def __init__(self,
in_channels,
out_channels,
growth_rate,
mesh_kernel_size,
max_iter=2,
bn_momentum=0.1):
super(EncoderMeshBlock, self).__init__()
self.maxIter = max_iter
factor = 4
self.MeshConvGroup = []
_in_channels_ = in_channels
for n in range(0, self.maxIter):
MeshConv1 = V2VConv3d(_in_channels_,
[factor * growth_rate, factor * growth_rate],
mesh_kernel_size,
dual_depth_multiplier=[1, 1],
bn_momentum=bn_momentum)
MeshConv2 = V2VConv3d(factor * growth_rate,
[factor * growth_rate, growth_rate],
mesh_kernel_size,
dual_depth_multiplier=[1, 1],
bn_momentum=bn_momentum)
self.MeshConvGroup.append(MeshConv1)
self.MeshConvGroup.append(MeshConv2)
_in_channels_ += growth_rate
self.Transit = PerItemConv3d(_in_channels_,
out_channels,
bn_momentum=bn_momentum)
def __init__(self, in_channels, out_channels, mesh_kernel_size,
scope='', with_bn=True, bn_momentum=0.9):
super(FirstBlock, self).__init__()
self.MLP0 = PerItemConv3d(in_channels, out_channels, with_bn=with_bn,
scope='mlp0', bn_momentum=bn_momentum)
self.F2V0 = F2VConv3d(out_channels, out_channels, mesh_kernel_size,
depth_multiplier=2, scope=scope+'/F2V0',
with_bn=with_bn, bn_momentum=bn_momentum)
def __init__(self,
in_channels,
out_channels,
growth_rate,
mesh_kernel_size,
radius,
max_iter=2,
max_nn=None,
bn_momentum=0.1):
super(DualBlock, self).__init__()
self.radius = radius
self.max_nn = max_nn
self.maxIter = max_iter
self.pcloud_kernel = [8, 4, 1] # keep as default
factor = 4
self.DualConvGroup = []
_in_channels_ = in_channels
for n in range(0, self.maxIter):
MeshConv1 = V2VConv3d(_in_channels_,
[factor * growth_rate, factor * growth_rate],
mesh_kernel_size,
dual_depth_multiplier=[1, 1],
bn_momentum=bn_momentum)
MeshConv2 = V2VConv3d(factor * growth_rate,
[factor * growth_rate, 2 * growth_rate],
mesh_kernel_size,
dual_depth_multiplier=[1, 1],
bn_momentum=bn_momentum)
Dense = PerItemConv3d(_in_channels_,
factor * growth_rate,
bn_momentum=bn_momentum)
PCloudConv1 = PCloudConv3d(factor * growth_rate,
2 * growth_rate,
np.prod(self.pcloud_kernel) + 1,
bn_momentum=bn_momentum)
self.DualConvGroup.append(MeshConv1)
self.DualConvGroup.append(MeshConv2)
self.DualConvGroup.append(Dense)
self.DualConvGroup.append(PCloudConv1)
_in_channels_ += (2 * growth_rate * 2)
self.Transit = PerItemConv3d(_in_channels_,
out_channels,
bn_momentum=bn_momentum)
def __init__(self, in_channels, out_channels, mesh_kernel_size, scope='',
with_bn=True, bn_momentum=0.9):
super(FirstBlockTexture, self).__init__()
GeometryInChannels, TextureInChannels = in_channels
self.F2F0 = F2FConv3d(TextureInChannels, out_channels, with_bn=with_bn,
scope='f2f0', bn_momentum=bn_momentum)
self.MLP0 = PerItemConv3d(GeometryInChannels, out_channels, with_bn=with_bn,
scope='mlp0', bn_momentum=bn_momentum)
self.F2V0 = F2VConv3d(out_channels*2, out_channels, mesh_kernel_size,
depth_multiplier=1, scope=scope+'/F2V0',
with_bn=with_bn, bn_momentum=bn_momentum)
def __init__(self,
num_class,
stride=None,
mix_components=27,
use_height=False,
pred_facet=False):
super(PicassoNetII, self).__init__()
if stride is None:
self.stride = [1.5, 1.5, 1.5, 1.5]
self.num_class = num_class
self.num_clusters = mix_components
self.useArea, self.wgtBnd = (
True, 1.) # recommended hyper-parameters for mesh decimation
self.temperature = 0.1
self.bn_momentum = 0.99
self.num_blocks = 5
self.Niters = [2, 2, 4, 4, 4]
self.radius = [0.2, 0.4, 0.8]
self.EncodeChannels = [32, 64, 96, 128, 192, 256]
self.useHeight = use_height
if self.useHeight:
geometry_in_channels = 12
else:
geometry_in_channels = 9
self.predFacet = pred_facet
self.min_nv = 80 # set the minimum vertex number for the coarsest layer
self.stride = np.float32(self.stride)
self.nv_limit = np.int32(
np.cumprod(self.stride[::-1])[::-1] * self.min_nv)
self.build_mesh_hierarchy = meshUtil.MeshHierarchy(
self.stride, self.nv_limit, self.useArea, self.wgtBnd)
self.cluster = []
for k in range(self.num_blocks + 1):
self.cluster.append(
BuildF2VCoeff(self.num_clusters,
tau=self.temperature,
scope='cluster%d' % k))
# FirstBlock
self.Conv0 = FirstBlock(geometry_in_channels,
self.EncodeChannels[0],
self.num_clusters,
bn_momentum=self.bn_momentum,
scope='conv0')
# setting of convolution blocks in the encoder
self.Block = []
growth_rate = 32
self.Block.append(
