-
-
Notifications
You must be signed in to change notification settings - Fork 4.7k
/
maxxvit.py
2402 lines (2056 loc) · 85.8 KB
/
maxxvit.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
""" MaxVit and CoAtNet Vision Transformer - CNN Hybrids in PyTorch
This is a from-scratch implementation of both CoAtNet and MaxVit in PyTorch.
99% of the implementation was done from papers, however last minute some adjustments were made
based on the (as yet unfinished?) public code release https://github.com/google-research/maxvit
There are multiple sets of models defined for both architectures. Typically, names with a
`_rw` suffix are my own original configs prior to referencing https://github.com/google-research/maxvit.
These configs work well and appear to be a bit faster / lower resource than the paper.
The models without extra prefix / suffix' (coatnet_0_224, maxvit_tiny_224, etc), are intended to
match paper, BUT, without any official pretrained weights it's difficult to confirm a 100% match.
Papers:
MaxViT: Multi-Axis Vision Transformer - https://arxiv.org/abs/2204.01697
@article{tu2022maxvit,
title={MaxViT: Multi-Axis Vision Transformer},
author={Tu, Zhengzhong and Talebi, Hossein and Zhang, Han and Yang, Feng and Milanfar, Peyman and Bovik, Alan and Li, Yinxiao},
journal={ECCV},
year={2022},
}
CoAtNet: Marrying Convolution and Attention for All Data Sizes - https://arxiv.org/abs/2106.04803
@article{DBLP:journals/corr/abs-2106-04803,
author = {Zihang Dai and Hanxiao Liu and Quoc V. Le and Mingxing Tan},
title = {CoAtNet: Marrying Convolution and Attention for All Data Sizes},
journal = {CoRR},
volume = {abs/2106.04803},
year = {2021}
}
Hacked together by / Copyright 2022, Ross Wightman
"""
import math
from collections import OrderedDict
from dataclasses import dataclass, replace, field
from functools import partial
from typing import Callable, Optional, Union, Tuple, List
import torch
from torch import nn
from torch.jit import Final
from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
from timm.layers import Mlp, ConvMlp, DropPath, LayerNorm, ClassifierHead, NormMlpClassifierHead
from timm.layers import create_attn, get_act_layer, get_norm_layer, get_norm_act_layer, create_conv2d, create_pool2d
from timm.layers import trunc_normal_tf_, to_2tuple, extend_tuple, make_divisible, _assert
from timm.layers import RelPosMlp, RelPosBias, RelPosBiasTf, use_fused_attn, resize_rel_pos_bias_table
from ._builder import build_model_with_cfg
from ._features import feature_take_indices
from ._features_fx import register_notrace_function
from ._manipulate import named_apply, checkpoint_seq
from ._registry import generate_default_cfgs, register_model
__all__ = ['MaxxVitCfg', 'MaxxVitConvCfg', 'MaxxVitTransformerCfg', 'MaxxVit']
@dataclass
class MaxxVitTransformerCfg:
dim_head: int = 32
head_first: bool = True # head ordering in qkv channel dim
expand_ratio: float = 4.0
expand_first: bool = True
shortcut_bias: bool = True
attn_bias: bool = True
attn_drop: float = 0.
proj_drop: float = 0.
