def __init__(self, in_channels, out_channels, config, D=3, **kwargs):
super(UNet2, self).__init__(in_channels, out_channels, config, D)
PLANES = self.PLANES
# Output of the first conv concated to conv6
self.conv1 = conv(in_channels,
PLANES[0],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn1 = ME.MinkowskiBatchNorm(PLANES[0])
self.block1 = UNBlocks(PLANES[0], PLANES[0], D)
self.down1 = conv(PLANES[0], PLANES[1], kernel_size=2, stride=2, D=D)
self.down1bn = ME.MinkowskiBatchNorm(PLANES[1])
self.up1 = conv_tr(PLANES[1],
PLANES[0],
kernel_size=2,
upsample_stride=2,
D=D)
self.block1up = UNBlocks(PLANES[0] * 2, PLANES[0], D)
self.block2 = UNBlocks(PLANES[1], PLANES[1], D)
self.down2 = conv(PLANES[1], PLANES[2], kernel_size=2, stride=2, D=D)
self.up2 = conv_tr(PLANES[2],
PLANES[1],
kernel_size=2,
upsample_stride=2,
D=D)
self.block2up = UNBlocks(PLANES[1] * 2, PLANES[1], D)
self.block3 = UNBlocks(PLANES[2], PLANES[2], D)
self.down3 = conv(PLANES[2], PLANES[3], kernel_size=2, stride=2, D=D)
self.up3 = conv_tr(PLANES[3],
PLANES[2],
kernel_size=2,
upsample_stride=2,
D=D)
self.block3up = UNBlocks(PLANES[2] * 2, PLANES[2], D)
self.block4 = UNBlocks(PLANES[3], PLANES[3], D)
self.relu = ME.MinkowskiReLU(inplace=True)
self.final = conv(PLANES[0],
out_channels,
kernel_size=1,
bias=True,
D=D)
def __init__(self,
inplanes,
intermediate_inplanes,
intermediate_outplanes,
outplanes,
intermediate_module,
BLOCK=None,
reps=1,
conv_type=ConvType.HYPERCUBE,
norm_type=NormType.BATCH_NORM,
bn_momentum=0.1,
D=3):
super(UBlock, self).__init__()
self.block = BLOCK(inplanes,
inplanes,
conv_type=conv_type,
bn_momentum=bn_momentum,
D=D)
self.down = conv(inplanes,
intermediate_inplanes,
kernel_size=space_n_time_m(2, 1, D),
stride=space_n_time_m(2, 1, D),
conv_type=conv_type,
D=D)
self.down_norm = get_norm(norm_type,
intermediate_inplanes,
D,
bn_momentum=bn_momentum)
self.intermediate = intermediate_module
self.up = conv_tr(intermediate_outplanes,
outplanes,
kernel_size=space_n_time_m(2, 1, D),
upsample_stride=space_n_time_m(2, 1, D),
conv_type=conv_type,
D=D)
self.up_norm = get_norm(norm_type,
outplanes,
D,
bn_momentum=bn_momentum)
self.reps = reps
for i in range(reps):
if i == 0:
downsample = nn.Sequential(
conv(inplanes + outplanes,
outplanes,
kernel_size=1,
bias=False,
D=D),
get_norm(norm_type, outplanes, D, bn_momentum=bn_momentum),
)
setattr(
self, f'end_blocks{i}',
BLOCK((inplanes + outplanes) if i == 0 else outplanes,
outplanes,
downsample=downsample if i == 0 else None,
conv_type=conv_type,
bn_momentum=bn_momentum,
D=D))
def network_initialization(self, in_channels, out_channels, config, D):
self.conv_feat1 = conv(in_channels[0], config.upsample_feat_size, 3, D=D)
self.conv_feat2 = conv(in_channels[1], config.upsample_feat_size, 3, D=D)
self.conv_feat3 = conv(in_channels[2], config.upsample_feat_size, 3, D=D)
self.conv_feat4 = conv(in_channels[3], config.upsample_feat_size, 3, D=D)
self.bn_feat1 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_feat2 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_feat3 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_feat4 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.conv_up2 = conv_tr(
config.upsample_feat_size, config.upsample_feat_size, kernel_size=2, upsample_stride=2,
dilation=1, bias=False, D=3)
self.conv_up3 = conv_tr(
config.upsample_feat_size, config.upsample_feat_size, kernel_size=2, upsample_stride=2,
dilation=1, bias=False, D=3)
self.conv_up4 = conv_tr(
config.upsample_feat_size, config.upsample_feat_size, kernel_size=2, upsample_stride=2,
dilation=1, bias=False, D=3)
self.conv_up5 = conv_tr(
config.upsample_feat_size, config.upsample_feat_size, kernel_size=2, upsample_stride=2,
