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 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())
def conv_bn_relu(filters, twod=True, bn=True):
if twod:
subsubnet.add(
Conv2D(filters=filters, kernel_size=(4, 10), padding='same'))
else:
subsubnet.add(
Conv1D(filters=filters, kernel_size=9, padding='valid'))
if bn: subsubnet.add(BatchNorm())
subsubnet.add(Activation('relu'))
def dcgan(mark):
# Initiate model
model = GAN(z_dim=100, sample_shape=[28, 28, 1], mark=mark, classes=10)
# Define generator
model.G.add(Linear(output_dim=7 * 7 * 128))
model.G.add(Reshape(shape=[7, 7, 128]))
model.G.add(BatchNorm())
model.G.add(Activation.ReLU())
model.G.add(Deconv2D(filters=128, kernel_size=5, strides=2,
padding='same'))
model.G.add(BatchNorm())
model.G.add(Activation.ReLU())
model.G.add(Deconv2D(filters=1, kernel_size=5, strides=2, padding='same'))
model.G.add(Activation('sigmoid'))
# model.G.add(Activation('tanh'))
# model.G.add(Rescale(from_scale=[-1., 1.], to_scale=[0., 1.]))
# Define discriminator
# model.D.add(Rescale(from_scale=[0., 1.], to_scale=[-1., 1.]))
model.D.add(Conv2D(filters=128, kernel_size=5, strides=2, padding='same'))
model.D.add(Activation.LeakyReLU())
model.D.add(Conv2D(filters=128, kernel_size=5, strides=2, padding='same'))
model.D.add(BatchNorm())
model.D.add(Activation.LeakyReLU())
model.D.add(Reshape(shape=[7 * 7 * 128]))
model.D.add(Linear(output_dim=1))
model.D.add(Activation('sigmoid'))
# Build model
optimizer = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5)
model.build(loss=pedia.cross_entropy,
G_optimizer=optimizer,
D_optimizer=optimizer)
return model
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'))
def dcgan(mark):
model = GAN(z_dim=100, sample_shape=[32, 32, 3], mark=mark, classes=10)
nch = 256
h = 5
reg = regularizers.L2(strength=1e-7)
# Define generator
model.G.add(Linear(output_dim=nch * 4 * 4, weight_regularizer=reg))
model.G.add(BatchNorm())
model.G.add(Reshape(shape=(4, 4, nch)))
model.G.add(
Deconv2D(filters=int(nch / 2),
kernel_size=5,
padding='same',
kernel_regularizer=reg))
model.G.add(BatchNorm())
model.G.add(Activation.LeakyReLU())
model.G.add(
Deconv2D(filters=int(nch / 2),
kernel_size=5,
strides=2,
padding='same',
kernel_regularizer=reg))
model.G.add(BatchNorm())
model.G.add(Activation.LeakyReLU())
model.G.add(
Deconv2D(filters=int(nch / 4),
kernel_size=5,
strides=2,
padding='same',
kernel_regularizer=reg))
model.G.add(BatchNorm())
model.G.add(Activation.LeakyReLU())
model.G.add(
Deconv2D(filters=3,
kernel_size=5,
strides=2,
padding='same',
kernel_regularizer=reg))
model.G.add(Activation('sigmoid'))
# ===========================================================================
# Define discriminator
model.D.add(
Conv2D(filters=int(nch / 4),
kernel_size=h,
strides=2,
padding='same',
kernel_regularizer=reg))
model.D.add(Activation.LeakyReLU())
model.D.add(
Conv2D(filters=int(nch / 2),
kernel_size=h,
strides=2,
padding='same',
kernel_regularizer=reg))
model.D.add(Activation.LeakyReLU())
model.D.add(
Conv2D(filters=nch,
kernel_size=h,
strides=2,
padding='same',
kernel_regularizer=reg))
model.D.add(Activation.LeakyReLU())
model.D.add(Flatten())
model.D.add(Linear(output_dim=1, weight_regularizer=reg))
model.D.add(Activation('sigmoid'))
# Build model
optimizer = tf.train.AdamOptimizer(learning_rate=0.0002, beta1=0.5)
model.build(loss=pedia.cross_entropy,
G_optimizer=optimizer,
D_optimizer=optimizer)
return model
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