File size: 8,799 Bytes
475d332
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from .transformer_utils import BaseTemperalPointModel
from copy import deepcopy
import torch
import einops
import torch.nn as nn
import torch.nn.functional as F
from torch import nn
from einops import rearrange
import pointops
from pointcept.models.utils import offset2batch, batch2offset
class PointBatchNorm(nn.Module):
    """

    Batch Normalization for Point Clouds data in shape of [B*N, C], [B*N, L, C]

    """

    def __init__(self, embed_channels):
        super().__init__()
        self.norm = nn.BatchNorm1d(embed_channels)

    def forward(self, input: torch.Tensor) -> torch.Tensor:
        if input.dim() == 3:
            return (
                self.norm(input.transpose(1, 2).contiguous())
                .transpose(1, 2)
                .contiguous()
            )
        elif input.dim() == 2:
            return self.norm(input)
        else:
            raise NotImplementedError
#https://github.com/Pointcept/Pointcept/blob/main/pointcept/models/point_transformer_v2/point_transformer_v2m2_base.py
class GroupedVectorAttention(nn.Module):
    def __init__(

        self,

        embed_channels,

        groups,

        attn_drop_rate=0.0,

        qkv_bias=True,

        pe_multiplier=False,

        pe_bias=True,

    ):
        super(GroupedVectorAttention, self).__init__()
        self.embed_channels = embed_channels
        self.groups = groups
        assert embed_channels % groups == 0
        self.attn_drop_rate = attn_drop_rate
        self.qkv_bias = qkv_bias
        self.pe_multiplier = pe_multiplier
        self.pe_bias = pe_bias

        self.linear_q = nn.Sequential(
            nn.Linear(embed_channels, embed_channels, bias=qkv_bias),
            PointBatchNorm(embed_channels),
            nn.ReLU(inplace=True),
        )
        self.linear_k = nn.Sequential(
            nn.Linear(embed_channels, embed_channels, bias=qkv_bias),
            PointBatchNorm(embed_channels),
            nn.ReLU(inplace=True),
        )

        self.linear_v = nn.Linear(embed_channels, embed_channels, bias=qkv_bias)

        if self.pe_multiplier:
            self.linear_p_multiplier = nn.Sequential(
                nn.Linear(3, embed_channels),
                PointBatchNorm(embed_channels),
                nn.ReLU(inplace=True),
                nn.Linear(embed_channels, embed_channels),
            )
        if self.pe_bias:
            self.linear_p_bias = nn.Sequential(
                nn.Linear(3, embed_channels),
                PointBatchNorm(embed_channels),
                nn.ReLU(inplace=True),
                nn.Linear(embed_channels, embed_channels),
            )
        self.weight_encoding = nn.Sequential(
            nn.Linear(embed_channels, groups),
            PointBatchNorm(groups),
            nn.ReLU(inplace=True),
            nn.Linear(groups, groups),
        )
        self.softmax = nn.Softmax(dim=1)
        self.attn_drop = nn.Dropout(attn_drop_rate)

    def forward(self, feat, coord, reference_index):
        query, key, value = (
            self.linear_q(feat),
            self.linear_k(feat),
            self.linear_v(feat),
        )
        key = pointops.grouping(reference_index, key, coord, with_xyz=True)
        value = pointops.grouping(reference_index, value, coord, with_xyz=False)
        pos, key = key[:, :, 0:3], key[:, :, 3:]
        relation_qk = key - query.unsqueeze(1)
        if self.pe_multiplier:
            pem = self.linear_p_multiplier(pos)
            relation_qk = relation_qk * pem
        if self.pe_bias:
            peb = self.linear_p_bias(pos)
            relation_qk = relation_qk + peb
            value = value + peb

        weight = self.weight_encoding(relation_qk)
        weight = self.attn_drop(self.softmax(weight))

        mask = torch.sign(reference_index + 1)
        weight = torch.einsum("n s g, n s -> n s g", weight, mask)
        value = einops.rearrange(value, "n ns (g i) -> n ns g i", g=self.groups)
        feat = torch.einsum("n s g i, n s g -> n g i", value, weight)
        feat = einops.rearrange(feat, "n g i -> n (g i)")
        return feat

