【MobileNetV3】MobileNetV3网络结构详解


在看本文前,强烈建议先看一下之前写的 MobilenetV2

1 MobileNetV3创新点

  • bottleneck结构变了
  • 让网络更宽、更深,宽多少?深多少?采用NAS(Neural Architecture Search)搜索得到
  • 重新设计耗时层结构(针对NAS搜索的结构进行设计,咱可以不管)

2 block变成了什么样

2.1 总体介绍

参考大佬的图片进行解读,Mobilenetv2中的block如下图所示
mobilenetv2
Mobilenetv3中的block如下图所示

Mobilenetv3
可以发现,Mobilenetv3的block中加入了SE模块,更换了激活函数
SE模块下一节讲。
此处更新的激活函数在图中用NL(非线性)统一表示,因为用到的激活函数不一样,主要有hardswish、relu两种。
最后那个1x1降维投影层用的是线性激活(f(x)=x),也可以理解为没用激活。

2.2 SE模块理解

SE(Squeeze-and-Excitation) 模块类似于一个注意力模块,以在Mobilenetv3中的应用为例进行理解,如下图所示。
SE模块

2.3 ReLu6和hardswish激活函数理解

ReLu6激活函数如下图所示,相当于加了个最大值6进行限制。
relu6

hardswish激活函数如下图所示,相当于分成3段进行限制。
采用hardswish,计算速度相对较快,对量化过程友好

hardswish
hardswish图像

3 网络总体结构

作者针对不同需求,通过NAS得到两种结构,一个是MobilenetV3-Large,结构如下图:

MobilenetV3-Large
图中部分参数解释:

  • Input表示输入尺寸
  • Operator中的NBN表示不使用BN,最后的conv2d 1×1相当于全连接层的作用
  • exp size表示bottleneck中的第一层1×1卷积升维,维度升到多少(第一个bottleneck没有1×1卷积升维操作)
  • out表示bottleneck输出的channel个数
  • SE表示是否使用SE模块
  • NL表示使用何种激活函数,HS表示HardSwish,RE表示ReLu
  • s表示步长(s=2,长宽变为原来一半)

另一个是MobilenetV3-Small,结构如下图:

MobilenetV3-Small

4 代码解读

直接看代码注释即可,可运行

from typing import Callable, List, Optional

import torch
from torch import nn, Tensor
from torch.nn import functional as F
from functools import partial


# ------------------------------------------------------#
#   这个函数的目的是确保Channel个数能被8整除。
#   离它最近的8的倍数
#	很多嵌入式设备做优化时都采用这个准则
# ------------------------------------------------------#
def _make_divisible(ch, divisor=8, min_ch=None):
    if min_ch is None:
        min_ch = divisor
    # int(v + divisor / 2) // divisor * divisor:四舍五入到8
    new_ch = max(min_ch, int(ch + divisor / 2) // divisor * divisor)
    # Make sure that round down does not go down by more than 10%.
    if new_ch < 0.9 * ch:
        new_ch += divisor
    return new_ch


# -------------------------------------------------------------#
#   Conv+BN+Acti经常会用到,组在一起
# -------------------------------------------------------------#
class ConvBNActivation(nn.Sequential):
    def __init__(self,
                 in_planes: int,
                 out_planes: int,
                 kernel_size: int = 3,
                 stride: int = 1,
                 groups: int = 1,
                 norm_layer: Optional[Callable[..., nn.Module]] = None,         # 卷积后的BN层
                 activation_layer: Optional[Callable[..., nn.Module]] = None):  # 激活函数
        padding = (kernel_size - 1) // 2
        if norm_layer is None:          # 没有传入,就默认使用BN
            norm_layer = nn.BatchNorm2d
        if activation_layer is None:
            activation_layer = nn.ReLU6
        super(ConvBNActivation, self).__init__(nn.Conv2d(in_channels=in_planes,
                                                         out_channels=out_planes,
                                                         kernel_size=kernel_size,
                                                         stride=stride,
                                                         padding=padding,
                                                         groups=groups,
                                                         bias=False),           # 后面会用到BN层,故不使用bias
                                               norm_layer(out_planes),
                                               activation_layer(inplace=True))


