def feature_extraction_block(input_):
weights = []
biases = []
layer, weight, bias = utils.convolution(input_,
shape=[3, 3, 1, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='fblock',
index=0)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
layer, weight, bias = utils.convolution(layer,
shape=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='fblock',
index=1)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
return layer, weights, biases
def inference(image):
weights = []
biases = []
with tf.name_scope('vdsr_layer_0'):
layer, weight, bias = utils.convolution(
image,
shape=[3, 3, 1, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='vdsr',
index=0)
layer = tf.nn.relu(layer)
weights.append(weight)
biases.append(bias)
for i in range(1, 19):
with tf.name_scope('vdsr_layer_%d' % i):
layer, weight, bias = utils.convolution(
layer,
shape=[3, 3, 64, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='vdsr',
index=i)
layer = tf.nn.relu(layer)
weights.append(weight)
biases.append(bias)
with tf.name_scope('vdsr_layer_19'):
layer, weight, bias = utils.convolution(
layer,
shape=[3, 3, 64, 1],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='vdsr',
index=19)
weights.append(weight)
biases.append(bias)
with tf.name_scope('vdsr_output'):
output = image + layer
output = tf.clip_by_value(output, 0., 1.)
return output, layer, weights, biases
def move(prior, motion, motionKernel):
validate_dist(prior)
validate_dist(motionKernel)
posterior = np.zeros_like(prior)
if np.array_equal(motion, [0, 0]):
posterior = prior
elif np.array_equal(motion, [0, 1]):
posterior = convolution(prior, np.fliplr(motionKernel))
elif np.array_equal(motion, [0, -1]):
posterior = convolution(prior, motionKernel)
elif np.array_equal(motion, [1, 0]):
posterior = convolution(prior, np.rot90(motionKernel, k=3))
elif np.array_equal(motion, [-1, 0]):
posterior = convolution(prior, np.rot90(motionKernel, k=1))
else:
raise Exception("Invalid motion command {}".format(motion))
return posterior
def compression_unit(input_, block_index):
weights = []
biases = []
layer, weight, bias = utils.convolution(input_,
shape=[1, 1, 80, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='compression_unit_%d' %
block_index,
index=0)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
return layer, weights, biases
def enhancement_unit(input_, block_index):
weights = []
biases = []
layer, weight, bias = utils.grouped_convolution(
input_,
shape=[3, 3, 64, 48],
strides=[1, 1, 1, 1],
padding='SAME',
group=4,
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' % block_index,
index=0)
layer = tf.nn.leaky_relu(layer, 0.05)
weights += weight
biases += bias
layer, weight, bias = utils.convolution(layer,
shape=[3, 3, 48, 32],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' %
block_index,
index=1)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
layer, weight, bias = utils.convolution(layer,
shape=[3, 3, 32, 64],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' %
block_index,
index=2)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
segment_1, segment_2 = tf.split(layer, [16, 48], axis=3)
concatenated = tf.concat([input_, segment_1], axis=3)
layer, weight, bias = utils.grouped_convolution(
segment_2,
shape=[3, 3, 48, 64],
strides=[1, 1, 1, 1],
padding='SAME',
group=4,
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' % block_index,
index=3)
layer = tf.nn.leaky_relu(layer, 0.05)
weights += weight
biases += bias
layer, weight, bias = utils.convolution(layer,
shape=[3, 3, 64, 48],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' %
block_index,
index=4)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
layer, weight, bias = utils.convolution(layer,
shape=[3, 3, 48, 80],
strides=[1, 1, 1, 1],
padding='SAME',
init_w=tf.initializers.he_normal(),
init_b=tf.initializers.zeros(),
name='enhancement_unit_%d' %
block_index,
index=5)
layer = tf.nn.leaky_relu(layer, 0.05)
weights.append(weight)
biases.append(bias)
output = layer + concatenated
return output, weights, biases
def __init__(self, vocab_size, input_dimension):
super(AcousticModel, self).__init__()
self.vocab_size = vocab_size
self.input_dimension = input_dimension
self.dropout = nn.Dropout(p=0.5)
self.num_rnn_layers = 4
conv1 = nn.Sequential()
conv1.add_module(
'conv1_conv1',
convolution(in_channels=1, filter_size=32, bias=False))
conv1.add_module('conv1_norm1', normalization(num_features=32))
conv1.add_module('conv1_relu1', nn.ReLU())
conv1.add_module('conv1_dropout1', nn.Dropout(p=0.1))
conv1.add_module('conv1_conv2',
convolution(in_channels=32, filter_size=32))
conv1.add_module('conv1_norm2', normalization(num_features=32))
conv1.add_module('conv1_relu2', nn.ReLU())
conv1.add_module('conv1_maxpool', maxpooling(2))
conv1.add_module('conv1_dropout2', nn.Dropout(p=0.1))
self.conv1 = conv1
conv2 = nn.Sequential()
conv2.add_module('conv2_conv1',
convolution(in_channels=32, filter_size=64))
conv2.add_module('conv2_norm1', normalization(num_features=64))
conv2.add_module('conv2_relu1', nn.ReLU())
conv2.add_module('conv2_dropout1', nn.Dropout(p=0.1))
conv2.add_module('conv2_conv2',
convolution(in_channels=64, filter_size=64))
conv2.add_module('conv2_norm2', normalization(num_features=64))
conv2.add_module('conv2_relu2', nn.ReLU())
conv2.add_module('conv2_maxpool', maxpooling(2))
conv2.add_module('conv2_dropout2', nn.Dropout(p=0.1))
self.conv2 = conv2
conv3 = nn.Sequential()
conv3.add_module('conv3_conv1',
convolution(in_channels=64, filter_size=128))
conv3.add_module('conv3_relu1', nn.ReLU())
conv3.add_module('conv3_dropout1', nn.Dropout(p=0.2))
conv3.add_module('conv3_norm1', normalization(num_features=128))
conv3.add_module('conv3_conv2',
convolution(in_channels=128, filter_size=128))
conv3.add_module('conv3_norm2', normalization(num_features=128))
conv3.add_module('conv3_relu2', nn.ReLU())
conv3.add_module('conv3_maxpool', maxpooling(2))
conv3.add_module('conv3_dropout2', nn.Dropout(p=0.2))
self.conv3 = conv3
conv4 = nn.Sequential()
conv4.add_module('conv4_conv1',
convolution(in_channels=128, filter_size=128))
conv4.add_module('conv4_norm1', normalization(num_features=128))
conv4.add_module('conv4_relu1', nn.ReLU())
conv4.add_module('conv4_dropout1', nn.Dropout(p=0.2))
conv4.add_module('conv4_conv2',
convolution(in_channels=128, filter_size=128))
conv4.add_module('conv4_relu2', nn.ReLU())
conv4.add_module('conv4_conv3',
convolution(in_channels=128, filter_size=128))
conv4.add_module('conv4_norm2', normalization(num_features=128))
conv4.add_module('conv4_relu3', nn.ReLU())
conv4.add_module('conv4_dropout2', nn.Dropout(p=0.2))
self.conv4 = conv4 # no maxpooling
self.fc_features = int(input_dimension / 8 *
128) # due to three times of pooling 2**3 = 8
self.fc1 = fclayer(in_features=self.fc_features, out_features=512)
self.fc2 = fclayer(in_features=512, out_features=vocab_size)
self.transformer = TransformerEncoder(num_layers=4,
model_size=512,
inner_size=2048,
key_size=64,
value_size=64,
num_head=8)