forked from open-mmlab/mmdetection3d
-
Notifications
You must be signed in to change notification settings - Fork 76
/
transfusion_head.py
1375 lines (1206 loc) · 66.2 KB
/
transfusion_head.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
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
import copy
import numpy as np
import torch
from mmcv.cnn import ConvModule, build_conv_layer, kaiming_init
from mmcv.runner import force_fp32
from torch import nn
import torch.nn.functional as F
from torch.nn.parameter import Parameter
from torch.nn import Linear
from torch.nn.init import xavier_uniform_, constant_
from mmdet3d.core import (circle_nms, draw_heatmap_gaussian, gaussian_radius,
xywhr2xyxyr, limit_period, PseudoSampler)
from mmdet3d.core.bbox.structures import rotation_3d_in_axis
from mmdet3d.core import Box3DMode, LiDARInstance3DBoxes
from mmdet3d.models import builder
from mmdet3d.models.builder import HEADS, build_loss
from mmdet3d.models.utils import clip_sigmoid
from mmdet3d.models.fusion_layers import apply_3d_transformation
from mmdet3d.ops.iou3d.iou3d_utils import nms_gpu
from mmdet.core import build_bbox_coder, multi_apply, build_assigner, build_sampler, AssignResult
from mmdet3d.ops.roiaware_pool3d import points_in_boxes_batch
class PositionEmbeddingLearned(nn.Module):
"""
Absolute pos embedding, learned.
"""
def __init__(self, input_channel, num_pos_feats=288):
super().__init__()
self.position_embedding_head = nn.Sequential(
nn.Conv1d(input_channel, num_pos_feats, kernel_size=1),
nn.BatchNorm1d(num_pos_feats),
nn.ReLU(inplace=True),
nn.Conv1d(num_pos_feats, num_pos_feats, kernel_size=1))
def forward(self, xyz):
xyz = xyz.transpose(1, 2).contiguous()
position_embedding = self.position_embedding_head(xyz)
return position_embedding
class TransformerDecoderLayer(nn.Module):
def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1, activation="relu",
self_posembed=None, cross_posembed=None, cross_only=False):
super().__init__()
self.cross_only = cross_only
if not self.cross_only:
self.self_attn = MultiheadAttention(d_model, nhead, dropout=dropout)
self.multihead_attn = MultiheadAttention(d_model, nhead, dropout=dropout)
# Implementation of Feedforward model
self.linear1 = nn.Linear(d_model, dim_feedforward)
self.dropout = nn.Dropout(dropout)
self.linear2 = nn.Linear(dim_feedforward, d_model)
self.norm1 = nn.LayerNorm(d_model)
self.norm2 = nn.LayerNorm(d_model)
self.norm3 = nn.LayerNorm(d_model)
self.dropout1 = nn.Dropout(dropout)
self.dropout2 = nn.Dropout(dropout)
self.dropout3 = nn.Dropout(dropout)
def _get_activation_fn(activation):
"""Return an activation function given a string"""
if activation == "relu":
return F.relu
if activation == "gelu":
return F.gelu
if activation == "glu":
return F.glu
raise RuntimeError(F"activation should be relu/gelu, not {activation}.")
self.activation = _get_activation_fn(activation)
self.self_posembed = self_posembed
self.cross_posembed = cross_posembed
def with_pos_embed(self, tensor, pos_embed):
return tensor if pos_embed is None else tensor + pos_embed
def forward(self, query, key, query_pos, key_pos, attn_mask=None):
"""
:param query: B C Pq
:param key: B C Pk
:param query_pos: B Pq 3/6
:param key_pos: B Pk 3/6
:param value_pos: [B Pq 3/6]
:return:
"""
# NxCxP to PxNxC
if self.self_posembed is not None:
query_pos_embed = self.self_posembed(query_pos).permute(2, 0, 1)
else:
query_pos_embed = None
if self.cross_posembed is not None:
key_pos_embed = self.cross_posembed(key_pos).permute(2, 0, 1)
else:
key_pos_embed = None
query = query.permute(2, 0, 1)
key = key.permute(2, 0, 1)
if not self.cross_only:
q = k = v = self.with_pos_embed(query, query_pos_embed)
query2 = self.self_attn(q, k, value=v)[0]
query = query + self.dropout1(query2)
query = self.norm1(query)
query2 = self.multihead_attn(query=self.with_pos_embed(query, query_pos_embed),
key=self.with_pos_embed(key, key_pos_embed),
value=self.with_pos_embed(key, key_pos_embed), attn_mask=attn_mask)[0]
query = query + self.dropout2(query2)
query = self.norm2(query)
query2 = self.linear2(self.dropout(self.activation(self.linear1(query))))
query = query + self.dropout3(query2)
query = self.norm3(query)
# NxCxP to PxNxC
query = query.permute(1, 2, 0)
return query
class MultiheadAttention(nn.Module):
r"""Allows the model to jointly attend to information
from different representation subspaces.