dilation=1, bias=False, D=3)
self.conv_up6 = conv_tr(
config.upsample_feat_size, config.upsample_feat_size, kernel_size=2, upsample_stride=2,
dilation=1, bias=False, D=3)
self.bn_up2 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_up3 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_up4 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_up5 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.bn_up6 = get_norm(
NormType.BATCH_NORM, config.upsample_feat_size, D=3, bn_momentum=config.bn_momentum)
self.conv_final = conv(config.upsample_feat_size, out_channels, 1, D=D)
self.relu = ME.MinkowskiReLU(inplace=False)
def network_initialization(self, in_channels, out_channels, config, D):
net_metadata = self.net_metadata
def space_n_time_m(n, m):
return n if D == 3 else [n, n, n, m]
if D == 4:
self.OUT_PIXEL_DIST = space_n_time_m(self.OUT_PIXEL_DIST, 1)
# Output of the first conv concated to conv6
self.inplanes = self.PLANES[0]
self.conv1 = conv(in_channels,
self.inplanes,
pixel_dist=1,
kernel_size=space_n_time_m(5, 1),
stride=1,
dilation=1,
bias=False,
D=D,
net_metadata=net_metadata)
self.bn1 = nn.BatchNorm1d(self.inplanes)
self.block1 = self._make_layer(self.BLOCK,
self.PLANES[0],
self.LAYERS[0],
pixel_dist=1)
self.conv2p1s2 = conv(
self.inplanes,
self.inplanes,
pixel_dist=1,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bn2 = nn.BatchNorm1d(self.inplanes)
self.block2 = self._make_layer(self.BLOCK,
self.PLANES[1],
self.LAYERS[1],
pixel_dist=space_n_time_m(2, 1))
self.convtr2p2s2 = conv_tr(
self.inplanes,
self.PLANES[1],
pixel_dist=space_n_time_m(2, 1),
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bntr2 = nn.BatchNorm1d(self.PLANES[1])
self.conv3p2s2 = conv(
self.inplanes,
self.inplanes,
pixel_dist=space_n_time_m(2, 1),
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bn3 = nn.BatchNorm1d(self.inplanes)
self.block3 = self._make_layer(self.BLOCK,
self.PLANES[2],
self.LAYERS[2],
pixel_dist=space_n_time_m(4, 1))
self.convtr3p4s4 = conv_tr(
self.inplanes,
self.PLANES[2],
pixel_dist=space_n_time_m(4, 1),
kernel_size=space_n_time_m(4, 1),
upsample_stride=space_n_time_m(4, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bntr3 = nn.BatchNorm1d(self.PLANES[2])
self.conv4p4s2 = conv(
self.inplanes,
self.inplanes,
pixel_dist=space_n_time_m(4, 1),
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bn4 = nn.BatchNorm1d(self.inplanes)
self.block4 = self._make_layer(self.BLOCK,
self.PLANES[3],
self.LAYERS[3],
pixel_dist=space_n_time_m(8, 1))
self.convtr4p8s8 = conv_tr(
self.inplanes,
self.PLANES[3],
pixel_dist=space_n_time_m(8, 1),
kernel_size=space_n_time_m(8, 1),
upsample_stride=space_n_time_m(8, 1),
dilation=1,
conv_type=ConvType.SPATIAL_HYPERCUBE_TEMPORAL_HYPERCROSS,
bias=False,
D=D,
net_metadata=net_metadata)
self.bntr4 = nn.BatchNorm1d(self.PLANES[3])
self.relu = nn.ReLU(inplace=True)
self.final = conv(sum(self.PLANES[1:4]) +
self.PLANES[0] * self.BLOCK.expansion,
out_channels,
pixel_dist=1,
kernel_size=1,
stride=1,
dilation=1,
bias=True,
D=D,
net_metadata=net_metadata)
def network_initialization(self, in_channels, out_channels, config, D):
# Setup net_metadata
dilations = self.DILATIONS
bn_momentum = config['bn_momentum']
def space_n_time_m(n, m):
return n if D == 3 else [n, n, n, m]
if D == 4:
self.OUT_PIXEL_DIST = space_n_time_m(self.OUT_PIXEL_DIST, 1)
# Output of the first conv concated to conv6
self.inplanes = self.INIT_DIM
self.conv0p1s1 = conv(in_channels,
self.inplanes,
kernel_size=space_n_time_m(
config['conv1_kernel_size'], 1),
stride=1,
