OrgLion-ML
/
ASRT_SpeechRecognition
mirror of https://github.com/nl8590687/ASRT_SpeechRecognition.git
1
0
Fork 0
Code Issues Releases Wiki Activity

feat: 添加SpeechModel251BN模型

This commit is contained in:
nl 2022-03-16 14:31:16 +08:00
parent 94b681fefc
commit 8e6a1bf0d8
1 changed files with 126 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

126
speech_model_zoo.py
View File

@ -76,6 +76,132 @@ def ctc_lambda_func(args):
y_pred = y_pred[:, :, :]
return K.ctc_batch_cost(labels, y_pred, input_length, label_length)
class SpeechModel251BN(BaseModel):
'''
定义CNN+CTC模型，使用函数式模型
输入层：200维的特征值序列，一条语音数据的最大长度设为1600（大约16s）\\
隐藏层：卷积池化层，卷积核大小为3x3，池化窗口大小为2 \\
隐藏层：全连接层 \\
输出层：全连接层，神经元数量为self.MS_OUTPUT_SIZE，使用softmax作为激活函数， \\
CTC层：使用CTC的loss作为损失函数，实现连接性时序多输出
参数： \\
input_shape: tuple，默认值(1600, 200, 1) \\
output_shape: tuple，默认值(200, 1428)
'''
def __init__(self, input_shape :tuple=(1600, 200, 1), output_size :int=1428) -> None:
super().__init__()
self.input_shape = input_shape
self._pool_size = 8
self.output_shape = (input_shape[0] // self._pool_size, output_size)
self._model_name = 'SpeechModel251bn'
self.model, self.model_base = self._define_model(self.input_shape, self.output_shape[1])
def _define_model(self, input_shape, output_size) -> tuple:
label_max_string_length = 64
input_data = Input(name='the_input', shape=input_shape)
layer_h = Conv2D(32, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv0')(input_data) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN0')(layer_h)
layer_h = Activation('relu', name='Act0')(layer_h)
layer_h = Conv2D(32, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv1')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN1')(layer_h)
layer_h = Activation('relu', name='Act1')(layer_h)
layer_h = MaxPooling2D(pool_size=2, strides=None, padding="valid")(layer_h) # 池化层
layer_h = Conv2D(64, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv2')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN2')(layer_h)
layer_h = Activation('relu', name='Act2')(layer_h)
layer_h = Conv2D(64, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv3')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN3')(layer_h)
layer_h = Activation('relu', name='Act3')(layer_h)
layer_h = MaxPooling2D(pool_size=2, strides=None, padding="valid")(layer_h) # 池化层
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv4')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN4')(layer_h)
layer_h = Activation('relu', name='Act4')(layer_h)
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv5')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN5')(layer_h)
layer_h = Activation('relu', name='Act5')(layer_h)
layer_h = MaxPooling2D(pool_size=2, strides=None, padding="valid")(layer_h) # 池化层
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv6')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN6')(layer_h)
layer_h = Activation('relu', name='Act6')(layer_h)
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv7')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN7')(layer_h)
layer_h = Activation('relu', name='Act7')(layer_h)
layer_h = MaxPooling2D(pool_size=1, strides=None, padding="valid")(layer_h) # 池化层
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv8')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN8')(layer_h)
layer_h = Activation('relu', name='Act8')(layer_h)
layer_h = Conv2D(128, (3,3), use_bias=True, padding='same', kernel_initializer='he_normal', name='Conv9')(layer_h) # 卷积层
layer_h = BatchNormalization(epsilon=0.0002, name='BN9')(layer_h)
layer_h = Activation('relu', name='Act9')(layer_h)
layer_h = MaxPooling2D(pool_size=1, strides=None, padding="valid")(layer_h) # 池化层
#test=Model(inputs = input_data, outputs = layer_h12)
#test.summary()
layer_h = Reshape((self.output_shape[0], input_shape[1] // self._pool_size * 128), name='Reshape0')(layer_h) #Reshape层
layer_h = Dense(128, activation="relu", use_bias=True, kernel_initializer='he_normal', name='Dense0')(layer_h) # 全连接层
layer_h = Dense(output_size, use_bias=True, kernel_initializer='he_normal', name='Dense1')(layer_h) # 全连接层
y_pred = Activation('softmax', name='Activation0')(layer_h)
model_base = Model(inputs = input_data, outputs = y_pred)
#model_data.summary()
labels = Input(name='the_labels', shape=[label_max_string_length], dtype='float32')
input_length = Input(name='input_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 = Lambda(ctc_lambda_func, output_shape=(1,), name='ctc')([y_pred, labels, input_length, label_length])
model = Model(inputs=[input_data, labels, input_length, label_length], outputs=loss_out)
return model, model_base
def get_loss_function(self) -> dict:
return {'ctc': lambda y_true, y_pred: y_pred}
def forward(self, data_input):
batch_size = 1
in_len = np.zeros((batch_size),dtype = np.int32)
in_len[0] = self.output_shape[0]
x_in = np.zeros((batch_size,) + self.input_shape, dtype=np.float)
for i in range(batch_size):
x_in[i,0:len(data_input)] = data_input
base_pred = self.model_base.predict(x = x_in)
r = K.ctc_decode(base_pred, in_len, greedy = True, beam_width=100, top_paths=1)
if(tf.__version__[0:2] == '1.'):
r1 = r[0][0].eval(session=tf.compat.v1.Session())
else:
r1 = r[0][0].numpy()
speech_result = ctc_decode_delete_tail_blank(r1[0])
return speech_result
class SpeechModel251(BaseModel):
'''
定义CNN+CTC模型，使用函数式模型