-
Notifications
You must be signed in to change notification settings - Fork 4
/
model.py
341 lines (299 loc) · 13.7 KB
/
model.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
from __future__ import print_function
from math import pi
import torch
import torch.nn as nn
import torch.nn.parallel
import torch.utils.data
from torch.autograd import Variable
import numpy as np
import torch.nn.functional as F
import sys
import pointnet_feat as pn
import softpool as sp
from other_models.MSN import msn
# import MSN.MDS.MDS_module as MDS_module
import MSN.expansion_penalty.expansion_penalty_module as expasion
# from MSN.MDS.MDS_module import MinimumDensitySampling as mds
import argparse
def dict2namespace(config):
namespace = argparse.Namespace()
for key, value in config.items():
if isinstance(value, dict):
new_value = dict2namespace(value)
else:
new_value = value
setattr(namespace, key, new_value)
return namespace
class Network(nn.Module):
def __init__(self,
npoints=8192,
n_regions=8,
dim_pn=512,
sp_points=1024,
model_lists=['softpool', 'msn', 'folding', 'grnet']):
super(Network, self).__init__()
self.do_segment = False
self.npoints_in = 2048
self.npoints = npoints
self.dim_pn = dim_pn
self.n_regions = n_regions
self.sp_points = sp_points
self.n_regions = n_regions
self.model_lists = model_lists
if ('softpool' in self.model_lists):
self.sp_enc = sp.Encoder_softpool(
regions=self.n_regions,
npoints=self.npoints_in,
sp_ratio=self.n_regions,
dim_pn=self.dim_pn)
self.sp_dec = sp.Decoder_softpool(
regions=self.n_regions,
npoints=self.npoints_in,
sp_ratio=self.n_regions,
dim_pn=self.dim_pn)
if ('folding' in self.model_lists):
self.pn_enc = nn.Sequential(
pn.PointNetFeat(npoints, 1024), nn.Linear(1024, dim_pn),
nn.BatchNorm1d(dim_pn), nn.ReLU())
self.decoder_fold = msn.PointGenCon(
bottleneck_size=2 + self.dim_pn)
if ('msn' in self.model_lists):
self.pn_enc = nn.Sequential(
pn.PointNetFeat(npoints, 1024), nn.Linear(1024, dim_pn),
nn.BatchNorm1d(dim_pn), nn.ReLU())
self.expansion = expasion.expansionPenaltyModule()
self.msn = msn.MSN()
if ('grnet' in self.model_lists):
from GRNet import grnet
self.grnet = grnet.GRNet()
if ('im_grnet' in self.model_lists):
# NOTE: 2D encoder
from pix2vox.encoder import Encoder
self.encoder_img = Encoder()
from GRNet import grnet
self.grnet = grnet.GRNet()
if ('pcn' in self.model_lists):
from pcn import PCN
self.pcn = PCN().cuda()
if ('disp3d' in self.model_lists):
import displace.encode as disp3d
self.disp_enc = disp3d.Encoder(support_num=10, neighbor_num=20)
# NOTE: initial setting for disp3d without image inputs
# degrees=[1, 2, 2, 2, 2, 4, 64],
from displace.decode import Decoder
self.disp_dec = Decoder(
features=[1024, 256, 256, 256, 128, 128, 128, 3],
degrees=[1, 2, 2, 2, 2, 4, 64],
support=10,
root_num=1)
if self.do_segment is True:
self.disp_seg = disp3d.Disp3D(
class_num=12, support_num=10, neighbor_num=20)
if ('im_disp3d' in self.model_lists):
from pix2vox.encoder import Encoder
self.encoder_img = Encoder()
# NOTE: initial setting for disp3d without image inputs
# degrees=[1, 2, 2, 2, 2, 4, 64],
from displace.decode import Decoder
self.disp_dec = Decoder(
features=[1024, 256, 256, 256, 128, 128, 128, 3],
degrees=[1, 2, 2, 2, 2, 4, 64],
support=10,
root_num=1)
if ('shapegf' in self.model_lists):
import yaml
import importlib
with open('other_models/shapegf/shapenet_recon.yaml', 'r') as f:
cfg = yaml.load(f)
cfg = dict2namespace(cfg)
cfg.log_name = "logs/val"
cfg.save_dir = "logs/val"
cfg.log_dir = "logs/val"
trainer_lib = importlib.import_module(cfg.trainer.type)
