-
Notifications
You must be signed in to change notification settings - Fork 7
/
model_cls.py
177 lines (141 loc) · 6.84 KB
/
model_cls.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
import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from torch import einsum
from model_util import knn_point, index_points
import sys
import os
BASE_DIR = os.path.dirname(os.path.abspath(__file__))
sys.path.append(BASE_DIR)
from pointnet2_ops_lib.pointnet2_ops import pointnet2_utils as pt
class MGR(nn.Module):
def __init__(self, C_in, C_out, dim_k=32, heads=8):
super().__init__()
self.heads = heads
self.k = dim_k
assert (C_out % heads) == 0, 'values dimension must be integer'
dim_v = C_out // heads
self.conv_q = nn.Conv1d(C_in, dim_k * heads, 1, bias=False)
self.conv_k = nn.Conv1d(C_in, dim_k, 1, bias=False)
self.conv_v = nn.Conv1d(C_in, dim_v, 1, bias=False)
self.norm_q = nn.BatchNorm1d(dim_k * heads)
self.norm_v = nn.BatchNorm1d(dim_v)
self.blocker = nn.BatchNorm1d(C_out)
self.skip = nn.Conv1d(C_out, C_out, 1)
# multi-dimensional adjacency matrix
self.A = nn.Parameter(torch.randn(dim_v, dim_v, dim_k), requires_grad=True)
def forward(self, x):
'''
:param x: [B, C_in, N]
:return: out: [B, C_out, N]
'''
query = self.conv_q(x) # [B, head * C_k, N]
key = self.conv_k(x) # [B, C_k, N]
value = self.conv_v(x) # [B, C_v, N]
# normalization
query = self.norm_q(query)
value = self.norm_v(value)
key = key.softmax(dim=-1)
query = rearrange(query, 'b (h k) n -> b h k n', h=self.heads) # [B, head, C_k, N]
k_v_attn = einsum('b k n, b v n -> b k v', key, value) # [B, C_k, C_v]
Yc = einsum('b h k n, b k v -> b n h v', query, k_v_attn) # [B, N, head, C_v]
G = einsum('b v n, w v k -> b n k w', value, self.A).contiguous() # A*x: [B, N, C_k, C_v]
value = rearrange(value, 'b v n -> b n (1) v').contiguous()
G = F.relu(G + value) # [B, N, C_k, C_v]
Yp = einsum('b h k n, b n k v -> b n h v', query, G) # [B, N, head, C_v]
out = Yc + Yp
out = rearrange(out, 'b n h v -> b n (h v)')
out = rearrange(out, 'b n c -> b c n')
out = self.blocker(self.skip(out))
return F.relu(out + x)
class transformer(nn.Module):
def __init__(self, C_in, C_out, n_samples=None, K=20, dim_k=32, heads=8, ch_raise=64):
super().__init__()
self.d = dim_k
assert (C_out % heads) == 0, 'values dimension must be integer'
dim_v = C_out // heads
self.n_samples = n_samples
self.K = K
self.heads = heads
C_in = C_in * 2 + dim_v
self.mlp = nn.Sequential(
nn.Conv2d(C_in, ch_raise, 1, bias=False),
nn.BatchNorm2d(ch_raise),
nn.ReLU(True),
nn.Conv2d(ch_raise, ch_raise, 1, bias=False),
nn.BatchNorm2d(ch_raise),
nn.ReLU(True))
self.mlp_v = nn.Conv1d(C_in, dim_v, 1, bias=False)
self.mlp_k = nn.Conv1d(C_in, dim_k, 1, bias=False)
self.mlp_q = nn.Conv1d(ch_raise, heads * dim_k, 1, bias=False)
self.mlp_h = nn.Conv2d(3, dim_v, 1, bias=False)
self.bn_value = nn.BatchNorm1d(dim_v)
self.bn_query = nn.BatchNorm1d(heads * dim_k)
def forward(self, xyz, feature):
bs = xyz.shape[0]
fps_idx = pt.furthest_point_sample(xyz.contiguous(), self.n_samples).long() # [B, S]
