File size: 20,447 Bytes
d7e58f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import copy

import pytest
import torch

from mmcv.cnn.bricks.drop import DropPath
from mmcv.cnn.bricks.transformer import (FFN, AdaptivePadding,
                                         BaseTransformerLayer,
                                         MultiheadAttention, PatchEmbed,
                                         PatchMerging,
                                         TransformerLayerSequence)
from mmcv.runner import ModuleList


def test_adaptive_padding():

    for padding in ('same', 'corner'):
        kernel_size = 16
        stride = 16
        dilation = 1
        input = torch.rand(1, 1, 15, 17)
        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        out = adap_pad(input)
        # padding to divisible by 16
        assert (out.shape[2], out.shape[3]) == (16, 32)
        input = torch.rand(1, 1, 16, 17)
        out = adap_pad(input)
        # padding to divisible by 16
        assert (out.shape[2], out.shape[3]) == (16, 32)

        kernel_size = (2, 2)
        stride = (2, 2)
        dilation = (1, 1)

        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        input = torch.rand(1, 1, 11, 13)
        out = adap_pad(input)
        # padding to divisible by 2
        assert (out.shape[2], out.shape[3]) == (12, 14)

        kernel_size = (2, 2)
        stride = (10, 10)
        dilation = (1, 1)

        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        input = torch.rand(1, 1, 10, 13)
        out = adap_pad(input)
        #  no padding
        assert (out.shape[2], out.shape[3]) == (10, 13)

        kernel_size = (11, 11)
        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        input = torch.rand(1, 1, 11, 13)
        out = adap_pad(input)
        #  all padding
        assert (out.shape[2], out.shape[3]) == (21, 21)

        # test padding as kernel is (7,9)
        input = torch.rand(1, 1, 11, 13)
        stride = (3, 4)
        kernel_size = (4, 5)
        dilation = (2, 2)
        # actually (7, 9)
        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        dilation_out = adap_pad(input)
        assert (dilation_out.shape[2], dilation_out.shape[3]) == (16, 21)
        kernel_size = (7, 9)
        dilation = (1, 1)
        adap_pad = AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=padding)
        kernel79_out = adap_pad(input)
        assert (kernel79_out.shape[2], kernel79_out.shape[3]) == (16, 21)
        assert kernel79_out.shape == dilation_out.shape

    # assert only support "same" "corner"
    with pytest.raises(AssertionError):
        AdaptivePadding(
            kernel_size=kernel_size,
            stride=stride,
            dilation=dilation,
            padding=1)


def test_patch_embed():
    B = 2
    H = 3
    W = 4
    C = 3
    embed_dims = 10
    kernel_size = 3
    stride = 1
    dummy_input = torch.rand(B, C, H, W)
    patch_merge_1 = PatchEmbed(
        in_channels=C,
        embed_dims=embed_dims,
        kernel_size=kernel_size,
        stride=stride,
        padding=0,
        dilation=1,
        norm_cfg=None)

    x1, shape = patch_merge_1(dummy_input)
    # test out shape
    assert x1.shape == (2, 2, 10)
    # test outsize is correct
    assert shape == (1, 2)
    # test L = out_h * out_w
    assert shape[0] * shape[1] == x1.shape[1]

    B = 2
    H = 10
    W = 10
    C = 3
    embed_dims = 10
    kernel_size = 5
    stride = 2
    dummy_input = torch.rand(B, C, H, W)
    # test dilation
    patch_merge_2 = PatchEmbed(
        in_channels=C,
        embed_dims=embed_dims,
        kernel_size=kernel_size,
        stride=stride,
        padding=0,
        dilation=2,
        norm_cfg=None,
    )

    x2, shape = patch_merge_2(dummy_input)
    # test out shape
    assert x2.shape == (2, 1, 10)
    # test outsize is correct
    assert shape == (1, 1)
    # test L = out_h * out_w
    assert shape[0] * shape[1] == x2.shape[1]

    stride = 2
    input_size = (10, 10)

    dummy_input = torch.rand(B, C, H, W)
    # test stride and norm
    patch_merge_3 = PatchEmbed(
        in_channels=C,
        embed_dims=embed_dims,
        kernel_size=kernel_size,
        stride=stride,
        padding=0,
        dilation=2,
        norm_cfg=dict(type='LN'),
        input_size=input_size)

    x3, shape = patch_merge_3(dummy_input)
    # test out shape
    assert x3.shape == (2, 1, 10)
    # test outsize is correct
    assert shape == (1, 1)
    # test L = out_h * out_w
    assert shape[0] * shape[1] == x3.shape[1]

