Keras api 提前知道：
- BatchNormalization, 用来加快每次迭代中的训练速度
Normalize the activations of the previous layer at each batch, i.e. applies a transformation that maintains the mean activation close to 0 and the activation standard deviation close to 1.
- TimeDistributed, 总的来说TimeDistributed层在每个时间步上均操作了Dense，比单一dense操作更能发现数据集中比较复杂的模式
简单的理解：
- keras中TimeDistributed的用法
更进一步的理解： - How to Use the TimeDistributed Layer for Long Short-Term Memory Networks in Python
- 1对应的翻译
- Bidrectional, keras封装了的双向包装函数。
Keras 相关导入
from keras import backend as K
from keras.models import Model
from keras.layers import (BatchNormalization, Conv1D, Conv2D, Dense, Input, Dropout,
TimeDistributed, Activation, Bidirectional, SimpleRNN, GRU, LSTM, MaxPooling1D, Flatten, MaxPooling2D)
RNN
def simple_rnn_model(input_dim, output_dim=29):
""" Build a recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add recurrent layer
simp_rnn = GRU(output_dim, return_sequences=True,
implementation=2, name='rnn')(input_data)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(simp_rnn)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
或者直接使用Keras SimpleRNN
rnn + timedistribute
def rnn_model(input_dim, units, activation, output_dim=29):
""" Build a recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add recurrent layer
simp_rnn = LSTM(units, activation=activation,
return_sequences=True, implementation=2, name='rnn')(input_data)
# TODO: Add batch normalization
bn_rnn = BatchNormalization()(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bn_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax', )(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
cnn+rnn+timedistribute
def cnn_output_length(input_length, filter_size, border_mode, stride,
dilation=1):
""" Compute the length of the output sequence after 1D convolution along
time. Note that this function is in line with the function used in
Convolution1D class from Keras.
Params:
input_length (int): Length of the input sequence.
filter_size (int): Width of the convolution kernel.
border_mode (str): Only support `same` or `valid`.
stride (int): Stride size used in 1D convolution.
dilation (int)
"""
if input_length is None:
return None
assert border_mode in {'same', 'valid', 'causal', 'full'}
dilated_filter_size = filter_size + (filter_size - 1) * (dilation - 1)
if border_mode == 'same':
output_length = input_length
elif border_mode == 'valid':
output_length = input_length - dilated_filter_size + 1
elif border_mode == 'causal':
output_length = input_length
elif border_mode == 'full':
output_length = input_length + dilated_filter_size - 1
return (output_length + stride - 1) // stride
def cnn_rnn_model(input_dim, filters, kernel_size, conv_stride,
conv_border_mode, units, output_dim=29):
""" Build a recurrent + convolutional network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# Add convolutional layer
conv_1d = Conv1D(filters, kernel_size,
strides=conv_stride,
padding=conv_border_mode,
activation='relu',
name='conv1d')(input_data)
# Add batch normalization
bn_cnn = BatchNormalization(name='bn_conv_1d')(conv_1d)
# Add a recurrent layer
simp_rnn = SimpleRNN(units, activation='relu',
return_sequences=True, implementation=2, name='rnn')(bn_cnn)
# TODO: Add batch normalization
bn_rnn = BatchNormalization()(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bn_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: cnn_output_length(
x, kernel_size, conv_border_mode, conv_stride)
print(model.summary())
return model
deep rnn + timedistribute
def deep_rnn_model(input_dim, units, recur_layers, output_dim=29):
""" Build a deep recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Add recurrent layers, each with batch normalization
# Add a recurrent layer
for i in range(recur_layers):
if i:
simp_rnn = GRU(units, return_sequences=True,
implementation=2)(simp_rnn)
else:
simp_rnn = GRU(units, return_sequences=True,
implementation=2)(input_data)
# TODO: Add batch normalization
bn_rnn = BatchNormalization()(simp_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bn_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model
bidirection rnn + timedistribute
def bidirectional_rnn_model(input_dim, units, output_dim=29):
""" Build a bidirectional recurrent network for speech
"""
# Main acoustic input
input_data = Input(name='the_input', shape=(None, input_dim))
# TODO: Add bidirectional recurrent layer
bidir_rnn = Bidirectional(GRU(units, return_sequences=True))(input_data)
bidir_rnn = BatchNormalization()(bidir_rnn)
# TODO: Add a TimeDistributed(Dense(output_dim)) layer
time_dense = TimeDistributed(Dense(output_dim))(bidir_rnn)
# Add softmax activation layer
y_pred = Activation('softmax', name='softmax')(time_dense)
# Specify the model
model = Model(inputs=input_data, outputs=y_pred)
model.output_length = lambda x: x
print(model.summary())
return model