File size: 34,953 Bytes
afbb92a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
'''
Source code from OPEN_CLIP project.
https://github.com/mlfoundations/open_clip/blob/main/LICENSE
'''

from collections import OrderedDict
import math
from typing import Callable, Optional, Sequence, Tuple
from functools import partial

import torch
from torch import nn
from torch.nn import functional as F
from torch.utils.checkpoint import checkpoint

from itertools import repeat
import collections.abc

# From PyTorch internals
def _ntuple(n):
    def parse(x):
        if isinstance(x, collections.abc.Iterable):
            return x
        return tuple(repeat(x, n))
    return parse

to_1tuple = _ntuple(1)
to_2tuple = _ntuple(2)
to_3tuple = _ntuple(3)
to_4tuple = _ntuple(4)
to_ntuple = lambda n, x: _ntuple(n)(x)

class LayerNormFp32(nn.LayerNorm):
    """Subclass torch's LayerNorm to handle fp16 (by casting to float32 and back)."""

    def forward(self, x: torch.Tensor):
        orig_type = x.dtype
        x = F.layer_norm(x.to(torch.float32), self.normalized_shape, self.weight, self.bias, self.eps)

        #x = F.layer_norm(x.to(torch.bfloat16), self.normalized_shape, self.weight, self.bias, self.eps)
        return x.to(orig_type)


class LayerNorm(nn.LayerNorm):
    """Subclass torch's LayerNorm (with cast back to input dtype)."""

    def forward(self, x: torch.Tensor):
        orig_type = x.dtype
        x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)
        return x.to(orig_type)


class QuickGELU(nn.Module):
    # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory
    def forward(self, x: torch.Tensor):
        return x * torch.sigmoid(1.702 * 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):
        return x.mul_(self.gamma) if self.inplace else x * self.gamma


class PatchDropout(nn.Module):
    """
    https://arxiv.org/abs/2212.00794
    """

    def __init__(self, prob, exclude_first_token=True):
        super().__init__()
        assert 0 <= prob < 1.
        self.prob = prob
        self.exclude_first_token = exclude_first_token  # exclude CLS token

    def forward(self, x):
        if not self.training or self.prob == 0.:
            return x

        if self.exclude_first_token:
            cls_tokens, x = x[:, :1], x[:, 1:]
        else:
            cls_tokens = torch.jit.annotate(torch.Tensor, x[:, :1])

        batch = x.size()[0]
        num_tokens = x.size()[1]

        batch_indices = torch.arange(batch)
        batch_indices = batch_indices[..., None]

        keep_prob = 1 - self.prob
        num_patches_keep = max(1, int(num_tokens * keep_prob))

        rand = torch.randn(batch, num_tokens)
        patch_indices_keep = rand.topk(num_patches_keep, dim=-1).indices

        x = x[batch_indices, patch_indices_keep]

        if self.exclude_first_token:
            x = torch.cat((cls_tokens, x), dim=1)

        return x


class Attention(nn.Module):
    def __init__(
            self,
            dim,
            num_heads=8,
            qkv_bias=True,
            scaled_cosine=False,
            scale_heads=False,
            logit_scale_max=math.log(1. / 0.01),
            attn_drop=0.,
            proj_drop=0.
    ):
        super().__init__()
        self.scaled_cosine = scaled_cosine
        self.scale_heads = scale_heads
        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.scale = self.head_dim ** -0.5
        self.logit_scale_max = logit_scale_max

        # keeping in_proj in this form (instead of nn.Linear) to match weight scheme of original
        self.in_proj_weight = nn.Parameter(torch.randn((dim * 3, dim)) * self.scale)
        if qkv_bias:
            self.in_proj_bias = nn.Parameter(torch.zeros(dim * 3))
        else:
            self.in_proj_bias = None

        if self.scaled_cosine:
            self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
        else:
            self.logit_scale = None
        self.attn_drop = nn.Dropout(attn_drop)
        if self.scale_heads:
            self.head_scale = nn.Parameter(torch.ones((num_heads, 1, 1)))
        else:
            self.head_scale = None
        self.out_proj = nn.Linear(dim, dim)
        self.out_drop = nn.Dropout(proj_drop)

