File size: 7,616 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
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
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.

import itertools
import unittest

import numpy as np
import torch
from pytorchvideo.layers.convolutions import (
    Conv2plus1d,
    ConvReduce3D,
    create_conv_2plus1d,
)
from torch import nn


class TestConvReduce3D(unittest.TestCase):
    def setUp(self):
        super().setUp()
        torch.set_rng_state(torch.manual_seed(42).get_state())

    def test_create_stack_conv(self):
        """
        Test ConvReduce3D.
        """
        for input_dim, output_dim in itertools.product((2, 4), (4, 8, 16)):
            model = ConvReduce3D(
                in_channels=input_dim,
                out_channels=output_dim,
                kernel_size=((1, 1, 1), (3, 3, 3), (1, 3, 3)),
                stride=((1, 1, 1), (1, 1, 1), None),
                padding=((0, 0, 0), (1, 1, 1), (0, 1, 1)),
                dilation=((2, 2, 2), (1, 1, 1), None),
                groups=(1, 2, None),
                bias=(True, False, None),
            )
            model_gt_list = [
                nn.Conv3d(
                    in_channels=input_dim,
                    out_channels=output_dim,
                    kernel_size=(1, 1, 1),
                    stride=(1, 1, 1),
                    padding=(0, 0, 0),
                    dilation=(2, 2, 2),
                    groups=1,
                    bias=True,
                ),
                nn.Conv3d(
                    in_channels=input_dim,
                    out_channels=output_dim,
                    kernel_size=(3, 3, 3),
                    stride=(1, 1, 1),
                    padding=(1, 1, 1),
                    dilation=(1, 1, 1),
                    groups=2,
                    bias=False,
                ),
                nn.Conv3d(
                    in_channels=input_dim,
                    out_channels=output_dim,
                    kernel_size=(1, 3, 3),
                    padding=(0, 1, 1),
                ),
            ]
            model.convs[0].load_state_dict(
                model_gt_list[0].state_dict(), strict=True
            )  # explicitly use strict mode.
            model.convs[1].load_state_dict(
                model_gt_list[1].state_dict(), strict=True
            )  # explicitly use strict mode.
            model.convs[2].load_state_dict(
                model_gt_list[2].state_dict(), strict=True
            )  # explicitly use strict mode.

            # Test forwarding.
            for tensor in TestConvReduce3D._get_inputs(input_dim):
                if tensor.shape[1] != input_dim:
                    with self.assertRaises(RuntimeError):
                        output_tensor = model(tensor)
                    continue
                else:
                    output_tensor = model(tensor)
                    output_gt = []
                    for ind in range(3):
                        output_gt.append(model_gt_list[ind](tensor))
                    output_tensor_gt = torch.stack(output_gt, dim=0).sum(
                        dim=0, keepdim=False
                    )

                self.assertEqual(
                    output_tensor.shape,
                    output_tensor_gt.shape,
                    "Output shape {} is different from expected shape {}".format(
                        output_tensor.shape, output_tensor_gt.shape
                    ),
                )

    @staticmethod
    def _get_inputs(input_dim: int = 3) -> torch.tensor:
        """
        Provide different tensors as test cases.

        Yield:
            (torch.tensor): tensor as test case input.
        """
        # Prepare random tensor as test cases.
        shapes = (
            # Forward succeeded.
            (1, input_dim, 3, 7, 7),
            (1, input_dim, 5, 7, 7),
            (1, input_dim, 7, 7, 7),
            (2, input_dim, 3, 7, 7),
            (4, input_dim, 3, 7, 7),
            (8, input_dim, 3, 7, 7),
            (2, input_dim, 3, 7, 14),
            (2, input_dim, 3, 14, 7),
            (2, input_dim, 3, 14, 14),
            # Forward failed.
            (8, input_dim * 2, 3, 7, 7),
            (8, input_dim * 4, 5, 7, 7),
        )
        for shape in shapes:
            yield torch.rand(shape)


class TestConv2plus1d(unittest.TestCase):
    def setUp(self):
        super().setUp()
        torch.set_rng_state(torch.manual_seed(42).get_state())

    def test_create_2plus1d_conv(self):
        """
        Test Conv2plus1d.
        """
        for input_dim, output_dim in itertools.product((2, 4), (4, 8, 16)):
            model = Conv2plus1d(
                conv_t=nn.Conv3d(
                    in_channels=input_dim,
                    out_channels=output_dim,
                    kernel_size=(3, 1, 1),
                    stride=(2, 1, 1),
                    padding=(1, 0, 0),
                    bias=False,
                ),
                norm=nn.BatchNorm3d(output_dim),
                activation=nn.ReLU(),
                conv_xy=nn.Conv3d(
                    in_channels=output_dim,
                    out_channels=output_dim,
                    kernel_size=(1, 3, 3),
                    stride=(1, 2, 2),
                    padding=(0, 1, 1),
                    bias=False,
                ),
            )

            model_gt = create_conv_2plus1d(
                in_channels=input_dim,
                out_channels=output_dim,
                kernel_size=(3, 3, 3),
                stride=(2, 2, 2),
                padding=(1, 1, 1),
                bias=False,
                norm=nn.BatchNorm3d,
                norm_eps=1e-5,
                norm_momentum=0.1,
                activation=nn.ReLU,
            )

            model.load_state_dict(
                model_gt.state_dict(), strict=True
            )  # explicitly use strict mode.

            # Test forwarding.
            for input_tensor in TestConv2plus1d._get_inputs():
                with torch.no_grad():
                    if input_tensor.shape[1] != input_dim:
                        with self.assertRaises(RuntimeError):
                            output_tensor = model(input_tensor)
                        continue
                    else:
                        output_tensor = model(input_tensor)
                        output_tensor_gt = model_gt(input_tensor)
                self.assertEqual(
                    output_tensor.shape,
                    output_tensor_gt.shape,
                    "Output shape {} is different from expected shape {}".format(
                        output_tensor.shape, output_tensor_gt.shape
                    ),
                )
                self.assertTrue(
                    np.allclose(output_tensor.numpy(), output_tensor_gt.numpy())
                )

    @staticmethod
    def _get_inputs(input_dim: int = 3) -> torch.tensor:
        """
        Provide different tensors as test cases.

        Yield:
            (torch.tensor): tensor as test case input.
        """
        # Prepare random tensor as test cases.
        shapes = (
            # Forward succeeded.
            (1, input_dim, 3, 7, 7),
            (1, input_dim, 5, 7, 7),
            (1, input_dim, 7, 7, 7),
            (2, input_dim, 3, 7, 7),
            (4, input_dim, 3, 7, 7),
            (8, input_dim, 3, 7, 7),
            (2, input_dim, 3, 7, 14),
            (2, input_dim, 3, 14, 7),
            (2, input_dim, 3, 14, 14),
            # Forward failed.
            (8, input_dim * 2, 3, 7, 7),
            (8, input_dim * 4, 5, 7, 7),
        )
        for shape in shapes:
            yield torch.rand(shape)