See reference: Attention Is All You Need
.. math::
\text{MultiHead}(Q, K, V) = \text{Concat}(head_1,\dots,head_h)W^O
\text{where} head_i = \text{Attention}(QW_i^Q, KW_i^K, VW_i^V)
Args:
embed_dim: total dimension of the model.
num_heads: parallel attention heads.
dropout: a Dropout layer on attn_output_weights. Default: 0.0.
bias: add bias as module parameter. Default: True.
add_bias_kv: add bias to the key and value sequences at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
kdim: total number of features in key. Default: None.
vdim: total number of features in key. Default: None.
Note: if kdim and vdim are None, they will be set to embed_dim such that
query, key, and value have the same number of features.
Examples::
>>> multihead_attn = nn.MultiheadAttention(embed_dim, num_heads)
>>> attn_output, attn_output_weights = multihead_attn(query, key, value)
"""
def __init__(self, embed_dim, num_heads, dropout=0., bias=True, add_bias_kv=False, add_zero_attn=False, kdim=None,
vdim=None):
super(MultiheadAttention, self).__init__()
self.embed_dim = embed_dim
self.kdim = kdim if kdim is not None else embed_dim
self.vdim = vdim if vdim is not None else embed_dim
self._qkv_same_embed_dim = self.kdim == embed_dim and self.vdim == embed_dim
self.num_heads = num_heads
self.dropout = dropout
self.head_dim = embed_dim // num_heads
assert self.head_dim * num_heads == self.embed_dim, "embed_dim must be divisible by num_heads"
self.in_proj_weight = Parameter(torch.empty(3 * embed_dim, embed_dim))
if self._qkv_same_embed_dim is False:
self.q_proj_weight = Parameter(torch.Tensor(embed_dim, embed_dim))
self.k_proj_weight = Parameter(torch.Tensor(embed_dim, self.kdim))
self.v_proj_weight = Parameter(torch.Tensor(embed_dim, self.vdim))
if bias:
self.in_proj_bias = Parameter(torch.empty(3 * embed_dim))
else:
self.register_parameter('in_proj_bias', None)
self.out_proj = Linear(embed_dim, embed_dim, bias=bias)
if add_bias_kv:
self.bias_k = Parameter(torch.empty(1, 1, embed_dim))
self.bias_v = Parameter(torch.empty(1, 1, embed_dim))
else:
self.bias_k = self.bias_v = None
self.add_zero_attn = add_zero_attn
self._reset_parameters()
def _reset_parameters(self):
if self._qkv_same_embed_dim:
xavier_uniform_(self.in_proj_weight)
else:
xavier_uniform_(self.q_proj_weight)
xavier_uniform_(self.k_proj_weight)
xavier_uniform_(self.v_proj_weight)
if self.in_proj_bias is not None:
constant_(self.in_proj_bias, 0.)
constant_(self.out_proj.bias, 0.)
if self.bias_k is not None:
xavier_normal_(self.bias_k)
if self.bias_v is not None:
xavier_normal_(self.bias_v)
def forward(self, query, key, value, key_padding_mask=None, need_weights=True, attn_mask=None):
r"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. This is an binary mask. When the value is True,
the corresponding value on the attention layer will be filled with -inf.
need_weights: output attn_output_weights.
attn_mask: mask that prevents attention to certain positions. This is an additive mask
(i.e. the values will be added to the attention layer).
Shape:
- Inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- key_padding_mask: :math:`(N, S)`, ByteTensor, where N is the batch size, S is the source sequence length.
- attn_mask: :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
- Outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
if hasattr(self, '_qkv_same_embed_dim') and self._qkv_same_embed_dim is False:
return multi_head_attention_forward(
query, key, value, self.embed_dim, self.num_heads,
self.in_proj_weight, self.in_proj_bias,
self.bias_k, self.bias_v, self.add_zero_attn,
self.dropout, self.out_proj.weight, self.out_proj.bias,
training=self.training,
key_padding_mask=key_padding_mask, need_weights=need_weights,
attn_mask=attn_mask, use_separate_proj_weight=True,
q_proj_weight=self.q_proj_weight, k_proj_weight=self.k_proj_weight,
v_proj_weight=self.v_proj_weight)
else:
if not hasattr(self, '_qkv_same_embed_dim'):
warnings.warn('A new version of MultiheadAttention module has been implemented. \
Please re-train your model with the new module',
UserWarning)
return multi_head_attention_forward(
query, key, value, self.embed_dim, self.num_heads,
self.in_proj_weight, self.in_proj_bias,
self.bias_k, self.bias_v, self.add_zero_attn,
self.dropout, self.out_proj.weight, self.out_proj.bias,
training=self.training,
key_padding_mask=key_padding_mask, need_weights=need_weights,
attn_mask=attn_mask)
def multi_head_attention_forward(query, # type: Tensor
key, # type: Tensor
value, # type: Tensor
embed_dim_to_check, # type: int
num_heads, # type: int
in_proj_weight, # type: Tensor
in_proj_bias, # type: Tensor
bias_k, # type: Optional[Tensor]
bias_v, # type: Optional[Tensor]
add_zero_attn, # type: bool
dropout_p, # type: float
out_proj_weight, # type: Tensor
out_proj_bias, # type: Tensor
training=True, # type: bool
key_padding_mask=None, # type: Optional[Tensor]
need_weights=True, # type: bool
attn_mask=None, # type: Optional[Tensor]
use_separate_proj_weight=False, # type: bool
q_proj_weight=None, # type: Optional[Tensor]
k_proj_weight=None, # type: Optional[Tensor]
v_proj_weight=None, # type: Optional[Tensor]
static_k=None, # type: Optional[Tensor]
static_v=None, # type: Optional[Tensor]
):
# type: (...) -> Tuple[Tensor, Optional[Tensor]]
r"""
Args:
query, key, value: map a query and a set of key-value pairs to an output.