# ------------------------------------------------------#
#   注意力模块:SE模块
#	就是两个FC层,节点个数、激活函数要注意要注意
# ------------------------------------------------------#
class SqueezeExcitation(nn.Module):
    # squeeze_factor: int = 4:第一个FC层节点个数是输入特征矩阵的1/4
    def __init__(self, input_c: int, squeeze_factor: int = 4):
        super(SqueezeExcitation, self).__init__()
        # 第一个FC层节点个数,也要是8的整数倍
        squeeze_c = _make_divisible(input_c // squeeze_factor, 8)
        # 通过卷积核大小为1x1的卷积替代FC层,作用相同
        self.fc1 = nn.Conv2d(input_c, squeeze_c, 1)
        self.fc2 = nn.Conv2d(squeeze_c, input_c, 1)

    def forward(self, x: Tensor) -> Tensor:
        # x有很多channel,通过output_size=(1, 1)实现每个channel变成1个数字
        scale = F.adaptive_avg_pool2d(x, output_size=(1, 1))
        scale = self.fc1(scale)
        scale = F.relu(scale, inplace=True)
        scale = self.fc2(scale)
        # 此处的scale就是第二个FC层输出的数据
        scale = F.hardsigmoid(scale, inplace=True)  
        return scale * x        # 和原输入相乘,得到SE模块的输出


# ------------------------------------------------------#
#   InvertedResidualConfig是参数配置文件
# ------------------------------------------------------#
class InvertedResidualConfig:
    def __init__(self,
                 input_c: int,          
                 kernel: int,
                 expanded_c: int,   # bottleneck中的第一层1x1卷积升维,维度升到多少
                 out_c: int,
                 use_se: bool,
                 activation: str,
                 stride: int,
                 width_multi: float):       # 和mobilenetv2中倍率因子相同,通过它得到每一层channels个数和基线的区别
        self.input_c = self.adjust_channels(input_c, width_multi)   # 倍率因子用在这儿了
        self.kernel = kernel
        self.expanded_c = self.adjust_channels(expanded_c, width_multi)
        self.out_c = self.adjust_channels(out_c, width_multi)
        self.use_se = use_se
        # activation == "HS",则self.use_hs==True
        self.use_hs = activation == "HS"  # whether using h-swish activation
        self.stride = stride

    # 静态方法
    @staticmethod
    def adjust_channels(channels: int, width_multi: float):
        return _make_divisible(channels * width_multi, 8)


class InvertedResidual(nn.Module):
    def __init__(self,
                 cnf: InvertedResidualConfig,       # cnf是个config文件,对应的格式就是上面介绍的InvertedResidualConfig类
                 norm_layer: Callable[..., nn.Module]):
        super(InvertedResidual, self).__init__()

        if cnf.stride not in [1, 2]:
            raise ValueError("illegal stride value.")

        # 是否使用shortcut连接
        self.use_res_connect = (cnf.stride == 1 and cnf.input_c == cnf.out_c)       

        layers: List[nn.Module] = []    # 定义一个空列表,里面元素类型为nn.module
        activation_layer = nn.Hardswish if cnf.use_hs else nn.ReLU

        # expand
        if cnf.expanded_c != cnf.input_c:       # 第一个bottleneck没有这个1x1卷积,故有这个if哦安短
            layers.append(ConvBNActivation(cnf.input_c,
                                           cnf.expanded_c,
                                           kernel_size=1,
                                           norm_layer=norm_layer,
                                           activation_layer=activation_layer))

        # depthwise
        layers.append(ConvBNActivation(cnf.expanded_c,      # 上一层1x1输出通道数为cnf.expanded_c
                                       cnf.expanded_c,
                                       kernel_size=cnf.kernel,
                                       stride=cnf.stride,
                                       groups=cnf.expanded_c,       # DW卷积
                                       norm_layer=norm_layer,
                                       activation_layer=activation_layer))

        if cnf.use_se:      # 是否使用se模块,只需要传入个input_channel
            layers.append(SqueezeExcitation(cnf.expanded_c))

        # project       降维1x1卷积层
        layers.append(ConvBNActivation(cnf.expanded_c,
                                       cnf.out_c,
                                       kernel_size=1,
                                       norm_layer=norm_layer,
                                       # nn.Identity是一个线性激活,没进行任何处理
                                       #    内部实现:直接return input
                                       activation_layer=nn.Identity))   

        self.block = nn.Sequential(*layers)
        self.out_channels = cnf.out_c
        self.is_strided = cnf.stride > 1

    def forward(self, x: Tensor) -> Tensor:
        result = self.block(x)
        if self.use_res_connect:
            result += x

        return result


# 继承来自nn.module类
class MobileNetV3(nn.Module):
    def __init__(self,
                 inverted_residual_setting: List[InvertedResidualConfig],   # 参数设置列表,列表里面每个元素类型是上面定义的那个类的形式
                 last_channel: int,         # 倒数第二层channel个数
                 num_classes: int = 1000,   # 需要分类的类别数
                 block: Optional[Callable[..., nn.Module]] = None,
                 norm_layer: Optional[Callable[..., nn.Module]] = None):
        super(MobileNetV3, self).__init__()