    def forward(self, x, attn_mask: Optional[torch.Tensor] = None):
        L, N, C = x.shape
        q, k, v = F.linear(x, self.in_proj_weight, self.in_proj_bias).chunk(3, dim=-1)
        q = q.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
        k = k.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)
        v = v.contiguous().view(L, N * self.num_heads, -1).transpose(0, 1)

        if self.logit_scale is not None:
            attn = torch.bmm(F.normalize(q, dim=-1), F.normalize(k, dim=-1).transpose(-1, -2))
            logit_scale = torch.clamp(self.logit_scale, max=self.logit_scale_max).exp()
            attn = attn.view(N, self.num_heads, L, L) * logit_scale
            attn = attn.view(-1, L, L)
        else:
            q = q * self.scale
            attn = torch.bmm(q, k.transpose(-1, -2))

        if attn_mask is not None:
            if attn_mask.dtype == torch.bool:
                new_attn_mask = torch.zeros_like(attn_mask, dtype=q.dtype)
                new_attn_mask.masked_fill_(attn_mask, float("-inf"))
                attn_mask = new_attn_mask
            attn += attn_mask

        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = torch.bmm(attn, v)
        if self.head_scale is not None:
            x = x.view(N, self.num_heads, L, C) * self.head_scale
            x = x.view(-1, L, C)
        x = x.transpose(0, 1).reshape(L, N, C)
        x = self.out_proj(x)
        x = self.out_drop(x)
        return x


class AttentionalPooler(nn.Module):
    def __init__(
            self,
            d_model: int,
            context_dim: int,
            n_head: int = 8,
            n_queries: int = 256,
            norm_layer: Callable = LayerNorm
    ):
        super().__init__()
        self.query = nn.Parameter(torch.randn(n_queries, d_model))
        self.attn = nn.MultiheadAttention(d_model, n_head, kdim=context_dim, vdim=context_dim)
        self.ln_q = norm_layer(d_model)
        self.ln_k = norm_layer(context_dim)

    def forward(self, x: torch.Tensor):
        x = self.ln_k(x).permute(1, 0, 2)  # NLD -> LND
        N = x.shape[1]
        q = self.ln_q(self.query)
        out = self.attn(q.unsqueeze(1).expand(-1, N, -1), x, x, need_weights=False)[0]
        return out.permute(1, 0, 2)  # LND -> NLD


class ResidualAttentionBlock(nn.Module):
    def __init__(
            self,
            d_model: int,
            n_head: int,
            mlp_ratio: float = 4.0,
            ls_init_value: float = None,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
            is_cross_attention: bool = False,
    ):
        super().__init__()

        self.ln_1 = norm_layer(d_model)
        self.attn = nn.MultiheadAttention(d_model, n_head)
        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()
        if is_cross_attention:
            self.ln_1_kv = norm_layer(d_model)

        self.ln_2 = norm_layer(d_model)
        mlp_width = int(d_model * mlp_ratio)
        self.mlp = nn.Sequential(OrderedDict([
            ("c_fc", nn.Linear(d_model, mlp_width)),
            ("gelu", act_layer()),
            ("c_proj", nn.Linear(mlp_width, d_model))
        ]))
        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()

    def attention(
            self,
            q_x: torch.Tensor,
            k_x: Optional[torch.Tensor] = None,
            v_x: Optional[torch.Tensor] = None,
            attn_mask: Optional[torch.Tensor] = None,
    ):
        k_x = k_x if k_x is not None else q_x
        v_x = v_x if v_x is not None else q_x

        attn_mask = attn_mask.to(q_x.dtype) if attn_mask is not None else None
        return self.attn(
            q_x, k_x, v_x, need_weights=False, attn_mask=attn_mask
        )[0]

    def forward(
            self,
            q_x: torch.Tensor,
            k_x: Optional[torch.Tensor] = None,
            v_x: Optional[torch.Tensor] = None,
            attn_mask: Optional[torch.Tensor] = None,
    ):
        k_x = self.ln_1_kv(k_x) if hasattr(self, "ln_1_kv") and k_x is not None else None
        v_x = self.ln_1_kv(v_x) if hasattr(self, "ln_1_kv") and v_x is not None else None

