def __init__(self): super(AudioEncoder, self).__init__() seq = SequentialMaker() seq.add_module( "conv_0", MaskedConv1d(Hyper.dim_f, Hyper.dim_d, 1, padding="causal")) seq.add_module("relu_0", nn.ReLU()) seq.add_module("drop_0", nn.Dropout(Hyper.dropout)) seq.add_module( "conv_1", MaskedConv1d(Hyper.dim_d, Hyper.dim_d, 1, padding="causal")) seq.add_module("relu_1", nn.ReLU()) seq.add_module("drop_1", nn.Dropout(Hyper.dropout)) seq.add_module( "relu_2", MaskedConv1d(Hyper.dim_d, Hyper.dim_d, 1, padding="causal")) seq.add_module("drop_2", nn.Dropout(Hyper.dropout)) i = 3 for _ in range(2): for j in range(4): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_d, kernel_size=3, dilation=3**j, padding="causal")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for k in range(2): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_d, kernel_size=3, dilation=3, padding="causal")) if k == 0: seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 self.seq_ = seq()
def print_shape(self, input_shape): print("text-encode {") SequentialMaker.print_shape(self.seq_, torch.LongTensor(np.zeros(input_shape)), intent_size=2) print("}")
def print_shape(self, input_shape): print("super-resolution {") SequentialMaker.print_shape(self.seq_, torch.FloatTensor(np.zeros(input_shape)), intent_size=2) print("}")
def __init__(self): super(SuperRes, self).__init__() seq = SequentialMaker() seq.add_module( "conv_0", MaskedConv1d(Hyper.dim_f, Hyper.dim_c, 1, padding="same")) seq.add_module("drop_0", nn.Dropout(Hyper.dropout)) i = 1 for _ in range(1): for j in range(2): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_c, kernel_size=3, dilation=3**j, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for _ in range(2): seq.add_module( "deconv_{}".format(i), Deconv1d(Hyper.dim_c, Hyper.dim_c, 2, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for j in range(2): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_c, kernel_size=3, dilation=3**j, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 seq.add_module( "conv_{}".format(i), MaskedConv1d(Hyper.dim_c, Hyper.dim_c * 2, 1, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for _ in range(2): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_c * 2, kernel_size=3, dilation=1, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 F = Hyper.audio_nfft // 2 + 1 seq.add_module("conv_{}".format(i), MaskedConv1d(Hyper.dim_c * 2, F, 1, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for _ in range(2): seq.add_module("conv_{}".format(i), MaskedConv1d(F, F, 1, padding="same")) seq.add_module("relu_{}".format(i), nn.ReLU()) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 seq.add_module("conv_{}".format(i), MaskedConv1d(F, F, 1, padding="same")) self.seq_ = seq() self.sigmoid_ = nn.Sigmoid()
def print_shape(self, input_shape): print("audio-decoder {") SequentialMaker.print_shape(self.seq_, torch.FloatTensor(np.zeros(input_shape)), intent_size=2) print("}")
def __init__(self): super(TextEncoder, self).__init__() seq = SequentialMaker() seq.add_module( "char-embed", CharEmbed(len(Hyper.vocab), Hyper.dim_e, Hyper.vocab.find('P'))) seq.add_module( "conv_0", MaskedConv1d(Hyper.dim_e, Hyper.dim_d * 2, 1, padding="same")) seq.add_module("relu_0", nn.ReLU()) seq.add_module("drop_0", nn.Dropout(Hyper.dropout)) seq.add_module( "conv_1", MaskedConv1d(Hyper.dim_d * 2, Hyper.dim_d * 2, 1, padding="same")) seq.add_module("drop_1", nn.Dropout(Hyper.dropout)) i = 2 for _ in range(2): for j in range(4): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_d * 2, kernel_size=3, dilation=3**j, padding="same")) seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 for j in [1, 0]: for k in range(2): seq.add_module( "highway-conv_{}".format(i), HighwayConv1d(Hyper.dim_d * 2, kernel_size=3**j, dilation=1, padding="same")) if not (j == 0 and k == 1): seq.add_module("drop_{}".format(i), nn.Dropout(Hyper.dropout)) i += 1 self.seq_ = seq()