pool_type: str = 'avg2'
rel_pos_type: str = 'bias'
rel_pos_dim: int = 512 # for relative position types w/ MLP
partition_ratio: int = 32
window_size: Optional[Tuple[int, int]] = None
grid_size: Optional[Tuple[int, int]] = None
no_block_attn: bool = False # disable window block attention for maxvit (ie only grid)
use_nchw_attn: bool = False # for MaxViT variants (not used for CoAt), keep tensors in NCHW order
init_values: Optional[float] = None
act_layer: str = 'gelu'
norm_layer: str = 'layernorm2d'
norm_layer_cl: str = 'layernorm'
norm_eps: float = 1e-6
def __post_init__(self):
if self.grid_size is not None:
self.grid_size = to_2tuple(self.grid_size)
if self.window_size is not None:
self.window_size = to_2tuple(self.window_size)
if self.grid_size is None:
self.grid_size = self.window_size
@dataclass
class MaxxVitConvCfg:
block_type: str = 'mbconv'
expand_ratio: float = 4.0
expand_output: bool = True # calculate expansion channels from output (vs input chs)
kernel_size: int = 3
group_size: int = 1 # 1 == depthwise
pre_norm_act: bool = False # activation after pre-norm
output_bias: bool = True # bias for shortcut + final 1x1 projection conv
stride_mode: str = 'dw' # stride done via one of 'pool', '1x1', 'dw'
pool_type: str = 'avg2'
downsample_pool_type: str = 'avg2'
padding: str = ''
attn_early: bool = False # apply attn between conv2 and norm2, instead of after norm2
attn_layer: str = 'se'
attn_act_layer: str = 'silu'
attn_ratio: float = 0.25
init_values: Optional[float] = 1e-6 # for ConvNeXt block, ignored by MBConv
act_layer: str = 'gelu'
norm_layer: str = ''
norm_layer_cl: str = ''
norm_eps: Optional[float] = None
def __post_init__(self):
# mbconv vs convnext blocks have different defaults, set in post_init to avoid explicit config args
assert self.block_type in ('mbconv', 'convnext')
use_mbconv = self.block_type == 'mbconv'
if not self.norm_layer:
self.norm_layer = 'batchnorm2d' if use_mbconv else 'layernorm2d'
if not self.norm_layer_cl and not use_mbconv:
self.norm_layer_cl = 'layernorm'
if self.norm_eps is None:
self.norm_eps = 1e-5 if use_mbconv else 1e-6
self.downsample_pool_type = self.downsample_pool_type or self.pool_type
@dataclass
class MaxxVitCfg:
embed_dim: Tuple[int, ...] = (96, 192, 384, 768)
depths: Tuple[int, ...] = (2, 3, 5, 2)
block_type: Tuple[Union[str, Tuple[str, ...]], ...] = ('C', 'C', 'T', 'T')
stem_width: Union[int, Tuple[int, int]] = 64
stem_bias: bool = False
conv_cfg: MaxxVitConvCfg = field(default_factory=MaxxVitConvCfg)
transformer_cfg: MaxxVitTransformerCfg = field(default_factory=MaxxVitTransformerCfg)
head_hidden_size: int = None
weight_init: str = 'vit_eff'
class Attention2d(nn.Module):
fused_attn: Final[bool]
""" multi-head attention for 2D NCHW tensors"""
def __init__(
self,
dim: int,
dim_out: Optional[int] = None,
dim_head: int = 32,
bias: bool = True,
expand_first: bool = True,
head_first: bool = True,
rel_pos_cls: Callable = None,
attn_drop: float = 0.,
proj_drop: float = 0.
):
super().__init__()
dim_out = dim_out or dim
dim_attn = dim_out if expand_first else dim
self.num_heads = dim_attn // dim_head
self.dim_head = dim_head
self.head_first = head_first
self.scale = dim_head ** -0.5
self.fused_attn = use_fused_attn()
self.qkv = nn.Conv2d(dim, dim_attn * 3, 1, bias=bias)
self.rel_pos = rel_pos_cls(num_heads=self.num_heads) if rel_pos_cls else None
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Conv2d(dim_attn, dim_out, 1, bias=bias)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
B, C, H, W = x.shape
if self.head_first:
q, k, v = self.qkv(x).view(B, self.num_heads, self.dim_head * 3, -1).chunk(3, dim=2)
else:
q, k, v = self.qkv(x).reshape(B, 3, self.num_heads, self.dim_head, -1).unbind(1)
if self.fused_attn:
attn_bias = None
if self.rel_pos is not None:
attn_bias = self.rel_pos.get_bias()
elif shared_rel_pos is not None:
attn_bias = shared_rel_pos
x = torch.nn.functional.scaled_dot_product_attention(
q.transpose(-1, -2).contiguous(),
k.transpose(-1, -2).contiguous(),
v.transpose(-1, -2).contiguous(),
attn_mask=attn_bias,
dropout_p=self.attn_drop.p if self.training else 0.,
).transpose(-1, -2).reshape(B, -1, H, W)
else:
q = q * self.scale
attn = q.transpose(-2, -1) @ k
if self.rel_pos is not None:
attn = self.rel_pos(attn)
elif shared_rel_pos is not None:
attn = attn + shared_rel_pos
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = (v @ attn.transpose(-2, -1)).view(B, -1, H, W)
x = self.proj(x)
x = self.proj_drop(x)
return x
class AttentionCl(nn.Module):
""" Channels-last multi-head attention (B, ..., C) """
fused_attn: Final[bool]
def __init__(
self,
dim: int,
dim_out: Optional[int] = None,
dim_head: int = 32,
bias: bool = True,
expand_first: bool = True,
head_first: bool = True,
rel_pos_cls: Callable = None,
attn_drop: float = 0.,
proj_drop: float = 0.