        if not inverted_residual_setting:
            raise ValueError("The inverted_residual_setting should not be empty.")
        elif not (isinstance(inverted_residual_setting, List) and
                  all([isinstance(s, InvertedResidualConfig) for s in inverted_residual_setting])):
            raise TypeError("The inverted_residual_setting should be List[InvertedResidualConfig]")

        if block is None:
            block = InvertedResidual

        # 将norm_layer设置为BN
        #   partial()给输入函数BN指定默认参数,简化之后的函数参数量
        if norm_layer is None:
            norm_layer = partial(nn.BatchNorm2d, eps=0.001, momentum=0.01)

        layers: List[nn.Module] = []

        # building first layer   就是普通的conv
        firstconv_output_c = inverted_residual_setting[0].input_c
        layers.append(ConvBNActivation(3,
                                       firstconv_output_c,
                                       kernel_size=3,
                                       stride=2,
                                       norm_layer=norm_layer,
                                       activation_layer=nn.Hardswish))
        # building inverted residual blocks
        for cnf in inverted_residual_setting:
            layers.append(block(cnf, norm_layer))

        # building last several layers
        lastconv_input_c = inverted_residual_setting[-1].out_c
        lastconv_output_c = 6 * lastconv_input_c            # small:96->576; Large:160->960
        layers.append(ConvBNActivation(lastconv_input_c,
                                       lastconv_output_c,
                                       kernel_size=1,
                                       norm_layer=norm_layer,
                                       activation_layer=nn.Hardswish))
        self.features = nn.Sequential(*layers)
        self.avgpool = nn.AdaptiveAvgPool2d(1)
        self.classifier = nn.Sequential(nn.Linear(lastconv_output_c, last_channel),
                                        nn.Hardswish(inplace=True),
                                        nn.Dropout(p=0.2, inplace=True),
                                        nn.Linear(last_channel, num_classes))

        # initial weights
        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)
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
                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)

    def _forward_impl(self, x: Tensor) -> Tensor:
        x = self.features(x)
        x = self.avgpool(x)     # 到这后面不再需要高和宽的维度了
        x = torch.flatten(x, 1) # 故进行展平处理
        x = self.classifier(x)

        return x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)


def mobilenet_v3_large(num_classes: int = 1000,
                       reduced_tail: bool = False) -> MobileNetV3:
    """
    Constructs a large MobileNetV3 architecture from
    "Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>.

    weights_link:
    https://download.pytorch.org/models/mobilenet_v3_large-8738ca79.pth

    Args:
        num_classes (int): number of classes
        reduced_tail (bool): 需要的话, 设为True, 可以进一步减小网络
            If True, reduces the channel counts of all feature layers
            between C4 and C5 by 2. It is used to reduce the channel redundancy in the
            backbone for Detection and Segmentation.
    """
    width_multi = 1.0       # 调整channel个数,默认1.0
    bneck_conf = partial(InvertedResidualConfig, width_multi=width_multi)   # partial()给输入函数指定默认参数
    # 给类里的方法传入参数      有了上面一行,这行有必要吗?
    adjust_channels = partial(InvertedResidualConfig.adjust_channels, width_multi=width_multi)

    reduce_divider = 2 if reduced_tail else 1

    inverted_residual_setting = [
        # input_c, kernel, expanded_c, out_c, use_se, activation, stride
        bneck_conf(16, 3, 16, 16, False, "RE", 1),
        bneck_conf(16, 3, 64, 24, False, "RE", 2),  # C1
        bneck_conf(24, 3, 72, 24, False, "RE", 1),
        bneck_conf(24, 5, 72, 40, True, "RE", 2),   # C2
        bneck_conf(40, 5, 120, 40, True, "RE", 1),
        bneck_conf(40, 5, 120, 40, True, "RE", 1),
        bneck_conf(40, 3, 240, 80, False, "HS", 2),  # C3
        bneck_conf(80, 3, 200, 80, False, "HS", 1),
        bneck_conf(80, 3, 184, 80, False, "HS", 1),
        bneck_conf(80, 3, 184, 80, False, "HS", 1),
        bneck_conf(80, 3, 480, 112, True, "HS", 1),
        bneck_conf(112, 3, 672, 112, True, "HS", 1),
        bneck_conf(112, 5, 672, 160 // reduce_divider, True, "HS", 2),  # C4
        bneck_conf(160 // reduce_divider, 5, 960 // reduce_divider, 160 // reduce_divider, True, "HS", 1),
        bneck_conf(160 // reduce_divider, 5, 960 // reduce_divider, 160 // reduce_divider, True, "HS", 1),
    ]
    last_channel = adjust_channels(1280 // reduce_divider)  # C5    # 倒数第二个全连接层节点个数