self.trainer = trainer_lib.Trainer(cfg)
self.enc = self.trainer.encoder
self.dec = self.trainer.score_net
self.sigma = self.trainer.sigmas
if ('vrcnet' in self.model_lists):
import munch
import yaml
args = munch.munchify(
yaml.safe_load(open('other_models/VRCNet/vrcnet.yaml')))
from VRCNet.vrcnet import Model
self.vrcnet = Model(args)
if self.do_segment is True:
import displace.encode as disp3d
self.disp_seg = disp3d.Disp3D(
class_num=12, support_num=10, neighbor_num=20)
if ('pointr' in self.model_lists):
import yaml
from pointr.PoinTr import PoinTr
self.pointr = PoinTr(
dict2namespace(
yaml.load(
open('other_models/pointr/PoinTr.yaml'))['model']))
# self.pointr = PoinTr(dict2namespace(yaml.load(open('other_models/pointr/ftrans.yaml'))['model']))
if ('im_pointr' in self.model_lists):
# NOTE: 2D encoder: ResNet
from pix2vox.encoder import Encoder
self.encoder_img = Encoder()
"""
# NOTE: 2D encoder: ConvTransformer
from pix2vox.model import Model_encoder, Bottleneck
self.encoder_img = Model_encoder(Bottleneck, [3, 4, 6, 3])
"""
# NOTE: 3D decoder
import yaml
from imgpointr.PoinTr import PoinTr
self.pointr = PoinTr(
dict2namespace(
yaml.load(
open('other_models/imgpointr/PoinTr.yaml'))['model']))
if ('snowflake' in self.model_lists):
from snowflake.model import SnowflakeNet
self.snowflake = SnowflakeNet(dim_feat=512, up_factors=[2, 2])
if ('im_snowflake' in self.model_lists):
# NOTE: 2D encoder
from pix2vox.encoder import Encoder
self.encoder_img = Encoder()
# self.pooler = nn.MaxPool2d(2, stride=2)
from snowflake.model import SnowflakeNet
# NOTE: 2.5D encoder + 3D decoder
self.snowflake = SnowflakeNet(
dim_feat=1024, up_factors=[2, 2], global_feat=True)
def forward(self, part, images=[]):
output = {
'softpool': [],
'msn': [],
'folding': [],
'grnet': [],
'shapegf': [],
'disp3d': [],
'pcn': [],
'pointr': []
}
if ('msn' in self.model_lists):
# transpose part when using displace
pn_feat = self.pn_enc(part)
[pcd_msn1, pcd_msn2, loss_mst, mean_mst_dis] = self.msn(
part, pn_feat)
if ('softpool' in self.model_lists):
input_chosen, feat_softpool = self.sp_enc(part=part)
pcd_softpool, pcd_fusion, loss_mst = self.sp_dec(
feature=feat_softpool, part=part)
if ('folding' in self.model_lists):
# transpose part when using displace
pn_feat = self.pn_enc(part)
mesh_grid = torch.meshgrid([
torch.linspace(0.0, 1.0, 64),
torch.linspace(0.0, 1.0, self.npoints // 64)
])
mesh_grid = torch.cat(
(torch.reshape(mesh_grid[0], (self.npoints, 1)),
torch.reshape(mesh_grid[1], (self.npoints, 1))),
dim=1)
mesh_grid = torch.transpose(mesh_grid, 0, 1).unsqueeze(0).repeat(
part.shape[0], 1, 1).cuda()
pn_feat = pn_feat.unsqueeze(2).expand(
part.size(0), self.dim_pn, self.npoints).contiguous()
y = torch.cat((mesh_grid, pn_feat), 1).contiguous()
pcd_fold_t = self.decoder_fold(y)
pcd_fold = pcd_fold_t.transpose(1, 2).contiguous()
if ('grnet' in self.model_lists):
[pcd_grnet_voxel, pcd_grnet_fine, voxels] = self.grnet(
part.transpose(1, 2))
if ('im_grnet' in self.model_lists):
imgs_feat = self.encoder_img(images)
# torch.Size([batch, 16, 256, 2, 2])
imgs_feat = torch.max(imgs_feat, dim=1, keepdim=True).values
imgs_feat = torch.reshape(imgs_feat, (part.shape[0], 1024))
[pcd_grnet_voxel, pcd_grnet_fine, voxels] = self.grnet(
partial_cloud=part.transpose(1, 2), global_feature=imgs_feat)
if ('pointcnn' in self.model_lists):
pcd_pcnn = self.pointcnn(part.transpose(1, 2))
if ('pcn' in self.model_lists):
pcn_coarse, pcn_fine = self.pcn(part)
if ('disp3d' in self.model_lists):