):
super().__init__()
dim_out = dim_out or dim
dim_attn = dim_out if expand_first and dim_out > dim else dim
assert dim_attn % dim_head == 0, 'attn dim should be divisible by head_dim'
self.num_heads = dim_attn // dim_head
self.dim_head = dim_head
self.head_first = head_first
self.scale = dim_head ** -0.5
self.fused_attn = use_fused_attn()
self.qkv = nn.Linear(dim, dim_attn * 3, bias=bias)
self.rel_pos = rel_pos_cls(num_heads=self.num_heads) if rel_pos_cls else None
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim_attn, dim_out, bias=bias)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
B = x.shape[0]
restore_shape = x.shape[:-1]
if self.head_first:
q, k, v = self.qkv(x).view(B, -1, self.num_heads, self.dim_head * 3).transpose(1, 2).chunk(3, dim=3)
else:
q, k, v = self.qkv(x).reshape(B, -1, 3, self.num_heads, self.dim_head).transpose(1, 3).unbind(2)
if self.fused_attn:
attn_bias = None
if self.rel_pos is not None:
attn_bias = self.rel_pos.get_bias()
elif shared_rel_pos is not None:
attn_bias = shared_rel_pos
x = torch.nn.functional.scaled_dot_product_attention(
q, k, v,
attn_mask=attn_bias,
dropout_p=self.attn_drop.p if self.training else 0.,
)
else:
q = q * self.scale
attn = q @ k.transpose(-2, -1)
if self.rel_pos is not None:
attn = self.rel_pos(attn, shared_rel_pos=shared_rel_pos)
elif shared_rel_pos is not None:
attn = attn + shared_rel_pos
attn = attn.softmax(dim=-1)
attn = self.attn_drop(attn)
x = attn @ v
x = x.transpose(1, 2).reshape(restore_shape + (-1,))
x = self.proj(x)
x = self.proj_drop(x)
return x
class LayerScale(nn.Module):
def __init__(self, dim, init_values=1e-5, inplace=False):
super().__init__()
self.inplace = inplace
self.gamma = nn.Parameter(init_values * torch.ones(dim))
def forward(self, x):
gamma = self.gamma
return x.mul_(gamma) if self.inplace else x * gamma
class LayerScale2d(nn.Module):
def __init__(self, dim, init_values=1e-5, inplace=False):
super().__init__()
self.inplace = inplace
self.gamma = nn.Parameter(init_values * torch.ones(dim))
def forward(self, x):
gamma = self.gamma.view(1, -1, 1, 1)
return x.mul_(gamma) if self.inplace else x * gamma
class Downsample2d(nn.Module):
""" A downsample pooling module supporting several maxpool and avgpool modes
* 'max' - MaxPool2d w/ kernel_size 3, stride 2, padding 1
* 'max2' - MaxPool2d w/ kernel_size = stride = 2
* 'avg' - AvgPool2d w/ kernel_size 3, stride 2, padding 1
* 'avg2' - AvgPool2d w/ kernel_size = stride = 2
"""
def __init__(
self,
dim: int,
dim_out: int,
pool_type: str = 'avg2',
padding: str = '',
bias: bool = True,
):
super().__init__()
assert pool_type in ('max', 'max2', 'avg', 'avg2')
if pool_type == 'max':
self.pool = create_pool2d('max', kernel_size=3, stride=2, padding=padding or 1)
elif pool_type == 'max2':
self.pool = create_pool2d('max', 2, padding=padding or 0) # kernel_size == stride == 2
elif pool_type == 'avg':
self.pool = create_pool2d(
'avg', kernel_size=3, stride=2, count_include_pad=False, padding=padding or 1)
else:
self.pool = create_pool2d('avg', 2, padding=padding or 0)
if dim != dim_out:
self.expand = nn.Conv2d(dim, dim_out, 1, bias=bias)
else:
self.expand = nn.Identity()
def forward(self, x):
x = self.pool(x) # spatial downsample
x = self.expand(x) # expand chs
return x
def _init_transformer(module, name, scheme=''):
if isinstance(module, (nn.Conv2d, nn.Linear)):
if scheme == 'normal':
nn.init.normal_(module.weight, std=.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif scheme == 'trunc_normal':
trunc_normal_tf_(module.weight, std=.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif scheme == 'xavier_normal':
nn.init.xavier_normal_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
else:
# vit like
nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
if 'mlp' in name:
nn.init.normal_(module.bias, std=1e-6)
else:
nn.init.zeros_(module.bias)
class TransformerBlock2d(nn.Module):
""" Transformer block with 2D downsampling
'2D' NCHW tensor layout
Some gains can be seen on GPU using a 1D / CL block, BUT w/ the need to switch back/forth to NCHW
for spatial pooling, the benefit is minimal so ended up using just this variant for CoAt configs.
