File size: 25,091 Bytes
483de47
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# --------------------------------------------------------
# TinyViT Model Architecture
# Copyright (c) 2022 Microsoft
# Adapted from LeViT and Swin Transformer
#   LeViT: (https://github.com/facebookresearch/levit)
#   Swin: (https://github.com/microsoft/swin-transformer)
# Build the TinyViT Model
# --------------------------------------------------------

import itertools
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
from timm.models.layers import DropPath as TimmDropPath,\
    to_2tuple, trunc_normal_
from timm.models.registry import register_model
from typing import Tuple


class Conv2d_BN(torch.nn.Sequential):
    def __init__(self, a, b, ks=1, stride=1, pad=0, dilation=1,
                 groups=1, bn_weight_init=1):
        super().__init__()
        self.add_module('c', torch.nn.Conv2d(
            a, b, ks, stride, pad, dilation, groups, bias=False))
        bn = torch.nn.BatchNorm2d(b)
        torch.nn.init.constant_(bn.weight, bn_weight_init)
        torch.nn.init.constant_(bn.bias, 0)
        self.add_module('bn', bn)

    @torch.no_grad()
    def fuse(self):
        c, bn = self._modules.values()
        w = bn.weight / (bn.running_var + bn.eps)**0.5
        w = c.weight * w[:, None, None, None]
        b = bn.bias - bn.running_mean * bn.weight / \
            (bn.running_var + bn.eps)**0.5
        m = torch.nn.Conv2d(w.size(1) * self.c.groups, w.size(
            0), w.shape[2:], stride=self.c.stride, padding=self.c.padding, dilation=self.c.dilation, groups=self.c.groups)
        m.weight.data.copy_(w)
        m.bias.data.copy_(b)
        return m


class DropPath(TimmDropPath):
    def __init__(self, drop_prob=None):
        super().__init__(drop_prob=drop_prob)
        self.drop_prob = drop_prob

    def __repr__(self):
        msg = super().__repr__()
        msg += f'(drop_prob={self.drop_prob})'
        return msg


class PatchEmbed(nn.Module):
    def __init__(self, in_chans, embed_dim, resolution, activation):
        super().__init__()
        img_size: Tuple[int, int] = to_2tuple(resolution)
        self.patches_resolution = (img_size[0] // 4, img_size[1] // 4)
        self.num_patches = self.patches_resolution[0] * \
            self.patches_resolution[1]
        self.in_chans = in_chans
        self.embed_dim = embed_dim
        n = embed_dim
        self.seq = nn.Sequential(
            Conv2d_BN(in_chans, n // 2, 3, 2, 1),
            activation(),
            Conv2d_BN(n // 2, n, 3, 2, 1),
        )

    def forward(self, x):
        return self.seq(x)


class MBConv(nn.Module):
    def __init__(self, in_chans, out_chans, expand_ratio,
                 activation, drop_path):
        super().__init__()
        self.in_chans = in_chans
        self.hidden_chans = int(in_chans * expand_ratio)
        self.out_chans = out_chans

        self.conv1 = Conv2d_BN(in_chans, self.hidden_chans, ks=1)
        self.act1 = activation()

        self.conv2 = Conv2d_BN(self.hidden_chans, self.hidden_chans,
                               ks=3, stride=1, pad=1, groups=self.hidden_chans)
        self.act2 = activation()

        self.conv3 = Conv2d_BN(
            self.hidden_chans, out_chans, ks=1, bn_weight_init=0.0)
        self.act3 = activation()

        self.drop_path = DropPath(
            drop_path) if drop_path > 0. else nn.Identity()

    def forward(self, x):
        shortcut = x

        x = self.conv1(x)
        x = self.act1(x)

        x = self.conv2(x)
        x = self.act2(x)

        x = self.conv3(x)

        x = self.drop_path(x)

        x += shortcut
        x = self.act3(x)

        return x


class PatchMerging(nn.Module):
    def __init__(self, input_resolution, dim, out_dim, activation):
        super().__init__()

