File size: 26,556 Bytes
9eb3654
 
 
 
 
 
e7d2488
 
9eb3654
 
e7d2488
9eb3654
e7d2488
9eb3654
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e7d2488
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
9eb3654
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# Adapted from  https://github.com/microsoft/unilm/tree/master/beit
# --------------------------------------------------------
import math
import os
from functools import partial
from itertools import repeat
import collections.abc
import torch
import torch.nn as nn
import warnings
import torch.nn.functional as F

from .transformer import PatchDropout
from .rope import VisionRotaryEmbedding, VisionRotaryEmbeddingFast

if os.getenv('ENV_TYPE') == 'deepspeed':
    try:
        from deepspeed.runtime.activation_checkpointing.checkpointing import checkpoint
    except:
        from torch.utils.checkpoint import checkpoint
else:
    from torch.utils.checkpoint import checkpoint

try:
    import xformers
    import xformers.ops as xops
    XFORMERS_IS_AVAILBLE = True
except:
    XFORMERS_IS_AVAILBLE = False


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

to_2tuple = _ntuple(2)

def _no_grad_trunc_normal_(tensor, mean, std, a, b):
    # Cut & paste from PyTorch official master until it's in a few official releases - RW
    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
    def norm_cdf(x):
        # Computes standard normal cumulative distribution function
        return (1. + math.erf(x / math.sqrt(2.))) / 2.

    if (mean < a - 2 * std) or (mean > b + 2 * std):
        warnings.warn("mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
                      "The distribution of values may be incorrect.",
                      stacklevel=2)

    with torch.no_grad():
        # Values are generated by using a truncated uniform distribution and
        # then using the inverse CDF for the normal distribution.
        # Get upper and lower cdf values
        l = norm_cdf((a - mean) / std)
        u = norm_cdf((b - mean) / std)

        # Uniformly fill tensor with values from [l, u], then translate to
        # [2l-1, 2u-1].
        tensor.uniform_(2 * l - 1, 2 * u - 1)

        # Use inverse cdf transform for normal distribution to get truncated
        # standard normal
        tensor.erfinv_()

        # Transform to proper mean, std
        tensor.mul_(std * math.sqrt(2.))
        tensor.add_(mean)

        # Clamp to ensure it's in the proper range
        tensor.clamp_(min=a, max=b)
        return tensor


def trunc_normal_(tensor, mean=0., std=1., a=-2., b=2.):
    # type: (Tensor, float, float, float, float) -> Tensor
    r"""Fills the input Tensor with values drawn from a truncated
    normal distribution. The values are effectively drawn from the
    normal distribution :math:`\mathcal{N}(\text{mean}, \text{std}^2)`
    with values outside :math:`[a, b]` redrawn until they are within
    the bounds. The method used for generating the random values works
    best when :math:`a \leq \text{mean} \leq b`.
    Args:
        tensor: an n-dimensional `torch.Tensor`
        mean: the mean of the normal distribution
        std: the standard deviation of the normal distribution
        a: the minimum cutoff value
        b: the maximum cutoff value
    Examples:
        >>> w = torch.empty(3, 5)
        >>> nn.init.trunc_normal_(w)
    """
    return _no_grad_trunc_normal_(tensor, mean, std, a, b)

def drop_path(x, drop_prob: float = 0., training: bool = False, scale_by_keep: bool = True):
    """Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).

    This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,
    the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...
    See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for
    changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use
    'survival rate' as the argument.

    """
    if drop_prob == 0. or not training:
        return x
    keep_prob = 1 - drop_prob
    shape = (x.shape[0],) + (1,) * (x.ndim - 1)  # work with diff dim tensors, not just 2D ConvNets
    random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
    if keep_prob > 0.0 and scale_by_keep:
        random_tensor.div_(keep_prob)
    return x * random_tensor


class DropPath(nn.Module):
    """Drop paths (Stochastic Depth) per sample  (when applied in main path of residual blocks).
    """
    def __init__(self, drop_prob=None):
        super(DropPath, self).__init__()
        self.drop_prob = drop_prob

    def forward(self, x):
        return drop_path(x, self.drop_prob, self.training)
    
    def extra_repr(self) -> str:
        return 'p={}'.format(self.drop_prob)


class Mlp(nn.Module):
    def __init__(
        self, 
        in_features, 
        hidden_features=None, 
        out_features=None, 
        act_layer=nn.GELU, 
        norm_layer=nn.LayerNorm, 
        drop=0.,
        subln=False,

        ):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()

        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()

