def __init__(self,
fc_stack_layers=2,
fc_stack_ch=256,
rnn_ch=512,
rnn_type='gru',
n_harmonics=100,
amp_scale_fn=ddsp.core.exp_sigmoid,
f0_depth=64,
hz_min=20.0,
hz_max=1200.0,
sample_rate=16000,
name='sinusoidal_to_harmonic_encoder'):
"""Constructor."""
super().__init__(name=name)
self.n_harmonics = n_harmonics
self.amp_scale_fn = amp_scale_fn
self.f0_depth = f0_depth
self.hz_min = hz_min
self.hz_max = hz_max
self.sample_rate = sample_rate
# Layers.
self.pre_rnn = nn.FcStack(fc_stack_ch, fc_stack_layers)
self.rnn = nn.Rnn(rnn_ch, rnn_type)
self.post_rnn = nn.FcStack(fc_stack_ch, fc_stack_layers)
self.amp_out = tfkl.Dense(1)
self.hd_out = tfkl.Dense(n_harmonics)
self.f0_out = tfkl.Dense(f0_depth)
def __init__(self,
rnn_channels=512,
rnn_type='gru',
ch=512,
layers_per_stack=3,
input_keys=('ld_scaled', 'f0_scaled', 'z'),
output_splits=(('amps', 1), ('harmonic_distribution', 40)),
**kwargs):
super().__init__(
input_keys=input_keys, output_splits=output_splits, **kwargs)
stack = lambda: nn.FcStack(ch, layers_per_stack)
# Layers.
self.input_stacks = [stack() for k in self.input_keys]
self.rnn = nn.Rnn(rnn_channels, rnn_type)
self.out_stack = stack()
self.dense_out = tfkl.Dense(self.n_out)
# Backwards compatability.
self.f_stack = self.input_stacks[0] if len(
self.input_stacks) >= 1 else None
self.l_stack = self.input_stacks[1] if len(
self.input_stacks) >= 2 else None
self.z_stack = self.input_stacks[2] if len(
self.input_stacks) >= 3 else None
def __init__(self,
fc_units=512,
hidden_layers=3,
input_keys=('object_embedding'),
output_splits=(('frequencies', 200), ('gains', 200), ('dampings', 200)),
**kwargs):
super().__init__(
input_keys=input_keys, output_splits=output_splits, **kwargs)
stack = lambda: nn.FcStack(fc_units, hidden_layers)
self.input_stacks = [stack() for k in self.input_keys]
self.out_stack = stack()
def __init__(self,
ch=512,
layers_per_stack=3,
input_keys=['f0_midi_scaled'],
output_splits=(('car_amp', 1), ('mod_amp', 1), ('mod_freq',
1)),
**kwargs):
super().__init__(input_keys=input_keys,
output_splits=output_splits,
**kwargs)
stack = lambda: nn.FcStack(ch, layers_per_stack)
# Layers.
self.input_stacks = [stack() for k in self.input_keys]
self.out_stack = stack()
def __init__(self,
rnn_channels=512,
rnn_type='gru',
ch=512,
layers_per_stack=3,
input_keys=('ld_scaled', 'f0_scaled', 'z'),
output_splits=(('amps', 1), ('harmonic_distribution', 40)),
**kwargs):
super().__init__(
input_keys=input_keys, output_splits=output_splits, **kwargs)
stack = lambda: nn.FcStack(ch, layers_per_stack)
# Layers.
self.input_stacks = [stack() for k in self.input_keys]
self.rnn = nn.Rnn(rnn_channels, rnn_type)
self.out_stack = stack()
def fc_stack(): return nn.FcStack(ch, layers_per_input_stack-1)