This impl was faster on TPU w/ PT XLA than the 1D experiment.
"""
def __init__(
self,
dim: int,
dim_out: int,
stride: int = 1,
rel_pos_cls: Callable = None,
cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path: float = 0.,
):
super().__init__()
norm_layer = partial(get_norm_layer(cfg.norm_layer), eps=cfg.norm_eps)
act_layer = get_act_layer(cfg.act_layer)
if stride == 2:
self.shortcut = Downsample2d(dim, dim_out, pool_type=cfg.pool_type, bias=cfg.shortcut_bias)
self.norm1 = nn.Sequential(OrderedDict([
('norm', norm_layer(dim)),
('down', Downsample2d(dim, dim, pool_type=cfg.pool_type)),
]))
else:
assert dim == dim_out
self.shortcut = nn.Identity()
self.norm1 = norm_layer(dim)
self.attn = Attention2d(
dim,
dim_out,
dim_head=cfg.dim_head,
expand_first=cfg.expand_first,
bias=cfg.attn_bias,
rel_pos_cls=rel_pos_cls,
attn_drop=cfg.attn_drop,
proj_drop=cfg.proj_drop
)
self.ls1 = LayerScale2d(dim_out, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim_out)
self.mlp = ConvMlp(
in_features=dim_out,
hidden_features=int(dim_out * cfg.expand_ratio),
act_layer=act_layer,
drop=cfg.proj_drop)
self.ls2 = LayerScale2d(dim_out, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def init_weights(self, scheme=''):
named_apply(partial(_init_transformer, scheme=scheme), self)
def forward(self, x, shared_rel_pos: Optional[torch.Tensor] = None):
x = self.shortcut(x) + self.drop_path1(self.ls1(self.attn(self.norm1(x), shared_rel_pos=shared_rel_pos)))
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
return x
def _init_conv(module, name, scheme=''):
if isinstance(module, nn.Conv2d):
if scheme == 'normal':
nn.init.normal_(module.weight, std=.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif scheme == 'trunc_normal':
trunc_normal_tf_(module.weight, std=.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif scheme == 'xavier_normal':
nn.init.xavier_normal_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
else:
# efficientnet like
fan_out = module.kernel_size[0] * module.kernel_size[1] * module.out_channels
fan_out //= module.groups
nn.init.normal_(module.weight, 0, math.sqrt(2.0 / fan_out))
if module.bias is not None:
nn.init.zeros_(module.bias)
def num_groups(group_size, channels):
if not group_size: # 0 or None
return 1 # normal conv with 1 group
else:
# NOTE group_size == 1 -> depthwise conv
assert channels % group_size == 0
return channels // group_size
class MbConvBlock(nn.Module):
""" Pre-Norm Conv Block - 1x1 - kxk - 1x1, w/ inverted bottleneck (expand)
"""
def __init__(
self,
in_chs: int,
out_chs: int,
stride: int = 1,
dilation: Tuple[int, int] = (1, 1),
cfg: MaxxVitConvCfg = MaxxVitConvCfg(),
drop_path: float = 0.
