def __init__(self):
super(ToyModel, self).__init__()
self.paconvs = nn.ModuleList()
self.paconvs.append(PAConv(8, 16, 8))
self.paconvs.append(PAConv(8, 16, 8, kernel_input='identity'))
self.paconvs.append(PAConvCUDA(8, 16, 8))
self.conv1 = nn.Conv1d(3, 8, 1)
def __init__(self,
num_point,
radii,
sample_nums,
mlp_channels,
paconv_num_kernels,
fps_mod=['D-FPS'],
fps_sample_range_list=[-1],
dilated_group=False,
norm_cfg=dict(type='BN2d', momentum=0.1),
use_xyz=True,
pool_mod='max',
normalize_xyz=False,
bias='auto',
paconv_kernel_input='w_neighbor',
scorenet_input='w_neighbor_dist',
scorenet_cfg=dict(mlp_channels=[16, 16, 16],
score_norm='softmax',
temp_factor=1.0,
last_bn=False)):
super(PAConvSAModuleMSG,
self).__init__(num_point=num_point,
radii=radii,
sample_nums=sample_nums,
mlp_channels=mlp_channels,
fps_mod=fps_mod,
fps_sample_range_list=fps_sample_range_list,
dilated_group=dilated_group,
use_xyz=use_xyz,
pool_mod=pool_mod,
normalize_xyz=normalize_xyz,
grouper_return_grouped_xyz=True)
assert len(paconv_num_kernels) == len(mlp_channels)
for i in range(len(mlp_channels)):
assert len(paconv_num_kernels[i]) == len(mlp_channels[i]) - 1, \
'PAConv number of weight kernels wrong'
# in PAConv, bias only exists in ScoreNet
scorenet_cfg['bias'] = bias
for i in range(len(self.mlp_channels)):
mlp_channel = self.mlp_channels[i]
if use_xyz:
mlp_channel[0] += 3
num_kernels = paconv_num_kernels[i]
mlp = nn.Sequential()
for i in range(len(mlp_channel) - 1):
mlp.add_module(
f'layer{i}',
PAConv(mlp_channel[i],
mlp_channel[i + 1],
num_kernels[i],
norm_cfg=norm_cfg,
kernel_input=paconv_kernel_input,
scorenet_input=scorenet_input,
scorenet_cfg=scorenet_cfg))
self.mlps.append(mlp)
def test_paconv():
B = 2
in_channels = 6
out_channels = 12
npoint = 4
K = 3
points_xyz = torch.randn(B, 3, npoint, K)
features = torch.randn(B, in_channels, npoint, K)
paconv = PAConv(in_channels, out_channels, 4)
with torch.no_grad():
new_features = paconv(points_xyz, features)
assert new_features.shape == torch.Size([B, out_channels, npoint, K])
def test_paconv():
B = 2
in_channels = 6
out_channels = 12
npoint = 4
K = 3
num_kernels = 4
points_xyz = torch.randn(B, 3, npoint, K)
features = torch.randn(B, in_channels, npoint, K)
paconv = PAConv(in_channels, out_channels, num_kernels)
assert paconv.weight_bank.shape == torch.Size(
[in_channels * 2, out_channels * num_kernels])
with torch.no_grad():
new_features, _ = paconv((features, points_xyz))
assert new_features.shape == torch.Size([B, out_channels, npoint, K])