        self.input_resolution = input_resolution
        self.dim = dim
        self.out_dim = out_dim
        self.act = activation()
        self.conv1 = Conv2d_BN(dim, out_dim, 1, 1, 0)
        stride_c=2
        if(out_dim==320 or out_dim==448 or out_dim==576):
            stride_c=1
        self.conv2 = Conv2d_BN(out_dim, out_dim, 3, stride_c, 1, groups=out_dim)
        self.conv3 = Conv2d_BN(out_dim, out_dim, 1, 1, 0)

    def forward(self, x):
        if x.ndim == 3:
            H, W = self.input_resolution
            B = len(x)
            # (B, C, H, W)
            x = x.view(B, H, W, -1).permute(0, 3, 1, 2)

        x = self.conv1(x)
        x = self.act(x)

        x = self.conv2(x)
        x = self.act(x)
        x = self.conv3(x)
        x = x.flatten(2).transpose(1, 2)
        return x


class ConvLayer(nn.Module):
    def __init__(self, dim, input_resolution, depth,
                 activation,
                 drop_path=0., downsample=None, use_checkpoint=False,
                 out_dim=None,
                 conv_expand_ratio=4.,
                 ):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList([
            MBConv(dim, dim, conv_expand_ratio, activation,
                   drop_path[i] if isinstance(drop_path, list) else drop_path,
                   )
            for i in range(depth)])

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(
                input_resolution, dim=dim, out_dim=out_dim, activation=activation)
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None,
                 out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.norm = nn.LayerNorm(in_features)
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.act = act_layer()
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.norm(x)

        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class Attention(torch.nn.Module):
    def __init__(self, dim, key_dim, num_heads=8,
                 attn_ratio=4,
                 resolution=(14, 14),
                 ):
        super().__init__()
        # (h, w)
        assert isinstance(resolution, tuple) and len(resolution) == 2
        self.num_heads = num_heads
        self.scale = key_dim ** -0.5
        self.key_dim = key_dim
        self.nh_kd = nh_kd = key_dim * num_heads
        self.d = int(attn_ratio * key_dim)
        self.dh = int(attn_ratio * key_dim) * num_heads
        self.attn_ratio = attn_ratio
        h = self.dh + nh_kd * 2

        self.norm = nn.LayerNorm(dim)
        self.qkv = nn.Linear(dim, h)
        self.proj = nn.Linear(self.dh, dim)

        points = list(itertools.product(
            range(resolution[0]), range(resolution[1])))
        N = len(points)
        attention_offsets = {}
        idxs = []
        for p1 in points:
            for p2 in points:
                offset = (abs(p1[0] - p2[0]), abs(p1[1] - p2[1]))
                if offset not in attention_offsets:
                    attention_offsets[offset] = len(attention_offsets)
                idxs.append(attention_offsets[offset])
        self.attention_biases = torch.nn.Parameter(
            torch.zeros(num_heads, len(attention_offsets)))
        self.register_buffer('attention_bias_idxs',
                             torch.LongTensor(idxs).view(N, N),
                             persistent=False)

    @torch.no_grad()
    def train(self, mode=True):
        super().train(mode)
        if mode and hasattr(self, 'ab'):
            del self.ab
        else:
            self.register_buffer('ab',
                                 self.attention_biases[:, self.attention_bias_idxs],
                                 persistent=False)

    def forward(self, x):  # x (B,N,C)
        B, N, _ = x.shape

        # Normalization
        x = self.norm(x)

        qkv = self.qkv(x)
        # (B, N, num_heads, d)
        q, k, v = qkv.view(B, N, self.num_heads, -
                           1).split([self.key_dim, self.key_dim, self.d], dim=3)
        # (B, num_heads, N, d)
        q = q.permute(0, 2, 1, 3)
        k = k.permute(0, 2, 1, 3)
        v = v.permute(0, 2, 1, 3)

        attn = (
            (q @ k.transpose(-2, -1)) * self.scale
            +
            (self.attention_biases[:, self.attention_bias_idxs]
             if self.training else self.ab)
        )
        attn = attn.softmax(dim=-1)
        x = (attn @ v).transpose(1, 2).reshape(B, N, self.dh)
        x = self.proj(x)
        return x


class TinyViTBlock(nn.Module):
    r""" TinyViT Block.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int, int]): Input resolution.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        drop (float, optional): Dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        local_conv_size (int): the kernel size of the convolution between
                               Attention and MLP. Default: 3
        activation: the activation function. Default: nn.GELU
    """