new_xyz = index_points(xyz, fps_idx) # [B, S, 3]
new_feature = index_points(feature, fps_idx).transpose(2, 1).contiguous() # [B, C, S]
knn_idx = knn_point(self.K, xyz, new_xyz) # [B, S, K]
neighbor_xyz = index_points(xyz, knn_idx) # [B, S, K, 3]
grouped_features = index_points(feature, knn_idx) # [B, S, K, C]
grouped_features = grouped_features.permute(0, 3, 1, 2).contiguous() # [B, C, S, K]
grouped_points_norm = grouped_features - new_feature.unsqueeze(-1).contiguous() # [B, C, S, K]
# relative spatial coordinates
relative_pos = neighbor_xyz - new_xyz.unsqueeze(-2).repeat(1, 1, self.K, 1) # [B, S, K, 3]
relative_pos = relative_pos.permute(0, 3, 1, 2).contiguous() # [B, 3, S, K]
pos_encoder = self.mlp_h(relative_pos)
feature = torch.cat([grouped_points_norm,
new_feature.unsqueeze(-1).repeat(1, 1, 1, self.K),
pos_encoder], dim=1) # [B, 2C_in + d, S, K]
feature_q = self.mlp(feature).max(-1)[0] # [B, C, S]
query = F.relu(self.bn_query(self.mlp_q(feature_q))) # [B, head * d, S]
query = rearrange(query, 'b (h d) n -> b h d n', b=bs, h=self.heads, d=self.d) # [B, head, d, S]
feature = feature.permute(0, 2, 1, 3).contiguous() # [B, S, 2C, K]
feature = feature.view(bs * self.n_samples, -1, self.K) # [B*S, 2C, K]
value = self.bn_value(self.mlp_v(feature)) # [B*S, v, K]
value = value.view(bs, self.n_samples, -1, self.K) # [B, S, v, K]
key = self.mlp_k(feature).softmax(dim=-1) # [B*S, d, K]
key = key.view(bs, self.n_samples, -1, self.K) # [B, S, d, K]
k_v_attn = einsum('b n d k, b n v k -> b d v n', key, value) # [bs, d, v, N]
out = einsum('b h d n, b d v n -> b h v n', query, k_v_attn.contiguous()) # [B, S, head, v]
out = rearrange(out.contiguous(), 'b h v n -> b (h v) n') # [B, C_out, S]
return new_xyz, out
class Model(nn.Module):
def __init__(self, args, trans_block, output_channels=2):
super().__init__()
self.use_norm = args.use_norm
# transformer layer
self.tf1 = trans_block(3, 128, n_samples=512, K=args.num_K[0], dim_k=args.dim_k, heads=args.head, ch_raise=64)
self.tf2 = trans_block(128, 256, n_samples=128, K=args.num_K[1], dim_k=args.dim_k, heads=args.head, ch_raise=256)
# multi-graph attention
self.attn = MGR(256, 256, dim_k=args.dim_k, heads=args.head)
self.conv_raise = nn.Sequential(
nn.Conv1d(256, 512, kernel_size=1, bias=False),
nn.BatchNorm1d(512),
nn.ReLU(True),
nn.Conv1d(512, args.emb_dims, kernel_size=1, bias=False),
nn.BatchNorm1d(args.emb_dims),
nn.ReLU(True))
self.cls = nn.Sequential(
nn.Linear(args.emb_dims, 512, bias=False),
nn.BatchNorm1d(512),
nn.ReLU(True),
nn.Dropout(p=args.dropout),
nn.Linear(512, output_channels))
def forward(self, x):
# input x: [B, N, 3+3]
xyz = x[..., :3]
if not self.use_norm:
feature = xyz
else:
assert x.size()[-1] == 6
feature = x[..., 3:]
xyz1, feature1 = self.tf1(xyz, feature)
feature1 = feature1.transpose(2, 1)
_, feature2 = self.tf2(xyz1, feature1)
feature3 = self.attn(feature2)
out = self.conv_raise(feature3)
out = self.cls(out.max(-1)[0])
return out