def block_inside(i, o, index, norm_cfg):
mn = spconv.SparseSequential()
reps = 2
mn.add(build_norm_layer(norm_cfg, i)[1])
mn.add(nn.ReLU())
mn.add(SubMConv3d(i, i, 3, indice_key="su3_{}".format(index), bias=False))
mn.add(build_norm_layer(norm_cfg, i)[1])
mn.add(nn.ReLU())
mn.add(SubMConv3d(i, o, 3, indice_key="su3_{}".format(index), bias=False))
# mn.add(build_norm_layer(norm_cfg, o)[1])
# mn.add(nn.ReLU())
return mn
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="SASpMiddleResNetFHD",
**kwargs):
super(SASpMiddleResNetFHD, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
# input: # [1600, 1200, 41]
self.middle_conv1 = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=False,
indice_key="res0"),
build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseBasicBlock(16, 16, norm_cfg=norm_cfg, indice_key="res0"),
SparseBasicBlock(16, 16, norm_cfg=norm_cfg, indice_key="res0"),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=False), # [1600, 1200, 41] -> [800, 600, 21]
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SparseBasicBlock(32, 32, norm_cfg=norm_cfg, indice_key="res1"),
SparseBasicBlock(32, 32, norm_cfg=norm_cfg, indice_key="res1"),
)
self.middle_conv2 = spconv.SparseSequential(
SparseConv3d(32, 64, 3, 2, padding=1,
bias=False), # [800, 600, 21] -> [400, 300, 11]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SparseBasicBlock(64, 64, norm_cfg=norm_cfg, indice_key="res2"),
SparseBasicBlock(64, 64, norm_cfg=norm_cfg, indice_key="res2"),
)
self.middle_conv3 = spconv.SparseSequential(
SparseConv3d(64, 128, 3, 2, padding=[0, 1, 1],
bias=False), # [400, 300, 11] -> [200, 150, 5]
build_norm_layer(norm_cfg, 128)[1],
nn.ReLU(),
SparseBasicBlock(128, 128, norm_cfg=norm_cfg, indice_key="res3"),
SparseBasicBlock(128, 128, norm_cfg=norm_cfg, indice_key="res3"),
)
self.end_conv = spconv.SparseSequential(
SparseConv3d(128, 128, (3, 1, 1), (2, 1, 1),
bias=False), # [200, 150, 5] -> [200, 150, 2]
build_norm_layer(norm_cfg, 128)[1],
nn.ReLU(),
)
self.point_fc = nn.Linear(224, 64, bias=False)
self.point_cls = nn.Linear(64, 1, bias=False)
self.point_reg = nn.Linear(64, 3, bias=False)
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="SpMiddleFHD",
**kwargs):
super(SpMiddleFHDNobn, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
self.middle_conv = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=True,
indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SubMConv3d(16, 16, 3, bias=True, indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=True), # [1600, 1200, 41] -> [800, 600, 21]
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SparseConv3d(32, 64, 3, 2, padding=1,
bias=True), # [800, 600, 21] -> [400, 300, 11]
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SparseConv3d(64, 64, 3, 2, padding=[0, 1, 1],
bias=True), # [400, 300, 11] -> [200, 150, 5]
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SparseConv3d(64, 64, (3, 1, 1), (2, 1, 1),
bias=True), # [200, 150, 5] -> [200, 150, 2]
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
)
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="SpMiddleFHD",
**kwargs):
super(SpMiddleFHD, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
self.middle_conv = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=False,
indice_key="subm0"),
build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SubMConv3d(16, 16, 3, bias=False, indice_key="subm0"),
build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=False), # [41, 1600, 1408] -> [21, 800, 704]
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=False),
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=False),
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SparseConv3d(32, 64, 3, 2, padding=1,
bias=False), # [21, 800, 704] -> [11, 400, 352]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SparseConv3d(64, 64, 3, 2, padding=[0, 1, 1],
bias=False), # [11, 400, 352] -> [5, 200, 176]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm3", bias=False),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SparseConv3d(64, 64, (3, 1, 1), (2, 1, 1),
