def conv_2d_test(th):
assert isinstance(th, Config)
# Initiate model
th.mark = 'cnn_2d' + th.mark
def data_dim(sample_rate=44100, duration=2, n_mfcc=40):
audio_length = sample_rate * duration
dim = (n_mfcc, 1 + int(np.floor(audio_length / 512)), 1)
return dim
dim = data_dim()
model = Classifier(mark=th.mark)
# Add input layer
model.add(Input(sample_shape=[dim[0], dim[1], 1]))
# Add hidden layers
model.add(Conv2D(32, (4, 10), padding='same'))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
# model.add(Dropout(0.7))
model.add(Conv2D(32, (4, 10), padding='same'))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
# model.add(Dropout(0.7))
model.add(Conv2D(32, (4, 10), padding='same'))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
# model.add(Dropout(0.7))
model.add(Conv2D(32, (4, 10), padding='same'))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
# model.add(Dropout(0.7))
model.add(Flatten())
model.add(Linear(output_dim=64))
model.add(BatchNorm())
model.add(Activation('relu'))
# Add output layer
model.add(Linear(output_dim=41))
model.add(Activation('softmax'))
# Build model
optimizer = tf.train.AdamOptimizer(learning_rate=th.learning_rate)
model.build(optimizer=optimizer)
return model
def _get_layers(self):
"""The architecture follows a common pattern used in related literatures,
that is, having a stem followed by stacks of cells. The image-like
tensors will be down sampled after each stack. Finally a global average
pooling layer is added."""
layers = []
# Add stem
layers.extend(self.conv_bn_relu(self.stem_channels, 3))
# Add stacks
output_channels = self.stem_channels
for i in range(self.num_stacks):
input_channels = output_channels
# Down sample except in first cell
if i > 0:
layers.append(MaxPool2D(2, 2))
# Double channel number
output_channels *= 2
for _ in range(self.cells_per_stack):
layers.append(self._get_cell(input_channels, output_channels))
input_channels = output_channels
# Add global average pooling layer
layers.append(GlobalAveragePooling2D())
return layers
def deep_conv(mark):
# Initiate predictor
model = Classifier(mark=mark)
model.add(Input(sample_shape=config.sample_shape))
def ConvBNReLU(filters, strength=1.0, bn=True):
model.add(
Conv2D(filters=filters,
kernel_size=5,
padding='same',
kernel_regularizer=regularizers.L2(strength=strength)))
if bn:
model.add(BatchNorm())
model.add(Activation('relu'))
# Conv layers
reg = 1e-5
ConvBNReLU(32, reg)
model.add(Dropout(0.5))
ConvBNReLU(32, reg)
model.add(MaxPool2D(2, 2, 'same'))
ConvBNReLU(64, reg)
model.add(Dropout(0.5))
ConvBNReLU(64, reg)
model.add(MaxPool2D(2, 2, 'same'))
ConvBNReLU(128, reg)
# FC layers
model.add(Flatten())
model.add(Linear(256))
# model.add(BatchNorm())
model.add(Activation('relu'))
model.add(Linear(256))
# model.add(BatchNorm())
model.add(Activation('relu'))
model.add(Linear(config.y_dim))
# Build model
model.build(optimizer=tf.train.AdamOptimizer(learning_rate=1e-4))
return model
def ka_convnet(mark):
model = Classifier(mark=mark)
model.add(Input(sample_shape=config.sample_shape))
strength = 1e-5
def ConvLayer(filters, bn=False):
model.add(
Conv2D(filters=filters,
kernel_size=5,
padding='same',
kernel_regularizer=regularizers.L2(strength=strength)))
if bn:
model.add(BatchNorm())
model.add(Activation.ReLU())
# Define structure
ConvLayer(32)
model.add(Dropout(0.5))
ConvLayer(32, False)
model.add(Dropout(0.5))
model.add(MaxPool2D(2, 2, 'same'))
ConvLayer(64, True)
model.add(Dropout(0.5))
model.add(MaxPool2D(2, 2, 'same'))
model.add(Flatten())
model.add(Linear(128))
model.add(Activation.ReLU())
# model.add(Dropout(0.5))
model.add(Linear(10))
# Build model
model.build(optimizer=tf.train.AdamOptimizer(learning_rate=1e-4))
return model
def maxpool_drop(pool_size, strides, twod=True, drop=True):
if twod:
subsubnet.add(MaxPool2D(pool_size=pool_size, strides=strides))
else:
subsubnet.add(MaxPool1D(pool_size=pool_size, strides=strides))
if drop: subsubnet.add(Dropout(th.raw_keep_prob))
def res_00(th):
assert isinstance(th, Config)
model = Classifier(mark=th.mark)
def data_dim(sample_rate=16000, duration=2, n_mfcc=50):
audio_length = sample_rate * duration
dim = (n_mfcc, 1 + int(np.floor(audio_length / 512)), 1)
return dim
dim = data_dim()
# Add hidden layers
subnet = model.add(inter_type=model.CONCAT)
# the net to process raw data
subsubnet = subnet.add()