        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        # x = self.drop(x)
        # commit this for the orignal BERT implement 
        x = self.ffn_ln(x)

        x = self.fc2(x)
        x = self.drop(x)
        return x

class SwiGLU(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.SiLU, drop=0., 
                norm_layer=nn.LayerNorm, subln=False):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features

        self.w1 = nn.Linear(in_features, hidden_features)
        self.w2 = nn.Linear(in_features, hidden_features)

        self.act = act_layer()
        self.ffn_ln = norm_layer(hidden_features) if subln else nn.Identity()
        self.w3 = nn.Linear(hidden_features, out_features)
        
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x1 = self.w1(x)
        x2 = self.w2(x)
        hidden = self.act(x1) * x2
        x = self.ffn_ln(hidden)
        x = self.w3(x)
        x = self.drop(x)
        return x

class Attention(nn.Module):
    def __init__(
            self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,
            proj_drop=0., window_size=None, attn_head_dim=None, xattn=False, rope=None, subln=False, norm_layer=nn.LayerNorm):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        if attn_head_dim is not None:
            head_dim = attn_head_dim
        all_head_dim = head_dim * self.num_heads
        self.scale = qk_scale or head_dim ** -0.5

        self.subln = subln
        if self.subln:
            self.q_proj = nn.Linear(dim, all_head_dim, bias=False)
            self.k_proj = nn.Linear(dim, all_head_dim, bias=False)
            self.v_proj = nn.Linear(dim, all_head_dim, bias=False)
        else:
            self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)

        if qkv_bias:
            self.q_bias = nn.Parameter(torch.zeros(all_head_dim))
            self.v_bias = nn.Parameter(torch.zeros(all_head_dim))
        else:
            self.q_bias = None
            self.v_bias = None

        if window_size:
            self.window_size = window_size
            self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
            self.relative_position_bias_table = nn.Parameter(
                torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
            # cls to token & token 2 cls & cls to cls

            # get pair-wise relative position index for each token inside the window
            coords_h = torch.arange(window_size[0])
            coords_w = torch.arange(window_size[1])
            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
            relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
            relative_coords[:, :, 1] += window_size[1] - 1
            relative_coords[:, :, 0] *= 2 * window_size[1] - 1
            relative_position_index = \
                torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)
            relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
            relative_position_index[0, 0:] = self.num_relative_distance - 3
            relative_position_index[0:, 0] = self.num_relative_distance - 2
            relative_position_index[0, 0] = self.num_relative_distance - 1

            self.register_buffer("relative_position_index", relative_position_index)
        else:
            self.window_size = None
            self.relative_position_bias_table = None
            self.relative_position_index = None

        self.attn_drop = nn.Dropout(attn_drop)
        self.inner_attn_ln = norm_layer(all_head_dim) if subln else nn.Identity()
        # self.proj = nn.Linear(all_head_dim, all_head_dim)
        self.proj = nn.Linear(all_head_dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.xattn = xattn
        self.xattn_drop = attn_drop

        self.rope = rope

    def forward(self, x, rel_pos_bias=None, attn_mask=None):
        B, N, C = x.shape
        if self.subln: 
            q = F.linear(input=x, weight=self.q_proj.weight, bias=self.q_bias)
            k = F.linear(input=x, weight=self.k_proj.weight, bias=None)
            v = F.linear(input=x, weight=self.v_proj.weight, bias=self.v_bias)

            q = q.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)     # B, num_heads, N, C
            k = k.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3)  
            v = v.reshape(B, N, self.num_heads, -1).permute(0, 2, 1, 3) 
        else: 

            qkv_bias = None
            if self.q_bias is not None:
                qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))
            
            qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
            qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)   # 3, B, num_heads, N, C
            q, k, v = qkv[0], qkv[1], qkv[2]

        if self.rope:
            # slightly fast impl
            q_t = q[:, :, 1:, :]
            ro_q_t = self.rope(q_t)
            q = torch.cat((q[:, :, :1, :], ro_q_t), -2).type_as(v)

            k_t = k[:, :, 1:, :]
            ro_k_t = self.rope(k_t)
            k = torch.cat((k[:, :, :1, :], ro_k_t), -2).type_as(v)

        if self.xattn:
            q = q.permute(0, 2, 1, 3)   # B, num_heads, N, C -> B, N, num_heads, C
            k = k.permute(0, 2, 1, 3)
            v = v.permute(0, 2, 1, 3)

            x = xops.memory_efficient_attention(
                q, k, v,
                p=self.xattn_drop,
                scale=self.scale,
                )
            x = x.reshape(B, N, -1)
            x = self.inner_attn_ln(x)
            x = self.proj(x)
            x = self.proj_drop(x)
        else:
            q = q * self.scale
            attn = (q @ k.transpose(-2, -1))

            if self.relative_position_bias_table is not None:
                relative_position_bias = \
                    self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
                        self.window_size[0] * self.window_size[1] + 1,
                        self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
                relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
                attn = attn + relative_position_bias.unsqueeze(0).type_as(attn)