        x = q_x + self.ls_1(self.attention(q_x=self.ln_1(q_x), k_x=k_x, v_x=v_x, attn_mask=attn_mask))
        x = x + self.ls_2(self.mlp(self.ln_2(x)))
        return x


class CustomResidualAttentionBlock(nn.Module):
    def __init__(
            self,
            d_model: int,
            n_head: int,
            mlp_ratio: float = 4.0,
            ls_init_value: float = None,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
            scale_cosine_attn: bool = False,
            scale_heads: bool = False,
            scale_attn: bool = False,
            scale_fc: bool = False,
    ):
        super().__init__()

        self.ln_1 = norm_layer(d_model)
        self.attn = Attention(
            d_model, n_head,
            scaled_cosine=scale_cosine_attn,
            scale_heads=scale_heads,
        )
        self.ln_attn = norm_layer(d_model) if scale_attn else nn.Identity()
        self.ls_1 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()

        self.ln_2 = norm_layer(d_model)
        mlp_width = int(d_model * mlp_ratio)
        self.mlp = nn.Sequential(OrderedDict([
            ("c_fc", nn.Linear(d_model, mlp_width)),
            ("gelu", act_layer()),
            ('ln', norm_layer(mlp_width) if scale_fc else nn.Identity()),
            ("c_proj", nn.Linear(mlp_width, d_model))
        ]))
        self.ls_2 = LayerScale(d_model, ls_init_value) if ls_init_value is not None else nn.Identity()

    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
        x = x + self.ls_1(self.ln_attn(self.attn(self.ln_1(x), attn_mask=attn_mask)))
        x = x + self.ls_2(self.mlp(self.ln_2(x)))
        return x


def _expand_token(token, batch_size: int):
    return token.view(1, 1, -1).expand(batch_size, -1, -1)


class Transformer(nn.Module):
    def __init__(
            self,
            width: int,
            layers: int,
            heads: int,
            mlp_ratio: float = 4.0,
            ls_init_value: float = None,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
    ):
        super().__init__()
        self.width = width
        self.layers = layers
        self.grad_checkpointing = False

        self.resblocks = nn.ModuleList([
            ResidualAttentionBlock(
                width, heads, mlp_ratio, ls_init_value=ls_init_value, act_layer=act_layer, norm_layer=norm_layer)
            for _ in range(layers)
        ])

    def get_cast_dtype(self) -> torch.dtype:
        if hasattr(self.resblocks[0].mlp.c_fc, 'int8_original_dtype'):
            return self.resblocks[0].mlp.c_fc.int8_original_dtype
        return self.resblocks[0].mlp.c_fc.weight.dtype

    def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):
        for r in self.resblocks:
            if self.grad_checkpointing and not torch.jit.is_scripting():
                # TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
                x = checkpoint(r, x, None, None, attn_mask)
            else:
                x = r(x, attn_mask=attn_mask)
        return x


class VisionTransformer(nn.Module):
    output_tokens: torch.jit.Final[bool]

    def __init__(
            self,
            in_channels:int,
            image_size: int,
            patch_size: int,
            width: int,
            layers: int,
            heads: int,
            mlp_ratio: float,
            ls_init_value: float = None,
            attentional_pool: bool = False,
            attn_pooler_queries: int = 256,
            attn_pooler_heads: int = 8,
            output_dim: int = 512,
            patch_dropout: float = 0.,
            no_ln_pre: bool = False,
            pos_embed_type: str = 'learnable',
            pool_type: str = 'tok',
            final_ln_after_pool: bool = False,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
            output_tokens: bool = False,
    ):
        super().__init__()
        assert pool_type in ('tok', 'avg', 'none')
        self.output_tokens = output_tokens
        image_height, image_width = self.image_size = to_2tuple(image_size)
        patch_height, patch_width = self.patch_size = to_2tuple(patch_size)
        self.grid_size = (image_height // patch_height, image_width // patch_width)
        self.final_ln_after_pool = final_ln_after_pool  # currently ignored w/ attn pool enabled
        self.output_dim = output_dim

        self.conv1 = nn.Conv2d(in_channels=in_channels, out_channels=width, kernel_size=patch_size, stride=patch_size, bias=False)