disp_feat, pcd_anchors = self.disp_enc(part.transpose(1, 2))
pcd_disp = self.disp_dec([disp_feat])
if self.do_segment is True:
seg_disp = self.disp_seg(pcd_disp)
if ('im_disp3d' in self.model_lists):
imgs_feat = self.encoder_img(images)
# torch.Size([batch, 16, 256, 4, 4])
imgs_feat = torch.max(imgs_feat, dim=1, keepdim=True).values
imgs_feat = torch.reshape(imgs_feat, (part.shape[0], 1, 1024))
pcd_disp = self.disp_dec([imgs_feat])
if ('vrcnet' in self.model_lists):
# pcd_vrcnet_coarse, pcd_vrcnet_fine, pcd_vrcnet_3, pcd_vrcnet_4 = self.vrcnet(part)['out1'], self.vrcnet(part)['out2'], self.vrcnet(part)['out3'], self.vrcnet(part)['out4']
pcd_vrcnet_coarse, pcd_vrcnet_fine, pcd_vrcnet_3, pcd_vrcnet_4 = self.vrcnet(
part)
if self.do_segment is True:
seg_vrcnet = self.disp_seg(pcd_vrcnet)
if ('pointr' in self.model_lists):
pcd_pointr = self.pointr(part.transpose(1, 2))
if ('im_pointr' in self.model_lists):
imgs_feats = []
num_frames = 16
"""
for i in range(num_frames):
imgs_feat = self.encoder_img(images[:,i,:,:,:].transpose(1, 3))
imgs_feats.append(imgs_feat)
imgs_feat = torch.stack(imgs_feats)
# shape [2, 1024, 4, 4]
imgs_feat = torch.reshape(imgs_feat, (part.shape[0], 256, num_frames * 4 * 4))
"""
# NOTE encoding with ResNet
imgs_feat = self.encoder_img(images)
# torch.Size([batch, 16, 256, 4, 4])
imgs_feat = imgs_feat.transpose(1, 2)
# NOTE reshape for PE feature
imgs_feat = torch.reshape(imgs_feat,
(part.shape[0], 256, num_frames * 2 * 2))
# torch.Size([batch, 15*16, 256])
pcd_pointr = self.pointr(imgs_feat.transpose(1, 2))
if ('snowflake' in self.model_lists):
pcd_snowflake = self.snowflake(point_cloud=part.transpose(1, 2))
if ('im_snowflake' in self.model_lists):
imgs_feat = self.encoder_img(images)
# torch.Size([batch, 16, 256, 2, 2])
imgs_feat = torch.max(imgs_feat, dim=1, keepdim=True).values
imgs_feat = torch.reshape(imgs_feat, (part.shape[0], 1, 1024))
# NOTE: depends whether use point cloud and images as input together
# pcd_snowflake = self.snowflake(global_feature=imgs_feat.transpose(1, 2))
pcd_snowflake = self.snowflake(
point_cloud=part.transpose(1, 2),
global_feature=imgs_feat.transpose(1, 2))
# pcd_snowflake_3d = self.snowflake(point_cloud=part.transpose(1, 2))
# start to organize
output['softpool'] = [
pcd_softpool, pcd_fusion, input_chosen, loss_mst
] if ('softpool' in self.model_lists) else []
output['msn'] = [pcd_msn1, pcd_msn2, loss_mst
] if ('msn' in self.model_lists) else []
output['folding'] = [pcd_fold
] if ('folding' in self.model_lists) else []
output['grnet'] = [pcd_grnet_voxel, pcd_grnet_fine, voxels
] if ('grnet' in self.model_lists) else []
output['im_grnet'] = [pcd_grnet_voxel, pcd_grnet_fine, voxels
] if ('im_grnet' in self.model_lists) else []
output['shapegf'] = self.trainer if (
'shapegf' in self.model_lists) else []
output['pcn'] = [pcn_coarse, pcn_fine
] if ('pcn' in self.model_lists) else []
output['disp3d'] = [pcd_disp
] if ('disp3d' in self.model_lists) else []
output['im_disp3d'] = [pcd_disp] if (
'im_disp3d' in self.model_lists) else []
output['vrcnet'] = [
pcd_vrcnet_coarse, pcd_vrcnet_fine, pcd_vrcnet_3, pcd_vrcnet_4
] if ('vrcnet' in self.model_lists) else []
output['pointr'] = [pcd_pointr[0], pcd_pointr[1], pcd_pointr[2]
] if ('pointr' in self.model_lists) else []
output['im_pointr'] = [pcd_pointr[0], pcd_pointr[1]
] if ('im_pointr' in self.model_lists) else []
output['snowflake'] = pcd_snowflake if (
'snowflake' in self.model_lists) else []
output['im_snowflake'] = [pcd_snowflake] if (
'im_snowflake' in self.model_lists) else []
return output