探究MobileNet論文

一、從論文下載開始

MobileNet論文是由Google Brain團隊提出，2017年發布的一篇基於深度學習的輕量級模型架構講稿，主要針對嵌入式設備和移動端設備進行優化，實現高效的圖像識別應用。

首先，在官方網站https://arxiv.org/abs/1704.04861 上下載閱讀論文。

$ wget https://arxiv.org/pdf/1704.04861.pdf
$ open 1704.04861.pdf

二、Mobilenet網路結構圖

Mobilenet的網路結構的核心思想是Depthwise Separable Convolution。深度可分離卷積包括兩個步驟，第一步是Depthwise Convolution，第二步是Pointwise Convolution。首先通過Depthwise Convolution計算每一個輸入通道的卷積層，得到的結果叫做Depthwise Features。接著將Depthwise Features送入Pointwise Convolution中，進行全局卷積操作，然後產生新的N個卷積特徵，輸出結果就是Depthwise Features與Pointwise Features的融合。

該網路結構圖如圖所示：

from IPython.display import Image
Image(filename='mobilenet_structure.png', width=400, height=400)

三、Mobilenet v2

Mobilenet v2是MobileNet v1的後續版本。隨著深度學習的快速發展，人們越來越需要更深、更準確的模型，所以Google在v1的基礎上，提出了一些新的設計思路，如Inverted Residuals、Linear Bottlenecks等。具體的網路結構與v1有很大的不同，v2的網路結構圖如下：

from IPython.display import Image
Image(filename='mobilenet_v2_structure.png', width=400, height=400)

四、實現MobileNet

下面是使用Keras框架實現MobileNet的代碼示例：

from keras.layers import Input, Conv2D, DepthwiseConv2D, Activation
from keras.models import Model

def conv_block(inputs, filters, kernel, strides):
    x = Conv2D(filters, kernel_size=kernel, strides=strides, padding='same')(inputs)
    x = Activation('relu')(x)
    return x

def depthwise_block(inputs, depth_multiplier, kernel, strides):
    x = DepthwiseConv2D(kernel_size=kernel, strides=strides, depth_multiplier=depth_multiplier, padding='same')(inputs)
    x = Activation('relu')(x)
    return x

def mobileNet(input_shape, num_classes):
    inp = Input(shape=input_shape)
    x = conv_block(inp, 32, 3, strides=(2, 2))
    x = depthwise_block(x, 1, 3, strides=(1, 1))
    x = conv_block(x, 64, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(2, 2))
    x = conv_block(x, 128, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(1, 1))
    x = conv_block(x, 128, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(2, 2))
    x = conv_block(x, 256, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(1, 1))
    x = conv_block(x, 256, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(2, 2))
    x = conv_block(x, 512, 1, strides=(1, 1))
    for i in range(5):
        x = depthwise_block(x, 1, 3, strides=(1, 1))
        x = conv_block(x, 512, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(2, 2))
    x = conv_block(x, 1024, 1, strides=(1, 1))
    x = depthwise_block(x, 1, 3, strides=(1, 1))
    x = conv_block(x, 1024, 1, strides=(1, 1))
    x = Conv2D(num_classes, kernel_size=1, strides=(1, 1), padding='same')(x)
    x = Activation('softmax')(x)
    model = Model(inp, x)
    return model

input_shape = (224, 224, 3)
num_classes = 10
model = mobileNet(input_shape, num_classes)
model.summary()