赞
踩
为了防止某位我的粉丝寒假没有办法正常工作,我赶紧看了看MobilenetV3。
最新的MobileNetV3的被写在了论文《Searching for MobileNetV3》中。
它是mobilnet的最新版,据说效果还是很好的。
作为一种轻量级网络,它的参数量还是一如既往的小。
它综合了以下四个特点:
1、MobileNetV1的深度可分离卷积(depthwise separable convolutions)。
2、MobileNetV2的具有线性瓶颈的逆残差结构(the inverted residual with linear bottleneck)。
3、轻量级的注意力模型。
4、利用h-swish代替swish函数。
https://github.com/bubbliiiing/classic-convolution-network
如何看懂这个表呢?我们从每一列出发:
第一列Input代表mobilenetV3每个特征层的shape变化;
第二列Operator代表每次特征层即将经历的block结构,我们可以看到在MobileNetV3中,特征提取经过了许多的bneck结构;
第三、四列分别代表了bneck内逆残差结构上升后的通道数、输入到bneck时特征层的通道数。
第五列SE代表了是否在这一层引入注意力机制。
第六列NL代表了激活函数的种类,HS代表h-swish,RE代表RELU。
第七列s代表了每一次block结构所用的步长。
bneck结构如下图所示:
它综合了以下四个特点:
a、MobileNetV2的具有线性瓶颈的逆残差结构(the inverted residual with linear bottleneck)。
即先利用1x1卷积进行升维度,再进行下面的操作,并具有残差边。
b、MobileNetV1的深度可分离卷积(depthwise separable convolutions)。
在输入1x1卷积进行升维度后,进行3x3深度可分离卷积。
c、轻量级的注意力模型。
这个注意力机制的作用方式是调整每个通道的权重。
d、利用h-swish代替swish函数。
在结构中使用了h-swishj激活函数,代替swish函数,减少运算量,提高性能。
由于keras代码没有预训练权重,所以只是把网络结构po出来。
from keras.layers import Conv2D, DepthwiseConv2D, Dense, GlobalAveragePooling2D,Input from keras.layers import Activation, BatchNormalization, Add, Multiply, Reshape from keras.models import Model from keras import backend as K alpha = 1 def relu6(x): # relu函数 return K.relu(x, max_value=6.0) def hard_swish(x): # 利用relu函数乘上x模拟sigmoid return x * K.relu(x + 3.0, max_value=6.0) / 6.0 def return_activation(x, nl): # 用于判断使用哪个激活函数 if nl == 'HS': x = Activation(hard_swish)(x) if nl == 'RE': x = Activation(relu6)(x) return x def conv_block(inputs, filters, kernel, strides, nl): # 一个卷积单元,也就是conv2d + batchnormalization + activation 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 return_activation(x, nl) def squeeze(inputs): # 注意力机制单元 input_channels = int(inputs.shape[-1]) x = GlobalAveragePooling2D()(inputs) x = Dense(int(input_channels/4))(x) x = Activation(relu6)(x) x = Dense(input_channels)(x) x = Activation(hard_swish)(x) x = Reshape((1, 1, input_channels))(x) x = Multiply()([inputs, x]) return x def bottleneck(inputs, filters, kernel, up_dim, stride, sq, nl): channel_axis = 1 if K.image_data_format() == 'channels_first' else -1 input_shape = K.int_shape(inputs) tchannel = int(up_dim) cchannel = int(alpha * filters) r = stride == 1 and input_shape[3] == filters # 1x1卷积调整通道数,通道数上升 x = conv_block(inputs, tchannel, (1, 1), (1, 1), nl) # 进行3x3深度可分离卷积 x = DepthwiseConv2D(kernel, strides=(stride, stride), depth_multiplier=1, padding='same')(x) x = BatchNormalization(axis=channel_axis)(x) x = return_activation(x, nl) # 引入注意力机制 if sq: x = squeeze(x) # 下降通道数 x = Conv2D(cchannel, (1, 1), strides=(1, 1), padding='same')(x) x = BatchNormalization(axis=channel_axis)(x) if r: x = Add()([x, inputs]) return x def MobileNetv3_large(shape = (224,224,3),n_class = 1000): inputs = Input(shape) # 224,224,3 -> 112,112,16 x = conv_block(inputs, 16, (3, 3), strides=(2, 2), nl='HS') x = bottleneck(x, 16, (3, 3), up_dim=16, stride=1, sq=False, nl='RE') # 112,112,16 -> 56,56,24 x = bottleneck(x, 24, (3, 3), up_dim=64, stride=2, sq=False, nl='RE') x = bottleneck(x, 24, (3, 3), up_dim=72, stride=1, sq=False, nl='RE') # 56,56,24 -> 28,28,40 x = bottleneck(x, 40, (5, 5), up_dim=72, stride=2, sq=True, nl='RE') x = bottleneck(x, 40, (5, 5), up_dim=120, stride=1, sq=True, nl='RE') x = bottleneck(x, 40, (5, 5), up_dim=120, stride=1, sq=True, nl='RE') # 28,28,40 -> 14,14,80 x = bottleneck(x, 80, (3, 3), up_dim=240, stride=2, sq=False, nl='HS') x = bottleneck(x, 80, (3, 3), up_dim=200, stride=1, sq=False, nl='HS') x = bottleneck(x, 80, (3, 3), up_dim=184, stride=1, sq=False, nl='HS') x = bottleneck(x, 80, (3, 3), up_dim=184, stride=1, sq=False, nl='HS') # 14,14,80 -> 14,14,112 x = bottleneck(x, 112, (3, 3), up_dim=480, stride=1, sq=True, nl='HS') x = bottleneck(x, 112, (3, 3), up_dim=672, stride=1, sq=True, nl='HS') # 14,14,112 -> 7,7,160 x = bottleneck(x, 160, (5, 5), up_dim=672, stride=2, sq=True, nl='HS') x = bottleneck(x, 160, (5, 5), up_dim=960, stride=1, sq=True, nl='HS') x = bottleneck(x, 160, (5, 5), up_dim=960, stride=1, sq=True, nl='HS') # 7,7,160 -> 7,7,960 x = 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 = return_activation(x, 'HS') x = Conv2D(n_class, (1, 1), padding='same', activation='softmax')(x) x = Reshape((n_class,))(x) model = Model(inputs, x) return model if __name__ == "__main__": model = MobileNetv3_large() model.summary()
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。