def upsample_conv1d(x,
num_filters,
filter_length,
stride,
use_resize_conv,
name_patt,
act='tanh',
use_weight_norm=False,
init=False):
act_func = masked.get_upsample_act(act)
conv_args = {
"x": x,
"num_filters": num_filters,
"filter_length": filter_length,
"stride": stride,
"activation": act_func,
"use_weight_norm": use_weight_norm,
"init": init
}
if use_resize_conv:
y = masked.resize_conv1d(**conv_args,
name=name_patt.format("resize_conv"))
else:
y = masked.trans_conv1d(**conv_args,
name=name_patt.format("trans_conv"))
return y
def upsample_conv1d(
x, num_filters, filter_length, stride, name_patt,
use_weight_norm=False, init=False):
conv_args = {"x": x,
"num_filters": num_filters,
"filter_length": filter_length,
"stride": stride,
"use_weight_norm": use_weight_norm,
"init": init}
if USE_RESIZE_CONV:
y = masked.resize_conv1d(
**conv_args, name=name_patt.format("resize_conv"))
else:
y = masked.trans_conv1d(
**conv_args, name=name_patt.format("trans_conv"))
return y
def _deconv_stack(inputs, width, config, name=''):
b, l, _ = inputs.get_shape().as_list()
frame_shift = int(np.prod([c[1] for c in config]))
mel_en = inputs
for i in range(len(config)):
fl, s = config[i]
if name:
tc_name = '{}/trans_conv_{:d}'.format(name, i + 1)
else:
tc_name = 'trans_conv_{:d}'.format(i + 1)
mel_en = masked.trans_conv1d(mel_en,
num_filters=width,
filter_length=fl,
stride=s,
name=tc_name)
mel_en.set_shape([b, l * frame_shift, width])
return mel_en