File size: 4,896 Bytes
d670799
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
import math
from copy import deepcopy
from unittest import TestCase

import torch

from mmaction.models import MViT


class TestMViT(TestCase):

    def setUp(self):
        self.cfg = dict(arch='tiny', drop_path_rate=0.1)

    def test_structure(self):
        # Test invalid default arch
        with self.assertRaisesRegex(AssertionError, 'not in default archs'):
            cfg = deepcopy(self.cfg)
            cfg['arch'] = 'unknown'
            MViT(**cfg)

        # Test invalid custom arch
        with self.assertRaisesRegex(AssertionError, 'Custom arch needs'):
            cfg = deepcopy(self.cfg)
            cfg['arch'] = {
                'num_layers': 24,
                'num_heads': 16,
                'feedforward_channels': 4096
            }
            MViT(**cfg)

        # Test custom arch
        cfg = deepcopy(self.cfg)
        cfg['arch'] = {
            'embed_dims': 96,
            'num_layers': 10,
            'num_heads': 1,
            'downscale_indices': [2, 5, 8]
        }
        stage_indices = [0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3]
        model = MViT(**cfg)
        self.assertEqual(model.embed_dims, 96)
        self.assertEqual(model.num_layers, 10)
        for i, block in enumerate(model.blocks):
            stage = stage_indices[i]
            self.assertEqual(block.out_dims, 96 * 2**(stage))

        # Test out_indices
        cfg = deepcopy(self.cfg)
        cfg['out_scales'] = {1: 1}
        with self.assertRaisesRegex(AssertionError, "get <class 'dict'>"):
            MViT(**cfg)
        cfg['out_scales'] = [0, 13]
        with self.assertRaisesRegex(AssertionError, 'Invalid out_scales 13'):
            MViT(**cfg)

        # Test model structure
        cfg = deepcopy(self.cfg)
        model = MViT(**cfg)
        stage_indices = [0, 1, 1, 2, 2, 2, 2, 2, 3, 3, 3]
        self.assertEqual(len(model.blocks), 10)
        dpr_inc = 0.1 / (10 - 1)
        dpr = 0
        for i, block in enumerate(model.blocks):
            stage = stage_indices[i]
            print(i, stage)
            self.assertEqual(block.attn.num_heads, 2**stage)
            if dpr > 0:
                self.assertAlmostEqual(block.drop_path.drop_prob, dpr)
            dpr += dpr_inc

    def test_init_weights(self):
        # test weight init cfg
        cfg = deepcopy(self.cfg)
        cfg['init_cfg'] = [
            dict(
                type='Kaiming',
                layer='Conv3d',
                mode='fan_in',
                nonlinearity='linear')
        ]
        cfg['use_abs_pos_embed'] = True
        model = MViT(**cfg)
        ori_weight = model.patch_embed.projection.weight.clone().detach()
        # The pos_embed is all zero before initialize
        self.assertTrue(torch.allclose(model.pos_embed, torch.tensor(0.)))

        model.init_weights()
        initialized_weight = model.patch_embed.projection.weight
        self.assertFalse(torch.allclose(ori_weight, initialized_weight))
        self.assertFalse(torch.allclose(model.pos_embed, torch.tensor(0.)))

    def test_forward(self):
        imgs = torch.randn(1, 3, 6, 64, 64)

        cfg = deepcopy(self.cfg)
        model = MViT(**cfg)
        outs = model(imgs)
        self.assertIsInstance(outs, tuple)
        self.assertEqual(len(outs), 1)
        patch_token, cls_token = outs[-1]
        self.assertEqual(patch_token.shape, (1, 768, 3, 2, 2))

        # Test forward with multi out scales
        cfg = deepcopy(self.cfg)
        cfg['out_scales'] = (0, 1, 2, 3)
        model = MViT(**cfg)
        outs = model(imgs)
        self.assertIsInstance(outs, tuple)
        self.assertEqual(len(outs), 4)
        for stage, out in enumerate(outs):
            stride = 2**stage
            patch_token, cls_token = out
            self.assertEqual(patch_token.shape,
                             (1, 96 * stride, 3, 16 // stride, 16 // stride))
            self.assertEqual(cls_token.shape, (1, 96 * stride))

        # Test forward with dynamic input size
        imgs1 = torch.randn(1, 3, 2, 64, 64)
        imgs2 = torch.randn(1, 3, 2, 96, 96)
        imgs3 = torch.randn(1, 3, 2, 96, 128)
        cfg = deepcopy(self.cfg)
        model = MViT(**cfg)
        for imgs in [imgs1, imgs2, imgs3]:
            outs = model(imgs)
            self.assertIsInstance(outs, tuple)
            self.assertEqual(len(outs), 1)
            patch_token, cls_token = outs[-1]
            expect_feat_shape = (math.ceil(imgs.shape[2] / 2),
                                 math.ceil(imgs.shape[3] / 32),
                                 math.ceil(imgs.shape[4] / 32))
            self.assertEqual(patch_token.shape, (1, 768, *expect_feat_shape))
            self.assertEqual(cls_token.shape, (1, 768))