        # class embeddings and positional embeddings
        scale = width ** -0.5
        self.class_embedding = nn.Parameter(scale * torch.randn(width))
        if pos_embed_type == 'learnable':
            self.positional_embedding = nn.Parameter(
                scale * torch.randn(self.grid_size[0] * self.grid_size[1] + 1, width))
        elif pos_embed_type == 'sin_cos_2d':
            # fixed sin-cos embedding
            assert self.grid_size[0] == self.grid_size[1], \
                'currently sin cos 2d pos embedding only supports square input'
            self.positional_embedding = nn.Parameter(
                torch.zeros(self.grid_size[0] * self.grid_size[1] + 1, width), requires_grad=False)
            pos_embed_type = get_2d_sincos_pos_embed(width, self.grid_size[0], cls_token=True)
            self.positional_embedding.data.copy_(torch.from_numpy(pos_embed_type).float())
        else:
            raise ValueError

        # setting a patch_dropout of 0. would mean it is disabled and this function would be the identity fn
        self.patch_dropout = PatchDropout(patch_dropout) if patch_dropout > 0. else nn.Identity()

        self.ln_pre = nn.Identity() if no_ln_pre else norm_layer(width)
        self.transformer = Transformer(
            width,
            layers,
            heads,
            mlp_ratio,
            ls_init_value=ls_init_value,
            act_layer=act_layer,
            norm_layer=norm_layer,
        )

        if attentional_pool:
            if isinstance(attentional_pool, str):
                self.attn_pool_type = attentional_pool
                self.pool_type = 'none'
                if attentional_pool in ('parallel', 'cascade'):
                    self.attn_pool = AttentionalPooler(
                        output_dim,
                        width,
                        n_head=attn_pooler_heads,
                        n_queries=attn_pooler_queries,
                    )
                    self.attn_pool_contrastive = AttentionalPooler(
                        output_dim,
                        width,
                        n_head=attn_pooler_heads,
                        n_queries=1,
                    )
                else:
                    assert False
            else:
                self.attn_pool_type = ''
                self.pool_type = pool_type
                self.attn_pool = AttentionalPooler(
                    output_dim,
                    width,
                    n_head=attn_pooler_heads,
                    n_queries=attn_pooler_queries,
                )
                self.attn_pool_contrastive = None
            pool_dim = output_dim
        else:
            self.attn_pool = None
            pool_dim = width
            self.pool_type = pool_type

        self.ln_post = norm_layer(pool_dim)
        self.proj = nn.Parameter(scale * torch.randn(pool_dim, output_dim))

        self.init_parameters()

    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
        for param in self.parameters():
            param.requires_grad = False

        if unlocked_groups != 0:
            groups = [
                [
                    self.conv1,
                    self.class_embedding,
                    self.positional_embedding,
                    self.ln_pre,
                ],
                *self.transformer.resblocks[:-1],
                [
                    self.transformer.resblocks[-1],
                    self.ln_post,
                ],
                self.proj,
            ]

            def _unlock(x):
                if isinstance(x, Sequence):
                    for g in x:
                        _unlock(g)
                else:
                    if isinstance(x, torch.nn.Parameter):
                        x.requires_grad = True
                    else:
                        for p in x.parameters():
                            p.requires_grad = True

            _unlock(groups[-unlocked_groups:])

    def init_parameters(self):
        # FIXME OpenAI CLIP did not define an init for the VisualTransformer
        # TODO experiment if default PyTorch init, below, or alternate init is best.

        # nn.init.normal_(self.class_embedding, std=self.scale)
        # nn.init.normal_(self.positional_embedding, std=self.scale)
        #
        # proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
        # attn_std = self.transformer.width ** -0.5
        # fc_std = (2 * self.transformer.width) ** -0.5
        # for block in self.transformer.resblocks:
        #     nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
        #     nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
        #     nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
        #     nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
        #
        # if self.text_projection is not None:
        #     nn.init.normal_(self.text_projection, std=self.scale)
        pass