    # test the init_out_size with nn.Unfold
    assert patch_merge_3.init_out_size[1] == (input_size[0] - 2 * 4 -
                                              1) // 2 + 1
    assert patch_merge_3.init_out_size[0] == (input_size[0] - 2 * 4 -
                                              1) // 2 + 1
    H = 11
    W = 12
    input_size = (H, W)
    dummy_input = torch.rand(B, C, H, W)
    # test stride and norm
    patch_merge_3 = PatchEmbed(
        in_channels=C,
        embed_dims=embed_dims,
        kernel_size=kernel_size,
        stride=stride,
        padding=0,
        dilation=2,
        norm_cfg=dict(type='LN'),
        input_size=input_size)

    _, shape = patch_merge_3(dummy_input)
    # when input_size equal to real input
    # the out_size should be equal to `init_out_size`
    assert shape == patch_merge_3.init_out_size

    input_size = (H, W)
    dummy_input = torch.rand(B, C, H, W)
    # test stride and norm
    patch_merge_3 = PatchEmbed(
        in_channels=C,
        embed_dims=embed_dims,
        kernel_size=kernel_size,
        stride=stride,
        padding=0,
        dilation=2,
        norm_cfg=dict(type='LN'),
        input_size=input_size)

    _, shape = patch_merge_3(dummy_input)
    # when input_size equal to real input
    # the out_size should be equal to `init_out_size`
    assert shape == patch_merge_3.init_out_size

    # test adap padding
    for padding in ('same', 'corner'):
        in_c = 2
        embed_dims = 3
        B = 2

        # test stride is 1
        input_size = (5, 5)
        kernel_size = (5, 5)
        stride = (1, 1)
        dilation = 1
        bias = False

        x = torch.rand(B, in_c, *input_size)
        patch_embed = PatchEmbed(
            in_channels=in_c,
            embed_dims=embed_dims,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_embed(x)
        assert x_out.size() == (B, 25, 3)
        assert out_size == (5, 5)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test kernel_size == stride
        input_size = (5, 5)
        kernel_size = (5, 5)
        stride = (5, 5)
        dilation = 1
        bias = False

        x = torch.rand(B, in_c, *input_size)
        patch_embed = PatchEmbed(
            in_channels=in_c,
            embed_dims=embed_dims,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_embed(x)
        assert x_out.size() == (B, 1, 3)
        assert out_size == (1, 1)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test kernel_size == stride
        input_size = (6, 5)
        kernel_size = (5, 5)
        stride = (5, 5)
        dilation = 1
        bias = False

        x = torch.rand(B, in_c, *input_size)
        patch_embed = PatchEmbed(
            in_channels=in_c,
            embed_dims=embed_dims,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_embed(x)
        assert x_out.size() == (B, 2, 3)
        assert out_size == (2, 1)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test different kernel_size with different stride
        input_size = (6, 5)
        kernel_size = (6, 2)
        stride = (6, 2)
        dilation = 1
        bias = False

        x = torch.rand(B, in_c, *input_size)
        patch_embed = PatchEmbed(
            in_channels=in_c,
            embed_dims=embed_dims,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_embed(x)
        assert x_out.size() == (B, 3, 3)
        assert out_size == (1, 3)
        assert x_out.size(1) == out_size[0] * out_size[1]


def test_patch_merging():

    # Test the model with int padding
    in_c = 3
    out_c = 4
    kernel_size = 3
    stride = 3
    padding = 1
    dilation = 1
    bias = False
    # test the case `pad_to_stride` is False
    patch_merge = PatchMerging(
        in_channels=in_c,
        out_channels=out_c,
        kernel_size=kernel_size,
        stride=stride,
        padding=padding,
        dilation=dilation,
        bias=bias)
    B, L, C = 1, 100, 3
    input_size = (10, 10)
    x = torch.rand(B, L, C)
    x_out, out_size = patch_merge(x, input_size)
    assert x_out.size() == (1, 16, 4)
    assert out_size == (4, 4)
    # assert out size is consistent with real output
    assert x_out.size(1) == out_size[0] * out_size[1]
    in_c = 4
    out_c = 5
    kernel_size = 6
    stride = 3
    padding = 2
    dilation = 2
    bias = False
    patch_merge = PatchMerging(
        in_channels=in_c,
        out_channels=out_c,
        kernel_size=kernel_size,
        stride=stride,
        padding=padding,
        dilation=dilation,
        bias=bias)
    B, L, C = 1, 100, 4
    input_size = (10, 10)
    x = torch.rand(B, L, C)
    x_out, out_size = patch_merge(x, input_size)
    assert x_out.size() == (1, 4, 5)
    assert out_size == (2, 2)
    # assert out size is consistent with real output
    assert x_out.size(1) == out_size[0] * out_size[1]

