mobilenet.py 6.79 KB
Newer Older
Francisco Massa's avatar
Francisco Massa committed
1
from torch import nn
2
3
4
5
6
7
8
9
10
from .utils import load_state_dict_from_url


__all__ = ['MobileNetV2', 'mobilenet_v2']


model_urls = {
    'mobilenet_v2': 'https://download.pytorch.org/models/mobilenet_v2-b0353104.pth',
}
Francisco Massa's avatar
Francisco Massa committed
11
12


13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
def _make_divisible(v, divisor, min_value=None):
    """
    This function is taken from the original tf repo.
    It ensures that all layers have a channel number that is divisible by 8
    It can be seen here:
    https://github.com/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet.py
    :param v:
    :param divisor:
    :param min_value:
    :return:
    """
    if min_value is None:
        min_value = divisor
    new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_v < 0.9 * v:
        new_v += divisor
    return new_v


Francisco Massa's avatar
Francisco Massa committed
33
class ConvBNReLU(nn.Sequential):
34
    def __init__(self, in_planes, out_planes, kernel_size=3, stride=1, groups=1, norm_layer=None):
Francisco Massa's avatar
Francisco Massa committed
35
        padding = (kernel_size - 1) // 2
36
37
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
Francisco Massa's avatar
Francisco Massa committed
38
39
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_planes, out_planes, kernel_size, stride, padding, groups=groups, bias=False),
40
            norm_layer(out_planes),
Francisco Massa's avatar
Francisco Massa committed
41
42
43
44
45
            nn.ReLU6(inplace=True)
        )


class InvertedResidual(nn.Module):
46
    def __init__(self, inp, oup, stride, expand_ratio, norm_layer=None):
Francisco Massa's avatar
Francisco Massa committed
47
48
49
50
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]

51
52
53
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d

Francisco Massa's avatar
Francisco Massa committed
54
55
56
57
58
59
        hidden_dim = int(round(inp * expand_ratio))
        self.use_res_connect = self.stride == 1 and inp == oup

        layers = []
        if expand_ratio != 1:
            # pw
60
            layers.append(ConvBNReLU(inp, hidden_dim, kernel_size=1, norm_layer=norm_layer))
Francisco Massa's avatar
Francisco Massa committed
61
62
        layers.extend([
            # dw
63
            ConvBNReLU(hidden_dim, hidden_dim, stride=stride, groups=hidden_dim, norm_layer=norm_layer),
Francisco Massa's avatar
Francisco Massa committed
64
65
            # pw-linear
            nn.Conv2d(hidden_dim, oup, 1, 1, 0, bias=False),
66
            norm_layer(oup),
Francisco Massa's avatar
Francisco Massa committed
67
68
69
70
71
72
73
74
75
76
77
        ])
        self.conv = nn.Sequential(*layers)

    def forward(self, x):
        if self.use_res_connect:
            return x + self.conv(x)
        else:
            return self.conv(x)


class MobileNetV2(nn.Module):
78
79
80
81
82
    def __init__(self,
                 num_classes=1000,
                 width_mult=1.0,
                 inverted_residual_setting=None,
                 round_nearest=8,
83
84
                 block=None,
                 norm_layer=None):
85
86
87
88
89
90
91
92
93
        """
        MobileNet V2 main class

        Args:
            num_classes (int): Number of classes
            width_mult (float): Width multiplier - adjusts number of channels in each layer by this amount
            inverted_residual_setting: Network structure
            round_nearest (int): Round the number of channels in each layer to be a multiple of this number
            Set to 1 to turn off rounding
94
            block: Module specifying inverted residual building block for mobilenet
95
            norm_layer: Module specifying the normalization layer to use
96

97
        """
Francisco Massa's avatar
Francisco Massa committed
98
        super(MobileNetV2, self).__init__()
99
100
101

        if block is None:
            block = InvertedResidual
102
103
104
105

        if norm_layer is None:
            norm_layer = nn.BatchNorm2d

Francisco Massa's avatar
Francisco Massa committed
106
107
        input_channel = 32
        last_channel = 1280
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124

        if inverted_residual_setting is None:
            inverted_residual_setting = [
                # t, c, n, s
                [1, 16, 1, 1],
                [6, 24, 2, 2],
                [6, 32, 3, 2],
                [6, 64, 4, 2],
                [6, 96, 3, 1],
                [6, 160, 3, 2],
                [6, 320, 1, 1],
            ]

        # only check the first element, assuming user knows t,c,n,s are required
        if len(inverted_residual_setting) == 0 or len(inverted_residual_setting[0]) != 4:
            raise ValueError("inverted_residual_setting should be non-empty "
                             "or a 4-element list, got {}".format(inverted_residual_setting))
Francisco Massa's avatar
Francisco Massa committed
125
126

        # building first layer
127
128
        input_channel = _make_divisible(input_channel * width_mult, round_nearest)
        self.last_channel = _make_divisible(last_channel * max(1.0, width_mult), round_nearest)
129
        features = [ConvBNReLU(3, input_channel, stride=2, norm_layer=norm_layer)]
Francisco Massa's avatar
Francisco Massa committed
130
131
        # building inverted residual blocks
        for t, c, n, s in inverted_residual_setting:
132
            output_channel = _make_divisible(c * width_mult, round_nearest)
Francisco Massa's avatar
Francisco Massa committed
133
134
            for i in range(n):
                stride = s if i == 0 else 1
135
                features.append(block(input_channel, output_channel, stride, expand_ratio=t, norm_layer=norm_layer))
Francisco Massa's avatar
Francisco Massa committed
136
137
                input_channel = output_channel
        # building last several layers
138
        features.append(ConvBNReLU(input_channel, self.last_channel, kernel_size=1, norm_layer=norm_layer))
Francisco Massa's avatar
Francisco Massa committed
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
        # make it nn.Sequential
        self.features = nn.Sequential(*features)

        # building classifier
        self.classifier = nn.Sequential(
            nn.Dropout(0.2),
            nn.Linear(self.last_channel, num_classes),
        )

        # weight initialization
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, mode='fan_out')
                if m.bias is not None:
                    nn.init.zeros_(m.bias)
154
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
Francisco Massa's avatar
Francisco Massa committed
155
156
157
158
159
160
                nn.init.ones_(m.weight)
                nn.init.zeros_(m.bias)
            elif isinstance(m, nn.Linear):
                nn.init.normal_(m.weight, 0, 0.01)
                nn.init.zeros_(m.bias)

161
162
163
    def _forward_impl(self, x):
        # This exists since TorchScript doesn't support inheritance, so the superclass method
        # (this one) needs to have a name other than `forward` that can be accessed in a subclass
Francisco Massa's avatar
Francisco Massa committed
164
        x = self.features(x)
165
166
        # Cannot use "squeeze" as batch-size can be 1 => must use reshape with x.shape[0]
        x = nn.functional.adaptive_avg_pool2d(x, 1).reshape(x.shape[0], -1)
Francisco Massa's avatar
Francisco Massa committed
167
168
169
        x = self.classifier(x)
        return x

170
171
    def forward(self, x):
        return self._forward_impl(x)
172

Francisco Massa's avatar
Francisco Massa committed
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
def mobilenet_v2(pretrained=False, progress=True, **kwargs):
    """
    Constructs a MobileNetV2 architecture from
    `"MobileNetV2: Inverted Residuals and Linear Bottlenecks" <https://arxiv.org/abs/1801.04381>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    model = MobileNetV2(**kwargs)
    if pretrained:
        state_dict = load_state_dict_from_url(model_urls['mobilenet_v2'],
                                              progress=progress)
        model.load_state_dict(state_dict)
    return model