def build_model(num_class, max_features=15000, dim=200, max_len=300, dropout_rate=0.2, gru_size=100):
model = Sequential()
model.add(Embedding(max_features+1, dim, input_length=max_len))
model.add(SpatialDropout1D(dropout_rate))
model.add(SeparableConv1D(32, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(SeparableConv1D(64, kernel_size=3, padding='same', activation='relu'))
model.add(MaxPooling1D(pool_size=2))
model.add(GRU(gru_size))
model.add(Dense(num_class, activation='sigmoid', kernel_initializer='normal'))
model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=[total_acc, binary_acc])
return model
def C(filters, kernel_size, input_dim, conv_border_mode, conv_stride,
dilation): # Layer C1: first layer
""" Build a deep network for speech
"""
# Main acoustic input
input_data = Input(shape=(None, input_dim))
# TODO: Specify the layers in your network
conv = SeparableConv1D(filters=filters,
kernel_size=kernel_size,
strides=conv_stride,
padding=conv_border_mode,
dilation_rate=dilation,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='c')(input_data)
conv_bn = BatchNormalization(name='batch_norm1')(conv)
relu = Activation('relu', name='relu1')(conv_bn)
# Specify the model
model = Model(inputs=input_data, outputs=relu)
# TODO: Specify model.output_length
#model.output_length = lambda x: cnn_output_length(x,kernel_size,conv_border_mode,conv_stride,dilation)
return model
def Pointwise(filters, kernel_size, input_dim, conv_border_mode, conv_stride,
dilation):
""" Build a deep network for speech
"""
# Main acoustic input
input_data = Input(shape=(None, input_dim))
# TODO: Specify the layers in your network
conv = SeparableConv1D(filters=filters,
kernel_size=kernel_size,
strides=1,
padding=conv_border_mode,
name='c')(input_data)
# Specify the model
model = Model(inputs=input_data, outputs=conv)
# TODO: Specify model.output_length
#model.output_length = lambda x: cnn_output_length(x,kernel_size,conv_border_mode,conv_stride,dilation)
return model
def create_channel(x, filter_size, feature_map):
"""
Creates a layer, working channel wise
Arguments:
x : Input for convoltuional channel
filter_size : Filter size for creating Conv1D
feature_map : Feature map
Returns:
x : Channel including (Conv1D + GlobalMaxPooling + Dense + Dropout)
"""
x = SeparableConv1D(feature_map,
kernel_size=filter_size,
activation='relu',
strides=1,
padding='same',
depth_multiplier=4)(x)
x = GlobalMaxPooling1D()(x)
x = Dense(hidden_units)(x)
x = Dropout(dropout_rate)(x)
return x
def tcsconv(filters, kernel_size,
input_dim): # Layer B: output_length = input_length
input_data = Input(shape=(None, input_dim))
x0 = SeparableConv1D(
filters,
kernel_size=1, # x0: pointwise + batchnorm
strides=1,
padding='same',
activation=None,
name='c1')(input_data)
x0 = BatchNormalization(name='batch_norm1')(x0)
# x1*1
x1 = SeparableConv1D(
filters,
kernel_size, # x1: normal separable conv + batchnorm + relu
strides=1,
padding='same',
data_format='channels_last',
dilation_rate=1,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='layer_separable_conv1')(input_data)
x1 = BatchNormalization(name='conv_batch_norm1')(x1)
x1 = Activation('relu', name='relu1')(x1)
# x1*2
x1 = SeparableConv1D(
filters,
kernel_size, # x1: normal separable conv + batchnorm + relu
strides=1,
padding='same',
data_format='channels_last',
dilation_rate=1,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='layer_separable_conv2')(x1)
x1 = BatchNormalization(name='conv_batch_norm2')(x1)
x1 = Activation('relu', name='relu2')(x1)
# x1*3
x1 = SeparableConv1D(
filters,
kernel_size, # x1: normal separable conv + batchnorm + relu
strides=1,
padding='same',
data_format='channels_last',
dilation_rate=1,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='layer_separable_conv3')(x1)
x1 = BatchNormalization(name='conv_batch_norm3')(x1)
x1 = Activation('relu', name='relu3')(x1)
# x1*4
x1 = SeparableConv1D(
filters,
kernel_size, # x1: normal separable conv + batchnorm + relu
strides=1,
padding='same',
data_format='channels_last',
dilation_rate=1,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='layer_separable_conv4')(x1)
x1 = BatchNormalization(name='conv_batch_norm4')(x1)
x1 = Activation('relu', name='relu4')(x1)
# x2*1
x2 = SeparableConv1D(
filters,
kernel_size, # x2: normal separable conv + batchnorm
strides=1,
padding='same',
data_format='channels_last',
dilation_rate=1,
depth_multiplier=1,
activation=None,
use_bias=True,
depthwise_initializer='glorot_uniform',
pointwise_initializer='glorot_uniform',
bias_initializer='zeros',
name='layer_separable_conv')(x1)
x2 = BatchNormalization(name='conv_batch_norm')(x2)
added = Add()([x0, x2])
out = Activation('relu', name='relu')(
added) # out: out(x3) := ( x1 ->x1 ->x1 ->x1 ->x2) + x0
tcsconv_B = Model(inputs=input_data, outputs=out)
#tcsconv_B.output_length = lambda x: input_length
return tcsconv_B