):
super(MbConvBlock, self).__init__()
norm_act_layer = partial(get_norm_act_layer(cfg.norm_layer, cfg.act_layer), eps=cfg.norm_eps)
mid_chs = make_divisible((out_chs if cfg.expand_output else in_chs) * cfg.expand_ratio)
groups = num_groups(cfg.group_size, mid_chs)
if stride == 2:
self.shortcut = Downsample2d(
in_chs, out_chs, pool_type=cfg.pool_type, bias=cfg.output_bias, padding=cfg.padding)
else:
self.shortcut = nn.Identity()
assert cfg.stride_mode in ('pool', '1x1', 'dw')
stride_pool, stride_1, stride_2 = 1, 1, 1
if cfg.stride_mode == 'pool':
# NOTE this is not described in paper, experiment to find faster option that doesn't stride in 1x1
stride_pool, dilation_2 = stride, dilation[1]
# FIXME handle dilation of avg pool
elif cfg.stride_mode == '1x1':
# NOTE I don't like this option described in paper, 1x1 w/ stride throws info away
stride_1, dilation_2 = stride, dilation[1]
else:
stride_2, dilation_2 = stride, dilation[0]
self.pre_norm = norm_act_layer(in_chs, apply_act=cfg.pre_norm_act)
if stride_pool > 1:
self.down = Downsample2d(in_chs, in_chs, pool_type=cfg.downsample_pool_type, padding=cfg.padding)
else:
self.down = nn.Identity()
self.conv1_1x1 = create_conv2d(in_chs, mid_chs, 1, stride=stride_1)
self.norm1 = norm_act_layer(mid_chs)
self.conv2_kxk = create_conv2d(
mid_chs, mid_chs, cfg.kernel_size,
stride=stride_2, dilation=dilation_2, groups=groups, padding=cfg.padding)
attn_kwargs = {}
if isinstance(cfg.attn_layer, str):
if cfg.attn_layer == 'se' or cfg.attn_layer == 'eca':
attn_kwargs['act_layer'] = cfg.attn_act_layer
attn_kwargs['rd_channels'] = int(cfg.attn_ratio * (out_chs if cfg.expand_output else mid_chs))
# two different orderings for SE and norm2 (due to some weights and trials using SE before norm2)
if cfg.attn_early:
self.se_early = create_attn(cfg.attn_layer, mid_chs, **attn_kwargs)
self.norm2 = norm_act_layer(mid_chs)
self.se = None
else:
self.se_early = None
self.norm2 = norm_act_layer(mid_chs)
self.se = create_attn(cfg.attn_layer, mid_chs, **attn_kwargs)
self.conv3_1x1 = create_conv2d(mid_chs, out_chs, 1, bias=cfg.output_bias)
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def init_weights(self, scheme=''):
named_apply(partial(_init_conv, scheme=scheme), self)
def forward(self, x):
shortcut = self.shortcut(x)
x = self.pre_norm(x)
x = self.down(x)
# 1x1 expansion conv & norm-act
x = self.conv1_1x1(x)
x = self.norm1(x)
# depthwise / grouped 3x3 conv w/ SE (or other) channel attention & norm-act
x = self.conv2_kxk(x)
if self.se_early is not None:
x = self.se_early(x)
x = self.norm2(x)
if self.se is not None:
x = self.se(x)
# 1x1 linear projection to output width
x = self.conv3_1x1(x)
x = self.drop_path(x) + shortcut
return x
class ConvNeXtBlock(nn.Module):
""" ConvNeXt Block
"""
def __init__(
self,
in_chs: int,
out_chs: Optional[int] = None,
kernel_size: int = 7,
stride: int = 1,
dilation: Tuple[int, int] = (1, 1),
cfg: MaxxVitConvCfg = MaxxVitConvCfg(),
conv_mlp: bool = True,
drop_path: float = 0.
):
super().__init__()
out_chs = out_chs or in_chs
act_layer = get_act_layer(cfg.act_layer)
if conv_mlp:
norm_layer = partial(get_norm_layer(cfg.norm_layer), eps=cfg.norm_eps)
mlp_layer = ConvMlp
else:
assert 'layernorm' in cfg.norm_layer
norm_layer = LayerNorm
mlp_layer = Mlp
self.use_conv_mlp = conv_mlp
if stride == 2:
self.shortcut = Downsample2d(in_chs, out_chs)
elif in_chs != out_chs:
self.shortcut = nn.Conv2d(in_chs, out_chs, kernel_size=1, bias=cfg.output_bias)
else:
self.shortcut = nn.Identity()
assert cfg.stride_mode in ('pool', 'dw')
