def __init__(self, wrnn_dims, fc_dims, cond_channels, global_cond_channels):
super().__init__()
conv_channels = 128
rnn_channels = 512
self.warmup_steps = 64
self.conv0 = Conv4(1, conv_channels, global_cond_channels)
self.conv1 = Conv4(conv_channels, conv_channels, global_cond_channels)
self.conv2 = Conv4(conv_channels, conv_channels, global_cond_channels)
self.rnn0 = RNN4(conv_channels + cond_channels, rnn_channels, self.warmup_steps, global_cond_channels)
self.rnn1 = RNN4(conv_channels + rnn_channels, rnn_channels, self.warmup_steps, global_cond_channels)
self.rnn2 = RNN4(conv_channels + rnn_channels, rnn_channels, self.warmup_steps, global_cond_channels)
self.wavernn = WaveRNN(wrnn_dims, fc_dims, rnn_channels + global_cond_channels, 0)
self.delay_c0 = 9
self.delay_c1 = self.delay_c0 + 9 * 4
self.delay_c2 = self.delay_c1 + 9 * 16
self.delay_r0 = self.delay_c2 + self.warmup_steps * 64
self.delay_r1 = self.delay_r0 + self.warmup_steps * 16
self.delay_r2 = self.delay_r1 + self.warmup_steps * 4
self.delay_wr = self.delay_r2 + self.warmup_steps
cond_delay = self.delay_wr - self.delay_c2
if cond_delay % 64 != 0:
raise RuntimeError(f'Overtone: bad cond delay: {cond_delay}')
self.cond_pad = cond_delay // 64
def __init__(self,
rnn_dims,
fc_dims,
pad,
upsample_factors,
feat_dims,
DEVICE="cuda"):
super().__init__()
self.n_classes = 256
self.upsample = UpsampleNetwork(feat_dims,
upsample_factors,
DEVICE=DEVICE)
self.wavernn = WaveRNN(rnn_dims, fc_dims, feat_dims, 0, DEVICE=DEVICE)
self.num_params()
self.DEVICE = DEVICE
def __init__(self,
quantization_channels=256,
gru_channels=896,
fc_channels=896,
lc_channels=80,
upsample_factor=(5, 5, 8),
use_gru_in_upsample=True):
super().__init__()
self.upsample = ConvInUpsampleNetwork(upsample_scales=upsample_factor,
upsample_activation="none",
upsample_activation_params={},
mode="nearest",
cin_channels=lc_channels,
use_gru=use_gru_in_upsample)
self.wavernn = WaveRNN(quantization_channels, gru_channels,
fc_channels, lc_channels)
def __init__(self,
quantization_channels=256,
gru_channels=896,
fc_channels=896,
lc_channels=80,
lc_out_channles=80,
upsample_factor=(5, 5, 8),
use_lstm=True,
lstm_layer=2,
upsample_method='duplicate'):
super().__init__()
self.frame_net = FrameRateNet(lc_channels, lc_out_channles)
self.upsample = UpsampleNet(input_size=lc_out_channles,
output_size=lc_out_channles,
upsample_factor=upsample_factor,
use_lstm=use_lstm,
lstm_layer=lstm_layer,
upsample_method=upsample_method)
self.wavernn = WaveRNN(quantization_channels, gru_channels,
fc_channels, lc_channels)
self.num_params()