See "Attention Is All You Need" for more details.
embed_dim_to_check: total dimension of the model.
num_heads: parallel attention heads.
in_proj_weight, in_proj_bias: input projection weight and bias.
bias_k, bias_v: bias of the key and value sequences to be added at dim=0.
add_zero_attn: add a new batch of zeros to the key and
value sequences at dim=1.
dropout_p: probability of an element to be zeroed.
out_proj_weight, out_proj_bias: the output projection weight and bias.
training: apply dropout if is ``True``.
key_padding_mask: if provided, specified padding elements in the key will
be ignored by the attention. This is an binary mask. When the value is True,
the corresponding value on the attention layer will be filled with -inf.
need_weights: output attn_output_weights.
attn_mask: mask that prevents attention to certain positions. This is an additive mask
(i.e. the values will be added to the attention layer).
use_separate_proj_weight: the function accept the proj. weights for query, key,
and value in differnt forms. If false, in_proj_weight will be used, which is
a combination of q_proj_weight, k_proj_weight, v_proj_weight.
q_proj_weight, k_proj_weight, v_proj_weight, in_proj_bias: input projection weight and bias.
static_k, static_v: static key and value used for attention operators.
Shape:
Inputs:
- query: :math:`(L, N, E)` where L is the target sequence length, N is the batch size, E is
the embedding dimension.
- key: :math:`(S, N, E)`, where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- value: :math:`(S, N, E)` where S is the source sequence length, N is the batch size, E is
the embedding dimension.
- key_padding_mask: :math:`(N, S)`, ByteTensor, where N is the batch size, S is the source sequence length.
- attn_mask: :math:`(L, S)` where L is the target sequence length, S is the source sequence length.
- static_k: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
- static_v: :math:`(N*num_heads, S, E/num_heads)`, where S is the source sequence length,
N is the batch size, E is the embedding dimension. E/num_heads is the head dimension.
Outputs:
- attn_output: :math:`(L, N, E)` where L is the target sequence length, N is the batch size,
E is the embedding dimension.
- attn_output_weights: :math:`(N, L, S)` where N is the batch size,
L is the target sequence length, S is the source sequence length.
"""
qkv_same = torch.equal(query, key) and torch.equal(key, value)
kv_same = torch.equal(key, value)
tgt_len, bsz, embed_dim = query.size()
assert embed_dim == embed_dim_to_check
assert list(query.size()) == [tgt_len, bsz, embed_dim]
assert key.size() == value.size()
head_dim = embed_dim // num_heads
assert head_dim * num_heads == embed_dim, "embed_dim must be divisible by num_heads"
scaling = float(head_dim) ** -0.5
if use_separate_proj_weight is not True:
if qkv_same:
# self-attention
q, k, v = F.linear(query, in_proj_weight, in_proj_bias).chunk(3, dim=-1)
elif kv_same:
# encoder-decoder attention
# This is inline in_proj function with in_proj_weight and in_proj_bias
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = F.linear(query, _w, _b)
if key is None:
assert value is None
k = None
v = None
else:
# This is inline in_proj function with in_proj_weight and in_proj_bias
_b = in_proj_bias
_start = embed_dim
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
k, v = F.linear(key, _w, _b).chunk(2, dim=-1)
else:
# This is inline in_proj function with in_proj_weight and in_proj_bias
_b = in_proj_bias
_start = 0
_end = embed_dim
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
q = F.linear(query, _w, _b)
# This is inline in_proj function with in_proj_weight and in_proj_bias
_b = in_proj_bias
_start = embed_dim
_end = embed_dim * 2
_w = in_proj_weight[_start:_end, :]
if _b is not None:
_b = _b[_start:_end]
k = F.linear(key, _w, _b)
# This is inline in_proj function with in_proj_weight and in_proj_bias
_b = in_proj_bias
_start = embed_dim * 2
_end = None
_w = in_proj_weight[_start:, :]
if _b is not None:
_b = _b[_start:]
v = F.linear(value, _w, _b)
else:
q_proj_weight_non_opt = torch.jit._unwrap_optional(q_proj_weight)
len1, len2 = q_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == query.size(-1)
k_proj_weight_non_opt = torch.jit._unwrap_optional(k_proj_weight)
len1, len2 = k_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == key.size(-1)
v_proj_weight_non_opt = torch.jit._unwrap_optional(v_proj_weight)
len1, len2 = v_proj_weight_non_opt.size()
assert len1 == embed_dim and len2 == value.size(-1)
if in_proj_bias is not None:
q = F.linear(query, q_proj_weight_non_opt, in_proj_bias[0:embed_dim])
k = F.linear(key, k_proj_weight_non_opt, in_proj_bias[embed_dim:(embed_dim * 2)])
v = F.linear(value, v_proj_weight_non_opt, in_proj_bias[(embed_dim * 2):])
else:
q = F.linear(query, q_proj_weight_non_opt, in_proj_bias)
k = F.linear(key, k_proj_weight_non_opt, in_proj_bias)
v = F.linear(value, v_proj_weight_non_opt, in_proj_bias)
q = q * scaling
if bias_k is not None and bias_v is not None:
if static_k is None and static_v is None:
k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
if attn_mask is not None:
attn_mask = torch.cat([attn_mask,
torch.zeros((attn_mask.size(0), 1),
dtype=attn_mask.dtype,
device=attn_mask.device)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat(
[key_padding_mask, torch.zeros((key_padding_mask.size(0), 1),
dtype=key_padding_mask.dtype,
device=key_padding_mask.device)], dim=1)
else:
assert static_k is None, "bias cannot be added to static key."