    def __init__(self, dim, input_resolution, num_heads, window_size=7,
                 mlp_ratio=4., drop=0., drop_path=0.,
                 local_conv_size=3,
                 activation=nn.GELU,
                 ):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        assert window_size > 0, 'window_size must be greater than 0'
        self.window_size = window_size
        self.mlp_ratio = mlp_ratio

        self.drop_path = DropPath(
            drop_path) if drop_path > 0. else nn.Identity()

        assert dim % num_heads == 0, 'dim must be divisible by num_heads'
        head_dim = dim // num_heads

        window_resolution = (window_size, window_size)
        self.attn = Attention(dim, head_dim, num_heads,
                              attn_ratio=1, resolution=window_resolution)

        mlp_hidden_dim = int(dim * mlp_ratio)
        mlp_activation = activation
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim,
                       act_layer=mlp_activation, drop=drop)

        pad = local_conv_size // 2
        self.local_conv = Conv2d_BN(
            dim, dim, ks=local_conv_size, stride=1, pad=pad, groups=dim)

    def forward(self, x):
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"
        res_x = x
        if H == self.window_size and W == self.window_size:
            x = self.attn(x)
        else:
            x = x.view(B, H, W, C)
            pad_b = (self.window_size - H %
                     self.window_size) % self.window_size
            pad_r = (self.window_size - W %
                     self.window_size) % self.window_size
            padding = pad_b > 0 or pad_r > 0

            if padding:
                x = F.pad(x, (0, 0, 0, pad_r, 0, pad_b))

            pH, pW = H + pad_b, W + pad_r
            nH = pH // self.window_size
            nW = pW // self.window_size
            # window partition
            x = x.view(B, nH, self.window_size, nW, self.window_size, C).transpose(2, 3).reshape(
                B * nH * nW, self.window_size * self.window_size, C)
            x = self.attn(x)
            # window reverse
            x = x.view(B, nH, nW, self.window_size, self.window_size,
                       C).transpose(2, 3).reshape(B, pH, pW, C)

            if padding:
                x = x[:, :H, :W].contiguous()

            x = x.view(B, L, C)

        x = res_x + self.drop_path(x)

        x = x.transpose(1, 2).reshape(B, C, H, W)
        x = self.local_conv(x)
        x = x.view(B, C, L).transpose(1, 2)

        x = x + self.drop_path(self.mlp(x))
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
               f"window_size={self.window_size}, mlp_ratio={self.mlp_ratio}"


class BasicLayer(nn.Module):
    """ A basic TinyViT layer for one stage.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        depth (int): Number of blocks.
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        drop (float, optional): Dropout rate. Default: 0.0
        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
        local_conv_size: the kernel size of the depthwise convolution between attention and MLP. Default: 3
        activation: the activation function. Default: nn.GELU
        out_dim: the output dimension of the layer. Default: dim
    """

    def __init__(self, dim, input_resolution, depth, num_heads, window_size,
                 mlp_ratio=4., drop=0.,
                 drop_path=0., downsample=None, use_checkpoint=False,
                 local_conv_size=3,
                 activation=nn.GELU,
                 out_dim=None,
                 ):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList([
            TinyViTBlock(dim=dim, input_resolution=input_resolution,
                         num_heads=num_heads, window_size=window_size,
                         mlp_ratio=mlp_ratio,
                         drop=drop,
                         drop_path=drop_path[i] if isinstance(
                             drop_path, list) else drop_path,
                         local_conv_size=local_conv_size,
                         activation=activation,
                         )
            for i in range(depth)])

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(
                input_resolution, dim=dim, out_dim=out_dim, activation=activation)
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"

class LayerNorm2d(nn.Module):
    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.ones(num_channels))
        self.bias = nn.Parameter(torch.zeros(num_channels))
        self.eps = eps

