def compile(self,feature_shape = (None,200,1),label_max_string_length = 32,ms_output_size = 1423):
audio_ipt = Input(name="audio_input",shape=feature_shape)
layer_h1 = self.cnn2d_cell(32, audio_ipt)
layer_h2 = self.cnn2d_cell(64, layer_h1)
layer_h3 = self.cnn2d_cell(128, layer_h2)
layer_h4 = self.cnn2d_cell(128, layer_h3, pool=False)
layer_h5 = self.cnn2d_cell(128, layer_h4, pool=False)
layer_h6 = self.reshape1dfrom2d(layer_h5)
layer_h6 = Dropout(0.2)(layer_h6)
layer_h7 = Dense(256,activation="relu",kernel_initializer="he_normal")(layer_h6)
layer_h7 = Dropout(0.2)(layer_h7)
y_pred = Dense(ms_output_size, activation='softmax')(layer_h7)
y_true = Input(name='label_inputs', shape=[label_max_string_length], dtype='float32')
audio_length = Input(name='audio_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
loss_out = CTC_Batch_Cost()([y_true, y_pred, audio_length, label_length])
train_model = Model([audio_ipt, y_true, audio_length, label_length], [loss_out])
train_model.compile(optimizer="adam", loss={"ctc": lambda y_true, y_pred: y_pred})
base_model = Model(audio_ipt,y_pred)
self.built(train_model,base_model)
def compile(self,feature_shape = (1024,200),label_max_string_length = 32,ms_output_size = 1423):
audio_ipt = Input(name="audio_input", shape=feature_shape)
layer_h1 = self.cnn1d_cell(32, audio_ipt,pool=True,reshape=False)
layer_h2 = self.cnn1d_cell(32, layer_h1,pool=True,reshape=False)
layer_h3 = self.cnn1d_cell(64, layer_h2,pool=True,reshape=False)
layer_h4 = self.cnn1d_cell(64, layer_h3, pool=False)
layer_h5 = self.cnn1d_cell(128, layer_h4, pool=False)
layer_h5 = self.cnn1d_cell(128, layer_h5, pool=False)
layer_h5 = self.cnn1d_cell(128, layer_h5, pool=False)
layer_h5 = self.cnn1d_cell(128, layer_h5, pool=False) # TODO 考虑多叠加几层
# 64print(layer_h5)
layer_h6 = Dropout(0.2)(layer_h5) # KL，双Dense
layer_h7 = Dense(512, activation="relu", kernel_initializer="he_normal")(layer_h6) # TODO 考虑在这里加Attention
layer_h7 = Dropout(0.2)(layer_h7)
attention_prob = Dense(units=512, activation='softmax', name='attention_vec')(layer_h7)
attention_mul = Multiply()([layer_h7, attention_prob])
layer_h8 = Dense(ms_output_size)(attention_mul)
y_pred = Activation(activation="softmax")(layer_h8)
y_true = Input(name='label_inputs', shape=[label_max_string_length], dtype='float32')
audio_length = Input(name='audio_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
loss_out = CTC_Batch_Cost()([y_true, y_pred, audio_length, label_length])
train_model = Model([audio_ipt, y_true, audio_length, label_length], [loss_out])
train_model.compile(optimizer="adam", loss={"ctc": lambda y_true, y_pred: y_pred})
base_model = Model(audio_ipt, y_pred)
self.built(train_model,base_model)
def compile(self,
feature_shape=(256, 128, 1),
label_max_string_length=32,
ms_output_size=1242):
'''
建立模型[batch,times,vector,1]
:param feature_shape: 音频特征形状[timestamp,feature_vec_dim]
:param label_max_string_length: 标签最大长度
:param ms_output_size: 输出范围
:return:
'''
ipt = Input(name='audio_input', shape=feature_shape)
# eb1 = self.encoder_block(audio_ipt,128,128,position_embedding=False)
# eb2 = self.encoder_block(eb1,128,128,position_embedding = False)
layer_h1 = Conv2D(32,
3,
use_bias=False,
activation='relu',
padding='same',
kernel_initializer='he_normal')(ipt)
layer_h1 = Dropout(rate=0.05)(layer_h1)
layer_h2 = Conv2D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h1)
layer_h2 = MaxPooling2D()(layer_h2) # 池化层
layer_h3 = Dropout(rate=0.05)(layer_h2)
layer_h4 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h3) # 卷积层
layer_h4 = Dropout(rate=0.1)(layer_h4)
layer_h5 = Conv2D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h4) # 卷积层
layer_h5 = MaxPooling2D()(layer_h5) # 池化层
layer_h6 = Dropout(rate=0.1)(layer_h5)
layer_h7 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h6) # 卷积层
layer_h7 = Dropout(rate=0.15)(layer_h7)
layer_h8 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h7) # 卷积层
layer_h8 = MaxPooling2D()(layer_h8) # 池化层
layer_h9 = Dropout(0.15)(layer_h8)