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.transformer.grad_checkpointing = enable

    def _global_pool(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
        if self.pool_type == 'avg':
            pooled, tokens = x[:, 1:].mean(dim=1), x[:, 1:]
        elif self.pool_type == 'tok':
            pooled, tokens = x[:, 0], x[:, 1:]
        else:
            pooled = tokens = x

        return pooled, tokens

    def forward(self, x: torch.Tensor):
        x = self.conv1(x)  # shape = [*, width, grid, grid]
        x = x.reshape(x.shape[0], x.shape[1], -1)  # shape = [*, width, grid ** 2]
        x = x.permute(0, 2, 1)  # shape = [*, grid ** 2, width]

        # class embeddings and positional embeddings
        x = torch.cat([_expand_token(self.class_embedding, x.shape[0]).to(x.dtype), x], dim=1)
        # shape = [*, grid ** 2 + 1, width]
        x = x + self.positional_embedding.to(x.dtype)

        x = self.patch_dropout(x)
        x = self.ln_pre(x)

        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x)
        x = x.permute(1, 0, 2)  # LND -> NLD

        if self.attn_pool is not None:
            if self.attn_pool_contrastive is not None:
                # This is untested, WIP pooling that should match paper
                x = self.ln_post(x)  # TBD LN first or separate one after each pool?
                tokens = self.attn_pool(x)
                if self.attn_pool_type == 'parallel':
                    pooled = self.attn_pool_contrastive(x)
                else:
                    assert self.attn_pool_type == 'cascade'
                    pooled = self.attn_pool_contrastive(tokens)
            else:
                # this is the original OpenCLIP CoCa setup, does not match paper
                x = self.attn_pool(x)
                x = self.ln_post(x)
                pooled, tokens = self._global_pool(x)
        elif self.final_ln_after_pool:
            pooled, tokens = self._global_pool(x)
            pooled = self.ln_post(pooled)
        else:
            x = self.ln_post(x)
            pooled, tokens = self._global_pool(x)

        if self.proj is not None:
            pooled = pooled @ self.proj

        if self.output_tokens:
            return pooled, tokens

        return pooled


def text_global_pool(x, text: Optional[torch.Tensor] = None, pool_type: str = 'argmax'):
    if pool_type == 'first':
        pooled, tokens = x[:, 0], x[:, 1:]
    elif pool_type == 'last':
        pooled, tokens = x[:, -1], x[:, :-1]
    elif pool_type == 'argmax':
        # take features from the eot embedding (eot_token is the highest number in each sequence)
        assert text is not None
        pooled, tokens = x[torch.arange(x.shape[0]), text.argmax(dim=-1)], x
    else:
        pooled = tokens = x

    return pooled, tokens


class TextTransformer(nn.Module):
    output_tokens: torch.jit.Final[bool]

    def __init__(
            self,
            context_length: int = 77,
            vocab_size: int = 49408,
            width: int = 512,
            heads: int = 8,
            layers: int = 12,
            mlp_ratio: float = 4.0,
            ls_init_value: float = None,
            output_dim: int = 512,
            embed_cls: bool = False,
            no_causal_mask: bool = False,
            pad_id: int = 0,
            pool_type: str = 'argmax',
            proj_bias: bool = False,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
            output_tokens: bool = False,
    ):
        super().__init__()
        assert pool_type in ('first', 'last', 'argmax', 'none')
        self.output_tokens = output_tokens
        self.num_pos = self.context_length = context_length
        self.vocab_size = vocab_size
        self.width = width
        self.output_dim = output_dim
        self.heads = heads
        self.pad_id = pad_id
        self.pool_type = pool_type

        self.token_embedding = nn.Embedding(vocab_size, width)
        if embed_cls:
            self.cls_emb = nn.Parameter(torch.empty(width))
            self.num_pos += 1
        else:
            self.cls_emb = None
        self.positional_embedding = nn.Parameter(torch.empty(self.num_pos, width))
        self.transformer = Transformer(
            width=width,
            layers=layers,
            heads=heads,
            mlp_ratio=mlp_ratio,
            ls_init_value=ls_init_value,
            act_layer=act_layer,
            norm_layer=norm_layer,
        )
        self.ln_final = norm_layer(width)

        if no_causal_mask:
            self.attn_mask = None
        else:
            self.register_buffer('attn_mask', self.build_causal_mask(), persistent=False)

        if proj_bias:
            self.text_projection = nn.Linear(width, output_dim)
        else:
            self.text_projection = nn.Parameter(torch.empty(width, output_dim))

        self.init_parameters()

    def init_parameters(self):
        nn.init.normal_(self.token_embedding.weight, std=0.02)
        nn.init.normal_(self.positional_embedding, std=0.01)
        if self.cls_emb is not None:
            nn.init.normal_(self.cls_emb, std=0.01)