class BlockSequence(nn.Module):
    def __init__(

        self,

        depth,

        embed_channels,

        groups,

        neighbours=16,

        qkv_bias=True,

        pe_multiplier=False,

        pe_bias=True,

        attn_drop_rate=0.0,

        drop_path_rate=0.0,

        enable_checkpoint=False,

    ):
        super(BlockSequence, self).__init__()

        if isinstance(drop_path_rate, list):
            drop_path_rates = drop_path_rate
            assert len(drop_path_rates) == depth
        elif isinstance(drop_path_rate, float):
            drop_path_rates = [deepcopy(drop_path_rate) for _ in range(depth)]
        else:
            drop_path_rates = [0.0 for _ in range(depth)]

        self.neighbours = neighbours
        self.blocks = nn.ModuleList()
        for i in range(depth):
            block = Block(
                embed_channels=embed_channels,
                groups=groups,
                qkv_bias=qkv_bias,
                pe_multiplier=pe_multiplier,
                pe_bias=pe_bias,
                attn_drop_rate=attn_drop_rate,
                drop_path_rate=drop_path_rates[i],
                enable_checkpoint=enable_checkpoint,
            )
            self.blocks.append(block)

    def forward(self, points):
        coord, feat, offset = points
        # reference index query of neighbourhood attention
        # for windows attention, modify reference index query method
        reference_index, _ = pointops.knn_query(self.neighbours, coord, offset)
        for block in self.blocks:
            points = block(points, reference_index)
        return points

class GVAPatchEmbed(nn.Module):
    def __init__(

        self,

        depth,

        in_channels,

        embed_channels,

        groups,

        neighbours=16,

        qkv_bias=True,

        pe_multiplier=False,

        pe_bias=True,

        attn_drop_rate=0.0,

        drop_path_rate=0.0,

        enable_checkpoint=False,

    ):
        super(GVAPatchEmbed, self).__init__()
        self.in_channels = in_channels
        self.embed_channels = embed_channels
        self.proj = nn.Sequential(
            nn.Linear(in_channels, embed_channels, bias=False),
            PointBatchNorm(embed_channels),
            nn.ReLU(inplace=True),
        )
        self.blocks = BlockSequence(
            depth=depth,
            embed_channels=embed_channels,
            groups=groups,
            neighbours=neighbours,
            qkv_bias=qkv_bias,
            pe_multiplier=pe_multiplier,
            pe_bias=pe_bias,
            attn_drop_rate=attn_drop_rate,
            drop_path_rate=drop_path_rate,
            enable_checkpoint=enable_checkpoint,
        )

    def forward(self, points):
        coord, feat, offset = points
        feat = self.proj(feat)
        return self.blocks([coord, feat, offset])


class Block(nn.Module):
    def __init__(

        self,

        embed_channels,

        groups,

        qkv_bias=True,

        pe_multiplier=False,

        pe_bias=True,

        attn_drop_rate=0.0,

        drop_path_rate=0.0,

        enable_checkpoint=False,

    ):
        super(Block, self).__init__()
        self.attn = GroupedVectorAttention(
            embed_channels=embed_channels,
            groups=groups,
            qkv_bias=qkv_bias,
            attn_drop_rate=attn_drop_rate,
            pe_multiplier=pe_multiplier,
            pe_bias=pe_bias,
        )
        self.fc1 = nn.Linear(embed_channels, embed_channels, bias=False)
        self.fc3 = nn.Linear(embed_channels, embed_channels, bias=False)
        self.norm1 = PointBatchNorm(embed_channels)
        self.norm2 = PointBatchNorm(embed_channels)
        self.norm3 = PointBatchNorm(embed_channels)
        self.act = nn.ReLU(inplace=True)
        self.enable_checkpoint = enable_checkpoint
        self.drop_path = (
            DropPath(drop_path_rate) if drop_path_rate > 0.0 else nn.Identity()
        )

    def forward(self, points, reference_index):
        coord, feat, offset = points
        identity = feat
        feat = self.act(self.norm1(self.fc1(feat)))
        feat = (
            self.attn(feat, coord, reference_index)
            if not self.enable_checkpoint
            else checkpoint(self.attn, feat, coord, reference_index)
        )
        feat = self.act(self.norm2(feat))
        feat = self.norm3(self.fc3(feat))
        feat = identity + self.drop_path(feat)
        feat = self.act(feat)
        return [coord, feat, offset]