forked from torch-points3d/torch-points3d
-
Notifications
You must be signed in to change notification settings - Fork 47
/
Copy pathpointnet2.py
191 lines (160 loc) · 6.34 KB
/
pointnet2.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import os
import sys
from omegaconf import DictConfig, OmegaConf
import logging
from torch_points3d.applications.modelfactory import ModelFactory
from torch_points3d.modules.pointnet2 import *
from torch_points3d.core.base_conv.dense import DenseFPModule
from torch_points3d.models.base_architectures.unet import UnwrappedUnetBasedModel
from torch_points3d.datasets.multiscale_data import MultiScaleBatch
from torch_points3d.core.common_modules.dense_modules import Conv1D
from torch_points3d.core.common_modules.base_modules import Seq
from .utils import extract_output_nc
CUR_FILE = os.path.realpath(__file__)
DIR_PATH = os.path.dirname(os.path.realpath(__file__))
PATH_TO_CONFIG = os.path.join(DIR_PATH, "conf/pointnet2")
log = logging.getLogger(__name__)
def PointNet2(
architecture: str = None,
input_nc: int = None,
num_layers: int = None,
config: DictConfig = None,
multiscale=False,
*args,
**kwargs
):
""" Create a PointNet2 backbone model based on the architecture proposed in
https://arxiv.org/abs/1706.02413
Parameters
----------
architecture : str, optional
Architecture of the model, choose from unet, encoder and decoder
input_nc : int, optional
Number of channels for the input
output_nc : int, optional
If specified, then we add a fully connected head at the end of the network to provide the requested dimension
num_layers : int, optional
Depth of the network
config : DictConfig, optional
Custom config, overrides the num_layers and architecture parameters
"""
factory = PointNet2Factory(
architecture=architecture,
num_layers=num_layers,
input_nc=input_nc,
multiscale=multiscale,
config=config,
**kwargs
)
return factory.build()
class PointNet2Factory(ModelFactory):
def _build_unet(self):
if self._config:
model_config = self._config
else:
path_to_model = os.path.join(
PATH_TO_CONFIG, "unet_{}_{}.yaml".format(self.num_layers, "ms" if self.kwargs["multiscale"] else "ss")
)
model_config = OmegaConf.load(path_to_model)
ModelFactory.resolve_model(model_config, self.num_features, self._kwargs)
modules_lib = sys.modules[__name__]
return PointNet2Unet(model_config, None, None, modules_lib, **self.kwargs)
def _build_encoder(self):
if self._config:
model_config = self._config
else:
path_to_model = os.path.join(
PATH_TO_CONFIG,
"encoder_{}_{}.yaml".format(self.num_layers, "ms" if self.kwargs["multiscale"] else "ss"),
)
model_config = OmegaConf.load(path_to_model)
ModelFactory.resolve_model(model_config, self.num_features, self._kwargs)
modules_lib = sys.modules[__name__]
return PointNet2Encoder(model_config, None, None, modules_lib, **self.kwargs)
class BasePointnet2(UnwrappedUnetBasedModel):
CONV_TYPE = "dense"
def __init__(self, model_config, model_type, dataset, modules, *args, **kwargs):
super(BasePointnet2, self).__init__(model_config, model_type, dataset, modules)
try:
default_output_nc = extract_output_nc(model_config)
except:
default_output_nc = -1
log.warning("Could not resolve number of output channels")
self._has_mlp_head = False
self._output_nc = default_output_nc
if "output_nc" in kwargs:
self._has_mlp_head = True
self._output_nc = kwargs["output_nc"]
self.mlp = Seq()
self.mlp.append(Conv1D(default_output_nc, self._output_nc, bn=True, bias=False))
@property
def has_mlp_head(self):
return self._has_mlp_head
@property
def output_nc(self):
return self._output_nc
def _set_input(self, data):
"""Unpack input data from the dataloader and perform necessary pre-processing steps.
"""
assert len(data.pos.shape) == 3
data = data.to(self.device)
if data.x is not None:
data.x = data.x.transpose(1, 2).contiguous()
else:
data.x = None
self.input = data
class PointNet2Encoder(BasePointnet2):
def forward(self, data, *args, **kwargs):
"""
Parameters:
-----------
data
A dictionary that contains the data itself and its metadata information. Should contain
x -- Features [B, N, C]
pos -- Points [B, N, 3]
"""
self._set_input(data)
data = self.input
stack_down = [data]
for i in range(len(self.down_modules) - 1):
data = self.down_modules[i](data)
stack_down.append(data)
data = self.down_modules[-1](data)
if not isinstance(self.inner_modules[0], Identity):
stack_down.append(data)
data = self.inner_modules[0](data)
if self.has_mlp_head:
data.x = self.mlp(data.x)
return data
class PointNet2Unet(BasePointnet2):
def forward(self, data, *args, **kwargs):
""" This method does a forward on the Unet assuming symmetrical skip connections
Input --- D1 -- D2 -- I -- U1 -- U2 -- U3 -- output
| | |________| | |
| |______________________| |
|___________________________________|
Parameters:
-----------
data
A dictionary that contains the data itself and its metadata information. Should contain
x -- Features [B, N, C]
pos -- Points [B, N, 3]
"""
self._set_input(data)
data = self.input
stack_down = [data]
for i in range(len(self.down_modules) - 1):
data = self.down_modules[i](data)
stack_down.append(data)
data = self.down_modules[-1](data)
if not isinstance(self.inner_modules[0], Identity):
stack_down.append(data)
data = self.inner_modules[0](data)
sampling_ids = self._collect_sampling_ids(stack_down)
for i in range(len(self.up_modules)):
data = self.up_modules[i]((data, stack_down.pop()))
for key, value in sampling_ids.items():
setattr(data, key, value)
if self.has_mlp_head:
data.x = self.mlp(data.x)
return data