MobileNetV2: 一个更快、更有效的移动视觉网络

2022-11-14 10:37:18

MobileNetV2：重新定义移动视觉

倒置残差：精简计算

传统的神经网络将快捷连接放在卷积层之前，而MobileNetV2反其道而行之，引入了一种创新的“倒置残差”结构。这种巧妙的转换将卷积层移至快捷连接之前，大幅减少了计算量，同时保持了网络的准确性。

代码示例：

import tensorflow as tf

# 创建一个倒置残差块
class InvertedResidualBlock(tf.keras.layers.Layer):

    def __init__(self, filters, strides, expansion_factor):
        super(InvertedResidualBlock, self).__init__()

        # 扩展维度
        self.expand_conv = tf.keras.layers.Conv2D(filters * expansion_factor, 1, strides=1, padding='same')
        self.expand_bn = tf.keras.layers.BatchNormalization()

        # 深度卷积
        self.depth_conv = tf.keras.layers.DepthwiseConv2D(filters, 3, strides=strides, padding='same')
        self.depth_bn = tf.keras.layers.BatchNormalization()

        # 压缩维度
        self.project_conv = tf.keras.layers.Conv2D(filters, 1, strides=1, padding='same')
        self.project_bn = tf.keras.layers.BatchNormalization()

        # 快捷连接
        if strides == 1:
            self.shortcut = tf.identity()
        else:
            self.shortcut = tf.keras.layers.Conv2D(filters, 1, strides=strides, padding='same')
            self.shortcut_bn = tf.keras.layers.BatchNormalization()

    def call(self, inputs):
        x = self.expand_conv(inputs)
        x = self.expand_bn(x)
        x = tf.nn.relu(x)

        x = self.depth_conv(x)
        x = self.depth_bn(x)
        x = tf.nn.relu(x)

        x = self.project_conv(x)
        x = self.project_bn(x)

        x = x + self.shortcut(inputs)
        return x

线性瓶颈：进一步精简

MobileNetV2还采用了“线性瓶颈”技术，在保持准确性的同时进一步减少了计算量。它利用一对1x1卷积层：第一个卷积层减少特征图的通道数，第二个卷积层恢复通道数。这种轻量级的结构显著减轻了模型的负担。

代码示例：

import tensorflow as tf

# 创建一个线性瓶颈层
class LinearBottleneck(tf.keras.layers.Layer):

    def __init__(self, filters, expansion_factor):
        super(LinearBottleneck, self).__init__()

        self.expand_conv = tf.keras.layers.Conv2D(filters * expansion_factor, 1, strides=1, padding='same')
        self.expand_bn = tf.keras.layers.BatchNormalization()

        self.project_conv = tf.keras.layers.Conv2D(filters, 1, strides=1, padding='same')
        self.project_bn = tf.keras.layers.BatchNormalization()

    def call(self, inputs):
        x = self.expand_conv(inputs)
        x = self.expand_bn(x)
        x = tf.nn.relu(x)

        x = self.project_conv(x)
        x = self.project_bn(x)

        return x