bias=False), # [5, 200, 176] -> [2, 200, 176]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
)
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="RCNNSpMiddleFHD",
**kwargs):
super(RCNNSpMiddleFHD, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
self.middle_conv = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=False,
indice_key="subm0"),
build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SubMConv3d(16, 16, 3, bias=False, indice_key="subm0"),
build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=False), # [32, 80, 41] -> [16, 40, 21]
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, bias=False, indice_key="subm1"),
build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
# SubMConv3d(32, 32, 3, bias=False, indice_key="subm1"),
# build_norm_layer(norm_cfg, 32)[1],
# nn.ReLU(),
SparseConv3d(32, 64, 3, 2, bias=False,
padding=1), # [16, 40, 21] -> [8, 20, 11]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, bias=False, indice_key="subm2"),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
# SubMConv3d(64, 64, 3, bias=False, indice_key="subm2"),
# build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
# SubMConv3d(64, 64, 3, bias=False, indice_key="subm2"),
# build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
SparseConv3d(64, 64, 3, 2, bias=False,
padding=[1, 1, 0]), # [8, 20, 11] -> [4, 10, 5]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, bias=False, indice_key="subm3"),
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
# SubMConv3d(64, 64, 3, bias=False, indice_key="subm3"),
# build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
# SubMConv3d(64, 64, 3, bias=False, indice_key="subm3"),
# build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
SparseConv3d(64, 64, (1, 1, 3), (1, 1, 2),
bias=False), # [4, 10, 5] -> [4, 10, 2]
build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
)
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="SpMiddleResNetFHD",
n_channels=[16, 32, 64, 128],
**kwargs):
super(SpMiddleResNetFHDv2, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
# input: # [1024, 1024, 56]
self.middle_conv = spconv.SparseSequential(
SubMConv3d(num_input_features,
n_channels[0],
3,
bias=False,
indice_key="res0"),
build_norm_layer(norm_cfg, n_channels[0])[1],
nn.ReLU(),
SparseBasicBlock(n_channels[0],
n_channels[0],
norm_cfg=norm_cfg,
indice_key="res0"),
SparseBasicBlock(n_channels[0],
n_channels[0],
norm_cfg=norm_cfg,
indice_key="res0"),
SparseConv3d(n_channels[0],
n_channels[1],
3,
2,
padding=1,
bias=False), # [1024, 1024, 56] -> [512, 512, 28]
build_norm_layer(norm_cfg, n_channels[1])[1],
nn.ReLU(),
SparseBasicBlock(n_channels[1],
n_channels[1],
norm_cfg=norm_cfg,
indice_key="res1"),
SparseBasicBlock(n_channels[1],
n_channels[1],
norm_cfg=norm_cfg,
indice_key="res1"),
# SparseBasicBlock(n_channels[1], n_channels[1], norm_cfg=norm_cfg, indice_key="res1"),
SparseConv3d(n_channels[1],
n_channels[2],
3,
2,
padding=1,
bias=False), # [512, 512, 28] -> [256, 256, 14]
build_norm_layer(norm_cfg, n_channels[2])[1],
nn.ReLU(),
SparseBasicBlock(n_channels[2],
n_channels[2],
norm_cfg=norm_cfg,
indice_key="res2"),
SparseBasicBlock(n_channels[2],
n_channels[2],
norm_cfg=norm_cfg,
indice_key="res2"),
# SparseBasicBlock(n_channels[2], n_channels[2], norm_cfg=norm_cfg, indice_key="res2"),
SparseConv3d(n_channels[2],
n_channels[3],
3,
2,
padding=1,
bias=False), # [256, 256, 14] -> [128, 128, 7]
build_norm_layer(norm_cfg, n_channels[3])[1],
nn.ReLU(),
SparseBasicBlock(n_channels[3],
n_channels[3],
norm_cfg=norm_cfg,
indice_key="res3"),
SparseBasicBlock(n_channels[3],
n_channels[3],
norm_cfg=norm_cfg,
indice_key="res3"),
# SparseBasicBlock(n_channels[3], n_channels[3], norm_cfg=norm_cfg, indice_key="res3"),
SparseConv3d(n_channels[3],
n_channels[3], (3, 1, 1), (2, 1, 1),
bias=False), # [128, 128, 7] -> [64, 64, 3]
build_norm_layer(norm_cfg, n_channels[3])[1],
nn.ReLU(),
)
def __init__(self,
num_input_features=128,
num_filters_down=[16, 32, 64, 64, 96, 128],
dimension_feature_map=[128, 128, 128, 128],
norm_cfg=None,
