File size: 6,139 Bytes
3133fdb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.

import itertools
import unittest
import warnings

import torch
from pytorchvideo.models.vision_transformers import (
    create_multiscale_vision_transformers,
)


class TestVisionTransformers(unittest.TestCase):
    def setUp(self):
        super().setUp()
        torch.set_rng_state(torch.manual_seed(42).get_state())

    def test_create_mvit(self):
        """
        Test MViT.
        """
        # Test MViT with 3D case.
        num_head = 100
        batch_size = 1
        fake_input = torch.rand(batch_size, 3, 4, 28, 28)
        model = create_multiscale_vision_transformers(
            spatial_size=28,
            temporal_size=4,
            patch_embed_dim=12,
            depth=1,
            head_num_classes=num_head,
            pool_kv_stride_adaptive=[1, 2, 2],
        )
        output = model(fake_input)
        gt_shape_tensor = torch.rand(batch_size, num_head)
        self.assertEqual(output.shape, gt_shape_tensor.shape)
        # Test MViT with 3D case with pool first.
        num_head = 100
        batch_size = 1
        fake_input = torch.rand(batch_size, 3, 4, 28, 28)
        model = create_multiscale_vision_transformers(
            spatial_size=28,
            temporal_size=4,
            patch_embed_dim=12,
            depth=1,
            head_num_classes=num_head,
            pool_first=True,
            pool_q_stride_size=[[0, 1, 2, 2]],
        )
        output = model(fake_input)
        gt_shape_tensor = torch.rand(batch_size, num_head)
        self.assertEqual(output.shape, gt_shape_tensor.shape)

        # Test MViT with 2D case for images.
        conv_patch_kernel = (7, 7)
        conv_patch_stride = (4, 4)
        conv_patch_padding = (3, 3)
        num_head = 100
        batch_size = 1
        fake_input = torch.rand(batch_size, 3, 28, 28)
        model = create_multiscale_vision_transformers(
            spatial_size=(28, 28),
            temporal_size=1,
            patch_embed_dim=12,
            depth=1,
            head_num_classes=num_head,
            use_2d_patch=True,
            conv_patch_embed_kernel=conv_patch_kernel,
            conv_patch_embed_stride=conv_patch_stride,
            conv_patch_embed_padding=conv_patch_padding,
        )
        output = model(fake_input)
        gt_shape_tensor = torch.rand(batch_size, num_head)
        self.assertEqual(output.shape, gt_shape_tensor.shape)

        # Test MViT without patch_embed.
        conv_patch_kernel = (7, 7)
        conv_patch_stride = (4, 4)
        conv_patch_padding = (3, 3)
        num_head = 100
        batch_size = 1
        fake_input = torch.rand(batch_size, 8, 12)
        model = create_multiscale_vision_transformers(
            spatial_size=(8, 1),
            temporal_size=1,
            patch_embed_dim=12,
            depth=1,
            enable_patch_embed=False,
            head_num_classes=num_head,
        )
        output = model(fake_input)
        gt_shape_tensor = torch.rand(batch_size, num_head)
        self.assertEqual(output.shape, gt_shape_tensor.shape)

        self.assertRaises(
            AssertionError,
            create_multiscale_vision_transformers,
            spatial_size=28,
            temporal_size=4,
            use_2d_patch=True,
        )

        self.assertRaises(
            AssertionError,
            create_multiscale_vision_transformers,
            spatial_size=28,
            temporal_size=1,
            pool_kv_stride_adaptive=[[2, 2, 2]],
            pool_kv_stride_size=[[1, 1, 2, 2]],
        )

        self.assertRaises(
            NotImplementedError,
            create_multiscale_vision_transformers,
            spatial_size=28,
            temporal_size=1,
            norm="fakenorm",
        )

    def test_mvit_is_torchscriptable(self):
        batch_size = 2
        num_head = 4
        spatial_size = (28, 28)
        temporal_size = 4
        depth = 2
        patch_embed_dim = 96

        # The following binary settings are covered by `test_layers_attention.py`:
        # `qkv_bias`, `depthwise_conv`, `separate_qkv`, `bias_on` `pool_first`
        # `residual_pool`
        true_false_opts = [
            "cls_embed_on",
            "sep_pos_embed",
            "enable_patch_embed",
            "enable_patch_embed_norm",
        ]

        # Loop over `2 ^ len(true_false_opts)` configurations
        for true_false_settings in itertools.product(
            *([[True, False]] * len(true_false_opts))
        ):
            named_tf_settings = dict(zip(true_false_opts, true_false_settings))

            model = create_multiscale_vision_transformers(
                spatial_size=spatial_size,
                temporal_size=temporal_size,
                depth=depth,
                head_num_classes=num_head,
                patch_embed_dim=patch_embed_dim,
                pool_kv_stride_adaptive=[1, 2, 2],
                **named_tf_settings,
                create_scriptable_model=False,
            ).eval()
            ts_model = torch.jit.script(model)

            input_shape = (
                (3, temporal_size, spatial_size[0], spatial_size[1])
                if named_tf_settings["enable_patch_embed"]
                else (
                    temporal_size * spatial_size[0] * spatial_size[1],
                    patch_embed_dim,
                )
            )
            fake_input = torch.rand(batch_size, *input_shape)

            expected = model(fake_input)
            actual = ts_model(fake_input)
            torch.testing.assert_allclose(expected, actual)

    def test_mvit_create_scriptable_model_is_deprecated(self):
        with warnings.catch_warnings(record=True) as w:
            warnings.simplefilter("always")
            create_multiscale_vision_transformers(
                spatial_size=28,
                temporal_size=4,
                norm="batchnorm",
                depth=2,
                head_num_classes=100,
                create_scriptable_model=True,
            )

        assert len(w) == 1
        assert issubclass(w[-1].category, DeprecationWarning)