def __init__(self, in_channels, out_channels, kernel_size=1, bias=True):
self.kernel_size = kernel_size
in_channels = in_channels
out_channels = _single(out_channels)
modules = []
for channels in out_channels:
modules.append(nn.Conv1d(in_channels, channels, 1, bias=bias))
modules.append(nn.BatchNorm1d(channels))
modules.append(nn.ReLU(True))
in_channels = channels
super(FeaturePropagation, self).__init__(*modules)
def __init__(self, in_channels, out_channels, bias=True):
self.in_channels = in_channels
self.out_channels = _single(out_channels)
self.bias = bias
in_channels = self.in_channels
modules = []
for channels in self.out_channels:
modules.append(nn.Conv1d(in_channels, channels, 1, bias=self.bias))
modules.append(nn.ReLU(True))
in_channels = channels
modules.append(nn.Conv1d(in_channels, 3, 1, bias=self.bias))
super(Folding3d, self).__init__(*modules)
def __init__(self, in_channels, out_channels, kernel_size=1, bias=True):
self.in_channels = in_channels
self.kernel_size = kernel_size
in_channels = in_channels * 2
out_channels = _single(out_channels)
modules = []
for channels in out_channels:
modules.append(nn.Conv2d(in_channels, channels, 1, bias=bias))
modules.append(nn.BatchNorm2d(channels))
modules.append(nn.LeakyReLU(0.2, True))
in_channels = channels
modules.append(nn.MaxPool2d([kernel_size, 1]))
super(EdgeConv, self).__init__(*modules)
def __init__(self, in_channels, out_channels, kernel_size=1, bias=True):
self.in_channels = in_channels
self.out_channels = _single(out_channels)
self.kernel_size = kernel_size
self.bias = bias
in_channels = self.in_channels
modules = []
for channels in self.out_channels:
modules.append(nn.Conv1d(in_channels, channels, 1, bias=self.bias))
modules.append(nn.BatchNorm1d(channels))
modules.append(nn.ReLU(True))
in_channels = channels
super(SetDeconv, self).__init__(*modules)
def __init__(self,
in_channels,
out_channels,
kernel_size=1,
bandwidth=1.0,
bias=True):
super(PointDeconv, self).__init__()
self.in_channels = in_channels
self.out_channels = _single(out_channels)
self.kernel_size = kernel_size
self.bandwidth = bandwidth
self.bias = bias
self.scale = nn.Sequential(
nn.Conv1d(1, 8, 1, bias=self.bias),
nn.BatchNorm1d(8),
nn.ReLU(True),
nn.Conv1d(8, 8, 1, bias=self.bias),
nn.BatchNorm1d(8),
nn.ReLU(True),
nn.Conv1d(8, 1, 1, bias=self.bias),
nn.Sigmoid(),
)
self.weight = nn.Sequential(
nn.Conv2d(3, 8, 1, bias=self.bias),
nn.BatchNorm2d(8),
nn.ReLU(True),
nn.Conv2d(8, 8, 1, bias=self.bias),
nn.BatchNorm2d(8),
nn.ReLU(True),
nn.Conv2d(8, 16, 1, bias=self.bias),
)
in_channels = self.in_channels
modules = []
for channels in self.out_channels[:-1]:
modules.append(nn.Conv2d(in_channels, channels, 1, bias=self.bias))
modules.append(nn.BatchNorm2d(channels))
modules.append(nn.ReLU(True))
in_channels = channels
self.mlp = nn.Sequential(*modules)
self.lin = nn.Sequential(
nn.Conv2d(in_channels,
self.out_channels[-1], [16, 1],
bias=self.bias),
nn.BatchNorm2d(self.out_channels[-1]),
nn.ReLU(True),
)
def __init__(
self,
in_channels,
out_channels,
num_samples=1,
kernel_size=1,
radius=1.0,
bias=True,
):
self.kernel_size = kernel_size
self.num_samples = num_samples
self.radius = radius
out_channels = _single(out_channels)
modules = []
for channels in out_channels:
modules.append(nn.Conv2d(in_channels, channels, 1, bias=bias))
modules.append(nn.BatchNorm2d(channels))
modules.append(nn.ReLU(True))
in_channels = channels
modules.append(nn.AdaptiveMaxPool2d([1, None]))
super(SetAbstraction, self).__init__(*modules)
def __init__(
self,
in_channels,
out_channels,
kernel_size=1,
stride=1,
radius=None,
bias=True,
):
self.in_channels = in_channels
self.out_channels = _single(out_channels)
self.kernel_size = kernel_size
self.stride = stride
self.radius = radius
self.bias = bias
in_channels = self.in_channels
modules = []
for channels in self.out_channels:
modules.append(nn.Conv2d(in_channels, channels, 1, bias=self.bias))
modules.append(nn.BatchNorm2d(channels))
modules.append(nn.ReLU(True))
in_channels = channels
modules.append(nn.AdaptiveMaxPool2d([1, None]))
super(SetConv, self).__init__(*modules)
