/
layer_edge.py
226 lines (198 loc) · 10.3 KB
/
layer_edge.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
import math
import torch
import torch.nn as nn
import util_edge as util
import numpy as np
import ghalton
class GridEncoder(nn.Module):
def __init__(self,args, prep, grid_size,ops=None):
super(self.__class__, self).__init__()
self.grid_size = grid_size
self.preprocessing = prep
self.args=args
self.ops=ops
def initialize_grid_ball(self, x,grid_size):
if x.dim() == 2:
x = x.unsqueeze(0)
assert (x.min() >= -self.args.normalize_ratio)
assert (x.max() <= self.args.normalize_ratio)
# bring vector into range -self.args.normalize_ratio - grid_size-self.args.normalize_ratio
if self.args.normalize_ratio==0.5:
reshaped = (x + self.args.normalize_ratio) * grid_size - self.args.normalize_ratio
else:
reshaped = (x + self.args.normalize_ratio)/2 * grid_size - self.args.normalize_ratio/2
ind1 = reshaped.floor().clamp(0.0, grid_size - 1)
ind2 = reshaped.ceil().clamp(0.0, grid_size - 1)
ind = [torch.cat([ind1[:, 0, :].unsqueeze(1), ind1[:, 1, :].unsqueeze(1), ind1[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind1[:, 0, :].unsqueeze(1), ind1[:, 1, :].unsqueeze(1), ind2[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind1[:, 0, :].unsqueeze(1), ind2[:, 1, :].unsqueeze(1), ind1[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind1[:, 0, :].unsqueeze(1), ind2[:, 1, :].unsqueeze(1), ind2[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind2[:, 0, :].unsqueeze(1), ind1[:, 1, :].unsqueeze(1), ind1[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind2[:, 0, :].unsqueeze(1), ind1[:, 1, :].unsqueeze(1), ind2[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind2[:, 0, :].unsqueeze(1), ind2[:, 1, :].unsqueeze(1), ind1[:, 2, :].unsqueeze(1)], dim=1),
torch.cat([ind2[:, 0, :].unsqueeze(1), ind2[:, 1, :].unsqueeze(1), ind2[:, 2, :].unsqueeze(1)], dim=1)]
ind = torch.stack(ind, dim=-1).cuda()
res = reshaped.unsqueeze(-1).repeat([1, 1, 1, 8]) - ind # B*3*2500*8
ind = ind[:, 0, :, :] * grid_size * grid_size + ind[:, 1, :, :] * grid_size + ind[:, 2, :, :]
ind = ind.long() # B*2500*8
# binary weight to check wether point is in gridball
dist = res.norm(dim=1).detach()
weight = (dist < 0.87).float().detach() # half the diagonal of a grid cube
return res, weight, ind
def propose_grid(self,res,indices,weight,grid_size,per_point_features=False):
b, _, n,_ = res.size()
c = res.shape[1]
weight = weight.unsqueeze(1).expand_as(res) # B*32*2500*8
res = res * weight # zero out weights of points outside of ball
# sum up features of points inside ball
x = torch.zeros(b, c, grid_size * grid_size * grid_size).to(res.device)
count = torch.zeros(b, c, grid_size * grid_size * grid_size).to(x)
indices = indices.view(b, -1).contiguous()
indices.clamp_(0, grid_size ** 3)
res = res.view(b, c, 8 * n).contiguous()
weight = weight.view(b, c, 8 * n).contiguous()
for i in range(b):
x[i].index_add_(1, indices[i], res[i]) # dim, index, tensor
count[i].index_add_(1, indices[i], weight[i])
# number of points should have no effect
count = torch.max(count, torch.tensor([1.0]).to(weight.device))
x /= count
x = x.view(b, -1, grid_size, grid_size, grid_size).contiguous() # b x c x grid_size x grid_size x grid_size
return x
def forward(self, x):
b, _, n = x.size()
x_all=[]
for i in range(int(np.log2(self.args.grid_size))):
p1_idx = self.ops.pointnet2.pointnet2_utils.furthest_point_sample(x.transpose(1, 2).contiguous(),
n // 2 ** i)
x = self.ops.pointnet2.pointnet2_utils.gather_operation(x, p1_idx)
grid_size = self.grid_size // 2 ** i
res, weight, indices = self.initialize_grid_ball(x, grid_size) # b x 3 x n x 8
res = self.preprocessing[i](res) # b x k x n x 8
x1 = self.propose_grid(res, indices, weight, grid_size)
x_all.append(x1)
return x_all
class EdgePointCloudGenerator(nn.Module):
def __init__(self, generator, rnd_dim=2, res=16,ops=None,normalize_ratio=0.5,args=None):
super(self.__class__, self).__init__()
self.base_dim = rnd_dim
self.generator = generator
self.ops=ops
grid = util.meshgrid(res)
if normalize_ratio == 0.5:
self.o = (((grid + normalize_ratio) / res) - normalize_ratio).view(3, -1).contiguous()
elif normalize_ratio == 1:
self.o = (((grid + normalize_ratio / 2) * 2 / res) - normalize_ratio).view(3, -1).contiguous()
self.s = res
self.args=args
