MobileNetV1网络结构的原理与Tensorflow2.0实现

介绍 MobileNet 之前&＃xff0c;我们需要先了解一下它的重要组件——深度级可分离卷积&＃xff08;Depthwise separable convolution&＃xff09;网络&＃xff0c;这种网络可以被进一步分解成两个部分&＃xff1a;对输入的每个通道使用单通道卷积核进行卷积&＃xff08;depthwise convolution&＃xff09;、使用 1x1 卷积核将输入通道数转变成输出通道数&＃xff08;pointwise convolution&＃xff09;。举例如下图所示&＃xff1a;

在第一部分中&＃xff0c;深度级卷积和普通卷积不同&＃xff0c;对于标准卷积其卷积核是用在所有的输入通道上的&＃xff0c;而深度级可分离卷积针对每个输入通道采用不同的卷积核&＃xff0c;即一个卷积核对应一个输入通道&＃xff0c;所以我们说这个操作是发生在深度级上的。

在第二部分中&＃xff0c;pointwise convolution 其实就和普通卷积操作相同&＃xff0c;只不过使用的卷积核尺寸是 1x1 的。

深度级可分离卷积和一个普通卷积的整体效果差不多&＃xff0c;但前者会大大减少计算量和模型参数量&＃xff0c;具体可参考卷积、可分离卷积的参数量与计算量的对比。

在 MobileNet 网络结构中&＃xff0c;有 13 个深度级可分离卷积网络被使用。

用函数实现

import tensorflow as tfdef conv_block(inputs, filters, kernel_size &＃61; (3,3), strides&＃61;(1,1)):x &＃61; tf.keras.layers.Conv2D(filters&＃61;filters, kernel_size&＃61;kernel_size, strides&＃61;strides, padding&＃61;&＃39;SAME&＃39;, use_bias&＃61;False)(inputs)x &＃61; tf.keras.layers.BatchNormalization()(x)out &＃61; tf.keras.layers.Activation(&＃39;relu&＃39;)(x)return outdef depthwise_conv_block(inputs,pointwise_conv_filters,strides&＃61;(1,1)):x &＃61; tf.keras.layers.DepthwiseConv2D(kernel_size&＃61;(3,3), strides&＃61;strides, padding&＃61;&＃39;SAME&＃39;,use_bias&＃61;False)(inputs)x &＃61; tf.keras.layers.BatchNormalization()(x)x &＃61; tf.keras.layers.Activation(&＃39;relu&＃39;)(x)x &＃61; tf.keras.layers.Conv2D(filters&＃61;pointwise_conv_filters, kernel_size&＃61;(1,1), padding&＃61;&＃39;SAME&＃39;, use_bias&＃61;False)(x)x &＃61; tf.keras.layers.BatchNormalization()(x)out &＃61; tf.keras.layers.Activation(&＃39;relu&＃39;)(x)return outdef mobilenet_v1(inputs,classes):# [32, 32, 3] &＃61;> [16, 16, 32]x &＃61; conv_block(inputs, 32, strides&＃61;(2,2))# [16, 16, 32] &＃61;> [16, 16, 64]x &＃61; depthwise_conv_block(x, 64)# [16, 16, 64] &＃61;> [8, 8, 128]x &＃61; depthwise_conv_block(x, 128, strides&＃61;(2,2))# [8, 8, 128] &＃61;> [8, 8, 128]x &＃61; depthwise_conv_block(x, 128)# [8, 8, 128] &＃61;> [4, 4, 256]x &＃61; depthwise_conv_block(x, 256, strides&＃61;(2, 2))# [4, 4, 256] &＃61;> [4, 4, 256]x &＃61; depthwise_conv_block(x, 256)# [4, 4, 256] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512, strides&＃61;(2, 2))# [2, 2, 512] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512)# [2, 2, 512] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512)# [2, 2, 512] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512)# [2, 2, 512] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512)# [2, 2, 512] &＃61;> [2, 2, 512]x &＃61; depthwise_conv_block(x, 512)# [2, 2, 512] &＃61;> [1, 1, 1024]x &＃61; depthwise_conv_block(x, 1024, strides&＃61;(2,2))# [1, 1, 1024] &＃61;> [1, 1, 1024]x &＃61; depthwise_conv_block(x, 1024)# [1, 1, 1024] &＃61;> (1024,)x &＃61; tf.keras.layers.GlobalAveragePooling2D()(x)# (1024,) &＃61;> (classes,)pred &＃61; tf.keras.layers.Dense(classes, activation&＃61;&＃39;softmax&＃39;)(x)return predinputs &＃61; np.zeros((10, 32, 32, 3), dtype&＃61;np.float32)
classes &＃61; 10
output &＃61; mobilenet_v1(inputs, classes)output.shape


