def __init__(self,
npoint=None,
radii=None,
nsample=None,
down_conv_nn=None,
bn=True,
activation=torch.nn.LeakyReLU(negative_slope=0.01),
use_xyz=True,
normalize_xyz=False,
**kwargs):
assert len(radii) == len(nsample) == len(down_conv_nn)
super(PointNetMSGDown,
self).__init__(DenseFPSSampler(num_to_sample=npoint),
DenseRadiusNeighbourFinder(radii, nsample),
**kwargs)
self.use_xyz = use_xyz
self.npoint = npoint
self.mlps = nn.ModuleList()
for i in range(len(radii)):
self.mlps.append(
MLP2D(down_conv_nn[i],
bn=bn,
activation=activation,
bias=False))
self.radii = radii
self.normalize_xyz = normalize_xyz
def __init__(self,
npoint=None,
radii=None,
nsample=None,
down_conv_nn=None,
channel_raising_nn=None,
bn=True,
bias=True,
use_xyz=True,
activation=nn.ReLU(),
**kwargs):
assert len(radii) == len(nsample)
if len(radii) != len(down_conv_nn):
log.warning(
"The down_conv_nn has a different size as radii. Make sure to have sharedMLP"
)
super(RSConvMSGDown,
self).__init__(DenseFPSSampler(num_to_sample=npoint),
DenseRadiusNeighbourFinder(radii, nsample),
**kwargs)
self.use_xyz = use_xyz
self.npoint = npoint
self.mlps = nn.ModuleList()
# https://github.com/Yochengliu/Relation-Shape-CNN/blob/6464eb8bb4efc686adec9da437112ef888e55684/utils/pointnet2_modules.py#L106
self.mlp_out = Seq(*[
Conv1d(channel_raising_nn[0],
channel_raising_nn[-1],
kernel_size=1,
stride=1,
bias=True),
nn.BatchNorm1d(channel_raising_nn[-1]),
activation,
])
for i in range(len(radii)):
mapper = RSConvMapper(down_conv_nn,
activation=activation,
use_xyz=self.use_xyz)
self.mlps.append(SharedRSConv(mapper, radii[i]))
self._mapper = mapper
def __init__(self,
npoint=None,
radii=None,
nsample=None,
down_conv_nn=None,
bn=True,
activation="LeakyReLU",
use_xyz=True,
**kwargs):
assert len(radii) == len(nsample) == len(down_conv_nn)
super(PointNetMSGDown,
self).__init__(DenseFPSSampler(num_to_sample=npoint),
DenseRadiusNeighbourFinder(radii, nsample),
**kwargs)
self.use_xyz = use_xyz
self.npoint = npoint
self.mlps = nn.ModuleList()
for i in range(len(radii)):
self.mlps.append(
pt_utils.SharedMLP(down_conv_nn[i],
bn=bn,
activation=get_activation(activation)))
def __init__(
self,
npoint=None,
radii=None,
nsample=None,
down_conv_nn=None,
channel_raising_nn=None,
bn=True,
bias=True,
use_xyz=True,
activation=nn.ReLU(),
**kwargs
):
assert len(radii) == len(nsample)
if len(radii) != len(down_conv_nn):
log.warning("The down_conv_nn has a different size as radii. Make sure of have SharedRSConv")
super(RSConvOriginalMSGDown, self).__init__(
DenseFPSSampler(num_to_sample=npoint), DenseRadiusNeighbourFinder(radii, nsample), **kwargs
)
self.use_xyz = use_xyz
self.mlps = nn.ModuleList()
self.mappings = nn.ModuleList()
self._first_layer = True if len(down_conv_nn) == 2 else False
if self._first_layer:
C_in, C_intermediate, C_out = down_conv_nn[0]
feat_in, f_out = down_conv_nn[-1]
xyz_raising = nn.Conv2d(
in_channels=feat_in, out_channels=f_out, kernel_size=(1, 1), stride=(1, 1), bias=bias,
)
nn.init.kaiming_normal_(xyz_raising.weight)
if bias:
nn.init.constant_(xyz_raising.bias, 0)
else:
C_in, C_intermediate, C_out = down_conv_nn
mapping_func1 = nn.Conv2d(
in_channels=C_in, out_channels=C_intermediate, kernel_size=(1, 1), stride=(1, 1), bias=bias,
)
mapping_func2 = nn.Conv2d(
in_channels=C_intermediate, out_channels=C_out, kernel_size=(1, 1), stride=(1, 1), bias=bias,
)
nn.init.kaiming_normal_(mapping_func1.weight)
nn.init.kaiming_normal_(mapping_func2.weight)
if bias:
nn.init.constant_(mapping_func1.bias, 0)
nn.init.constant_(mapping_func2.bias, 0)
# channel raising mapping
cr_mapping = nn.Conv1d(
in_channels=channel_raising_nn[0], out_channels=channel_raising_nn[1], kernel_size=1, stride=1, bias=bias,
)
nn.init.kaiming_normal_(cr_mapping.weight)
nn.init.constant_(cr_mapping.bias, 0)
if self._first_layer:
mapping = [mapping_func1, mapping_func2, cr_mapping, xyz_raising]
elif npoint is not None:
mapping = [mapping_func1, mapping_func2, cr_mapping]
for m in mapping:
self.mappings.append(m)
for radius in radii:
self.mlps.append(OriginalRSConv(mapping=mapping, first_layer=self._first_layer, radius=radius))