stride_pool, stride_dw = 1, 1
# FIXME handle dilation?
if cfg.stride_mode == 'pool':
stride_pool = stride
else:
stride_dw = stride
if stride_pool == 2:
self.down = Downsample2d(in_chs, in_chs, pool_type=cfg.downsample_pool_type)
else:
self.down = nn.Identity()
self.conv_dw = create_conv2d(
in_chs, out_chs, kernel_size=kernel_size, stride=stride_dw, dilation=dilation[1],
depthwise=True, bias=cfg.output_bias)
self.norm = norm_layer(out_chs)
self.mlp = mlp_layer(out_chs, int(cfg.expand_ratio * out_chs), bias=cfg.output_bias, act_layer=act_layer)
if conv_mlp:
self.ls = LayerScale2d(out_chs, cfg.init_values) if cfg.init_values else nn.Identity()
else:
self.ls = LayerScale(out_chs, cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def forward(self, x):
shortcut = self.shortcut(x)
x = self.down(x)
x = self.conv_dw(x)
if self.use_conv_mlp:
x = self.norm(x)
x = self.mlp(x)
x = self.ls(x)
else:
x = x.permute(0, 2, 3, 1)
x = self.norm(x)
x = self.mlp(x)
x = self.ls(x)
x = x.permute(0, 3, 1, 2)
x = self.drop_path(x) + shortcut
return x
def window_partition(x, window_size: List[int]):
B, H, W, C = x.shape
_assert(H % window_size[0] == 0, f'height ({H}) must be divisible by window ({window_size[0]})')
_assert(W % window_size[1] == 0, f'width ({W}) must be divisible by window ({window_size[1]})')
x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
return windows
@register_notrace_function # reason: int argument is a Proxy
def window_reverse(windows, window_size: List[int], img_size: List[int]):
H, W = img_size
C = windows.shape[-1]
x = windows.view(-1, H // window_size[0], W // window_size[1], window_size[0], window_size[1], C)
x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, H, W, C)
return x
def grid_partition(x, grid_size: List[int]):
B, H, W, C = x.shape
_assert(H % grid_size[0] == 0, f'height {H} must be divisible by grid {grid_size[0]}')
_assert(W % grid_size[1] == 0, f'width {W} must be divisible by grid {grid_size[1]}')
x = x.view(B, grid_size[0], H // grid_size[0], grid_size[1], W // grid_size[1], C)
windows = x.permute(0, 2, 4, 1, 3, 5).contiguous().view(-1, grid_size[0], grid_size[1], C)
return windows
@register_notrace_function # reason: int argument is a Proxy
def grid_reverse(windows, grid_size: List[int], img_size: List[int]):
H, W = img_size
C = windows.shape[-1]
x = windows.view(-1, H // grid_size[0], W // grid_size[1], grid_size[0], grid_size[1], C)
x = x.permute(0, 3, 1, 4, 2, 5).contiguous().view(-1, H, W, C)
return x
def get_rel_pos_cls(cfg: MaxxVitTransformerCfg, window_size):
rel_pos_cls = None
if cfg.rel_pos_type == 'mlp':
rel_pos_cls = partial(RelPosMlp, window_size=window_size, hidden_dim=cfg.rel_pos_dim)
elif cfg.rel_pos_type == 'bias':
rel_pos_cls = partial(RelPosBias, window_size=window_size)
elif cfg.rel_pos_type == 'bias_tf':
rel_pos_cls = partial(RelPosBiasTf, window_size=window_size)
return rel_pos_cls
class PartitionAttentionCl(nn.Module):
""" Grid or Block partition + Attn + FFN.
NxC 'channels last' tensor layout.
"""
def __init__(
self,
dim: int,
partition_type: str = 'block',
cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path: float = 0.,
):
super().__init__()
norm_layer = partial(get_norm_layer(cfg.norm_layer_cl), eps=cfg.norm_eps) # NOTE this block is channels-last
act_layer = get_act_layer(cfg.act_layer)
self.partition_block = partition_type == 'block'
self.partition_size = to_2tuple(cfg.window_size if self.partition_block else cfg.grid_size)
rel_pos_cls = get_rel_pos_cls(cfg, self.partition_size)
self.norm1 = norm_layer(dim)
self.attn = AttentionCl(
dim,
dim,
dim_head=cfg.dim_head,
bias=cfg.attn_bias,
head_first=cfg.head_first,
rel_pos_cls=rel_pos_cls,
attn_drop=cfg.attn_drop,
proj_drop=cfg.proj_drop,
)
self.ls1 = LayerScale(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
self.mlp = Mlp(
in_features=dim,
hidden_features=int(dim * cfg.expand_ratio),
act_layer=act_layer,
drop=cfg.proj_drop)
self.ls2 = LayerScale(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def _partition_attn(self, x):
img_size = x.shape[1:3]
if self.partition_block:
partitioned = window_partition(x, self.partition_size)
else:
partitioned = grid_partition(x, self.partition_size)
partitioned = self.attn(partitioned)
if self.partition_block:
x = window_reverse(partitioned, self.partition_size, img_size)
else:
x = grid_reverse(partitioned, self.partition_size, img_size)
return x
def forward(self, x):
x = x + self.drop_path1(self.ls1(self._partition_attn(self.norm1(x))))
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
return x
class ParallelPartitionAttention(nn.Module):
""" Experimental. Grid and Block partition + single FFN
NxC tensor layout.