def forward_fixed_pattern(self, x, dens, n,ratio,global_feature=None,local_feature=None):
b, c, g, _, _ = x.shape
grid_o = self.o.to(x.device) # 3d meshgrid : 3* ()
N = util.densSample(dens, n)
N = N.view(b, -1).contiguous()
x = x.view(b, c, -1).contiguous()
b_rnd = torch.tensor(ghalton.GeneralizedHalton(2).get(n), dtype=torch.float32).t()
b_rnd = b_rnd.to(x) * ratio - ratio / 2
b_rnd = b_rnd.unsqueeze(0).repeat(b, 1, 1)
N = N.view(-1).contiguous()
x = x.permute(1, 0, 2).contiguous().view(c, -1).contiguous()
ind = (N > 0).nonzero().squeeze(-1)
x_ind = torch.repeat_interleave(x[:, ind], N[ind].long(), dim=-1)
x_ind = x_ind.view(c, b, n).contiguous().permute(1, 0, 2).contiguous()
b_inp = torch.cat([x_ind, b_rnd], dim=1)
grid_o = grid_o.unsqueeze(0).repeat(b, 1, 1).permute(1, 0, 2).contiguous().view(3, -1).contiguous()
o_ind = torch.repeat_interleave(grid_o[:, ind], N[ind].long(), dim=-1)
o_ind = o_ind.view(3, b, n).contiguous().permute(1, 0, 2).contiguous()
out = self.generator(b_inp)
norm = out.norm(dim=1)
reg = (norm - (math.sqrt(3) / (self.s))).clamp(0) # twice the size needed to cover a gridcell
return out + o_ind, reg
class PointCloudGenerator(nn.Module):
def __init__(self, generator, rnd_dim=2, res=16,ops=None,normalize_ratio=0.5,args=None):
super(self.__class__, self).__init__()
self.base_dim = rnd_dim
self.generator = generator
self.ops=ops
grid = util.meshgrid(res)
if normalize_ratio == 0.5:
self.o = (((grid + normalize_ratio) / res) - normalize_ratio).view(3, -1).contiguous()
elif normalize_ratio == 1:
self.o = (((grid + normalize_ratio/2)*2 / res) - normalize_ratio).view(3, -1).contiguous()
self.s = res
self.args=args
def forward_fixed_pattern(self, x, dens, n,ratio,global_feature=None,local_feature=None):
b, c, g, _, _ = x.shape
grid_o = self.o.to(x.device) # 3d meshgrid : 3* ()
N = util.densSample(dens, n)
N = N.view(b, -1).contiguous()
x = x.view(b, c, -1).contiguous()
b_rnd = torch.tensor(ghalton.GeneralizedHalton(2).get(n), dtype=torch.float32).t()
b_rnd = b_rnd.to(x) * ratio - ratio / 2
b_rnd = b_rnd.unsqueeze(0).repeat(b, 1, 1)
N = N.view(-1).contiguous()
x = x.permute(1, 0, 2).contiguous().view(c, -1).contiguous()
ind = (N > 0).nonzero().squeeze(-1)
x_ind = torch.repeat_interleave(x[:, ind], N[ind].long(), dim=-1)
x_ind = x_ind.view(c, b, n).contiguous().permute(1, 0, 2).contiguous()
b_inp = torch.cat([x_ind, b_rnd], dim=1)
grid_o = grid_o.unsqueeze(0).repeat(b, 1, 1).permute(1, 0, 2).contiguous().view(3, -1).contiguous()
o_ind = torch.repeat_interleave(grid_o[:, ind], N[ind].long(), dim=-1)
o_ind = o_ind.view(3, b, n).contiguous().permute(1, 0, 2).contiguous()
out = self.generator(b_inp)
norm = out.norm(dim=1)
reg = (norm - (math.sqrt(3) / (self.s))).clamp(0) # twice the size needed to cover a gridcell
return out + o_ind, reg
class AdaptiveDecoder(nn.Module):
def __init__(self, decoder, n_classes=None,args=None, max_layer=None):
super(self.__class__, self).__init__()
self.args=args
self.decoder = decoder
self.norm_indices = []
self.conditional = n_classes is not None
if self.args.grid_size==32:
self.norm_indices = [0, 3, 8, 12, 17, 21, 26, 30, 35, 39]
self.slices=[128 * 2,128 * 2,
256 * 2,128 * 2,
192 * 2,64 * 2,
129 * 2,32 * 2,
65 * 2,62 * 2]
elif self.args.grid_size==64:
self.norm_indices=[0,3,8,12,17,21,26,30,35,39,44,48]
self.slices=[128 * 2,128 * 2,
256 * 2,128 * 2,
192 * 2,128 * 2,
192 * 2,64 * 2,
97 * 2,32 * 2,
65 * 2,62 * 2]
self.max_layer = len(self.norm_indices)
def forward(self, w,x, contour_16=None,contour_32=None):
size = 0
if self.args.grid_size == 32:
jj=[0,8,17,26,35]
elif self.args.grid_size == 64:
jj = [0, 8, 17, 26, 35,44]
for i, l in enumerate(self.decoder):
if i==0:
x1 = x[int(np.log2(self.args.grid_size))-1-i]
else:
x1=torch.cat((x1,x[int(np.log2(self.args.grid_size))-1-i]),dim=1)
if self.args.grid_size == 32:
if i == 3:
x1 = torch.cat((x1, contour_16), dim=1)
if i == 4:
x1 = torch.cat((x1, contour_32), dim=1)
elif self.args.grid_size == 64:
if i == 4:
x1 = torch.cat((x1, contour_16), dim=1)
if i == 5:
x1 = torch.cat((x1, contour_32), dim=1)
jjj = 0
for j,ll in enumerate(l):
x1=ll(x1)
if j+jj[i] in self.norm_indices:
s = w[:, size:size + self.slices[i*2+jjj], None, None, None]
size += self.slices[i*2+jjj]
x1 = x1 * s[:, :self.slices[i*2+jjj] // 2]
x1 = x1 + s[:, self.slices[i*2+jjj] // 2:]
jjj += 1
return x1