        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
        attn_std = self.transformer.width ** -0.5
        fc_std = (2 * self.transformer.width) ** -0.5
        for block in self.transformer.resblocks:
            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)

        if self.text_projection is not None:
            if isinstance(self.text_projection, nn.Linear):
                nn.init.normal_(self.text_projection.weight, std=self.transformer.width ** -0.5)
                if self.text_projection.bias is not None:
                    nn.init.zeros_(self.text_projection.bias)
            else:
                nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.transformer.grad_checkpointing = enable

    def build_causal_mask(self):
        # lazily create causal attention mask, with full attention between the tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(self.num_pos, self.num_pos)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    def build_cls_mask(self, text, cast_dtype: torch.dtype):
        cls_mask = (text != self.pad_id).unsqueeze(1)
        cls_mask = F.pad(cls_mask, (1, 0, cls_mask.shape[2], 0), value=True)
        additive_mask = torch.empty(cls_mask.shape, dtype=cast_dtype, device=cls_mask.device)
        additive_mask.fill_(0)
        additive_mask.masked_fill_(~cls_mask, float("-inf"))
        additive_mask = torch.repeat_interleave(additive_mask, self.heads, 0)
        return additive_mask

    def forward(self, text):
        cast_dtype = self.transformer.get_cast_dtype()
        seq_len = text.shape[1]

        x = self.token_embedding(text).to(cast_dtype)  # [batch_size, n_ctx, d_model]
        attn_mask = self.attn_mask
        if self.cls_emb is not None:
            seq_len += 1
            x = torch.cat([x, _expand_token(self.cls_emb, x.shape[0])], dim=1)
            cls_mask = self.build_cls_mask(text, cast_dtype)
            if attn_mask is not None:
                attn_mask = attn_mask[None, :seq_len, :seq_len] + cls_mask[:, :seq_len, :seq_len]

        x = x + self.positional_embedding[:seq_len].to(cast_dtype)
        x = x.permute(1, 0, 2)  # NLD -> LND
        x = self.transformer(x, attn_mask=attn_mask)
        x = x.permute(1, 0, 2)  # LND -> NLD

        # x.shape = [batch_size, n_ctx, transformer.width]
        if self.cls_emb is not None:
            # presence of appended cls embed (CoCa) overrides pool_type, always take last token
            pooled, tokens = text_global_pool(x, pool_type='last')
            pooled = self.ln_final(pooled)  # final LN applied after pooling in this case
        else:
            x = self.ln_final(x)
            pooled, tokens = text_global_pool(x, text, pool_type=self.pool_type)

        if self.text_projection is not None:
            if isinstance(self.text_projection, nn.Linear):
                pooled = self.text_projection(pooled)
            else:
                pooled = pooled @ self.text_projection

        if self.output_tokens:
            return pooled, tokens

        return pooled


class MultimodalTransformer(Transformer):
    def __init__(
            self,
            width: int,
            layers: int,
            heads: int,
            context_length: int = 77,
            mlp_ratio: float = 4.0,
            ls_init_value: float = None,
            act_layer: Callable = nn.GELU,
            norm_layer: Callable = LayerNorm,
            output_dim: int = 512,
    ):

        super().__init__(
            width=width,
            layers=layers,
            heads=heads,
            mlp_ratio=mlp_ratio,
            ls_init_value=ls_init_value,
            act_layer=act_layer,
            norm_layer=norm_layer,
        )
        self.context_length = context_length
        self.cross_attn = nn.ModuleList([
            ResidualAttentionBlock(
                width,
                heads,
                mlp_ratio,
                ls_init_value=ls_init_value,
                act_layer=act_layer,
                norm_layer=norm_layer,
                is_cross_attention=True,
            )
            for _ in range(layers)
        ])

        self.register_buffer('attn_mask', self.build_attention_mask(), persistent=False)

