out = input_.sg_aconv1d_gpus(size=opt.size,
rate=opt.rate,
causal=opt.causal,
act='leaky_relu',
ln=opt.causal,
name="aconv_" + opt.name)
# dimension recover and residual connection
out = out.sg_conv1d_gpus(size=1, dim=in_dim,
name="convo_" + opt.name) + tensor
return out
# inject residual multiplicative block
tf.sg_inject_func(sg_res_block)
@tf.sg_layer_func_gpus
def sg_quasi_conv1d(tensor, opt):
opt += tf.sg_opt(is_enc=False, causal=True)
# Split into H and H_zfo
H = tensor[:Hp.batch_size]
H_z = tensor[Hp.batch_size:2 * Hp.batch_size]
H_f = tensor[2 * Hp.batch_size:3 * Hp.batch_size]
H_o = tensor[3 * Hp.batch_size:]
if opt.is_enc:
H_z, H_f, H_o = 0, 0, 0
Args:
tensor: A `Tensor` (automatically passed by decorator).
opt:
bn: Boolean. If True, batch normalization is applied.
ln: Boolean. If True, layer normalization is applied.
dout: A float of range [0, 100). A dropout rate. Default is 0.
act: A name of activation function. e.g., `sigmoid`, `tanh`, etc.
Returns:
The same tensor as `tensor`.
"""
return tensor
# inject the custom identity function
tf.sg_inject_func(identity)
# residual block
@tf.sg_sugar_func
def sg_res_block(tensor, opt):
# default rate
opt += tf.sg_opt(size=3, rate=1, causal=False, is_first=False, dout=0)
# input dimension
in_dim = tensor.get_shape().as_list()[-1]
with tf.sg_context(name='block_%d_%d' % (opt.block, opt.rate)):
# reduce dimension
input_ = (tensor.sg_bypass(act='relu',
ln=(not opt.is_first),
O = O.sg_bypass(act="sigmoid") # (16, 300, 320)
# Masking
M = tf.sign(tf.abs(H))[:, :, :1] # (16, 300, 1) float32. 0 or 1
Z *= M # broadcasting
F *= M # broadcasting
O *= M # broadcasting
# Concat
ZFO = tf.concat(axis=0, values=[Z, F, O])
return ZFO # (16*3, 150, 320)
# injection
tf.sg_inject_func(sg_quasi_conv1d)
@tf.sg_rnn_layer_func
def sg_quasi_rnn(tensor, opt):
# Split
if opt.att:
H, Z, F, O = tf.split(axis=0, num_or_size_splits=4,
value=tensor) # (16, 150, 320) for all
else:
Z, F, O = tf.split(axis=0, num_or_size_splits=3,
value=tensor) # (16, 150, 320) for all
# step func
def step(z, f, o, c):
'''
# reduce dimension
input_ = (tensor
.sg_bypass(act='relu', ln=(not opt.is_first), name='bypass') # do not
.sg_conv1d(size=1, dim=in_dim/2, act='relu', ln=True, name='conv_in'))
# 1xk conv dilated
out = (input_
.sg_aconv1d(size=opt.size, rate=opt.rate, causal=opt.causal, act='relu', ln=True, name='aconv'))
# dimension recover and residual connection
out = out.sg_conv1d(size=1, dim=in_dim, name='conv_out') + tensor
return out
# inject residual multiplicative block
tf.sg_inject_func(sg_res_block)
#
# encode graph ( atrous convolution )
#
def encode(x):
with tf.sg_context(name='encoder'):
res = x
# loop dilated conv block
for i in range(num_blocks):
res = (res
.sg_res_block(size=5, block=i, rate=1, is_first=True)
.sg_res_block(size=5, block=i, rate=2)
.sg_res_block(size=5, block=i, rate=4)
Args:
tensor: A `Tensor` (automatically passed by decorator).
opt:
bn: Boolean. If True, batch normalization is applied.
ln: Boolean. If True, layer normalization is applied.
dout: A float of range [0, 100). A dropout rate. Default is 0.
act: A name of activation function. e.g., `sigmoid`, `tanh`, etc.
Returns:
The same tensor as `tensor`.
"""
return tensor
# inject the custom identity function
tf.sg_inject_func(identity)
# residual block
@tf.sg_sugar_func
def sg_res_block(tensor, opt):
# default rate
opt += tf.sg_opt(size=3, rate=1, causal=False, is_first=False, dout=0)
# input dimension
in_dim = tensor.get_shape().as_list()[-1]
with tf.sg_context(name='block_%d_%d' % (opt.block, opt.rate)):
# reduce dimension
input_ = (tensor
.sg_bypass(act='relu', ln=(not opt.is_first), name='bypass')