assert static_v is None, "bias cannot be added to static value."
else:
assert bias_k is None
assert bias_v is None
q = q.contiguous().view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
if k is not None:
k = k.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if v is not None:
v = v.contiguous().view(-1, bsz * num_heads, head_dim).transpose(0, 1)
if static_k is not None:
assert static_k.size(0) == bsz * num_heads
assert static_k.size(2) == head_dim
k = static_k
if static_v is not None:
assert static_v.size(0) == bsz * num_heads
assert static_v.size(2) == head_dim
v = static_v
src_len = k.size(1)
if key_padding_mask is not None:
assert key_padding_mask.size(0) == bsz
assert key_padding_mask.size(1) == src_len
if add_zero_attn:
src_len += 1
k = torch.cat([k, torch.zeros((k.size(0), 1) + k.size()[2:], dtype=k.dtype, device=k.device)], dim=1)
v = torch.cat([v, torch.zeros((v.size(0), 1) + v.size()[2:], dtype=v.dtype, device=v.device)], dim=1)
if attn_mask is not None:
attn_mask = torch.cat([attn_mask, torch.zeros((attn_mask.size(0), 1),
dtype=attn_mask.dtype,
device=attn_mask.device)], dim=1)
if key_padding_mask is not None:
key_padding_mask = torch.cat(
[key_padding_mask, torch.zeros((key_padding_mask.size(0), 1),
dtype=key_padding_mask.dtype,
device=key_padding_mask.device)], dim=1)
attn_output_weights = torch.bmm(q, k.transpose(1, 2))
assert list(attn_output_weights.size()) == [bsz * num_heads, tgt_len, src_len]
if attn_mask is not None:
attn_mask = attn_mask.unsqueeze(0)
attn_output_weights += attn_mask
if key_padding_mask is not None:
attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
attn_output_weights = attn_output_weights.masked_fill(
key_padding_mask.unsqueeze(1).unsqueeze(2),
float('-inf'),
)
attn_output_weights = attn_output_weights.view(bsz * num_heads, tgt_len, src_len)
attn_output_weights = F.softmax(
attn_output_weights, dim=-1)
attn_output_weights = F.dropout(attn_output_weights, p=dropout_p, training=training)
attn_output = torch.bmm(attn_output_weights, v)
assert list(attn_output.size()) == [bsz * num_heads, tgt_len, head_dim]
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)
attn_output = F.linear(attn_output, out_proj_weight, out_proj_bias)
if need_weights:
# average attention weights over heads
attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
return attn_output, attn_output_weights.sum(dim=1) / num_heads
else:
return attn_output, None
class FFN(nn.Module):
def __init__(self,
in_channels,
heads,
head_conv=64,
final_kernel=1,
init_bias=-2.19,
conv_cfg=dict(type='Conv1d'),
norm_cfg=dict(type='BN1d'),
bias='auto',
**kwargs):
super(FFN, self).__init__()
self.heads = heads
self.init_bias = init_bias
for head in self.heads:
classes, num_conv = self.heads[head]
conv_layers = []
c_in = in_channels
for i in range(num_conv - 1):
conv_layers.append(
ConvModule(
c_in,
head_conv,
kernel_size=final_kernel,
stride=1,
padding=final_kernel // 2,
bias=bias,
conv_cfg=conv_cfg,
norm_cfg=norm_cfg))
c_in = head_conv
conv_layers.append(
build_conv_layer(
conv_cfg,
head_conv,
classes,
kernel_size=final_kernel,
stride=1,
padding=final_kernel // 2,
bias=True))
conv_layers = nn.Sequential(*conv_layers)
self.__setattr__(head, conv_layers)
def init_weights(self):
"""Initialize weights."""
for head in self.heads:
if head == 'heatmap':
self.__getattr__(head)[-1].bias.data.fill_(self.init_bias)
else:
for m in self.__getattr__(head).modules():
if isinstance(m, nn.Conv2d):
kaiming_init(m)
def forward(self, x):
"""Forward function for SepHead.