name="tDBN_2",
**kwargs):
super(tDBN_2_bv, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
middle_layers = spconv.SparseSequential()
input_filters_layers = num_filters_down[:
4] # feature channels in the raw data
num_filter_fpn = num_filters_down[-4:]
dimension_feature_map = dimension_feature_map
dimension_kernel_size = [15, 7, 3, 1]
# the main network framework
self.block_input = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=True,
indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SubMConv3d(16, 16, 3, bias=True, indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=True), # [1600, 1200, 41] -> [800, 600, 21]
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SparseConv3d(32, 64, 3, 2, padding=[0, 2, 2],
bias=True), # [800, 600, 21] -> [400, 300, 11]
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
# SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# # build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
# SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# # build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
# SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# # build_norm_layer(norm_cfg, 64)[1],
# nn.ReLU(),
)
# self.block0 = middle_layers
m = spconv.SparseSequential()
reps = 2
# residual_use = False # using residual block or not
residual_use = True # using residual block or not
_index = 0
for _ in range(reps):
block(m,
num_filter_fpn[0],
num_filter_fpn[0],
index=_index,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
self.x0_in = m
for k in range(1, 4):
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[k - 1],
num_filter_fpn[k],
3,
stride=2,
bias=False,
indice_key="conv{}".format(k)))
m.add(build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
m.add(nn.ReLU())
for _ in range(reps):
if k == 4:
block(m,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
dimension=2,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
else:
block(m,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
dimension=3,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 1:
self.x1_in = m
elif k == 2:
self.x2_in = m
elif k == 3:
self.x3_in = m
for k in range(2, -1, -1):
m = spconv.SparseSequential()
m.add(
SparseInverseConv3d(num_filter_fpn[k + 1],
num_filter_fpn[k],
3,
"conv{}".format(k + 1),
bias=False) # no stride inputs
)
m.add(build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
m.add(nn.ReLU())
if k == 2:
self.upsample32 = m
elif k == 1:
self.upsample21 = m
elif k == 0:
self.upsample10 = m
m = spconv.SparseSequential()
m.add(JoinTable())
for i in range(reps):
block(m,
num_filter_fpn[k] * (2 if i == 0 else 1),
num_filter_fpn[k],
index=(k + 100),
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 2:
self.concate2 = m
elif k == 1:
self.concate1 = m
elif k == 0:
self.concate0 = m
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[k],
dimension_feature_map[k],
(dimension_kernel_size[k], 1, 1), (2, 1, 1),
bias=False))
m.add(build_norm_layer(norm_cfg, dimension_feature_map[k])[1])
m.add(nn.ReLU())
if k == 2:
self.feature_map2 = m
elif k == 1:
self.feature_map1 = m
elif k == 0:
self.feature_map0 = m
def __init__(self,
num_input_features=4,
num_filters_down=[16, 32, 64, 64, 96, 128],
dimension_feature_map=[128, 128, 128, 128],
norm_cfg=None,
name="tDBN_1",
**kwargs):
super(tDBN_1_bv, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
middle_layers = spconv.SparseSequential()
input_filters_layers = num_filters_down[:
4] # feature channels in the raw data
num_filter_fpn = num_filters_down[-4:]
dimension_feature_map = dimension_feature_map
dimension_kernel_size = [15, 7, 3, 1]
self.block0 = spconv.SparseSequential(
SubMConv3d(num_input_features,
16,
3,
bias=True,
indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SubMConv3d(16, 16, 3, bias=True, indice_key="subm0"),
# build_norm_layer(norm_cfg, 16)[1],
nn.ReLU(),
SparseConv3d(16, 32, 3, 2, padding=1,
bias=True), # [1600, 1200, 41] -> [800, 600, 21]
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SubMConv3d(32, 32, 3, indice_key="subm1", bias=True),
# build_norm_layer(norm_cfg, 32)[1],
nn.ReLU(),