# subsubnet.add(Input(sample_shape=[32000, 1], name='raw_data'))
subsubnet.add(Input(sample_shape=[32000, 1]))
subsubnet.add(Conv1D(filters=16, kernel_size=9, padding='valid'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(Conv1D(filters=16, kernel_size=9, padding='valid'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool1D(pool_size=16, strides=16))
subsubnet.add(Dropout(th.raw_keep_prob))
subsubnet.add(Conv1D(filters=32, kernel_size=3, padding='valid'))
subsubnet.add(Activation('relu'))
subsubnet.add(Conv1D(filters=32, kernel_size=3, padding='valid'))
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool1D(pool_size=4, strides=4))
subsubnet.add(Dropout(th.raw_keep_prob))
subsubnet.add(Conv1D(filters=32, kernel_size=3, padding='valid'))
subsubnet.add(Activation('relu'))
subsubnet.add(Conv1D(filters=32, kernel_size=3, padding='valid'))
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool1D(pool_size=4, strides=4))
subsubnet.add(Conv1D(filters=256, kernel_size=3, padding='valid'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(Conv1D(filters=256, kernel_size=3, padding='valid'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(GlobalMaxPooling1D())
# the net to process mfcc features
subsubnet = subnet.add()
subsubnet.add(Input(sample_shape=[dim[0], dim[1], 1], name='mfcc'))
subsubnet.add(Conv2D(32, (4, 10), padding='same'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(th.mfcc_keep_prob))
net = subsubnet.add(ResidualNet())
net.add(Conv2D(32, (4, 10), padding='same'))
net.add(BatchNorm())
net.add_shortcut()
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(th.mfcc_keep_prob))
#
net = subsubnet.add(ResidualNet())
net.add(Conv2D(32, (4, 10), padding='same'))
net.add(BatchNorm())
net.add_shortcut()
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(th.mfcc_keep_prob))
net = subsubnet.add(ResidualNet())
net.add(Conv2D(32, (4, 10), padding='same'))
net.add(BatchNorm())
net.add_shortcut()
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(th.mfcc_keep_prob))
net = subsubnet.add(ResidualNet())
net.add(Conv2D(32, (4, 10), padding='same'))
net.add(BatchNorm())
net.add_shortcut()
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(th.mfcc_keep_prob))
subsubnet.add(Flatten())
subsubnet.add(Dropout(th.concat_keep_prob))
model.add(Linear(output_dim=128))
model.add(BatchNorm())
model.add(Activation('relu'))
#
model.add(Linear(output_dim=64))
model.add(BatchNorm())
model.add(Activation('relu'))
# Add output layer
model.add(Linear(output_dim=41))
model.add(Activation('softmax'))
# Build model
optimizer = tf.train.AdamOptimizer(learning_rate=th.learning_rate)
model.build(optimizer=optimizer)
return model
def multinput_mlp(th):
assert isinstance(th, Config)
model = Classifier(mark=th.mark)
def data_dim(sample_rate=16000, duration=2, n_mfcc=50):
audio_length = sample_rate * duration
dim = (n_mfcc, 1 + int(np.floor(audio_length / 512)), 1)
return dim
dim = data_dim()
# Add hidden layers
subnet = model.add(inter_type=model.CONCAT)
subsubnet = subnet.add()
subsubnet.add(Input(sample_shape=[32000, 1]))
subsubnet.add(Linear(output_dim=512))
subsubnet.add(Activation('relu'))
subsubnet.add(Linear(output_dim=256))
subsubnet.add(Activation('relu'))
subsubnet = subnet.add()
subsubnet.add(Input(sample_shape=[dim[0], dim[1], 1], name='mfcc'))
subsubnet.add(Conv2D(32, (4, 10), padding='same'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(0.8))
# subsubnet.add(Conv2D(32, (4, 10), padding='same'))
# subsubnet.add(BatchNorm())
# subsubnet.add(Activation('relu'))
# subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
# subsubnet.add(Dropout(0.8))
subsubnet.add(Conv2D(32, (4, 10), padding='same'))
subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
subsubnet.add(MaxPool2D(pool_size=(2, 2), strides=(2, 2)))
subsubnet.add(Dropout(0.7))
subsubnet.add(Flatten())
model.add(Linear(output_dim=128))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(Linear(output_dim=64))
model.add(BatchNorm())
model.add(Activation('relu'))
model.add(Linear(output_dim=64))
model.add(BatchNorm())
model.add(Activation('relu'))
# Add output layer
model.add(Linear(output_dim=41))
model.add(Activation('softmax'))
# Build model
optimizer = tf.train.AdamOptimizer(learning_rate=th.learning_rate)
model.build(optimizer=optimizer)
return model