Args:
x (torch.Tensor): Input feature map with the shape of
[B, 512, 128, 128].
Returns:
dict[str: torch.Tensor]: contains the following keys:
-reg (torch.Tensor): 2D regression value with the \
shape of [B, 2, H, W].
-height (torch.Tensor): Height value with the \
shape of [B, 1, H, W].
-dim (torch.Tensor): Size value with the shape \
of [B, 3, H, W].
-rot (torch.Tensor): Rotation value with the \
shape of [B, 1, H, W].
-vel (torch.Tensor): Velocity value with the \
shape of [B, 2, H, W].
-heatmap (torch.Tensor): Heatmap with the shape of \
[B, N, H, W].
"""
ret_dict = dict()
for head in self.heads:
ret_dict[head] = self.__getattr__(head)(x)
return ret_dict
@HEADS.register_module()
class TransFusionHead(nn.Module):
def __init__(self,
fuse_img=False,
num_views=0,
in_channels_img=64,
out_size_factor_img=4,
num_proposals=128,
auxiliary=True,
in_channels=128 * 3,
hidden_channel=128,
num_classes=4,
# config for Transformer
num_decoder_layers=3,
num_heads=8,
learnable_query_pos=False,
initialize_by_heatmap=False,
nms_kernel_size=1,
ffn_channel=256,
dropout=0.1,
bn_momentum=0.1,
activation='relu',
# config for FFN
common_heads=dict(),
num_heatmap_convs=2,
conv_cfg=dict(type='Conv1d'),
norm_cfg=dict(type='BN1d'),
bias='auto',
# loss
loss_cls=dict(type='GaussianFocalLoss', reduction='mean'),
loss_iou=dict(type='VarifocalLoss', use_sigmoid=True, iou_weighted=True, reduction='mean'),
loss_bbox=dict(type='L1Loss', reduction='mean'),
loss_heatmap=dict(type='GaussianFocalLoss', reduction='mean'),
# others
train_cfg=None,
test_cfg=None,
bbox_coder=None,
):
super(TransFusionHead, self).__init__()
self.num_classes = num_classes
self.num_proposals = num_proposals
self.auxiliary = auxiliary
self.in_channels = in_channels
self.num_heads = num_heads
self.num_decoder_layers = num_decoder_layers
self.bn_momentum = bn_momentum
self.learnable_query_pos = learnable_query_pos
self.initialize_by_heatmap = initialize_by_heatmap
self.nms_kernel_size = nms_kernel_size
if self.initialize_by_heatmap is True:
assert self.learnable_query_pos is False, "initialized by heatmap is conflicting with learnable query position"
self.train_cfg = train_cfg
self.test_cfg = test_cfg
self.use_sigmoid_cls = loss_cls.get('use_sigmoid', False)
if not self.use_sigmoid_cls:
self.num_classes += 1
self.loss_cls = build_loss(loss_cls)
self.loss_bbox = build_loss(loss_bbox)
self.loss_iou = build_loss(loss_iou)
self.loss_heatmap = build_loss(loss_heatmap)
self.bbox_coder = build_bbox_coder(bbox_coder)
self.sampling = False
# a shared convolution
self.shared_conv = build_conv_layer(
dict(type='Conv2d'),
in_channels,
hidden_channel,
kernel_size=3,
padding=1,
bias=bias,
)
if self.initialize_by_heatmap:
layers = []
layers.append(ConvModule(
hidden_channel,
hidden_channel,
kernel_size=3,
padding=1,
bias=bias,
conv_cfg=dict(type='Conv2d'),
norm_cfg=dict(type='BN2d'),
))
layers.append(build_conv_layer(
dict(type='Conv2d'),
hidden_channel,
num_classes,
kernel_size=3,
padding=1,
bias=bias,
))
self.heatmap_head = nn.Sequential(*layers)
self.class_encoding = nn.Conv1d(num_classes, hidden_channel, 1)
else:
# query feature
self.query_feat = nn.Parameter(torch.randn(1, hidden_channel, self.num_proposals))
self.query_pos = nn.Parameter(torch.rand([1, self.num_proposals, 2]), requires_grad=learnable_query_pos)
# transformer decoder layers for object query with LiDAR feature
self.decoder = nn.ModuleList()
for i in range(self.num_decoder_layers):
self.decoder.append(
TransformerDecoderLayer(
hidden_channel, num_heads, ffn_channel, dropout, activation,
