[嵌入式]MobileNet的理解与实现

MobileNet出现原因

在真实应用场景或嵌入式设备中，需要延迟较低，响应速度较快的模型。

实现目的的两个方向：

• 对训练好的复杂模型进行压缩得到小模型。
• 直接设计小模型并进行训练。

MobileNet属于第二种实现方式。

MobileNet网络结构

以上为MobileNet的整体结构，总体是由Conv结构于Conv dw结构堆叠而成。

以上为Conv结构。

以上为Conv dw结构。

MobileNet首先是经过一个3 × 3的标准卷积，然后堆叠depthwise separable convolution结构，最后通过平均池化将feature变成1 × 1，进行全连接，最后连接上一个softmax。

相比其他网络，其虽然准确率有所降低，但参数量明显降低。

以下是我仿照实现的MobileNet：

def relu6(x):
    return K.relu(x=x, max_value=6)

def conv(inputs, filters, kernel_size=3, strides=1, layer_num=0):
    x = Conv2D(filters=filters, kernel_size=kernel_size, strides=strides, padding='same', use_bias=False, name="conv_%d"%layer_num)(inputs)
    x = BatchNormalization(name="conv_batchnormalization_%d"%layer_num)(x)
    x = Activation(relu6, name="conv_activation_%d"%layer_num)(x)
    return x, layer_num + 1

def conv_ds(inputs, pointwise_conv_filters, depth_multiplier=1, strides=1, layer_num=0):
    x = DepthwiseConv2D(kernel_size=3, depth_multiplier=depth_multiplier, strides=strides, padding='same', use_bias=False, name="dp_conv_%d"%layer_num)(inputs)
    x = BatchNormalization(name="dp_conv_batchnormalization_%d"%layer_num)(x)
    x = Activation(relu6, name="dp_conv_activation_%d"%layer_num)(x)

    x, layer_num = conv(inputs=x, filters=pointwise_conv_filters, kernel_size=1, strides=1, layer_num=layer_num + 1)
    return x, layer_num

def MobileNet(inputs, embedding=128, dropout_keep_prob=0.4, depth_multiplier=1, layer_num=0):
    # 160,160,3 -> 80,80,32
    x, layer_num = conv(inputs, filters=32, strides=2, layer_num=layer_num)

    # 80,80,32 -> 80,80,64
    x, layer_num = conv_ds(x, pointwise_conv_filters=64, depth_multiplier=depth_multiplier, layer_num=layer_num)

    # 80,80,64 -> 40,40,128
    x, layer_num = conv_ds(x, pointwise_conv_filters=128, depth_multiplier=depth_multiplier, strides=2, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=128, depth_multiplier=depth_multiplier, layer_num=layer_num)

    # 40,40,128 -> 20,20,256
    x, layer_num = conv_ds(x, pointwise_conv_filters=256, depth_multiplier=depth_multiplier, strides=2, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=256, depth_multiplier=depth_multiplier, layer_num=layer_num)

    # 20,20,256 -> 10,10,512
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, strides=2, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=512, depth_multiplier=depth_multiplier, layer_num=layer_num)

    # 10,10,512 -> 5,5,1024
    x, layer_num = conv_ds(x, pointwise_conv_filters=1024, depth_multiplier=depth_multiplier, strides=2, layer_num=layer_num)
    x, layer_num = conv_ds(x, pointwise_conv_filters=1024, depth_multiplier=depth_multiplier, layer_num=layer_num)

    # cbam = cbam_block(x)
    # x = Concatenate(axis=3, name="attention_concatenate_%d"%layer_num)([x, cbam])
    # layer_num = layer_num + 1

    # 1024 Pooling
    x = GlobalAveragePooling2D(name="global_average_pooling_%d"%layer_num)(x)
    # dropout
    x = Dropout(1 - dropout_keep_prob, name="mobile_net_dropout_%d"%(layer_num+1))(x)

    # 全连接
    x = Dense(embedding, use_bias=False, name="mobilenet_dense_%d"%(layer_num+2))(x)
    x = BatchNormalization(momentum=0.99, epsilon=10e-6, scale=False, name="mobile_net_batchnormal_%d"%(layer_num+3))(x)

    # 创建模型
    model = Model(inputs, x)

    return model