-
Notifications
You must be signed in to change notification settings - Fork 4
/
pcn.py
159 lines (134 loc) · 6.02 KB
/
pcn.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
import open3d
import torch
import torch.nn as nn
import torch.nn.init as init
import torch.nn.functional as F
import numpy as np
from time import time
# from emd import earth_mover_distance
# from chamfer_distance import ChamferDistance
# chamfer_dist = ChamferDistance()
class PCN(nn.Module):
def __init__(self):
super(PCN, self).__init__()
self.num_coarse = 1024
self.grid_size = 4
self.grid_scale = 0.05
self.num_fine = self.grid_size**2 * self.num_coarse
self.npts = [1]
#alpha = [10000, 20000, 50000],[0.01, 0.1, 0.5, 1.0]
#### ENCODER
## first mlp
mlps1 = [128, 256]
first_mlp_list = []
in_features = 3
for m in range(0, len(mlps1) - 1):
first_mlp_list.append(nn.Conv1d(in_features, mlps1[m], 1))
first_mlp_list.append(nn.ReLU())
in_features = mlps1[m]
first_mlp_list.append(nn.Conv1d(in_features, mlps1[-1], 1))
self.first_mpl = nn.Sequential(*first_mlp_list)
## Second mlp
mlps2 = [512, 1024]
second_mlp_list = []
in_features = 512
for m in range(0, len(mlps2) - 1):
second_mlp_list.append(nn.Conv1d(in_features, mlps2[m], 1))
second_mlp_list.append(nn.ReLU())
in_features = mlps2[m]
second_mlp_list.append(nn.Conv1d(in_features, mlps2[-1], 1))
self.second_mpl = nn.Sequential(*second_mlp_list)
#### DECODER
coarse1 = [1024, 1024, self.num_coarse * 3]
in_features = 1024
decoder_list = []
for m in range(0, len(coarse1) - 1):
decoder_list.append(nn.Linear(in_features, coarse1[m]))
in_features = coarse1[m]
decoder_list.append(nn.Linear(in_features, coarse1[-1]))
self.decoder = nn.Sequential(*decoder_list)
## FOLDING
mlpsfold = [512, 512, 3]
fold_mlp_list = []
in_features = 1029
for m in range(0, len(mlpsfold) - 1):
fold_mlp_list.append(nn.Conv1d(in_features, mlpsfold[m], 1))
fold_mlp_list.append(nn.ReLU())
in_features = mlpsfold[m]
fold_mlp_list.append(nn.Conv1d(in_features, mlpsfold[-1], 1))
self.fold_mpl = nn.Sequential(*fold_mlp_list)
def point_maxpool(self, features, npts, keepdims=True):
# splitted = torch.split(features,npts[0],dim=1)
# outputs = [torch.max(f,dim=2,keepdims=keepdims)[0] for f in splitted]
# return torch.cat(outputs,dim=0)
return torch.max(features, dim=2, keepdims=keepdims)[0]
def point_unpool(self, features, npts):
# features = torch.split(features,features.shape[0],dim=0)
# outputs = [f.repeat([1,npts[i],1]) for i,f in enumerate(features)]
# return torch.cat(outputs,dim=1)
return features.repeat([1, 1, 256])
def forward(self, xyz):
#####ENCODER
features = self.first_mpl(xyz)
features_global = self.point_maxpool(
features.permute(0, 2, 1), self.npts, keepdims=True)
features_global = self.point_unpool(features_global, self.npts)
features = torch.cat(
[features, features_global.permute(0, 2, 1)], dim=1)
features = self.second_mpl(features)
# features = self.point_maxpool(features.permute(0,2,1),self.npts).squeeze(2)
features = self.point_maxpool(features, self.npts).squeeze(2)
##DECODER
coarse = self.decoder(features)
coarse = coarse.view(-1, self.num_coarse, 3)
##FOLDING
grid_row = torch.linspace(-0.05, 0.05, self.grid_size).cuda()
grid_column = torch.linspace(-0.05, 0.05, self.grid_size).cuda()
grid = torch.meshgrid(grid_row, grid_column)
grid = torch.reshape(torch.stack(grid, dim=2), (-1, 2)).unsqueeze(0)
grid_feat = grid.repeat([features.shape[0], self.num_coarse, 1])
# print("grid_Feat",grid_feat.shape)
point_feat = coarse.unsqueeze(2).repeat([1, 1, self.grid_size**2, 1])
point_feat = torch.reshape(point_feat, [-1, self.num_fine, 3])
# print("point_Feat",point_feat.shape)
global_feat = features.unsqueeze(1).repeat([1, self.num_fine, 1])
# print("global_Feat",global_feat.shape)
feat = torch.cat([grid_feat, point_feat, global_feat], dim=2)
center = coarse.unsqueeze(2).repeat([1, 1, self.grid_size**2, 1])
center = torch.reshape(center, [-1, self.num_fine, 3])
fine = self.fold_mpl(feat.permute(0, 2, 1))
# print("fine shape",fine.shape," center shape",center.shape)
fine = fine.permute(0, 2, 1) + center
return coarse, fine
def create_loss(self, coarse, fine, gt, alpha):
gt_ds = gt[:, :coarse.shape[1], :]
loss_coarse = earth_mover_distance(coarse, gt_ds, transpose=False)
dist1, dist2 = chamfer_dist(fine, gt)
loss_fine = (torch.mean(dist1)) + (torch.mean(dist2))
loss = loss_coarse + alpha * loss_fine
return loss
if __name__ == '__main__':
# alpha [ 0.01,0.1,0.5,1.0]
for i in range(10):
xyz = torch.rand(1, 1024, 3).cuda()
pcd1 = open3d.PointCloud()
pcd1.points = open3d.Vector3dVector(xyz.data.cpu().numpy()[0])
pcd1.colors = open3d.Vector3dVector(
np.ones((1024, 3)) * [0.00, 0.53, 0.90])
colors = torch.rand(1, 2048, 3).cuda()
net = PCN()
net.cuda()
coarse, fine = net(xyz)
net.create_loss(coarse, fine, xyz, 1.0)
pcd = open3d.PointCloud()
pcd.points = open3d.Vector3dVector(coarse.data.cpu().numpy()[0] +
np.array([1.0, 0.0, 0.0]))
pcd.colors = open3d.Vector3dVector(
np.ones((1024, 3)) * [0.76, 0.23, 0.14])
pcd2 = open3d.PointCloud()
pcd2.points = open3d.Vector3dVector(fine.data.cpu().numpy()[0] +
np.array([-1.0, 0.0, 0.0]))
pcd2.colors = open3d.Vector3dVector(
np.ones((fine.shape[1], 3)) * [0.16, 0.53, 0.44])
open3d.draw_geometries([pcd, pcd1, pcd2])
exit()