def compile(self,
feature_shape=(None, 200),
label_max_string_length=64,
ms_output_size=1438,
h_dim=128):
ipt = Input(shape=feature_shape, name="mfcc_input")
out = self.wave_conv1d_block(ipt, filters=h_dim)
skip = None
for _ in range(3):
for rate in [1, 2, 4, 8, 16]:
out, s = self.wave_residual_block(out,
kernal_size=7,
rate=rate,
filters=h_dim)
if skip is None:
skip = s
else:
skip = Add()([skip, s])
logit = self.wave_conv1d_block(skip, filters=64)
y_pred = self.wave_conv1d_block(
logit,
filters=ms_output_size,
bias=True,
activation="softmax",
)
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')
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=(1024, 200),
label_max_string_length=32,
ms_output_size=1423):
audio_ipt = Input(name="audio_input", shape=feature_shape)
parent_out = self.cnn1d_cell(32, audio_ipt, pool=True)
parent_out = self.cnn1d_cell(64, parent_out, pool=True)
parent_out = self.cnn1d_cell(64, parent_out, pool=True)
layer_h1 = self.conv1d_layers(parent_out, 64, 8)
layer_h2 = self.cnn1d_cell(64, layer_h1, pool=False)
layer_h3 = Add()([parent_out, layer_h2])
# 64print(layer_h5)
layer_h6 = Dropout(0.2)(layer_h3) # KL，双Dense
layer_h7 = Dense(256,
activation="relu",
kernel_initializer="he_normal")(
layer_h6) # TODO 考虑在这里加Attention
layer_h7 = Dropout(0.2)(layer_h7)
layer_h8 = Dense(ms_output_size)(layer_h7)
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)