    # Test with adaptive padding
    for padding in ('same', 'corner'):
        in_c = 2
        out_c = 3
        B = 2

        # test stride is 1
        input_size = (5, 5)
        kernel_size = (5, 5)
        stride = (1, 1)
        dilation = 1
        bias = False
        L = input_size[0] * input_size[1]

        x = torch.rand(B, L, in_c)
        patch_merge = PatchMerging(
            in_channels=in_c,
            out_channels=out_c,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_merge(x, input_size)
        assert x_out.size() == (B, 25, 3)
        assert out_size == (5, 5)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test kernel_size == stride
        input_size = (5, 5)
        kernel_size = (5, 5)
        stride = (5, 5)
        dilation = 1
        bias = False
        L = input_size[0] * input_size[1]

        x = torch.rand(B, L, in_c)
        patch_merge = PatchMerging(
            in_channels=in_c,
            out_channels=out_c,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_merge(x, input_size)
        assert x_out.size() == (B, 1, 3)
        assert out_size == (1, 1)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test kernel_size == stride
        input_size = (6, 5)
        kernel_size = (5, 5)
        stride = (5, 5)
        dilation = 1
        bias = False
        L = input_size[0] * input_size[1]

        x = torch.rand(B, L, in_c)
        patch_merge = PatchMerging(
            in_channels=in_c,
            out_channels=out_c,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_merge(x, input_size)
        assert x_out.size() == (B, 2, 3)
        assert out_size == (2, 1)
        assert x_out.size(1) == out_size[0] * out_size[1]

        # test different kernel_size with different stride
        input_size = (6, 5)
        kernel_size = (6, 2)
        stride = (6, 2)
        dilation = 1
        bias = False
        L = input_size[0] * input_size[1]

        x = torch.rand(B, L, in_c)
        patch_merge = PatchMerging(
            in_channels=in_c,
            out_channels=out_c,
            kernel_size=kernel_size,
            stride=stride,
            padding=padding,
            dilation=dilation,
            bias=bias)

        x_out, out_size = patch_merge(x, input_size)
        assert x_out.size() == (B, 3, 3)
        assert out_size == (1, 3)
        assert x_out.size(1) == out_size[0] * out_size[1]


def test_multiheadattention():
    MultiheadAttention(
        embed_dims=5,
        num_heads=5,
        attn_drop=0,
        proj_drop=0,
        dropout_layer=dict(type='Dropout', drop_prob=0.),
        batch_first=True)
    batch_dim = 2
    embed_dim = 5
    num_query = 100
    attn_batch_first = MultiheadAttention(
        embed_dims=5,
        num_heads=5,
        attn_drop=0,
        proj_drop=0,
        dropout_layer=dict(type='DropPath', drop_prob=0.),
        batch_first=True)

    attn_query_first = MultiheadAttention(
        embed_dims=5,
        num_heads=5,
        attn_drop=0,
        proj_drop=0,
        dropout_layer=dict(type='DropPath', drop_prob=0.),
        batch_first=False)

    param_dict = dict(attn_query_first.named_parameters())
    for n, v in attn_batch_first.named_parameters():
        param_dict[n].data = v.data

    input_batch_first = torch.rand(batch_dim, num_query, embed_dim)
    input_query_first = input_batch_first.transpose(0, 1)

    assert torch.allclose(
        attn_query_first(input_query_first).sum(),
        attn_batch_first(input_batch_first).sum())

    key_batch_first = torch.rand(batch_dim, num_query, embed_dim)
    key_query_first = key_batch_first.transpose(0, 1)

    assert torch.allclose(
        attn_query_first(input_query_first, key_query_first).sum(),
        attn_batch_first(input_batch_first, key_batch_first).sum())

    identity = torch.ones_like(input_query_first)

    # check deprecated arguments can be used normally

    assert torch.allclose(
        attn_query_first(
            input_query_first, key_query_first, residual=identity).sum(),
        attn_batch_first(input_batch_first, key_batch_first).sum() +
        identity.sum() - input_batch_first.sum())

    assert torch.allclose(
        attn_query_first(
            input_query_first, key_query_first, identity=identity).sum(),
        attn_batch_first(input_batch_first, key_batch_first).sum() +
        identity.sum() - input_batch_first.sum())

    attn_query_first(
        input_query_first, key_query_first, identity=identity).sum(),


def test_ffn():
    with pytest.raises(AssertionError):
        # num_fcs should be no less than 2
        FFN(num_fcs=1)
    FFN(dropout=0, add_residual=True)
    ffn = FFN(dropout=0, add_identity=True)

    input_tensor = torch.rand(2, 20, 256)
    input_tensor_nbc = input_tensor.transpose(0, 1)
    assert torch.allclose(ffn(input_tensor).sum(), ffn(input_tensor_nbc).sum())
    residual = torch.rand_like(input_tensor)
    torch.allclose(
        ffn(input_tensor, residual=residual).sum(),
        ffn(input_tensor).sum() + residual.sum() - input_tensor.sum())

    torch.allclose(
        ffn(input_tensor, identity=residual).sum(),
        ffn(input_tensor).sum() + residual.sum() - input_tensor.sum())