self_posembed=PositionEmbeddingLearned(2, hidden_channel),
cross_posembed=PositionEmbeddingLearned(2, hidden_channel),
))
# Prediction Head
self.prediction_heads = nn.ModuleList()
for i in range(self.num_decoder_layers):
heads = copy.deepcopy(common_heads)
heads.update(dict(heatmap=(self.num_classes, num_heatmap_convs)))
self.prediction_heads.append(FFN(hidden_channel, heads, conv_cfg=conv_cfg, norm_cfg=norm_cfg, bias=bias))
self.fuse_img = fuse_img
if self.fuse_img:
self.num_views = num_views
self.out_size_factor_img = out_size_factor_img
self.shared_conv_img = build_conv_layer(
dict(type='Conv2d'),
in_channels_img, # channel of img feature map
hidden_channel,
kernel_size=3,
padding=1,
bias=bias,
)
self.heatmap_head_img = copy.deepcopy(self.heatmap_head)
# transformer decoder layers for img fusion
self.decoder.append(
TransformerDecoderLayer(
hidden_channel, num_heads, ffn_channel, dropout, activation,
self_posembed=PositionEmbeddingLearned(2, hidden_channel),
cross_posembed=PositionEmbeddingLearned(2, hidden_channel),
))
# cross-attention only layers for projecting img feature onto BEV
for i in range(num_views):
self.decoder.append(
TransformerDecoderLayer(
hidden_channel, num_heads, ffn_channel, dropout, activation,
self_posembed=PositionEmbeddingLearned(2, hidden_channel),
cross_posembed=PositionEmbeddingLearned(2, hidden_channel),
cross_only=True,
))
self.fc = nn.Sequential(*[nn.Conv1d(hidden_channel, hidden_channel, kernel_size=1)])
heads = copy.deepcopy(common_heads)
heads.update(dict(heatmap=(self.num_classes, num_heatmap_convs)))
self.prediction_heads.append(FFN(hidden_channel * 2, heads, conv_cfg=conv_cfg, norm_cfg=norm_cfg, bias=bias))
self.init_weights()
self._init_assigner_sampler()
# Position Embedding for Cross-Attention, which is re-used during training
x_size = self.test_cfg['grid_size'][0] // self.test_cfg['out_size_factor']
y_size = self.test_cfg['grid_size'][1] // self.test_cfg['out_size_factor']
self.bev_pos = self.create_2D_grid(x_size, y_size)
self.img_feat_pos = None
self.img_feat_collapsed_pos = None
def create_2D_grid(self, x_size, y_size):
meshgrid = [[0, x_size - 1, x_size], [0, y_size - 1, y_size]]
batch_y, batch_x = torch.meshgrid(*[torch.linspace(it[0], it[1], it[2]) for it in meshgrid])
batch_x = batch_x + 0.5
batch_y = batch_y + 0.5
coord_base = torch.cat([batch_x[None], batch_y[None]], dim=0)[None]
coord_base = coord_base.view(1, 2, -1).permute(0, 2, 1)
return coord_base
def init_weights(self):
# initialize transformer
for m in self.decoder.parameters():
if m.dim() > 1:
nn.init.xavier_uniform_(m)
if hasattr(self, 'query'):
nn.init.xavier_normal_(self.query)
self.init_bn_momentum()
def init_bn_momentum(self):
for m in self.modules():
if isinstance(m, (nn.BatchNorm2d, nn.BatchNorm1d)):
m.momentum = self.bn_momentum
def _init_assigner_sampler(self):
"""Initialize the target assigner and sampler of the head."""
if self.train_cfg is None:
return
if self.sampling:
self.bbox_sampler = build_sampler(self.train_cfg.sampler)
else:
self.bbox_sampler = PseudoSampler()
if isinstance(self.train_cfg.assigner, dict):
self.bbox_assigner = build_assigner(self.train_cfg.assigner)
elif isinstance(self.train_cfg.assigner, list):
self.bbox_assigner = [
build_assigner(res) for res in self.train_cfg.assigner
]
def forward_single(self, inputs, img_inputs, img_metas):
"""Forward function for CenterPoint.
Args:
inputs (torch.Tensor): Input feature map with the shape of
[B, 512, 128(H), 128(W)]. (consistent with L748)
Returns:
list[dict]: Output results for tasks.