EncoderMeshBlock(self.EncodeChannels[0],
self.EncodeChannels[1],
growth_rate,
self.num_clusters,
max_iter=self.Niters[0],
bn_momentum=self.bn_momentum,
scope='block1'))
self.Block.append(
EncoderMeshBlock(self.EncodeChannels[1],
self.EncodeChannels[2],
growth_rate,
self.num_clusters,
max_iter=self.Niters[1],
bn_momentum=self.bn_momentum,
scope='block2'))
self.Block.append(
DualBlock(self.EncodeChannels[2],
self.EncodeChannels[3],
growth_rate,
self.num_clusters,
radius=self.radius[0],
max_iter=self.Niters[2],
bn_momentum=self.bn_momentum,
scope='block3'))
self.Block.append(
DualBlock(self.EncodeChannels[3],
self.EncodeChannels[4],
growth_rate,
self.num_clusters,
radius=self.radius[1],
max_iter=self.Niters[3],
bn_momentum=self.bn_momentum,
scope='block4'))
self.Block.append(
DualBlock(self.EncodeChannels[4],
self.EncodeChannels[5],
growth_rate,
self.num_clusters,
radius=self.radius[2],
max_iter=self.Niters[4],
bn_momentum=self.bn_momentum,
scope='block5'))
# reducing channels of low-level features in the encoder
self.Decode = []
self.DecodeChannels = [self.EncodeChannels[-1], 128, 128, 96, 96, 96]
for k in range(self.num_blocks):
m = self.num_blocks - k - 1
self.Decode.append(
PerItemConv3d(self.EncodeChannels[m] + self.DecodeChannels[k],
self.DecodeChannels[k + 1],
bn_momentum=self.bn_momentum,
scope='decode%d' % m))
print(self.EncodeChannels[m] + self.DecodeChannels[k],
self.DecodeChannels[k + 1])
# mesh pooling, unpooling configuration
self.Mesh_Pool = mesh_pool(method='max')
self.Mesh_Unpool = mesh_unpool()
# classifier
if self.predFacet:
self.V2F0 = V2FConv3d(96, 96, scope='V2F0_pred')
self.Predict = PerItemConv3d(96,
self.num_class,
scope='prediction',
with_bn=False,
activation_fn=None)
def __init__(self,
num_class,
stride=None,
mix_components=27,
use_height=False):
super(PicassoNetII, self).__init__()
if stride is None:
self.stride = [1.5, 1.5, 1.5, 1.5]
else:
assert (len(stride) == 4)
self.stride = stride
self.num_class = num_class
self.num_clusters = mix_components
self.useArea, self.wgtBnd = (
True, 1.) # recommended hyper-parameters for mesh decimation
self.temperature = 0.1
self.bn_momentum = 0.01
self.num_blocks = 5
self.Niters = [2, 2, 4, 4, 4]
self.radius = [0.2, 0.4, 0.8]
self.EncodeChannels = [32, 64, 96, 128, 192, 256]
self.useHeight = use_height
if self.useHeight:
geometry_in_channels = 12
else:
geometry_in_channels = 9
self.min_nv = 80 # set the minimum vertex number for the coarsest layer
self.stride = np.float32(self.stride)
self.nv_limit = np.int32(
np.cumprod(self.stride[::-1])[::-1] * self.min_nv)
self.build_mesh_hierarchy = meshUtil.MeshHierarchy(
self.stride, self.nv_limit, self.useArea, self.wgtBnd)
self.cluster = []
for k in range(self.num_blocks):
self.cluster.append(
BuildF2VCoeff(self.num_clusters, tau=self.temperature))
# FirstBlock
self.Conv0 = FirstBlock(geometry_in_channels,
self.EncodeChannels[0],
self.num_clusters,
bn_momentum=self.bn_momentum)
# setting of convolution blocks in the encoder
self.Block = []
growth_rate = 32
self.Block.append(
EncoderMeshBlock(self.EncodeChannels[0],
self.EncodeChannels[1],
growth_rate,
self.num_clusters,
max_iter=self.Niters[0],
bn_momentum=self.bn_momentum))
self.Block.append(
EncoderMeshBlock(self.EncodeChannels[1],
self.EncodeChannels[2],
growth_rate,
self.num_clusters,
max_iter=self.Niters[1],
bn_momentum=self.bn_momentum))
self.Block.append(
DualBlock(self.EncodeChannels[2],
self.EncodeChannels[3],
growth_rate,
self.num_clusters,
radius=self.radius[0],
max_iter=self.Niters[2],
bn_momentum=self.bn_momentum))
self.Block.append(
DualBlock(self.EncodeChannels[3],
self.EncodeChannels[4],
growth_rate,
self.num_clusters,
radius=self.radius[1],
max_iter=self.Niters[3],
bn_momentum=self.bn_momentum))
self.Block.append(
DualBlock(self.EncodeChannels[4],
self.EncodeChannels[5],
growth_rate,
self.num_clusters,
radius=self.radius[2],
max_iter=self.Niters[4],
bn_momentum=self.bn_momentum))
# mesh pooling, unpooling configuration
self.Mesh_Pool = mesh_pool(method='max')
self.Mesh_Unpool = mesh_unpool()
self.globalPool = global_mesh_pool(method='avg')
# classifier
self.FC1 = PerItemConv3d(self.EncodeChannels[-1], 128)
self.DP = nn.Dropout(0.2)
self.FC2 = PerItemConv3d(128,
self.num_class,
with_bn=False,
activation_fn=None)