"""
def __init__(
self,
dim: int,
cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path: float = 0.,
):
super().__init__()
assert dim % 2 == 0
norm_layer = partial(get_norm_layer(cfg.norm_layer_cl), eps=cfg.norm_eps) # NOTE this block is channels-last
act_layer = get_act_layer(cfg.act_layer)
assert cfg.window_size == cfg.grid_size
self.partition_size = to_2tuple(cfg.window_size)
rel_pos_cls = get_rel_pos_cls(cfg, self.partition_size)
self.norm1 = norm_layer(dim)
self.attn_block = AttentionCl(
dim,
dim // 2,
dim_head=cfg.dim_head,
bias=cfg.attn_bias,
head_first=cfg.head_first,
rel_pos_cls=rel_pos_cls,
attn_drop=cfg.attn_drop,
proj_drop=cfg.proj_drop,
)
self.attn_grid = AttentionCl(
dim,
dim // 2,
dim_head=cfg.dim_head,
bias=cfg.attn_bias,
head_first=cfg.head_first,
rel_pos_cls=rel_pos_cls,
attn_drop=cfg.attn_drop,
proj_drop=cfg.proj_drop,
)
self.ls1 = LayerScale(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
self.mlp = Mlp(
in_features=dim,
hidden_features=int(dim * cfg.expand_ratio),
out_features=dim,
act_layer=act_layer,
drop=cfg.proj_drop)
self.ls2 = LayerScale(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def _partition_attn(self, x):
img_size = x.shape[1:3]
partitioned_block = window_partition(x, self.partition_size)
partitioned_block = self.attn_block(partitioned_block)
x_window = window_reverse(partitioned_block, self.partition_size, img_size)
partitioned_grid = grid_partition(x, self.partition_size)
partitioned_grid = self.attn_grid(partitioned_grid)
x_grid = grid_reverse(partitioned_grid, self.partition_size, img_size)
return torch.cat([x_window, x_grid], dim=-1)
def forward(self, x):
x = x + self.drop_path1(self.ls1(self._partition_attn(self.norm1(x))))
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
return x
def window_partition_nchw(x, window_size: List[int]):
B, C, H, W = x.shape
_assert(H % window_size[0] == 0, f'height ({H}) must be divisible by window ({window_size[0]})')
_assert(W % window_size[1] == 0, f'width ({W}) must be divisible by window ({window_size[1]})')
x = x.view(B, C, H // window_size[0], window_size[0], W // window_size[1], window_size[1])
windows = x.permute(0, 2, 4, 1, 3, 5).contiguous().view(-1, C, window_size[0], window_size[1])
return windows
@register_notrace_function # reason: int argument is a Proxy
def window_reverse_nchw(windows, window_size: List[int], img_size: List[int]):
H, W = img_size
C = windows.shape[1]
x = windows.view(-1, H // window_size[0], W // window_size[1], C, window_size[0], window_size[1])
x = x.permute(0, 3, 1, 4, 2, 5).contiguous().view(-1, C, H, W)
return x
def grid_partition_nchw(x, grid_size: List[int]):
B, C, H, W = x.shape
_assert(H % grid_size[0] == 0, f'height {H} must be divisible by grid {grid_size[0]}')
_assert(W % grid_size[1] == 0, f'width {W} must be divisible by grid {grid_size[1]}')
x = x.view(B, C, grid_size[0], H // grid_size[0], grid_size[1], W // grid_size[1])
windows = x.permute(0, 3, 5, 1, 2, 4).contiguous().view(-1, C, grid_size[0], grid_size[1])
return windows
@register_notrace_function # reason: int argument is a Proxy
def grid_reverse_nchw(windows, grid_size: List[int], img_size: List[int]):
H, W = img_size
C = windows.shape[1]
x = windows.view(-1, H // grid_size[0], W // grid_size[1], C, grid_size[0], grid_size[1])
x = x.permute(0, 3, 4, 1, 5, 2).contiguous().view(-1, C, H, W)
return x
class PartitionAttention2d(nn.Module):
""" Grid or Block partition + Attn + FFN
'2D' NCHW tensor layout.