@pytest.mark.skipif(not torch.cuda.is_available(), reason='Cuda not available')
def test_basetransformerlayer_cuda():
    # To test if the BaseTransformerLayer's behaviour remains
    # consistent after being deepcopied
    operation_order = ('self_attn', 'ffn')
    baselayer = BaseTransformerLayer(
        operation_order=operation_order,
        batch_first=True,
        attn_cfgs=dict(
            type='MultiheadAttention',
            embed_dims=256,
            num_heads=8,
        ),
    )
    baselayers = ModuleList([copy.deepcopy(baselayer) for _ in range(2)])
    baselayers.to('cuda')
    x = torch.rand(2, 10, 256).cuda()
    for m in baselayers:
        x = m(x)
        assert x.shape == torch.Size([2, 10, 256])


@pytest.mark.parametrize('embed_dims', [False, 256])
def test_basetransformerlayer(embed_dims):
    attn_cfgs = dict(type='MultiheadAttention', embed_dims=256, num_heads=8),
    if embed_dims:
        ffn_cfgs = dict(
            type='FFN',
            embed_dims=embed_dims,
            feedforward_channels=1024,
            num_fcs=2,
            ffn_drop=0.,
            act_cfg=dict(type='ReLU', inplace=True),
        )
    else:
        ffn_cfgs = dict(
            type='FFN',
            feedforward_channels=1024,
            num_fcs=2,
            ffn_drop=0.,
            act_cfg=dict(type='ReLU', inplace=True),
        )

    feedforward_channels = 2048
    ffn_dropout = 0.1
    operation_order = ('self_attn', 'norm', 'ffn', 'norm')

    # test deprecated_args
    baselayer = BaseTransformerLayer(
        attn_cfgs=attn_cfgs,
        ffn_cfgs=ffn_cfgs,
        feedforward_channels=feedforward_channels,
        ffn_dropout=ffn_dropout,
        operation_order=operation_order)
    assert baselayer.batch_first is False
    assert baselayer.ffns[0].feedforward_channels == feedforward_channels

    attn_cfgs = dict(type='MultiheadAttention', num_heads=8, embed_dims=256),
    feedforward_channels = 2048
    ffn_dropout = 0.1
    operation_order = ('self_attn', 'norm', 'ffn', 'norm')
    baselayer = BaseTransformerLayer(
        attn_cfgs=attn_cfgs,
        feedforward_channels=feedforward_channels,
        ffn_dropout=ffn_dropout,
        operation_order=operation_order,
        batch_first=True)
    assert baselayer.attentions[0].batch_first
    in_tensor = torch.rand(2, 10, 256)
    baselayer(in_tensor)


def test_transformerlayersequence():
    squeue = TransformerLayerSequence(
        num_layers=6,
        transformerlayers=dict(
            type='BaseTransformerLayer',
            attn_cfgs=[
                dict(
                    type='MultiheadAttention',
                    embed_dims=256,
                    num_heads=8,
                    dropout=0.1),
                dict(type='MultiheadAttention', embed_dims=256, num_heads=4)
            ],
            feedforward_channels=1024,
            ffn_dropout=0.1,
            operation_order=('self_attn', 'norm', 'cross_attn', 'norm', 'ffn',
                             'norm')))
    assert len(squeue.layers) == 6
    assert squeue.pre_norm is False
    with pytest.raises(AssertionError):
        # if transformerlayers is a list, len(transformerlayers)
        # should be equal to num_layers
        TransformerLayerSequence(
            num_layers=6,
            transformerlayers=[
                dict(
                    type='BaseTransformerLayer',
                    attn_cfgs=[
                        dict(
                            type='MultiheadAttention',
                            embed_dims=256,
                            num_heads=8,
                            dropout=0.1),
                        dict(type='MultiheadAttention', embed_dims=256)
                    ],
                    feedforward_channels=1024,
                    ffn_dropout=0.1,
                    operation_order=('self_attn', 'norm', 'cross_attn', 'norm',
                                     'ffn', 'norm'))
            ])


def test_drop_path():
    drop_path = DropPath(drop_prob=0)
    test_in = torch.rand(2, 3, 4, 5)
    assert test_in is drop_path(test_in)

    drop_path = DropPath(drop_prob=0.1)
    drop_path.training = False
    test_in = torch.rand(2, 3, 4, 5)
    assert test_in is drop_path(test_in)
    drop_path.training = True
    assert test_in is not drop_path(test_in)