            if rel_pos_bias is not None:
                attn = attn + rel_pos_bias.type_as(attn)

            if attn_mask is not None:
                attn_mask = attn_mask.bool()
                attn = attn.masked_fill(~attn_mask[:, None, None, :], float("-inf"))
            
            attn = attn.softmax(dim=-1)
            attn = self.attn_drop(attn)

            x = (attn @ v).transpose(1, 2).reshape(B, N, -1)
            x = self.inner_attn_ln(x)
            x = self.proj(x)
            x = self.proj_drop(x)
        return x


class Block(nn.Module):

    def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 window_size=None, attn_head_dim=None, xattn=False, rope=None, postnorm=False,
                 subln=False, naiveswiglu=False):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = Attention(
            dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,
            attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim,
            xattn=xattn, rope=rope, subln=subln, norm_layer=norm_layer)
        # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here
        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)

        if naiveswiglu:
            self.mlp = SwiGLU(
                in_features=dim, 
                hidden_features=mlp_hidden_dim, 
                subln=subln,
                norm_layer=norm_layer,
            )
        else:
            self.mlp = Mlp(
                in_features=dim, 
                hidden_features=mlp_hidden_dim, 
                act_layer=act_layer,
                subln=subln,
                drop=drop
            )

        if init_values is not None and init_values > 0:
            self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
            self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)
        else:
            self.gamma_1, self.gamma_2 = None, None

        self.postnorm = postnorm

    def forward(self, x, rel_pos_bias=None, attn_mask=None):
        if self.gamma_1 is None:
            if self.postnorm:
                x = x + self.drop_path(self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
                x = x + self.drop_path(self.norm2(self.mlp(x)))
            else:
                x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
                x = x + self.drop_path(self.mlp(self.norm2(x)))
        else:
            if self.postnorm:
                x = x + self.drop_path(self.gamma_1 * self.norm1(self.attn(x, rel_pos_bias=rel_pos_bias, attn_mask=attn_mask)))
                x = x + self.drop_path(self.gamma_2 * self.norm2(self.mlp(x)))
            else:
                x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias, attn_mask=attn_mask))
                x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
        return x


class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
        self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])
        self.img_size = img_size
        self.patch_size = patch_size
        self.num_patches = num_patches

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)

    def forward(self, x, **kwargs):
        B, C, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)
        return x


class RelativePositionBias(nn.Module):

    def __init__(self, window_size, num_heads):
        super().__init__()
        self.window_size = window_size
        self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros(self.num_relative_distance, num_heads))  # 2*Wh-1 * 2*Ww-1, nH
        # cls to token & token 2 cls & cls to cls

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(window_size[0])
        coords_w = torch.arange(window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * window_size[1] - 1
        relative_position_index = \
            torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)
        relative_position_index[1:, 1:] = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        relative_position_index[0, 0:] = self.num_relative_distance - 3
        relative_position_index[0:, 0] = self.num_relative_distance - 2
        relative_position_index[0, 0] = self.num_relative_distance - 1

        self.register_buffer("relative_position_index", relative_position_index)

    def forward(self):
        relative_position_bias = \
            self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
                self.window_size[0] * self.window_size[1] + 1,
                self.window_size[0] * self.window_size[1] + 1, -1)  # Wh*Ww,Wh*Ww,nH
        return relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww


class EVAVisionTransformer(nn.Module):
    """ Vision Transformer with support for patch or hybrid CNN input stage
    """
    def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,
                 num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,
                 drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None, patch_dropout=0.,
                 use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False, rope=False,
                 use_mean_pooling=True, init_scale=0.001, grad_checkpointing=False, xattn=False, postnorm=False,
                 pt_hw_seq_len=16, intp_freq=False, naiveswiglu=False, subln=False):
        super().__init__()

        if not XFORMERS_IS_AVAILBLE:
            xattn = False

        self.image_size = img_size
        self.num_classes = num_classes
        self.num_features = self.embed_dim = embed_dim  # num_features for consistency with other models

        self.patch_embed = PatchEmbed(
            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
        num_patches = self.patch_embed.num_patches

        self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        # self.mask_token = nn.Parameter(torch.zeros(1, 1, embed_dim))
        if use_abs_pos_emb:
            self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))
        else:
            self.pos_embed = None
        self.pos_drop = nn.Dropout(p=drop_rate)

        if use_shared_rel_pos_bias:
            self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)
        else:
            self.rel_pos_bias = None

