作為移動端輕量級網(wǎng)絡(luò)的代表，MobileNet一直是大家關(guān)注的焦點。最近，Google提出了新一代的MobileNetV3網(wǎng)絡(luò)。這一代MobileNet結(jié)合了AutoML和人工調(diào)整，帶來了更加高效的性能。

Paper ：Searching for MobileNetV3
Github ：https://github.com/xiaochus/MobileNetV3

改進

1. MobileNetV3的網(wǎng)絡(luò)模塊結(jié)構(gòu)延續(xù)了MobileNetV1的深度可分離卷積和MobileNetV2的bottleneck with residual 結(jié)構(gòu)。在此基礎(chǔ)上，還加入了SENet中的基于squeeze and excitation結(jié)構(gòu)的輕量級注意力模型。

squeeze

2. MobileNetV3的結(jié)構(gòu)是通過AutoML技術(shù)生成的。在網(wǎng)絡(luò)結(jié)構(gòu)搜索中，作者結(jié)合兩種技術(shù)：資源受限的NAS與NetAdapt，前者用于在計算和參數(shù)量受限的前提下搜索網(wǎng)絡(luò)的各個模塊，所以稱之為模塊級的搜索（Block-wise Search） ，后者用于對各個模塊確定之后網(wǎng)絡(luò)層的微調(diào)。

3. MobileNetV2在預(yù)測部分使用了一個Bottleneck結(jié)構(gòu)來提取特征，這種結(jié)構(gòu)帶來了額外的計算開銷。MobileNetV3中使用兩個1*1的卷積來代替了這個操作。

head

4. 使用swish激活函數(shù)代替ReLU能夠有效提高網(wǎng)絡(luò)的精度，但是swish的計算量太大了。因此作者對swish進行了數(shù)值近似，提出h-swish（hard version of swish）函數(shù)。

activation

approximation

result

網(wǎng)絡(luò)結(jié)構(gòu)

Large MobileNetV3：

MV3-large

Small MobileNetV3：

MV3-small

效果

與前一代MobieNetV2比較，V3-Large取得了最高的精度，V3-Small 取得了V2近似的精度，速度卻快很多。

compare1

compare2

與其他移動端網(wǎng)絡(luò)比較，MobieNetV3也有著良好的性能。

compare3

Keras實現(xiàn)

首先實現(xiàn)一個基礎(chǔ)類，定義了MobileNetV3的激活函數(shù)和各種基本結(jié)構(gòu)。

          
            """MobileNet v3 models for Keras.
# Reference
    [Searching for MobileNetV3](https://arxiv.org/abs/1905.02244?context=cs)
"""


from keras.layers import Conv2D, DepthwiseConv2D, Dense, GlobalAveragePooling2D
from keras.layers import Activation, BatchNormalization, Add, Lambda

from keras import backend as K


class MobileNetBase:
    def __init__(self, shape, n_class):
        self.shape = shape
        self.n_class = n_class

    def _relu6(self, x):
        """Relu 6
        """
        return K.relu(x, max_value=6.0)

    def _hard_swish(self, x):
        """Hard swish
        """
        return x * K.relu(x + 3.0, max_value=6.0) / 6.0

    def _return_activation(self, x, nl):
        """Convolution Block
        This function defines a activation choice.

        # Arguments
            x: Tensor, input tensor of conv layer.
            nl: String, nonlinearity activation type.

        # Returns
            Output tensor.
        """
        if nl == 'HS':
            x = Activation(self._hard_swish)(x)
        if nl == 'RE':
            x = Activation(self._relu6)(x)

        return x

    def _conv_block(self, inputs, filters, kernel, strides, nl):
        """Convolution Block
        This function defines a 2D convolution operation with BN and activation.

        # Arguments
            inputs: Tensor, input tensor of conv layer.
            filters: Integer, the dimensionality of the output space.
            kernel: An integer or tuple/list of 2 integers, specifying the
                width and height of the 2D convolution window.
            strides: An integer or tuple/list of 2 integers,
                specifying the strides of the convolution along the width and height.
                Can be a single integer to specify the same value for
                all spatial dimensions.
            nl: String, nonlinearity activation type.

