MnasNet 神经网络简介与代码实战

1.介绍

    MnasNet是由Google大脑AutoML组所提出的，以运算时长为优化目标来搜索网络结构，同时模型要能平衡精确度和运算时长。更加详细的介绍可以参见：MnasNet: Platform-Aware Neural Architecture Search for Mobile

 

2.模型结构

  下图中的可分离卷积可参考我的一篇博客（MobileNets V1神经网络简介与代码实战）

 

3.模型特点

（1）在depthwise convolution中，5X5卷积核的效果要优于两个3X3的卷积核
（2）层分级的重要性。很多轻量化模型重复 block 架构，只改变滤波器尺寸和空间维度。论文提出的层级搜索空间允许模型的各个 block 包括不同的卷积层。通过比较 MnasNet 的各种变体（即单独重复使用各个 block），在准确率和实时性方面难以达到平衡，验证了层分级的重要性。

 

4.代码实现 pytorch

def Conv_3x3(inp, oup, stride):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 3, stride, 1, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU6(inplace=True)
    )


def Conv_1x1(inp, oup):
    return nn.Sequential(
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
        nn.ReLU6(inplace=True)
    )

def SepConv_3x3(inp, oup): #input=32, output=16
    return nn.Sequential(
        # dw
        nn.Conv2d(inp, inp , 3, 1, 1, groups=inp, bias=False),
        nn.BatchNorm2d(inp),
        nn.ReLU6(inplace=True),
        # pw-linear
        nn.Conv2d(inp, oup, 1, 1, 0, bias=False),
        nn.BatchNorm2d(oup),
    )


class InvertedResidual(nn.Module):
    def __init__(self, inp, oup, stride, expand_ratio, kernel):
        super(InvertedResidual, self).__init__()
        self.stride = stride
        assert stride in [1, 2]

        self.use_res_connect = self.stride == 1 and inp == oup

        self.conv = nn.Sequential(
            # pw
            nn.Conv2d(inp, inp * expand_ratio, 1, 1, 0, bias=False),
            nn.BatchNorm2d(inp * expand_ratio),
            nn.ReLU6(inplace=True),
            # dw
            nn.Conv2d(inp * expand_ratio, inp * expand_ratio, kernel, stride, kernel // 2, groups=inp * expand_ratio, bias=False),
            nn.BatchNorm2d(inp * expand_ratio),
            nn.ReLU6(inplace=True),
            # pw-linear
            nn.Conv2d(inp * expand_ratio, oup, 1, 1, 0, bias=False),
            nn.BatchNorm2d(oup),
        )

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


class MnasNet(nn.Module):
    def __init__(self, n_class=1000, input_size=224, width_mult=1.):
        super(MnasNet, self).__init__()

        # setting of inverted residual blocks
        self.interverted_residual_setting = [
            # t, c, n, s, k
            [3, 24,  3, 2, 3],  # -> 56x56
            [3, 40,  3, 2, 5],  # -> 28x28
            [6, 80,  3, 2, 5],  # -> 14x14
            [6, 96,  2, 1, 3],  # -> 14x14
            [6, 192, 4, 2, 5],  # -> 7x7
            [6, 320, 1, 1, 3],  # -> 7x7
        ]

        assert input_size % 32 == 0
        input_channel = int(32 * width_mult)
        self.last_channel = int(1280 * width_mult) if width_mult > 1.0 else 1280

        # building first two layer
        self.features = [Conv_3x3(3, input_channel, 2), SepConv_3x3(input_channel, 16)]
        input_channel = 16

        # building inverted residual blocks (MBConv)
        for t, c, n, s, k in self.interverted_residual_setting:
            output_channel = int(c * width_mult)
            for i in range(n):
                if i == 0:
                    self.features.append(InvertedResidual(input_channel, output_channel, s, t, k))
                else:
                    self.features.append(InvertedResidual(input_channel, output_channel, 1, t, k))
                input_channel = output_channel

        # building last several layers
        self.features.append(Conv_1x1(input_channel, self.last_channel))
        self.features.append(nn.AdaptiveAvgPool2d(1))

        # make it nn.Sequential
        self.features = nn.Sequential(*self.features)

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

        self._initialize_weights()

    def forward(self, x):
        x = self.features(x)
        x = x.view(-1, self.last_channel)
        x = self.classifier(x)
        return x

    def _initialize_weights(self):
        for m in self.modules():
            if isinstance(m, nn.Conv2d):
                n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
                m.weight.data.normal_(0, math.sqrt(2. / n))
                if m.bias is not None:
                    m.bias.data.zero_()
            elif isinstance(m, nn.BatchNorm2d):
                m.weight.data.fill_(1)
                m.bias.data.zero_()
            elif isinstance(m, nn.Linear):
                n = m.weight.size(1)
                m.weight.data.normal_(0, 0.01)
                m.bias.data.zero_()