"""
def __init__(
self,
dim: int,
partition_type: str = 'block',
cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path: float = 0.,
):
super().__init__()
norm_layer = partial(get_norm_layer(cfg.norm_layer), eps=cfg.norm_eps) # NOTE this block is channels-last
act_layer = get_act_layer(cfg.act_layer)
self.partition_block = partition_type == 'block'
self.partition_size = to_2tuple(cfg.window_size if self.partition_block else cfg.grid_size)
rel_pos_cls = get_rel_pos_cls(cfg, self.partition_size)
self.norm1 = norm_layer(dim)
self.attn = Attention2d(
dim,
dim,
dim_head=cfg.dim_head,
bias=cfg.attn_bias,
head_first=cfg.head_first,
rel_pos_cls=rel_pos_cls,
attn_drop=cfg.attn_drop,
proj_drop=cfg.proj_drop,
)
self.ls1 = LayerScale2d(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
self.norm2 = norm_layer(dim)
self.mlp = ConvMlp(
in_features=dim,
hidden_features=int(dim * cfg.expand_ratio),
act_layer=act_layer,
drop=cfg.proj_drop)
self.ls2 = LayerScale2d(dim, init_values=cfg.init_values) if cfg.init_values else nn.Identity()
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
def _partition_attn(self, x):
img_size = x.shape[-2:]
if self.partition_block:
partitioned = window_partition_nchw(x, self.partition_size)
else:
partitioned = grid_partition_nchw(x, self.partition_size)
partitioned = self.attn(partitioned)
if self.partition_block:
x = window_reverse_nchw(partitioned, self.partition_size, img_size)
else:
x = grid_reverse_nchw(partitioned, self.partition_size, img_size)
return x
def forward(self, x):
x = x + self.drop_path1(self.ls1(self._partition_attn(self.norm1(x))))
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
return x
class MaxxVitBlock(nn.Module):
""" MaxVit conv, window partition + FFN , grid partition + FFN
"""
def __init__(
self,
dim: int,
dim_out: int,
stride: int = 1,
conv_cfg: MaxxVitConvCfg = MaxxVitConvCfg(),
transformer_cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path: float = 0.,
):
super().__init__()
self.nchw_attn = transformer_cfg.use_nchw_attn
conv_cls = ConvNeXtBlock if conv_cfg.block_type == 'convnext' else MbConvBlock
self.conv = conv_cls(dim, dim_out, stride=stride, cfg=conv_cfg, drop_path=drop_path)
attn_kwargs = dict(dim=dim_out, cfg=transformer_cfg, drop_path=drop_path)
partition_layer = PartitionAttention2d if self.nchw_attn else PartitionAttentionCl
self.attn_block = None if transformer_cfg.no_block_attn else partition_layer(**attn_kwargs)
self.attn_grid = partition_layer(partition_type='grid', **attn_kwargs)
def init_weights(self, scheme=''):
if self.attn_block is not None:
named_apply(partial(_init_transformer, scheme=scheme), self.attn_block)
named_apply(partial(_init_transformer, scheme=scheme), self.attn_grid)
named_apply(partial(_init_conv, scheme=scheme), self.conv)
def forward(self, x):
# NCHW format
x = self.conv(x)
if not self.nchw_attn:
x = x.permute(0, 2, 3, 1) # to NHWC (channels-last)
if self.attn_block is not None:
x = self.attn_block(x)
x = self.attn_grid(x)
if not self.nchw_attn:
x = x.permute(0, 3, 1, 2) # back to NCHW
return x
class ParallelMaxxVitBlock(nn.Module):
""" MaxVit block with parallel cat(window + grid), one FF
Experimental timm block.
"""
def __init__(
self,
dim,
dim_out,
stride=1,
num_conv=2,
conv_cfg: MaxxVitConvCfg = MaxxVitConvCfg(),
transformer_cfg: MaxxVitTransformerCfg = MaxxVitTransformerCfg(),
drop_path=0.,
):
super().__init__()
conv_cls = ConvNeXtBlock if conv_cfg.block_type == 'convnext' else MbConvBlock
if num_conv > 1:
convs = [conv_cls(dim, dim_out, stride=stride, cfg=conv_cfg, drop_path=drop_path)]
convs += [conv_cls(dim_out, dim_out, cfg=conv_cfg, drop_path=drop_path)] * (num_conv - 1)
self.conv = nn.Sequential(*convs)
else:
self.conv = conv_cls(dim, dim_out, stride=stride, cfg=conv_cfg, drop_path=drop_path)
self.attn = ParallelPartitionAttention(dim=dim_out, cfg=transformer_cfg, drop_path=drop_path)
def init_weights(self, scheme=''):
named_apply(partial(_init_transformer, scheme=scheme), self.attn)
named_apply(partial(_init_conv, scheme=scheme), self.conv)
def forward(self, x):
x = self.conv(x)
x = x.permute(0, 2, 3, 1)
x = self.attn(x)
x = x.permute(0, 3, 1, 2)
return x
class MaxxVitStage(nn.Module):