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        u = x.mean(1, keepdim=True)
        s = (x - u).pow(2).mean(1, keepdim=True)
        x = (x - u) / torch.sqrt(s + self.eps)
        x = self.weight[:, None, None] * x + self.bias[:, None, None]
        return x
class TinyViT(nn.Module):
    def __init__(self, img_size=224, in_chans=3, num_classes=1000,
                 embed_dims=[96, 192, 384, 768], depths=[2, 2, 6, 2],
                 num_heads=[3, 6, 12, 24],
                 window_sizes=[7, 7, 14, 7],
                 mlp_ratio=4.,
                 drop_rate=0.,
                 drop_path_rate=0.1,
                 use_checkpoint=False,
                 mbconv_expand_ratio=4.0,
                 local_conv_size=3,
                 layer_lr_decay=1.0,
                 ):
        super().__init__()
        self.img_size=img_size
        self.num_classes = num_classes
        self.depths = depths
        self.num_layers = len(depths)
        self.mlp_ratio = mlp_ratio

        activation = nn.GELU

        self.patch_embed = PatchEmbed(in_chans=in_chans,
                                      embed_dim=embed_dims[0],
                                      resolution=img_size,
                                      activation=activation)

        patches_resolution = self.patch_embed.patches_resolution
        self.patches_resolution = patches_resolution

        # stochastic depth
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate,
                                                sum(depths))]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            kwargs = dict(dim=embed_dims[i_layer],
                        input_resolution=(patches_resolution[0] // (2 ** (i_layer-1 if i_layer == 3 else i_layer)),
                                patches_resolution[1] // (2 ** (i_layer-1 if i_layer == 3 else i_layer))),
                        #   input_resolution=(patches_resolution[0] // (2 ** i_layer),
                        #                     patches_resolution[1] // (2 ** i_layer)),
                          depth=depths[i_layer],
                          drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
                          downsample=PatchMerging if (
                              i_layer < self.num_layers - 1) else None,
                          use_checkpoint=use_checkpoint,
                          out_dim=embed_dims[min(
                              i_layer + 1, len(embed_dims) - 1)],
                          activation=activation,
                          )
            if i_layer == 0:
                layer = ConvLayer(
                    conv_expand_ratio=mbconv_expand_ratio,
                    **kwargs,
                )
            else:
                layer = BasicLayer(
                    num_heads=num_heads[i_layer],
                    window_size=window_sizes[i_layer],
                    mlp_ratio=self.mlp_ratio,
                    drop=drop_rate,
                    local_conv_size=local_conv_size,
                    **kwargs)
            self.layers.append(layer)

        # Classifier head
        self.norm_head = nn.LayerNorm(embed_dims[-1])
        self.head = nn.Linear(
            embed_dims[-1], num_classes) if num_classes > 0 else torch.nn.Identity()

        # init weights
        self.apply(self._init_weights)
        self.set_layer_lr_decay(layer_lr_decay)
        self.neck = nn.Sequential(
            nn.Conv2d(
                embed_dims[-1],
                256,
                kernel_size=1,
                bias=False,
            ),
            LayerNorm2d(256),
            nn.Conv2d(
                256,
                256,
                kernel_size=3,
                padding=1,
                bias=False,
            ),
            LayerNorm2d(256),
        )
    def set_layer_lr_decay(self, layer_lr_decay):
        decay_rate = layer_lr_decay

        # layers -> blocks (depth)
        depth = sum(self.depths)
        lr_scales = [decay_rate ** (depth - i - 1) for i in range(depth)]
        #print("LR SCALES:", lr_scales)

        def _set_lr_scale(m, scale):
            for p in m.parameters():
                p.lr_scale = scale

        self.patch_embed.apply(lambda x: _set_lr_scale(x, lr_scales[0]))
        i = 0
        for layer in self.layers:
            for block in layer.blocks:
                block.apply(lambda x: _set_lr_scale(x, lr_scales[i]))
                i += 1
            if layer.downsample is not None:
                layer.downsample.apply(
                    lambda x: _set_lr_scale(x, lr_scales[i - 1]))
        assert i == depth
        for m in [self.norm_head, self.head]:
            m.apply(lambda x: _set_lr_scale(x, lr_scales[-1]))

        for k, p in self.named_parameters():
            p.param_name = k

        def _check_lr_scale(m):
            for p in m.parameters():
                assert hasattr(p, 'lr_scale'), p.param_name

        self.apply(_check_lr_scale)

    def _init_weights(self, m):
        if isinstance(m, nn.Linear):
            trunc_normal_(m.weight, std=.02)
            if isinstance(m, nn.Linear) and 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)