        self.ln_final = norm_layer(width)
        self.text_projection = nn.Parameter(torch.empty(width, output_dim))

    def init_parameters(self):
        proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)
        attn_std = self.transformer.width ** -0.5
        fc_std = (2 * self.transformer.width) ** -0.5
        for block in self.transformer.resblocks:
            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)
        for block in self.transformer.cross_attn:
            nn.init.normal_(block.attn.in_proj_weight, std=attn_std)
            nn.init.normal_(block.attn.out_proj.weight, std=proj_std)
            nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)
            nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)

        if self.text_projection is not None:
            nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)

    def build_attention_mask(self):
        # lazily create causal attention mask, with full attention between the tokens
        # pytorch uses additive attention mask; fill with -inf
        mask = torch.empty(self.context_length, self.context_length)
        mask.fill_(float("-inf"))
        mask.triu_(1)  # zero out the lower diagonal
        return mask

    def forward(self, image_embs, text_embs):
        text_embs = text_embs.permute(1, 0, 2)  # NLD -> LNDsq
        image_embs = image_embs.permute(1, 0, 2)  # NLD -> LND
        seq_len = text_embs.shape[0]

        for resblock, cross_attn in zip(self.resblocks, self.cross_attn):
            if self.grad_checkpointing and not torch.jit.is_scripting():
                # TODO: handle kwargs https://github.com/pytorch/pytorch/issues/79887#issuecomment-1161758372
                text_embs = checkpoint(resblock, text_embs, None, None, self.attn_mask[:seq_len, :seq_len])
                text_embs = checkpoint(cross_attn, text_embs, image_embs, image_embs, None)
            else:
                text_embs = resblock(text_embs, attn_mask=self.attn_mask[:seq_len, :seq_len])
                text_embs = cross_attn(text_embs, k_x=image_embs, v_x=image_embs)

        x = text_embs.permute(1, 0, 2)  # LND -> NLD
        x = self.ln_final(x)

        if self.text_projection is not None:
            x = x @ self.text_projection

        return x

    @torch.jit.ignore
    def set_grad_checkpointing(self, enable=True):
        self.grad_checkpointing = enable


# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.

# This source code is licensed under the license found in the
# LICENSE file in the root directory of this source tree.
# --------------------------------------------------------
# Position embedding utils
# --------------------------------------------------------

import numpy as np

import torch

# --------------------------------------------------------
# 2D sine-cosine position embedding
# References:
# Transformer: https://github.com/tensorflow/models/blob/master/official/nlp/transformer/model_utils.py
# MoCo v3: https://github.com/facebookresearch/moco-v3
# --------------------------------------------------------
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
    """
    grid_size: int of the grid height and width
    return:
    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid_h = np.arange(grid_size, dtype=np.float32)
    grid_w = np.arange(grid_size, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)

    grid = grid.reshape([2, 1, grid_size, grid_size])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token:
        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)

    emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position
    pos: a list of positions to be encoded: size (M,)
    out: (M, D)
    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=float)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    emb_sin = np.sin(out) # (M, D/2)
    emb_cos = np.cos(out) # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
    return emb


# --------------------------------------------------------
# Interpolate position embeddings for high-resolution
# References:
# DeiT: https://github.com/facebookresearch/deit
# --------------------------------------------------------
def interpolate_pos_embed(model, checkpoint_model):
    if 'pos_embed' in checkpoint_model:
        pos_embed_checkpoint = checkpoint_model['pos_embed']
        embedding_size = pos_embed_checkpoint.shape[-1]
        num_patches = model.patch_embed.num_patches
        num_extra_tokens = model.pos_embed.shape[-2] - num_patches
        # height (== width) for the checkpoint position embedding
        orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)
        # height (== width) for the new position embedding
        new_size = int(num_patches ** 0.5)
        # class_token and dist_token are kept unchanged
        if orig_size != new_size:
            print("Position interpolate from %dx%d to %dx%d" % (orig_size, orig_size, new_size, new_size))
            extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]
            # only the position tokens are interpolated
            pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]
            pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
            pos_tokens = torch.nn.functional.interpolate(
                pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
            pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)
            new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)
            checkpoint_model['pos_embed'] = new_pos_embed