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn0 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.conv1p1s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn1 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block1 = self._make_layer(self.BLOCK,
self.PLANES[0],
self.LAYERS[0],
dilation=dilations[0],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv2p2s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn2 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block2 = self._make_layer(self.BLOCK,
self.PLANES[1],
self.LAYERS[1],
dilation=dilations[1],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv3p4s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn3 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block3 = self._make_layer(self.BLOCK,
self.PLANES[2],
self.LAYERS[2],
dilation=dilations[2],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv4p8s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn4 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block4 = self._make_layer(self.BLOCK,
self.PLANES[3],
self.LAYERS[3],
dilation=dilations[3],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.convtr4p16s2 = conv_tr(self.inplanes,
self.PLANES[4],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr4 = get_norm(self.NORM_TYPE,
self.PLANES[4],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[4] + self.PLANES[2] * self.BLOCK.expansion
self.block5 = self._make_layer(self.BLOCK,
self.PLANES[4],
self.LAYERS[4],
dilation=dilations[4],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.convtr5p8s2 = conv_tr(self.inplanes,
self.PLANES[5],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr5 = get_norm(self.NORM_TYPE,
self.PLANES[5],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[5] + self.PLANES[1] * self.BLOCK.expansion
self.block6 = self._make_layer(self.BLOCK,
self.PLANES[5],
self.LAYERS[5],
dilation=dilations[5],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.convtr6p4s2 = conv_tr(self.inplanes,
self.PLANES[6],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr6 = get_norm(self.NORM_TYPE,
self.PLANES[6],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[6] + self.PLANES[0] * self.BLOCK.expansion
self.block7 = self._make_layer(self.BLOCK,
self.PLANES[6],
self.LAYERS[6],
dilation=dilations[6],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.convtr7p2s2 = conv_tr(self.inplanes,
self.PLANES[7],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr7 = get_norm(self.NORM_TYPE,
self.PLANES[7],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[7] + self.INIT_DIM
self.block8 = self._make_layer(self.BLOCK,
self.PLANES[7],
self.LAYERS[7],
dilation=dilations[7],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.final = conv(self.PLANES[7],
out_channels,
kernel_size=1,
stride=1,
bias=True,
D=D)
self.relu = MinkowskiReLU(inplace=True)
def network_initialization(self, in_channels, out_channels, config, D):
# Setup net_metadata
dilations = self.DILATIONS
bn_momentum = config.bn_momentum
def space_n_time_m(n, m):
return n if D == 3 else [n, n, n, m]
if D == 4:
self.OUT_PIXEL_DIST = space_n_time_m(self.OUT_PIXEL_DIST, 1)
# Output of the first conv concated to conv6
self.inplanes = self.INIT_DIM
self.conv1p1s1 = conv(in_channels,
self.inplanes,
kernel_size=space_n_time_m(
config.conv1_kernel_size, 1),
stride=1,
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn1 = get_norm(self.NORM_TYPE,
self.PLANES[0],
D,
bn_momentum=bn_momentum)
self.block1 = self._make_layer(self.BLOCK,
self.PLANES[0],
self.LAYERS[0],
dilation=dilations[0],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv2p1s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn2 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block2 = self._make_layer(self.BLOCK,
self.PLANES[1],
self.LAYERS[1],