SparseConv3d(32, 64, 3, 2, padding=[0, 2, 2],
bias=True), # [800, 600, 21] -> [400, 300, 11]
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
SubMConv3d(64, 64, 3, indice_key="subm2", bias=True),
# build_norm_layer(norm_cfg, 64)[1],
nn.ReLU(),
)
# self.block0 = middle_layers
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[0],
dimension_feature_map[0],
(dimension_kernel_size[0], 1, 1), (2, 1, 1),
bias=False))
m.add(build_norm_layer(norm_cfg, dimension_feature_map[0])[1])
m.add(nn.ReLU())
# dense function add later
self.feature_map0 = m
reps = 2
residual_use = True # using residual block or not
m = spconv.SparseSequential()
## block1-3 and feature map1-3
for k in range(1, 4):
m.add(
SparseConv3d(num_filter_fpn[k - 1],
num_filter_fpn[k],
3,
2,
bias=False))
m.add(build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
m.add(nn.ReLU())
# dense function add later
for _ in range(reps):
block(m,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 1:
self.block1 = m
elif k == 2:
self.block2 = m
elif k == 3:
self.block3 = m
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[k],
dimension_feature_map[k],
(dimension_kernel_size[k], 1, 1), (2, 1, 1),
bias=False))
m.add(build_norm_layer(norm_cfg, dimension_feature_map[k])[1])
m.add(nn.ReLU())
# adding dense in forward step
if k == 1:
self.feature_map1 = m
elif k == 2:
self.feature_map2 = m
elif k == 3:
self.feature_map3 = None # convert the sparse data into dense one
# Sequential(scn.SparseToDense(3, dimension_feature_map[k])) ## last one is the 2D instead of 3D
m = spconv.SparseSequential()
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="tDBN_2",
**kwargs):
super(tDBN_2, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
middle_layers = spconv.SparseSequential()
num_filter_fpn = [32, 64, 96, 128]
dimension_feature_map = [128, 128, 128, 128]
dimension_kernel_size = [15, 7, 3, 1]
# the main network framework
m = spconv.SparseSequential()
# --------------------------------------
## block1 and feature map 0, convert from voxel into 3D tensor
# --------------------------------------
for i, o in [[num_input_features, num_filter_fpn[0]]
]: # , [num_filter_fpn[0], num_filter_fpn[0]]]:
m.add(SubMConv3d(i, o, 3, indice_key="subm_0", bias=False))
m.add(build_norm_layer(norm_cfg, o)[1])
m.add(nn.ReLU())
self.block_input = m
m = spconv.SparseSequential()
reps = 2
# residual_use = False # using residual block or not
residual_use = True # using residual block or not
_index = 0
for _ in range(reps):
block(m,
num_filter_fpn[0],
num_filter_fpn[0],
index=_index,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
self.x0_in = m
for k in range(1, 4):
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[k - 1],
num_filter_fpn[k],
3,
stride=2,
bias=False,
indice_key="conv{}".format(k)))
m.add(build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
m.add(nn.ReLU())
for _ in range(reps):
if k == 4:
block(m,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
dimension=2,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
else:
block(m,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
dimension=3,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 1:
self.x1_in = m
elif k == 2:
self.x2_in = m
elif k == 3:
self.x3_in = m
for k in range(2, -1, -1):
m = spconv.SparseSequential()
m.add(
SparseInverseConv3d(num_filter_fpn[k + 1],
num_filter_fpn[k],
3,
"conv{}".format(k + 1),
bias=False) # no stride inputs
)
m.add(build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
m.add(nn.ReLU())
if k == 2:
self.upsample32 = m
elif k == 1:
self.upsample21 = m
elif k == 0:
self.upsample10 = m
m = spconv.SparseSequential()
m.add(JoinTable())
for i in range(reps):
block(m,
num_filter_fpn[k] * (2 if i == 0 else 1),
num_filter_fpn[k],
index=(k + 100),
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 2:
self.concate2 = m
elif k == 1:
self.concate1 = m
elif k == 0:
self.concate0 = m
m = spconv.SparseSequential()
m.add(
SparseConv3d(num_filter_fpn[k],
dimension_feature_map[k],
(dimension_kernel_size[k], 1, 1), (2, 1, 1),
bias=False))
