def model(self):
conv0 = layers.conv1d(
inputs=self.embed,
filters=100,
k_size=3,
stride=1,
padding="SAME",
scope_name="conv0",
)
relu0 = layers.relu(inputs=conv0, scope_name="relu0")
pool0 = layers.one_maxpool(inputs=relu0,
padding="VALID",
scope_name="pool0")
flatten0 = layers.flatten(inputs=pool0, scope_name="flatten0")
conv1 = layers.conv1d(
inputs=self.embed,
filters=100,
k_size=4,
stride=1,
padding="SAME",
scope_name="conv1",
)
relu1 = layers.relu(inputs=conv1, scope_name="relu0")
pool1 = layers.one_maxpool(inputs=relu1,
padding="VALID",
scope_name="pool1")
flatten1 = layers.flatten(inputs=pool1, scope_name="flatten1")
conv2 = layers.conv1d(
inputs=self.embed,
filters=100,
k_size=5,
stride=1,
padding="SAME",
scope_name="conv2",
)
relu2 = layers.relu(inputs=conv2, scope_name="relu0")
pool2 = layers.one_maxpool(inputs=relu2,
padding="VALID",
scope_name="pool2")
flatten2 = layers.flatten(inputs=pool2, scope_name="flatten2")
concat0 = layers.concatinate([flatten0, flatten1, flatten2],
scope_name="concat0")
dropout0 = layers.Dropout(inputs=concat0,
rate=1 - self.keep_prob,
scope_name="dropout0")
self.logits = layers.fully_connected(inputs=dropout0,
out_dim=self.n_classes,
scope_name="fc0")
def __init__(self,
num_time_samples,
num_channels=1,
num_classes=256,
num_blocks=2,
num_layers=14,
num_hidden=128,
gpu_fraction=1.0):
self.num_time_samples = num_time_samples
self.num_channels = num_channels
self.num_classes = num_classes
self.num_blocks = num_blocks
self.num_layers = num_layers
self.num_hidden = num_hidden
self.gpu_fraction = gpu_fraction
inputs = tf.compat.v1.placeholder(
tf.float32, shape=(None, num_time_samples, num_channels))
targets = tf.compat.v1.placeholder(
tf.int32, shape=(None, num_time_samples))
h = inputs
hs = []
for b in range(num_blocks):
for i in range(num_layers):
rate = 2**i
name = 'b{}-l{}'.format(b, i)
h = dilated_conv1d(h, num_hidden, rate=rate, name=name)
hs.append(h)
outputs = conv1d(h,
num_classes,
filter_width=1,
gain=1.0,
activation=None,
bias=True)
costs = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=outputs, labels=targets)
cost = tf.compat.v1.reduce_mean(costs)
train_step = tf.compat.v1.train.AdamOptimizer(
learning_rate=0.001).minimize(cost)
# gpu_options = tf.GPUOptions(
# per_process_gpu_memory_fraction=gpu_fraction)
# sess = tf.compat.v1.Session(
# config=tf.ConfigProto(gpu_options=gpu_options))
sess = tf.compat.v1.Session()
sess.run(tf.compat.v1.global_variables_initializer())
self.inputs = inputs
self.targets = targets
self.outputs = outputs
self.hs = hs
self.costs = costs
self.cost = cost
self.train_step = train_step
self.sess = sess
def __init__(self, num_time_samples, num_channels, gpu_fraction):
inputs = tf.placeholder(tf.float32,
shape=(None, num_time_samples, num_channels))
targets = tf.placeholder(tf.int32, shape=(None, num_time_samples))
h = inputs
for b in range(2):
for i in range(14):
rate = 2**i
name = 'b{}-l{}'.format(b, i)
h = dilated_conv1d(h, 128, rate=rate, name=name)
outputs = conv1d(h,
256,
filter_width=1,
gain=1.0,
activation=None,
bias=True)
cost = tf.reduce_mean(
tf.nn.sparse_softmax_cross_entropy_with_logits(outputs, targets))
train_step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(cost)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=gpu_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
sess.run(tf.initialize_all_variables())
self.inputs = inputs
self.targets = targets
self.outputs = outputs
self.cost = cost
self.train_step = train_step
self.sess = sess
def audioencoder(input_tensor, dropout_rate, num_hidden_layers):
L1 = conv1d(input_tensor=input_tensor,
filters=num_hidden_layers,
kernel_size=1,
strides=1,
padding="CAUSAL",
dilation_rate=1,
activation=tf.nn.relu,
dropout_rate=dropout_rate)
L2 = conv1d(input_tensor=L1,
filters=None,
kernel_size=1,
strides=1,
padding="CAUSAL",
dilation_rate=1,
activation=tf.nn.relu,
dropout_rate=dropout_rate)
L3 = conv1d(input_tensor=L2,
filters=None,
kernel_size=1,
strides=1,
padding="CAUSAL",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate)
L4 = highwaynet(input_tensor=L3,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(1))
L5 = highwaynet(input_tensor=L4,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(2))
L6 = highwaynet(input_tensor=L5,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=9,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(3))
L7 = highwaynet(input_tensor=L6,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=27,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(4))