"""
batch_size = inputs.shape[0]
lidar_feat = self.shared_conv(inputs)
#################################
# image to BEV
#################################
lidar_feat_flatten = lidar_feat.view(batch_size, lidar_feat.shape[1], -1) # [BS, C, H*W]
bev_pos = self.bev_pos.repeat(batch_size, 1, 1).to(lidar_feat.device)
if self.fuse_img:
img_feat = self.shared_conv_img(img_inputs) # [BS, n_views, H, W, C]
img_h, img_w, num_channel = img_inputs.shape[-2], img_inputs.shape[-1], img_feat.shape[1]
img_feat = img_feat.permute(0, 4, 2, 1, 3)
img_feat = img_feat.view(batch_size, num_channel, img_h, img_w * self.num_views) # [BS, C, H, n_views*W]
img_feat_collapsed = img_feat.max(2).values
img_feat_collapsed = self.fc(img_feat_collapsed).view(batch_size, num_channel, img_w * self.num_views)
# positional encoding for image guided query initialization
if self.img_feat_collapsed_pos is None:
img_feat_collapsed_pos = self.img_feat_collapsed_po = self.create_2D_grid(1, img_feat_collapsed.shape[-1]).to(img_feat.device)
else:
img_feat_collapsed_pos = self.img_feat_collapsed_pos
bev_feat = lidar_feat_flatten
for idx_view in range(self.num_views):
bev_feat = self.decoder[2 + idx_view](bev_feat, img_feat_collapsed[..., img_w * idx_view:img_w * (idx_view + 1)], bev_pos, img_feat_collapsed_pos[:, img_w * idx_view:img_w * (idx_view + 1)])
#################################
# image guided query initialization
#################################
if self.initialize_by_heatmap:
dense_heatmap = self.heatmap_head(lidar_feat)
dense_heatmap_img = None
if self.fuse_img:
dense_heatmap_img = self.heatmap_head_img(bev_feat.view(lidar_feat.shape)) # [BS, num_classes, H, W]
heatmap = (dense_heatmap.detach().sigmoid() + dense_heatmap_img.detach().sigmoid()) / 2
else:
heatmap = dense_heatmap.detach().sigmoid()
padding = self.nms_kernel_size // 2
local_max = torch.zeros_like(heatmap)
# equals to nms radius = voxel_size * out_size_factor * kenel_size
local_max_inner = F.max_pool2d(heatmap, kernel_size=self.nms_kernel_size, stride=1, padding=0)
local_max[:, :, padding:(-padding), padding:(-padding)] = local_max_inner
## for Pedestrian & Traffic_cone in nuScenes
if self.test_cfg['dataset'] == 'nuScenes':
local_max[:, 8, ] = F.max_pool2d(heatmap[:, 8], kernel_size=1, stride=1, padding=0)
local_max[:, 9, ] = F.max_pool2d(heatmap[:, 9], kernel_size=1, stride=1, padding=0)
elif self.test_cfg['dataset'] == 'Waymo': # for Pedestrian & Cyclist in Waymo
local_max[:, 1, ] = F.max_pool2d(heatmap[:, 1], kernel_size=1, stride=1, padding=0)
local_max[:, 2, ] = F.max_pool2d(heatmap[:, 2], kernel_size=1, stride=1, padding=0)
heatmap = heatmap * (heatmap == local_max)
heatmap = heatmap.view(batch_size, heatmap.shape[1], -1)
# top #num_proposals among all classes
top_proposals = heatmap.view(batch_size, -1).argsort(dim=-1, descending=True)[..., :self.num_proposals]
top_proposals_class = top_proposals // heatmap.shape[-1]
top_proposals_index = top_proposals % heatmap.shape[-1]
query_feat = lidar_feat_flatten.gather(index=top_proposals_index[:, None, :].expand(-1, lidar_feat_flatten.shape[1], -1), dim=-1)
self.query_labels = top_proposals_class
# add category embedding
one_hot = F.one_hot(top_proposals_class, num_classes=self.num_classes).permute(0, 2, 1)
query_cat_encoding = self.class_encoding(one_hot.float())
query_feat += query_cat_encoding
query_pos = bev_pos.gather(index=top_proposals_index[:, None, :].permute(0, 2, 1).expand(-1, -1, bev_pos.shape[-1]), dim=1)
else:
query_feat = self.query_feat.repeat(batch_size, 1, 1) # [BS, C, num_proposals]
base_xyz = self.query_pos.repeat(batch_size, 1, 1).to(lidar_feat.device) # [BS, num_proposals, 2]
#################################
# transformer decoder layer (LiDAR feature as K,V)
#################################
ret_dicts = []
for i in range(self.num_decoder_layers):
prefix = 'last_' if (i == self.num_decoder_layers - 1) else f'{i}head_'
# Transformer Decoder Layer
# :param query: B C Pq :param query_pos: B Pq 3/6
query_feat = self.decoder[i](query_feat, lidar_feat_flatten, query_pos, bev_pos)
# Prediction
res_layer = self.prediction_heads[i](query_feat)
res_layer['center'] = res_layer['center'] + query_pos.permute(0, 2, 1)
first_res_layer = res_layer
if not self.fuse_img:
ret_dicts.append(res_layer)
# for next level positional embedding
query_pos = res_layer['center'].detach().clone().permute(0, 2, 1)
#################################
# transformer decoder layer (img feature as K,V)
#################################
if self.fuse_img:
# positional encoding for image fusion
img_feat_flatten = img_feat.view(batch_size, self.num_views, img_feat.shape[1], -1) # [BS, n_views, C, H*W]
if self.img_feat_pos is None:
(h, w) = img_inputs.shape[-2], img_inputs.shape[-1]
img_feat_pos = self.img_feat_pos = self.create_2D_grid(h, w).to(img_feat_flatten.device)
else:
img_feat_pos = self.img_feat_pos
prev_query_feat = query_feat.detach().clone()
query_feat = torch.zeros_like(query_feat) # create new container for img query feature
query_pos_realmetric = query_pos.permute(0, 2, 1) * self.test_cfg['out_size_factor'] * self.test_cfg['voxel_size'][0] + self.test_cfg['pc_range'][0]
query_pos_3d = torch.cat([query_pos_realmetric, res_layer['height']], dim=1).detach().clone()
if 'vel' in res_layer:
vel = copy.deepcopy(res_layer['vel'].detach())
else:
vel = None
pred_boxes = self.bbox_coder.decode(
copy.deepcopy(res_layer['heatmap'].detach()),
copy.deepcopy(res_layer['rot'].detach()),
copy.deepcopy(res_layer['dim'].detach()),
copy.deepcopy(res_layer['center'].detach()),
copy.deepcopy(res_layer['height'].detach()),
vel,
)
on_the_image_mask = torch.ones([batch_size, self.num_proposals]).to(query_pos_3d.device) * -1
for sample_idx in range(batch_size if self.fuse_img else 0):
lidar2img_rt = query_pos_3d.new_tensor(img_metas[sample_idx]['lidar2img'])
img_scale_factor = (
query_pos_3d.new_tensor(img_metas[sample_idx]['scale_factor'][:2]
if 'scale_factor' in img_metas[sample_idx].keys() else [1.0, 1.0]))
img_flip = img_metas[sample_idx]['flip'] if 'flip' in img_metas[sample_idx].keys() else False
img_crop_offset = (
query_pos_3d.new_tensor(img_metas[sample_idx]['img_crop_offset'])
if 'img_crop_offset' in img_metas[sample_idx].keys() else 0)
img_shape = img_metas[sample_idx]['img_shape'][:2]
img_pad_shape = img_metas[sample_idx]['input_shape'][:2]
boxes = LiDARInstance3DBoxes(pred_boxes[sample_idx]['bboxes'][:, :7], box_dim=7)
query_pos_3d_with_corners = torch.cat([query_pos_3d[sample_idx], boxes.corners.permute(2, 0, 1).view(3, -1)], dim=-1) # [3, num_proposals] + [3, num_proposals*8]
# transform point clouds back to original coordinate system by reverting the data augmentation
if batch_size == 1: # skip during inference to save time
points = query_pos_3d_with_corners.T
else:
points = apply_3d_transformation(query_pos_3d_with_corners.T, 'LIDAR', img_metas[sample_idx], reverse=True).detach()
num_points = points.shape[0]
for view_idx in range(self.num_views):
pts_4d = torch.cat([points, points.new_ones(size=(num_points, 1))], dim=-1)
pts_2d = pts_4d @ lidar2img_rt[view_idx].t()
pts_2d[:, 2] = torch.clamp(pts_2d[:, 2], min=1e-5)
pts_2d[:, 0] /= pts_2d[:, 2]
pts_2d[:, 1] /= pts_2d[:, 2]
# img transformation: scale -> crop -> flip
# the image is resized by img_scale_factor
img_coors = pts_2d[:, 0:2] * img_scale_factor # Nx2
img_coors -= img_crop_offset
# grid sample, the valid grid range should be in [-1,1]
coor_x, coor_y = torch.split(img_coors, 1, dim=1) # each is Nx1
if img_flip:
# by default we take it as horizontal flip
# use img_shape before padding for flip
orig_h, orig_w = img_shape
coor_x = orig_w - coor_x
coor_x, coor_corner_x = coor_x[0:self.num_proposals, :], coor_x[self.num_proposals:, :]
coor_y, coor_corner_y = coor_y[0:self.num_proposals, :], coor_y[self.num_proposals:, :]
coor_corner_x = coor_corner_x.reshape(self.num_proposals, 8, 1)
coor_corner_y = coor_corner_y.reshape(self.num_proposals, 8, 1)
coor_corner_xy = torch.cat([coor_corner_x, coor_corner_y], dim=-1)
h, w = img_pad_shape
on_the_image = (coor_x > 0) * (coor_x < w) * (coor_y > 0) * (coor_y < h)
on_the_image = on_the_image.squeeze()
# skip the following computation if no object query fall on current image
if on_the_image.sum() <= 1:
continue
on_the_image_mask[sample_idx, on_the_image] = view_idx
# add spatial constraint
center_ys = (coor_y[on_the_image] / self.out_size_factor_img)
center_xs = (coor_x[on_the_image] / self.out_size_factor_img)
centers = torch.cat([center_xs, center_ys], dim=-1).int() # center on the feature map
corners = (coor_corner_xy[on_the_image].max(1).values - coor_corner_xy[on_the_image].min(1).values) / self.out_size_factor_img
radius = torch.ceil(corners.norm(dim=-1, p=2) / 2).int() # radius of the minimum circumscribed circle of the wireframe
sigma = (radius * 2 + 1) / 6.0
distance = (centers[:, None, :] - (img_feat_pos - 0.5)).norm(dim=-1) ** 2
gaussian_mask = (-distance / (2 * sigma[:, None] ** 2)).exp()
gaussian_mask[gaussian_mask < torch.finfo(torch.float32).eps] = 0
attn_mask = gaussian_mask
query_feat_view = prev_query_feat[sample_idx, :, on_the_image]
query_pos_view = torch.cat([center_xs, center_ys], dim=-1)
query_feat_view = self.decoder[self.num_decoder_layers](query_feat_view[None], img_feat_flatten[sample_idx:sample_idx + 1, view_idx], query_pos_view[None], img_feat_pos, attn_mask=attn_mask.log())