        if rope:
            half_head_dim = embed_dim // num_heads // 2
            hw_seq_len = img_size // patch_size
            self.rope = VisionRotaryEmbeddingFast(
                dim=half_head_dim,
                pt_seq_len=pt_hw_seq_len,
                ft_seq_len=hw_seq_len if intp_freq else None,
                # patch_dropout=patch_dropout
            )
        else: 
            self.rope = None

        self.naiveswiglu = naiveswiglu

        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)]  # stochastic depth decay rule
        self.use_rel_pos_bias = use_rel_pos_bias
        self.blocks = nn.ModuleList([
            Block(
                dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,
                drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,
                init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None,
                xattn=xattn, rope=self.rope, postnorm=postnorm, subln=subln, naiveswiglu=naiveswiglu)
            for i in range(depth)])
        self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)
        self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None
        self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()

        if self.pos_embed is not None:
            trunc_normal_(self.pos_embed, std=.02)

        trunc_normal_(self.cls_token, std=.02)
        # trunc_normal_(self.mask_token, std=.02)

        self.apply(self._init_weights)
        self.fix_init_weight()

        if isinstance(self.head, nn.Linear):
            trunc_normal_(self.head.weight, std=.02)
            self.head.weight.data.mul_(init_scale)
            self.head.bias.data.mul_(init_scale)

        # 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.grad_checkpointing = grad_checkpointing

    def fix_init_weight(self):
        def rescale(param, layer_id):
            param.div_(math.sqrt(2.0 * layer_id))

        for layer_id, layer in enumerate(self.blocks):
            rescale(layer.attn.proj.weight.data, layer_id + 1)
            if self.naiveswiglu:
                rescale(layer.mlp.w3.weight.data, layer_id + 1)
            else:
                rescale(layer.mlp.fc2.weight.data, layer_id + 1)

    def get_cast_dtype(self) -> torch.dtype:
        return self.blocks[0].mlp.fc2.weight.dtype

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if m.bias is not None:
                nn.init.constant_(m.bias, 0)
        elif isinstance(m, nn.LayerNorm):
            nn.init.constant_(m.bias, 0)
            nn.init.constant_(m.weight, 1.0)

    def get_num_layers(self):
        return len(self.blocks)
    
    def lock(self, unlocked_groups=0, freeze_bn_stats=False):
        assert unlocked_groups == 0, 'partial locking not currently supported for this model'
        for param in self.parameters():
            param.requires_grad = False

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

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'pos_embed', 'cls_token'}

    def get_classifier(self):
        return self.head

    def reset_classifier(self, num_classes, global_pool=''):
        self.num_classes = num_classes
        self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()

    def forward_features(self, x, return_all_features=False, return_hidden=False, shuffle=False):
        
        x = self.patch_embed(x)
        batch_size, seq_len, _ = x.size()

        if shuffle:
            idx = torch.randperm(x.shape[1]) + 1
            zero = torch.LongTensor([0, ])
            idx = torch.cat([zero, idx])
            pos_embed = self.pos_embed[:, idx]

        cls_tokens = self.cls_token.expand(batch_size, -1, -1)  # stole cls_tokens impl from Phil Wang, thanks
        x = torch.cat((cls_tokens, x), dim=1)
        if shuffle:
            x = x + pos_embed
        elif self.pos_embed is not None:
            x = x + self.pos_embed
        x = self.pos_drop(x)

        # a patch_dropout of 0. would mean it is disabled and this function would do nothing but return what was passed in
        if os.getenv('RoPE') == '1':
            if self.training and not isinstance(self.patch_dropout, nn.Identity):
                x, patch_indices_keep = self.patch_dropout(x)
                self.rope.forward = partial(self.rope.forward, patch_indices_keep=patch_indices_keep)
            else:
                self.rope.forward = partial(self.rope.forward, patch_indices_keep=None)
                x = self.patch_dropout(x)
        else:
            x = self.patch_dropout(x)

        rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None
        hidden_states = []
        for idx, blk in enumerate(self.blocks):
            if (0 < idx <= 20) and (idx % 4 == 0) and return_hidden:
                hidden_states.append(x)
            if self.grad_checkpointing:
                x = checkpoint(blk, x, (rel_pos_bias,))
            else:
                x = blk(x, rel_pos_bias=rel_pos_bias)

        if not return_all_features:
            x = self.norm(x)
            if self.fc_norm is not None:
                return self.fc_norm(x.mean(1)), hidden_states
            else:
                return x[:, 0], hidden_states
        return x

    def forward(self, x, return_all_features=False, return_hidden=False, shuffle=False):
        if return_all_features:
            return self.forward_features(x, return_all_features, return_hidden, shuffle)
        x, hidden_states = self.forward_features(x, return_all_features, return_hidden, shuffle)
        x = self.head(x)
        if return_hidden:
            return x, hidden_states
        return x