dilation=dilations[1],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv3p2s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn3 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block3 = self._make_layer(self.BLOCK,
self.PLANES[2],
self.LAYERS[2],
dilation=dilations[2],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.conv4p4s2 = conv(self.inplanes,
self.inplanes,
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn4 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block4 = self._make_layer(self.BLOCK,
self.PLANES[3],
self.LAYERS[3],
dilation=dilations[3],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.pool_tr4 = ME.MinkowskiPoolingTranspose(kernel_size=8,
stride=8,
dimension=D)
out_pool4 = self.inplanes
self.convtr4p8s2 = conv_tr(self.inplanes,
self.PLANES[4],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr4 = get_norm(self.NORM_TYPE,
self.PLANES[4],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[4] + self.PLANES[2] * self.BLOCK.expansion
self.block5 = self._make_layer(self.BLOCK,
self.PLANES[4],
self.LAYERS[4],
dilation=dilations[4],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.pool_tr5 = ME.MinkowskiPoolingTranspose(kernel_size=4,
stride=4,
dimension=D)
out_pool5 = self.inplanes
self.convtr5p4s2 = conv_tr(self.inplanes,
self.PLANES[5],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr5 = get_norm(self.NORM_TYPE,
self.PLANES[5],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[5] + self.PLANES[1] * self.BLOCK.expansion
self.block6 = self._make_layer(self.BLOCK,
self.PLANES[5],
self.LAYERS[5],
dilation=dilations[5],
norm_type=self.NORM_TYPE,
bn_momentum=bn_momentum)
self.pool_tr6 = ME.MinkowskiPoolingTranspose(kernel_size=2,
stride=2,
dimension=D)
out_pool6 = self.inplanes
self.convtr6p2s2 = conv_tr(self.inplanes,
self.PLANES[6],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr6 = get_norm(self.NORM_TYPE,
self.PLANES[6],
D,
bn_momentum=bn_momentum)
self.relu = MinkowskiReLU(inplace=True)
self.final = nn.Sequential(
conv(out_pool5 + out_pool6 + self.PLANES[6] +
self.PLANES[0] * self.BLOCK.expansion,
512,
kernel_size=1,
bias=False,
D=D), ME.MinkowskiBatchNorm(512), ME.MinkowskiReLU(),
conv(512, out_channels, kernel_size=1, bias=True, D=D))
def network_initialization(self, in_channels, out_channels, config, D):
if not isinstance(self.BLOCK, list): # if single type
self.BLOCK = [self.BLOCK] * len(self.PLANES)
# print('IN CHANNEL inside the model {}'.format(in_channels))
# Setup net_metadata
dilations = self.DILATIONS
bn_momentum = config.bn_momentum
block_noexpansion = True
def space_n_time_m(n, m):
return n if D == 3 else [n, n, n, m]
if D == 4:
self.OUT_PIXEL_DIST = space_n_time_m(self.OUT_PIXEL_DIST, 1)
if config.xyz_input is not None:
if config.xyz_input:
in_channels = in_channels + 3
if config.dataset == 'SemanticKITTI':
in_channels = 4
if config.dataset == 'S3DIS':
in_channels = 9
# Output of the first conv concated to conv6
self.inplanes = self.INIT_DIM
self.conv0p1s1 = conv(in_channels,
self.inplanes,
kernel_size=space_n_time_m(
config.conv1_kernel_size, 1),
stride=1,
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bn0 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.conv1p1s2 = conv(
self.inplanes,
self.inplanes if not block_noexpansion else self.PLANES[0],
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.inplanes = self.inplanes if not block_noexpansion else self.PLANES[
0]
self.bn1 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block1 = self._make_layer(
self.BLOCK[0],
self.PLANES[0],
self.LAYERS[0],
dilation=dilations[0],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.conv2p2s2 = conv(
self.inplanes,