m.add(build_norm_layer(norm_cfg, dimension_feature_map[k])[1])
m.add(nn.ReLU())
if k == 2:
self.feature_map2 = m
elif k == 1:
self.feature_map1 = m
elif k == 0:
self.feature_map0 = m
def __init__(self,
num_input_features=128,
norm_cfg=None,
name="tDBN_1",
**kwargs):
super(tDBN_1, self).__init__()
self.name = name
self.dcn = None
self.zero_init_residual = False
if norm_cfg is None:
norm_cfg = dict(type="BN1d", eps=1e-3, momentum=0.01)
middle_layers = spconv.SparseSequential()
num_filter_fpn = [32, 64, 96, 128]
dimension_feature_map = [128, 128, 128, 128]
dimension_kernel_size = [15, 7, 3, 1]
for i, o in [[num_input_features, num_filter_fpn[0]]]:
middle_layers.add(
SubMConv3d(i, o, 3, indice_key="subm_0", bias=False))
middle_layers.add(build_norm_layer(norm_cfg, o)[1])
middle_layers.add(nn.ReLU())
self.block0 = middle_layers
middle_layers = spconv.SparseSequential()
middle_layers.add(
SparseConv3d(num_filter_fpn[0],
dimension_feature_map[0],
(dimension_kernel_size[0], 1, 1), (2, 1, 1),
bias=False))
middle_layers.add(
build_norm_layer(norm_cfg, dimension_feature_map[0])[1])
middle_layers.add(nn.ReLU())
# dense function add later
self.feature_map0 = middle_layers
reps = 2
residual_use = True
middle_layers = spconv.SparseSequential()
## block1-3 and feature map1-3
for k in range(1, 4):
if k == 3:
middle_layers.add(
SparseConv3d(num_filter_fpn[k - 1],
num_filter_fpn[k],
3,
2,
bias=False))
middle_layers.add(
build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
middle_layers.add(nn.ReLU())
# dense function add later
else:
middle_layers.add(
SparseConv3d(num_filter_fpn[k - 1],
num_filter_fpn[k],
3,
2,
bias=False))
middle_layers.add(
build_norm_layer(norm_cfg, num_filter_fpn[k])[1])
middle_layers.add(nn.ReLU())
# 128*7*199*175 recurrent
for _ in range(reps):
block(middle_layers,
num_filter_fpn[k],
num_filter_fpn[k],
index=k,
residual_blocks=residual_use,
norm_cfg=norm_cfg)
if k == 1:
self.block1 = middle_layers
elif k == 2:
self.block2 = middle_layers
elif k == 3:
self.block3 = middle_layers
middle_layers = spconv.SparseSequential()
middle_layers.add(
SparseConv3d(num_filter_fpn[k],
dimension_feature_map[k],
(dimension_kernel_size[k], 1, 1), (2, 1, 1),
bias=False))
middle_layers.add(
build_norm_layer(norm_cfg, dimension_feature_map[k])[1])
middle_layers.add(nn.ReLU())
# adding dense in forward step
if k == 1:
self.feature_map1 = middle_layers
elif k == 2:
self.feature_map2 = middle_layers
elif k == 3:
self.feature_map3 = None # convert the sparse data into dense one
# Sequential(scn.SparseToDense(3, dimension_feature_map[k])) ## last one is the 2D instead of 3D
middle_layers = spconv.SparseSequential()
def block(m,
i,
o,
dimension=3,
index=0,
residual_blocks=False,
norm_cfg=None): # default using residual_block
if dimension == 3: ## 3x3x3 convlution
if residual_blocks: #ResNet style blocks
m.add(
SubMConv3d(i,
o,
3,
indice_key="su3_{}".format(index),
bias=False))
m.add(ConcatTable().add(Identity()).add(
block_inside(o, o, index, norm_cfg)))
m.add(AddTable())
#m.add(SubMConv3d(2*o, o, 3, indice_key="su3_{}".format(index), bias = False))
m.add(build_norm_layer(norm_cfg, o)[1])
m.add(nn.ReLU())
else:
m.add(
SubMConv3d(i,
i,
3,
indice_key="su3_{}".format(index),
bias=False))
m.add(build_norm_layer(norm_cfg, i)[1])
m.add(nn.ReLU())
m.add(
SubMConv3d(i,
o,
3,
indice_key="su3_{}".format(index),
bias=False))
m.add(build_norm_layer(norm_cfg, o)[1])
m.add(nn.ReLU())
else: # 2x2x2 convoltion
if residual_blocks: #ResNet style blocks
m.add(ConcatTable().add(Identity()).add(
SubMConv2d(i,
o,
3,
indice_key="su2_{}".format(index),
bias=False)).add(build_norm_layer(
norm_cfg, o)[1]).add(nn.ReLU()).add(
SubMConv2d(
i,
o,
3,
indice_key="su2_{}".format(index),
bias=False)).add(
build_norm_layer(
norm_cfg, o)[1]).add(
nn.ReLU())).add(AddTable())
else:
m.add(spconv.SparseSequential().add(
SubMConv2d(i,
o,
3,
indice_key="su2_{}".format(index),
bias=False)).add(build_norm_layer(
norm_cfg, o)[1]).add(nn.ReLU()).add(
SubMConv2d(
i,
o,
3,
indice_key="su2_{}".format(index),
bias=False)).add(
build_norm_layer(
norm_cfg, o)[1]).add(nn.ReLU()))