    return MobileNetV3(inverted_residual_setting=inverted_residual_setting,
                       last_channel=last_channel,
                       num_classes=num_classes)


def mobilenet_v3_small(num_classes: int = 1000,
                       reduced_tail: bool = False) -> MobileNetV3:
    """
    Constructs a large MobileNetV3 architecture from
    "Searching for MobileNetV3" <https://arxiv.org/abs/1905.02244>.

    weights_link:
    https://download.pytorch.org/models/mobilenet_v3_small-047dcff4.pth

    Args:
        num_classes (int): number of classes
        reduced_tail (bool): If True, reduces the channel counts of all feature layers
            between C4 and C5 by 2. It is used to reduce the channel redundancy in the
            backbone for Detection and Segmentation.
    """
    width_multi = 1.0
    bneck_conf = partial(InvertedResidualConfig, width_multi=width_multi)
    adjust_channels = partial(InvertedResidualConfig.adjust_channels, width_multi=width_multi)

    reduce_divider = 2 if reduced_tail else 1

    inverted_residual_setting = [
        # input_c, kernel, expanded_c, out_c, use_se, activation, stride
        bneck_conf(16, 3, 16, 16, True, "RE", 2),  # C1
        bneck_conf(16, 3, 72, 24, False, "RE", 2),  # C2
        bneck_conf(24, 3, 88, 24, False, "RE", 1),
        bneck_conf(24, 5, 96, 40, True, "HS", 2),  # C3
        bneck_conf(40, 5, 240, 40, True, "HS", 1),
        bneck_conf(40, 5, 240, 40, True, "HS", 1),
        bneck_conf(40, 5, 120, 48, True, "HS", 1),
        bneck_conf(48, 5, 144, 48, True, "HS", 1),
        bneck_conf(48, 5, 288, 96 // reduce_divider, True, "HS", 2),  # C4
        bneck_conf(96 // reduce_divider, 5, 576 // reduce_divider, 96 // reduce_divider, True, "HS", 1),
        bneck_conf(96 // reduce_divider, 5, 576 // reduce_divider, 96 // reduce_divider, True, "HS", 1)
    ]
    last_channel = adjust_channels(1024 // reduce_divider)  # C5

    return MobileNetV3(inverted_residual_setting=inverted_residual_setting,
                       last_channel=last_channel,
                       num_classes=num_classes)

if __name__ == "__main__":
    model = mobilenet_v3_small()
    print(model)

    from torchsummaryX import summary
    summary(model, torch.randn(1,3,224,224))

输出:

MobileNetV3(
  (features): Sequential(
    (0): ConvBNActivation(
      (0): Conv2d(3, 16, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
      (1): BatchNorm2d(16, eps=0.001, momentum=0.01, affine=True, track_running_stats=True)
      (2): Hardswish()
    )
    (1): InvertedResidual(
...
(classifier): Sequential(
    (0): Linear(in_features=576, out_features=1024, bias=True)
    (1): Hardswish()
    (2): Dropout(p=0.2, inplace=True)
    (3): Linear(in_features=1024, out_features=1000, bias=True)
  )
)
================================================================================================
                                           Kernel Shape       Output Shape  \
Layer
0_features.0.Conv2d_0                     [3, 16, 3, 3]  [1, 16, 112, 112]
1_features.0.BatchNorm2d_1                         [16]  [1, 16, 112, 112]
...
123_classifier.Dropout_2                      -          -
124_classifier.Linear_3                  1.025M     1.024M
------------------------------------------------------------------------------------------------
                          Totals
Total params           2.542856M
Trainable params       2.542856M
Non-trainable params         0.0
Mult-Adds             56.516456M

5 感谢链接

https://www.bilibili.com/video/BV1GK4y1p7uE/?spm_id_from=333.788
https://blog.csdn.net/m0_48742971/article/details/123438626
https://www.bilibili.com/video/BV1zT4y1P7pd/?spm_id_from=333.788

文章出处登录后可见!

已经登录?立即刷新

共计人评分,平均

到目前为止还没有投票!成为第一位评论此文章。

(0)
乘风的头像乘风管理团队
上一篇 2023年3月5日 下午9:40
下一篇 2023年3月5日

相关推荐