self.inplanes if not block_noexpansion else self.PLANES[1],
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.inplanes = self.inplanes if not block_noexpansion else self.PLANES[
1]
self.bn2 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block2 = self._make_layer(
self.BLOCK[1],
self.PLANES[1],
self.LAYERS[1],
dilation=dilations[1],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.conv3p4s2 = conv(
self.inplanes,
self.inplanes if not block_noexpansion else self.PLANES[2],
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.inplanes = self.inplanes if not block_noexpansion else self.PLANES[
2]
self.bn3 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block3 = self._make_layer(
self.BLOCK[2],
self.PLANES[2],
self.LAYERS[2],
dilation=dilations[2],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.conv4p8s2 = conv(
self.inplanes,
self.inplanes if not block_noexpansion else self.PLANES[3],
kernel_size=space_n_time_m(2, 1),
stride=space_n_time_m(2, 1),
dilation=1,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.inplanes = self.inplanes if not block_noexpansion else self.PLANES[
3]
self.bn4 = get_norm(self.NORM_TYPE,
self.inplanes,
D,
bn_momentum=bn_momentum)
self.block4 = self._make_layer(
self.BLOCK[3],
self.PLANES[3],
self.LAYERS[3],
dilation=dilations[3],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.convtr4p16s2 = conv_tr(self.inplanes,
self.PLANES[4],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr4 = get_norm(self.NORM_TYPE,
self.PLANES[4],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[
4] + self.PLANES[2] * self.BLOCK[4].expansion
self.block5 = self._make_layer(
self.BLOCK[4],
self.PLANES[4],
self.LAYERS[4],
dilation=dilations[4],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.convtr5p8s2 = conv_tr(self.inplanes,
self.PLANES[5],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr5 = get_norm(self.NORM_TYPE,
self.PLANES[5],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[
5] + self.PLANES[1] * self.BLOCK[5].expansion
self.block6 = self._make_layer(
self.BLOCK[5],
self.PLANES[5],
self.LAYERS[5],
dilation=dilations[5],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.convtr6p4s2 = conv_tr(self.inplanes,
self.PLANES[6],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr6 = get_norm(self.NORM_TYPE,
self.PLANES[6],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[
6] + self.PLANES[0] * self.BLOCK[6].expansion
self.block7 = self._make_layer(
self.BLOCK[6],
self.PLANES[6],
self.LAYERS[6],
dilation=dilations[6],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.convtr7p2s2 = conv_tr(self.inplanes,
self.PLANES[7],
kernel_size=space_n_time_m(2, 1),
upsample_stride=space_n_time_m(2, 1),
dilation=1,
bias=False,
conv_type=self.NON_BLOCK_CONV_TYPE,
D=D)
self.bntr7 = get_norm(self.NORM_TYPE,
self.PLANES[7],
D,
bn_momentum=bn_momentum)
self.inplanes = self.PLANES[7] + self.INIT_DIM
self.block8 = self._make_layer(
self.BLOCK[7],
self.PLANES[7],
self.LAYERS[7],
dilation=dilations[7],
norm_type=self.NORM_TYPE,
nonlinearity_type=self.config.nonlinearity,
bn_momentum=bn_momentum)
self.final = conv(self.PLANES[7],
out_channels,
kernel_size=1,
stride=1,
bias=True,
D=D)
if config.enable_point_branch:
self.point_transform_mlp = nn.ModuleList([
nn.Sequential(
nn.Linear(self.INIT_DIM, self.PLANES[3]),
# nn.BatchNorm1d(self.PLANES[3]),
nn.ReLU(True)),
nn.Sequential(
nn.Linear(self.PLANES[3], self.PLANES[5]),
#nn.BatchNorm1d(self.PLANES[5]),
nn.ReLU(True)),
nn.Sequential(
nn.Linear(self.PLANES[5], self.PLANES[7]),
# nn.BatchNorm1d(self.PLANES[7]),
nn.ReLU(True))
])
self.downsample16x = nn.Sequential(
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3),