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {'attention_biases'}

    def forward_features(self, x):
        # x: (N, C, H, W)
        x = self.patch_embed(x)

        x = self.layers[0](x)
        start_i = 1

        interm_embeddings=[]
        for i in range(start_i, len(self.layers)):
            layer = self.layers[i]
            x = layer(x)
            # print('x shape:', x.shape, '---i:', i)
            if i == 1:
                interm_embeddings.append(x.view(x.shape[0], 64, 64, -1))

        B,_,C=x.size()
        x = x.view(B, 64, 64, C)
        x=x.permute(0, 3, 1, 2)
        x=self.neck(x)
        return x, interm_embeddings

    def forward(self, x):
        x, interm_embeddings = self.forward_features(x)
        #x = self.norm_head(x)
        #x = self.head(x)
        # print('come to here is correct'* 3)
        return x, interm_embeddings


_checkpoint_url_format = \
    'https://github.com/wkcn/TinyViT-model-zoo/releases/download/checkpoints/{}.pth'
_provided_checkpoints = {
    'tiny_vit_5m_224': 'tiny_vit_5m_22kto1k_distill',
    'tiny_vit_11m_224': 'tiny_vit_11m_22kto1k_distill',
    'tiny_vit_21m_224': 'tiny_vit_21m_22kto1k_distill',
    'tiny_vit_21m_384': 'tiny_vit_21m_22kto1k_384_distill',
    'tiny_vit_21m_512': 'tiny_vit_21m_22kto1k_512_distill',
}


def register_tiny_vit_model(fn):
    '''Register a TinyViT model
    It is a wrapper of `register_model` with loading the pretrained checkpoint.
    '''
    def fn_wrapper(pretrained=False, **kwargs):
        model = fn()
        if pretrained:
            model_name = fn.__name__
            assert model_name in _provided_checkpoints, \
                f'Sorry that the checkpoint `{model_name}` is not provided yet.'
            url = _checkpoint_url_format.format(
                _provided_checkpoints[model_name])
            checkpoint = torch.hub.load_state_dict_from_url(
                url=url,
                map_location='cpu', check_hash=False,
            )
            model.load_state_dict(checkpoint['model'])

        return model

    # rename the name of fn_wrapper
    fn_wrapper.__name__ = fn.__name__
    return register_model(fn_wrapper)


@register_tiny_vit_model
def tiny_vit_5m_224(pretrained=False, num_classes=1000, drop_path_rate=0.0):
    return TinyViT(
        num_classes=num_classes,
        embed_dims=[64, 128, 160, 320],
        depths=[2, 2, 6, 2],
        num_heads=[2, 4, 5, 10],
        window_sizes=[7, 7, 14, 7],
        drop_path_rate=drop_path_rate,
    )


@register_tiny_vit_model
def tiny_vit_11m_224(pretrained=False, num_classes=1000, drop_path_rate=0.1):
    return TinyViT(
        num_classes=num_classes,
        embed_dims=[64, 128, 256, 448],
        depths=[2, 2, 6, 2],
        num_heads=[2, 4, 8, 14],
        window_sizes=[7, 7, 14, 7],
        drop_path_rate=drop_path_rate,
    )


@register_tiny_vit_model
def tiny_vit_21m_224(pretrained=False, num_classes=1000, drop_path_rate=0.2):
    return TinyViT(
        num_classes=num_classes,
        embed_dims=[96, 192, 384, 576],
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 18],
        window_sizes=[7, 7, 14, 7],
        drop_path_rate=drop_path_rate,
    )


@register_tiny_vit_model
def tiny_vit_21m_384(pretrained=False, num_classes=1000, drop_path_rate=0.1):
    return TinyViT(
        img_size=384,
        num_classes=num_classes,
        embed_dims=[96, 192, 384, 576],
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 18],
        window_sizes=[12, 12, 24, 12],
        drop_path_rate=drop_path_rate,
    )


@register_tiny_vit_model
def tiny_vit_21m_512(pretrained=False, num_classes=1000, drop_path_rate=0.1):
    return TinyViT(
        img_size=512,
        num_classes=num_classes,
        embed_dims=[96, 192, 384, 576],
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 18],
        window_sizes=[16, 16, 32, 16],
        drop_path_rate=drop_path_rate,
    )