TensorShape([10, 10])


用类实现

import tensorflow as tfclass conv_block(tf.keras.Model):def __init__(self, filters, kernel_size &＃61; (3,3), strides&＃61;(1,1)):super().__init__()self.listLayers &＃61; []self.listLayers.append(tf.keras.layers.Conv2D(filters&＃61;filters, kernel_size&＃61;kernel_size, strides&＃61;strides, padding&＃61;&＃39;SAME&＃39;, use_bias&＃61;False))self.listLayers.append(tf.keras.layers.BatchNormalization())self.listLayers.append(tf.keras.layers.Activation(&＃39;relu&＃39;))def call(self, x):for layer in self.listLayers.layers:x &＃61; layer(x)return xclass depthwise_conv_block(tf.keras.Model):def __init__(self, pointwise_conv_filters, strides&＃61;(1,1)):super().__init__()self.listLayers &＃61; []self.listLayers.append(tf.keras.layers.DepthwiseConv2D(kernel_size&＃61;(3,3), strides&＃61;strides, padding&＃61;&＃39;SAME&＃39;,use_bias&＃61;False))self.listLayers.append(tf.keras.layers.BatchNormalization())self.listLayers.append(tf.keras.layers.Activation(&＃39;relu&＃39;))self.listLayers.append(tf.keras.layers.Conv2D(filters&＃61;pointwise_conv_filters, kernel_size&＃61;(1,1), padding&＃61;&＃39;SAME&＃39;, use_bias&＃61;False))self.listLayers.append(tf.keras.layers.BatchNormalization())self.listLayers.append(tf.keras.layers.Activation(&＃39;relu&＃39;))def call(self, x):for layer in self.listLayers.layers:x &＃61; layer(x)return xclass mobilenet_v1(tf.keras.Model):def __init__(self, classes):super().__init__()self.listLayers &＃61; []self.listLayers.append(conv_block(32, strides&＃61;(2,2)))self.listLayers.append(depthwise_conv_block(64))self.listLayers.append(depthwise_conv_block(128, strides&＃61;(2,2)))self.listLayers.append(depthwise_conv_block(128))self.listLayers.append(depthwise_conv_block(256, strides&＃61;(2, 2)))self.listLayers.append(depthwise_conv_block(256))self.listLayers.append(depthwise_conv_block(512, strides&＃61;(2, 2)))self.listLayers.append(depthwise_conv_block(512))self.listLayers.append(depthwise_conv_block(512))self.listLayers.append(depthwise_conv_block(512))self.listLayers.append(depthwise_conv_block(512))self.listLayers.append(depthwise_conv_block(512))self.listLayers.append(depthwise_conv_block(1024, strides&＃61;(2,2)))self.listLayers.append(depthwise_conv_block(1024))self.listLayers.append(tf.keras.layers.GlobalAveragePooling2D())self.listLayers.append(tf.keras.layers.Dense(classes, activation&＃61;&＃39;softmax&＃39;))def call(self, x):for layer in self.listLayers.layers:x &＃61; layer(x)return xinputs &＃61; np.zeros((10, 32, 32, 3), dtype&＃61;np.float32)
model &＃61; mobilenet_v1(10)
model(inputs).shape


TensorShape([10, 10])