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3),
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3),
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3))
self.downsample4x = nn.Sequential(
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3),
ME.MinkowskiMaxPooling(kernel_size=2, stride=2, dimension=3))
self.dropout = nn.Dropout(0.1, True)
# self.dropout = nn.Dropout(0.3, True)
self.interpolate = ME.MinkowskiInterpolation(
return_kernel_map=False, return_weights=False)
def __init__(self,
in_channels,
out_channels,
config,
D=3,
return_feat=False,
**kwargs):
super(UNet, self).__init__(in_channels, out_channels, config, D)
self.in_channels = in_channels
self.return_feat = return_feat
PLANES = self.PLANES
# Output of the first conv concated to conv6
self.conv_down1 = conv(in_channels,
PLANES[0],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down1 = ME.MinkowskiBatchNorm(PLANES[0])
self.down1 = conv(PLANES[0], PLANES[1], kernel_size=2, stride=2, D=D)
self.conv_down2 = conv(PLANES[1],
PLANES[1],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down2 = ME.MinkowskiBatchNorm(PLANES[1])
self.down2 = conv(PLANES[1], PLANES[2], kernel_size=2, stride=2, D=D)
self.conv_down3 = conv(PLANES[2],
PLANES[2],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down3 = ME.MinkowskiBatchNorm(PLANES[2])
self.down3 = conv(PLANES[2], PLANES[3], kernel_size=2, stride=2, D=D)
self.conv_down4 = conv(PLANES[3],
PLANES[3],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down4 = ME.MinkowskiBatchNorm(PLANES[3])
self.down4 = conv(PLANES[3], PLANES[4], kernel_size=2, stride=2, D=D)
self.conv_down5 = conv(PLANES[4],
PLANES[4],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down5 = ME.MinkowskiBatchNorm(PLANES[4])
self.down5 = conv(PLANES[4], PLANES[5], kernel_size=2, stride=2, D=D)
self.conv_down6 = conv(PLANES[5],
PLANES[5],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_down6 = ME.MinkowskiBatchNorm(PLANES[5])
self.down6 = conv(PLANES[5], PLANES[6], kernel_size=2, stride=2, D=D)
self.conv7 = conv(PLANES[6],
PLANES[6],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn7 = ME.MinkowskiBatchNorm(PLANES[6])
self.up6 = conv_tr(PLANES[6],
PLANES[5],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up6 = conv(PLANES[5] * 2,
PLANES[5],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up6 = ME.MinkowskiBatchNorm(PLANES[5])
self.up5 = conv_tr(PLANES[5],
PLANES[4],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up5 = conv(PLANES[4] * 2,
PLANES[4],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up5 = ME.MinkowskiBatchNorm(PLANES[4])
self.up4 = conv_tr(PLANES[4],
PLANES[3],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up4 = conv(PLANES[3] * 2,
PLANES[3],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up4 = ME.MinkowskiBatchNorm(PLANES[3])
self.up3 = conv_tr(PLANES[3],
PLANES[2],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up3 = conv(PLANES[2] * 2,
PLANES[2],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up3 = ME.MinkowskiBatchNorm(PLANES[2])
self.up2 = conv_tr(PLANES[2],
PLANES[1],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up2 = conv(PLANES[1] * 2,
PLANES[1],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up2 = ME.MinkowskiBatchNorm(PLANES[1])
self.up1 = conv_tr(PLANES[1],
PLANES[0],
kernel_size=2,
upsample_stride=2,
D=D)
self.conv_up1 = conv(PLANES[0] * 2,
PLANES[0],
kernel_size=3,
stride=1,
bias=False,
D=D)
self.bn_up1 = ME.MinkowskiBatchNorm(PLANES[0])
self.mask_feat = conv(PLANES[0],
self.in_channels,
kernel_size=1,
bias=True,
D=D)
self.final = conv(PLANES[0],
out_channels,
kernel_size=1,
bias=True,
D=D)
self.relu = ME.MinkowskiReLU(inplace=True)