        # Returns
            Output tensor.
        """

        channel_axis = 1 if K.image_data_format() == 'channels_first' else -1

        x = Conv2D(filters, kernel, padding='same', strides=strides)(inputs)
        x = BatchNormalization(axis=channel_axis)(x)

        return self._return_activation(x, nl)

    def _squeeze(self, inputs):
        """Squeeze and Excitation.
        This function defines a squeeze structure.

        # Arguments
            inputs: Tensor, input tensor of conv layer.
        """
        input_channels = int(inputs.shape[-1])

        x = GlobalAveragePooling2D()(inputs)
        x = Dense(input_channels, activation='relu')(x)
        x = Dense(input_channels, activation='hard_sigmoid')(x)

        return x

    def _bottleneck(self, inputs, filters, kernel, e, s, squeeze, nl):
        """Bottleneck
        This function defines a basic bottleneck structure.

        # Arguments
            inputs: Tensor, input tensor of conv layer.
            filters: Integer, the dimensionality of the output space.
            kernel: An integer or tuple/list of 2 integers, specifying the
                width and height of the 2D convolution window.
            e: Integer, expansion factor.
                t is always applied to the input size.
            s: An integer or tuple/list of 2 integers,specifying the strides
                of the convolution along the width and height.Can be a single
                integer to specify the same value for all spatial dimensions.
            squeeze: Boolean, Whether to use the squeeze.
            nl: String, nonlinearity activation type.

        # Returns
            Output tensor.
        """

        channel_axis = 1 if K.image_data_format() == 'channels_first' else -1
        input_shape = K.int_shape(inputs)
        tchannel = input_shape[channel_axis] * e
        r = s == 1 and input_shape[3] == filters

        x = self._conv_block(inputs, tchannel, (1, 1), (1, 1), nl)

        x = DepthwiseConv2D(kernel, strides=(s, s), depth_multiplier=1, padding='same')(x)
        x = BatchNormalization(axis=channel_axis)(x)

        if squeeze:
            x = Lambda(lambda x: x * self._squeeze(x))(x)

        x = self._return_activation(x, nl)

        x = Conv2D(filters, (1, 1), strides=(1, 1), padding='same')(x)
        x = BatchNormalization(axis=channel_axis)(x)

        if r:
            x = Add()([x, inputs])

        return x

    def build(self):
        pass


          
        

MobileNet V3 Large：

          
            """MobileNet v3 Large models for Keras.
# Reference
    [Searching for MobileNetV3](https://arxiv.org/abs/1905.02244?context=cs)
"""


from keras.models import Model
from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Reshape
from keras.utils.vis_utils import plot_model

from model.mobilenet_base import MobileNetBase


class MobileNetV3_Large(MobileNetBase):
    def __init__(self, shape, n_class):
        """Init.

        # Arguments
            input_shape: An integer or tuple/list of 3 integers, shape
                of input tensor.
            n_class: Integer, number of classes.

        # Returns
            MobileNetv2 model.
        """
        super(MobileNetV3_Large, self).__init__(shape, n_class)

    def build(self, plot=False):
        """build MobileNetV3 Large.

        # Arguments
            plot: Boolean, weather to plot model.

        # Returns
            model: Model, model.
        """
        inputs = Input(shape=self.shape)

        x = self._conv_block(inputs, 16, (3, 3), strides=(2, 2), nl='HS')