layer_h10 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h9) # 卷积层
layer_h10 = Dropout(0.2)(layer_h10)
layer_h11 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h10) # 卷积层
layer_h11 = MaxPooling2D(pool_size=1, strides=None,
padding="valid")(layer_h11) # 池化层
layer_h12 = Dropout(0.2)(layer_h11)
layer_h13 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h12) # 卷积层
layer_h13 = Dropout(0.3)(layer_h13)
layer_h14 = Conv2D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h13) # 卷积层
layer_h14 = MaxPooling2D(pool_size=1, strides=None,
padding="valid")(layer_h14) # 池化层
layer_h16 = self.reshape1dfrom2d(layer_h14)
layer_h16 = Dropout(0.3)(layer_h16)
layer_h17 = Dense(128,
activation="relu",
kernel_initializer='he_normal')(layer_h16) # 全连接层
layer_h17 = Dropout(0.3)(layer_h17)
layer_h18 = Dense(ms_output_size,
kernel_initializer='he_normal')(layer_h17) # 全连接层
y_pred = Activation('softmax', name='Activation0')(layer_h18)
model_data = Model(inputs=ipt, outputs=y_pred)
# model_data.summary()
label_ipt = Input(name='label_inputs',
shape=[label_max_string_length],
dtype='float32')
audio_length = Input(name='audio_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
# Keras doesn't currently support loss funcs with extra parameters
# so CTC loss is implemented in a lambda layer
loss_out = CTC_Batch_Cost()(
[label_ipt, y_pred, audio_length, label_length])
train_model = Model([ipt, label_ipt, audio_length, label_length],
[loss_out])
train_model.compile(optimizer="adam",
loss={
"ctc": lambda y_true, y_pred: y_pred
})
base_model = Model(ipt, y_pred)
self.built(train_model, base_model)
def compile(self,
feature_shape=(256, 128),
label_max_string_length=32,
ms_output_size=1423):
self.ms_output_size = ms_output_size
ipt = Input(shape=feature_shape, name="audio_input")
layer_h1 = Bidirectional(LSTM(32, return_sequences=True),
merge_mode="concat")(ipt)
layer_h2 = Reshape(
(int(feature_shape[0] / 2), feature_shape[1]))(layer_h1)
layer_h3 = Bidirectional(LSTM(32, return_sequences=True),
merge_mode="concat")(layer_h2)
layer_h4 = Reshape(
(int(feature_shape[0] / 4), feature_shape[1]))(layer_h3)
layer_h4 = Conv1D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h4) # 卷积层
layer_h4 = Dropout(rate=0.1)(layer_h4)
layer_h5 = Conv1D(32,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h4) # 卷积层
layer_h6 = Dropout(rate=0.1)(layer_h5)
layer_h7 = Conv1D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h6) # 卷积层
layer_h7 = Dropout(rate=0.15)(layer_h7)
layer_h8 = Conv1D(64,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h7) # 卷积层
# layer_h8 = MaxPooling1D()(layer_h8) # 池化层
layer_h9 = Dropout(0.15)(layer_h8)
layer_h10 = Conv1D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h9) # 卷积层
layer_h10 = Dropout(0.2)(layer_h10)
layer_h11 = Conv1D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h10) # 卷积层
# layer_h11 = MaxPooling1D(pool_size=1, strides=None, padding="valid")(layer_h11) # 池化层
layer_h12 = Dropout(0.2)(layer_h11)
layer_h13 = Conv1D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h12) # 卷积层
layer_h13 = Dropout(0.3)(layer_h13)
layer_h14 = Conv1D(128,
3,
activation='relu',
padding='same',
kernel_initializer='he_normal')(layer_h13) # 卷积层
print(layer_h14)
layer_h17 = Dense(512,
activation="relu",
kernel_initializer='he_normal')(layer_h14) # 全连接层
layer_h17 = Dropout(0.3)(layer_h17)
layer_h18 = Dense(ms_output_size,
kernel_initializer='he_normal')(layer_h17) # 全连接层
y_true = Input(name='label_inputs',
shape=[label_max_string_length],
dtype='float32')
audio_length = Input(name='audio_length', shape=[1], dtype='int64')
label_length = Input(name='label_length', shape=[1], dtype='int64')
print(layer_h18)
y_pred = Activation('softmax', name='Activation0')(layer_h18)
loss_out = CTC_Batch_Cost()(
[y_true, y_pred, audio_length, label_length])
train_model = Model([ipt, y_true, audio_length, label_length],
[loss_out])
train_model.compile(optimizer="adam",
loss={
"ctc": lambda y_true, y_pred: y_pred
})
train_model.summary()
base_model = Model(ipt, y_pred)
self.built(train_model, base_model)