def get_local_map(y,
ndilation,
nchannel,
filter_len,
dilate_init=1,
dilate_scale=2,
dtype=0):
'''
Args :
y - 2D tensor [batch_size, time_len]
ndilation - int
nchannel - int
filter_len - int
dilate_init - int defaults to be 1
dilate_scale - int defaults to be 2
dtype - int defaults to be 0
Return :
dtype - 0
local_map - 4D tensor [batch_size, time_len, ndilationxnchannel, 1]
dtype - 1
local_map - 4D tensor [batch_size, time_len, nchannel, ndilation]
dtype - 2
local_map - 4D tensor [batch_size*ndilation, time_len, nchannel, 1]
'''
batch_size, _ = get_shape(y)
y_r = tf.expand_dims(y, axis=-1)
local_map = []
for i in range(ndilation):
local_map.append(dilated_conv1d(y_r, filter_shape=[filter_len, 1, nchannel],\
stride=1, activation=tf.nn.relu, padding=False,\
dilation_rate=dilate_init*int(dilate_scale**i), scope = "dilated_conv%d"%i))
min_size = min([item.get_shape().as_list()[1] for item in local_map])
for i in range(ndilation):
local_map[i] = local_map[i][:, :min_size, :]
if dtype == 0:
local_map = tf.reshape(tf.concat(local_map, axis=2),
[batch_size, min_size, ndilation * nchannel, 1])
elif dtype == 1:
local_map = tf.reshape(tf.concat(local_map, axis=2),
[batch_size, min_size, ndilation, nchannel])
local_map = tf.transpose(
local_map,
[0, 1, 3, 2]) # [batch_size, min_size, nchannel, ndilation]
elif dtype == 2:
local_map = tf.reshape(tf.concat(local_map, axis=2),
[batch_size, min_size, ndilation, nchannel])
local_map = tf.transpose(
local_map,
[0, 2, 1, 3]) # [batch_size, ndilation, min_size, nchannel]
local_map = tf.reshape(
local_map, [batch_size * ndilation, min_size, nchannel, 1
]) # [batch_size*ndilation, min_size, nchannel, 1]
return local_map
def encode_layer(self, input_, dilation, layer_no):
options = self.options
relu1 = tf.nn.relu(input_, name = 'enc_relu1_layer{}'.format(layer_no))
conv1 = ops.conv1d(relu1, options['residual_channels'], name = 'enc_conv1d_1_layer{}'.format(layer_no))
relu2 = tf.nn.relu(conv1, name = 'enc_relu2_layer{}'.format(layer_no))
dilated_conv = ops.dilated_conv1d(relu2, options['residual_channels'],
dilation, options['encoder_filter_width'],
causal = False,
name = "enc_dilated_conv_layer{}".format(layer_no)
)
relu3 = tf.nn.relu(dilated_conv, name = 'enc_relu1_layer{}'.format(layer_no))
conv2 = ops.conv1d(relu3, 2 * options['residual_channels'], name = 'enc_conv1d_2_layer{}'.format(layer_no))
return input_ + conv2