        x = self._bottleneck(x, 16, (3, 3), e=16, s=1, squeeze=False, nl='RE')
        x = self._bottleneck(x, 24, (3, 3), e=64, s=2, squeeze=False, nl='RE')
        x = self._bottleneck(x, 24, (3, 3), e=72, s=1, squeeze=False, nl='RE')
        x = self._bottleneck(x, 40, (5, 5), e=72, s=2, squeeze=True, nl='RE')
        x = self._bottleneck(x, 40, (5, 5), e=120, s=1, squeeze=True, nl='RE')
        x = self._bottleneck(x, 40, (5, 5), e=120, s=1, squeeze=True, nl='RE')
        x = self._bottleneck(x, 80, (3, 3), e=240, s=2, squeeze=False, nl='HS')
        x = self._bottleneck(x, 80, (3, 3), e=200, s=1, squeeze=False, nl='HS')
        x = self._bottleneck(x, 80, (3, 3), e=184, s=1, squeeze=False, nl='HS')
        x = self._bottleneck(x, 80, (3, 3), e=184, s=1, squeeze=False, nl='HS')
        x = self._bottleneck(x, 112, (3, 3), e=480, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 112, (3, 3), e=672, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 160, (5, 5), e=672, s=2, squeeze=True, nl='HS')
        x = self._bottleneck(x, 160, (5, 5), e=960, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 160, (5, 5), e=960, s=1, squeeze=True, nl='HS')

        x = self._conv_block(x, 960, (1, 1), strides=(1, 1), nl='HS')
        x = GlobalAveragePooling2D()(x)
        x = Reshape((1, 1, 960))(x)

        x = Conv2D(1280, (1, 1), padding='same')(x)
        x = self._return_activation(x, 'HS')
        x = Conv2D(self.n_class, (1, 1), padding='same', activation='softmax')(x)

        output = Reshape((self.n_class,))(x)

        model = Model(inputs, output)

        if plot:
            plot_model(model, to_file='images/MobileNetv3_large.png', show_shapes=True)

        return model


          
        

MobileNet V3 Small：

          
            """MobileNet v3 small models for Keras.
# Reference
    [Searching for MobileNetV3](https://arxiv.org/abs/1905.02244?context=cs)
"""


from keras.models import Model
from keras.layers import Input, Conv2D, GlobalAveragePooling2D, Reshape
from keras.utils.vis_utils import plot_model

from model.mobilenet_base import MobileNetBase


class MobileNetV3_Small(MobileNetBase):
    def __init__(self, shape, n_class):
        """Init.

        # Arguments
            input_shape: An integer or tuple/list of 3 integers, shape
                of input tensor.
            n_class: Integer, number of classes.

        # Returns
            MobileNetv2 model.
        """
        super(MobileNetV3_Small, self).__init__(shape, n_class)


    def build(self, plot=False):
        """build MobileNetV3 Small.

        # Arguments
            plot: Boolean, weather to plot model.

        # Returns
            model: Model, model.
        """
        inputs = Input(shape=self.shape)

        x = self._conv_block(inputs, 16, (3, 3), strides=(2, 2), nl='HS')

        x = self._bottleneck(x, 16, (3, 3), e=16, s=2, squeeze=True, nl='RE')
        x = self._bottleneck(x, 24, (3, 3), e=72, s=2, squeeze=False, nl='RE')
        x = self._bottleneck(x, 24, (3, 3), e=88, s=1, squeeze=False, nl='RE')
        x = self._bottleneck(x, 40, (5, 5), e=96, s=2, squeeze=True, nl='HS')
        x = self._bottleneck(x, 40, (5, 5), e=240, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 40, (5, 5), e=240, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 48, (5, 5), e=120, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 48, (5, 5), e=144, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 96, (5, 5), e=288, s=2, squeeze=True, nl='HS')
        x = self._bottleneck(x, 96, (5, 5), e=576, s=1, squeeze=True, nl='HS')
        x = self._bottleneck(x, 96, (5, 5), e=576, s=1, squeeze=True, nl='HS')

        x = self._conv_block(x, 576, (1, 1), strides=(1, 1), nl='HS')
        x = GlobalAveragePooling2D()(x)
        x = Reshape((1, 1, 576))(x)

        x = Conv2D(1280, (1, 1), padding='same')(x)
        x = self._return_activation(x, 'HS')
        x = Conv2D(self.n_class, (1, 1), padding='same', activation='softmax')(x)

        output = Reshape((self.n_class,))(x)

        model = Model(inputs, output)

        if plot:
            plot_model(model, to_file='images/MobileNetv3_small.png', show_shapes=True)

        return model