Spaces:
Sleeping
Sleeping
File size: 4,870 Bytes
14d1720 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 |
import torch
import torch.nn as nn
class Discriminator(nn.Module):
def __init__(self, ndf = 16, n_layers = 3, downsampling_factor = 4, disc_out = 512):
super(Discriminator, self).__init__()
discriminator = nn.ModuleDict()
discriminator["layer_0"] = nn.Sequential(
nn.ReflectionPad1d(7),
nn.utils.weight_norm(nn.Conv1d(1, ndf, kernel_size=15, stride=1)),
nn.LeakyReLU(0.2, True),
)
nf = ndf
stride = downsampling_factor
for n in range(1, n_layers + 1):
nf_prev = nf
nf = min(nf * stride, disc_out)
discriminator["layer_%d" % n] = nn.Sequential(
nn.utils.weight_norm(nn.Conv1d(
nf_prev,
nf,
kernel_size=stride * 10 + 1,
stride=stride,
padding=stride * 5,
groups=nf_prev // 4,
)),
nn.LeakyReLU(0.2, True),
)
nf = min(nf * 2, disc_out)
discriminator["layer_%d" % (n_layers + 1)] = nn.Sequential(
nn.utils.weight_norm(nn.Conv1d(nf, disc_out, kernel_size=5, stride=1, padding=2)),
nn.LeakyReLU(0.2, True),
)
discriminator["layer_%d" % (n_layers + 2)] = nn.utils.weight_norm(nn.Conv1d(
nf, 1, kernel_size=3, stride=1, padding=1
))
self.discriminator = discriminator
def forward(self, x):
'''
returns: (list of 6 features, discriminator score)
we directly predict score without last sigmoid function
since we're using Least Squares GAN (https://arxiv.org/abs/1611.04076)
'''
features = list()
for key, module in self.discriminator.items():
x = module(x)
features.append(x)
return features[:-1], features[-1]
# JCU Discriminator
class JCU_Discriminator(nn.Module):
def __init__(self):
super(JCU_Discriminator, self).__init__()
self.mel_conv = nn.Sequential(
nn.ReflectionPad1d(3),
nn.utils.weight_norm(nn.Conv1d(80, 128, kernel_size=2, stride=1)),
nn.LeakyReLU(0.2, True),
)
x_conv = [nn.ReflectionPad1d(7),
nn.utils.weight_norm(nn.Conv1d(1, 16, kernel_size=7, stride=1)),
nn.LeakyReLU(0.2, True),
]
x_conv += [
nn.utils.weight_norm(nn.Conv1d(
16,
64,
kernel_size=41,
stride=4,
padding=4 * 5,
groups=16 // 4,
)
),
nn.LeakyReLU(0.2),
]
x_conv += [
nn.utils.weight_norm(nn.Conv1d(
64,
128,
kernel_size=21,
stride=2,
padding=2 * 5,
groups=64 // 4,
)
),
nn.LeakyReLU(0.2),
]
self.x_conv = nn.Sequential(*x_conv)
self.mel_conv2 = nn.Sequential(
nn.utils.weight_norm(nn.Conv1d(128, 128, kernel_size=5, stride=1, padding=2)),
nn.LeakyReLU(0.2, True),
)
self.mel_conv3 = nn.utils.weight_norm(nn.Conv1d(
128, 1, kernel_size=3, stride=1, padding=1
))
self.x_conv2 = nn.Sequential(
nn.utils.weight_norm(nn.Conv1d(128, 128, kernel_size=5, stride=1, padding=2)),
nn.LeakyReLU(0.2, True),
)
self.x_conv3 = nn.utils.weight_norm(nn.Conv1d(
128, 1, kernel_size=3, stride=1, padding=1
))
def forward(self, x, mel):
out = self.mel_conv(mel)
out1 = self.x_conv(x)
out = torch.cat([out, out1], dim=2)
out = self.mel_conv2(out)
cond_out = self.mel_conv3(out)
out1 = self.x_conv2(out1)
uncond_out = self.x_conv3(out1)
return uncond_out, cond_out
if __name__ == '__main__':
model = Discriminator()
'''
Length of features : 5
Length of score : 3
torch.Size([3, 16, 25600])
torch.Size([3, 64, 6400])
torch.Size([3, 256, 1600])
torch.Size([3, 512, 400])
torch.Size([3, 512, 400])
torch.Size([3, 1, 400]) -> score
'''
x = torch.randn(3, 1, 25600)
print(x.shape)
features, score = model(x)
print("Length of features : ", len(features))
print("Length of score : ", len(score))
for feat in features:
print(feat.shape)
print(score.shape)
pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad)
print(pytorch_total_params)
|