L8 = highwaynet(input_tensor=L7,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(5))
L9 = highwaynet(input_tensor=L8,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(6))
L10 = highwaynet(input_tensor=L9,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=9,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(7))
L11 = highwaynet(input_tensor=L10,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=27,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(8))
L12 = highwaynet(input_tensor=L11,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(9))
L13 = highwaynet(input_tensor=L12,
filters=None,
kernel_size=3,
strides=1,
padding="CAUSAL",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="audioencoder_highwaynet_Block" + str(10))
return L13
def textencoder(embeding_tensor, dropout_rate, num_hidden_layers):
L1 = conv1d(input_tensor=embeding_tensor,
filters=num_hidden_layers * 2,
kernel_size=1,
strides=1,
padding="SAME",
dilation_rate=1,
activation=tf.nn.relu,
dropout_rate=dropout_rate)
L2 = conv1d(input_tensor=L1,
filters=None,
kernel_size=1,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate)
L3 = highwaynet(input_tensor=L2,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(1))
L4 = highwaynet(input_tensor=L3,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(2))
L5 = highwaynet(input_tensor=L4,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=9,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(3))
L6 = highwaynet(input_tensor=L5,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=27,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(4))
L7 = highwaynet(input_tensor=L6,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(5))
L8 = highwaynet(input_tensor=L7,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=3,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(6))
L9 = highwaynet(input_tensor=L8,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=9,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(7))
L10 = highwaynet(input_tensor=L9,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=27,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(8))
L11 = highwaynet(input_tensor=L10,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(9))
L12 = highwaynet(input_tensor=L11,
filters=None,
kernel_size=3,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(10))
L13 = highwaynet(input_tensor=L12,
filters=None,
kernel_size=1,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(11))
L14 = highwaynet(input_tensor=L13,
filters=None,
kernel_size=1,
strides=1,
padding="SAME",
dilation_rate=1,
activation=None,
dropout_rate=dropout_rate,
scope_name="textencoder_highwaynet_Block" + str(12))
K, V = tf.split(L14, 2, axis=-1)
return K, V
def __init__(
self,
# num_time_samples,
num_channels=1,
num_classes=256,
num_blocks=2,
num_layers=14,
num_hidden=128,
gpu_fraction=1.0,
prob_model_type='softmax',
):
# self.num_time_samples = num_time_samples
self.num_channels = num_channels
self.num_classes = num_classes
self.num_blocks = num_blocks
self.num_layers = num_layers
self.num_hidden = num_hidden
self.gpu_fraction = gpu_fraction
self.prob_model_type = prob_model_type
inputs = tf.placeholder(tf.float32, shape=(None, None, num_channels))
# targets = tf.placeholder(tf.int32, shape=(None, num_time_samples))
h = inputs
hs = []
for b in range(num_blocks):
for i in range(num_layers):
rate = 2**i
name = 'b{}-l{}'.format(b, i)
h = dilated_conv1d(h, num_hidden, rate=rate, name=name)
hs.append(h)
outputs = conv1d(h,
num_classes,
filter_width=1,
gain=1.0,
activation=None,
bias=True)
raw_outputs_shape = tf.shape(outputs)
outputs = tf.reshape(outputs, [-1, num_classes])
if prob_model_type == 'softmax':
self.prob_model = MultinomialLayer(outputs,
num_classes,
num_classes,
one_hot=False)
elif prob_model_type == 'sdp':
self.prob_model = LocallySmoothedMultiscaleLayer(outputs,
num_classes,
num_classes,
one_hot=False,
k=1,
lam=0)
train_step = tf.train.AdamOptimizer(learning_rate=0.001).minimize(
self.prob_model.train_loss)
gpu_options = tf.GPUOptions(
per_process_gpu_memory_fraction=gpu_fraction)
sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_options))
sess.run(tf.global_variables_initializer())
self.inputs = inputs
# self.targets = targets
self.outputs = outputs
self.hs = hs
# self.costs = costs
self.cost = self.prob_model.test_loss
self.train_step = train_step
self.sess = sess