scope='layer_embedding')
cell1 = rnn.GRUCell(16, activation=tf.nn.relu)
cell2 = rnn.GRUCell(8, activation=tf.nn.relu)
cell3 = rnn.GRUCell(4, activation=tf.nn.relu)
cell = rnn.MultiRNNCell([cell1, cell2, cell3], state_is_tuple=True)
initial_state = cell.zero_state(batch_size, tf.float32)
outputs, last_state = tf.nn.dynamic_rnn(cell,
inputs,
initial_state=initial_state,
scope='rnnlm')
predict = slim.fully_connected(outputs[:, -1, :], 2,
activation_fn=None) # [-1,2]