def __init__(self, channel):
super(ResidualModule, self).__init__()
self.conv1 = nn.Conv2d(in_channels=channel,
out_channels=channel,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.conv2 = nn.Conv2d(in_channels=channel,
out_channels=channel,
kernel_size=3,
stride=1,
padding=1,
bias=False)
self.bn = nn.InstanceNorm2d(channel,
eps=1e-5,
momentum=0.01,
affine=True)
self.cbam = CBAM(channel)
self.relu = nn.ReLU(inplace=True)
self.apply(weights_init)
def __init__(self, in_channel, out_channel):
super(DeConvPitchPadding, self).__init__()
self.deConv1 = nn.ConvTranspose2d(in_channels=in_channel,
out_channels=out_channel,
kernel_size=4,
stride=2,
padding=1,
output_padding=(0, 1),
bias=True)
self.deConv2 = nn.ConvTranspose2d(in_channels=in_channel,
out_channels=out_channel,
kernel_size=4,
stride=2,
padding=1,
output_padding=(0, 1),
bias=True)
self.conv = nn.Conv2d(in_channels=in_channel,
out_channels=out_channel,
kernel_size=1,
stride=1,
bias=False)
self.bn1 = nn.InstanceNorm2d(out_channel,
eps=1e-5,
momentum=0.01,
affine=True)
self.bn2 = nn.InstanceNorm2d(out_channel,
eps=1e-5,
momentum=0.01,
affine=True)
self.bn3 = nn.InstanceNorm2d(out_channel,
eps=1e-5,
momentum=0.01,
affine=True)
self.cbam1 = CBAM(out_channel)
self.cbam2 = CBAM(out_channel)
self.relu = nn.ReLU(inplace=True)
self.apply(weights_init)
def __init__(self, layers): # [1024, 512, 256, 128, 64]
super().__init__()
self.leaky = nn.LeakyReLU(inplace=True)
self.relu = nn.ReLU(inplace=True)
self.sigmoid = nn.Sigmoid()
self.dropout = nn.Dropout(p=0.3)
self.bar_linear = nn.Linear(1152 * 2, 1152)
self.phrase_linear = nn.Linear(1152 * 2, 1152)
self.time = TimePitchModule()
self.pitch = PitchTimeModule()
self.fit1 = nn.Conv2d(in_channels=2048,
out_channels=1024,
kernel_size=1,
stride=1,
bias=False)
self.bn = nn.InstanceNorm2d(1024, eps=1e-5, momentum=0.01, affine=True)
self.fit2 = nn.Conv2d(in_channels=64,
out_channels=1,
kernel_size=1,
stride=1,
bias=False)
self.layers = []
for i in range(1, len(layers)):
if i < 3:
self.layers.append(DeConvPitchPadding(layers[i - 1],
layers[i]))
else:
self.layers.append(DeConvModule(layers[i - 1], layers[i]))
self.layers = nn.ModuleList(self.layers)
self.cbam = CBAM(1024)
self.position_embedding = nn.Embedding(332, 1152)
nn.init.uniform_(self.position_embedding.weight, -1.0, 1.0)
self.apply(weights_init)
def __init__(self):
super(PitchTimeModule, self).__init__()
self.pitch = nn.ConvTranspose2d(in_channels=2304,
out_channels=1024,
kernel_size=(1, 3),
stride=(1, 3),
bias=False)
self.time = nn.ConvTranspose2d(in_channels=1024,
out_channels=1024,
kernel_size=(6, 1),
stride=(6, 1),
bias=False)
self.bn = nn.InstanceNorm2d(1024, eps=1e-5, momentum=0.01, affine=True)
self.cbam = CBAM(1024)
self.relu = nn.ReLU(inplace=True)
self.apply(weights_init)
def __init__(self):
super(TimePitchModule, self).__init__()
self.time = nn.Conv2d(in_channels=1,
out_channels=32,
kernel_size=(4, 1),
stride=(2, 1),
padding=[1, 0],
bias=False)
self.pitch = nn.Conv2d(in_channels=32,
out_channels=32,
kernel_size=(1, 4),
stride=(1, 2),
padding=[0, 1],
bias=False)
self.bn = nn.InstanceNorm2d(32, eps=1e-5, momentum=0.01, affine=True)
self.cbam = CBAM(32)
self.leaky = nn.LeakyReLU(inplace=True)
self.apply(weights_init)