ResNeXt

ResNeXt

背景介绍

  ResNeXt:是2017年发表于CVPR的一个模型，是ResNet网络的升级版本。和Inception-ResNet类似，Inception-ResNet可以认为是Inception模型的基础上吸收ResNet残差思想，而ResNext则可以认为是ResNet模型的基础上吸收Inception分块合并思想。

ResNeXt特点

  网络结构和ResNet相同，根据ResNet50，则可以修改为ResNeXt50，根据ResNet101，则可以修改为ResNeXt101，等等
  引入Inception模型分块合并思想，将ResNet中Conv Block和Identity Block中的普通卷积变成GroupConv分组卷积。提出了**cardinality(基数)**名词，基数为32，相当于分组卷积的组数为32，最后将32组卷积结果合并

Group Convolution

  Group Convolution(分组卷积)：传统卷积是采用一种卷积全连接的思想，特征图中的每一个像素点都结合了图像中所有通道的信息。而分组卷积特征图像每一个像素点只利用到一部分原始图像的通道。
  主要作用是大大降低网络的参数量。如果一个64x64x256的图像，经过5x5的卷积核后变为64x64x256的图像，经过普通卷积的参数量为256x(256x5x5+1)=1638656，而分成32组的分组卷积的参数量为256x(8*5x5+1)=51456，参数量缩小了约32倍，当组数变成通道数时，则类似于Depthwise Convolution深度卷积

ResNeXt50图像分析

TensorFlow2.0实现

from functools import reduce
import tensorflow as tf
import tensorflow.keras as keras


def compose(*funcs):
    if funcs:
        return reduce(lambda f, g: lambda *a, **kw: g(f(*a, **kw)), funcs)
    else:
        raise ValueError('Composition of empty sequence not supported.')


class GroupConv(keras.layers.Layer):
    def __init__(self, filters, kernel_size, strides, g_num, name='groupconv'):
        super(GroupConv, self).__init__()
        self._name = name
        self.g_num = g_num
        self.groupconv = [keras.layers.Conv2D(filters // g_num, kernel_size, strides, padding='same', name='{}{}'.format(name, i + 1)) for i in range(g_num)]
        self.concatenate = keras.layers.Concatenate(name='{}_concatenate'.format(name))

    def call(self, inputs, **kwargs):
        x_split = tf.split(inputs, self.g_num, axis=-1)
        x = [self.groupconv[i](x_split[i]) for i in range(self.g_num)]
        output = self.concatenate(x)

        return output


class Conv_Bn_Relu(keras.layers.Layer):
    def __init__(self, filters, kernel_size, strides, padding, name):
        super(Conv_Bn_Relu, self).__init__()
        self._name = name
        if name.find('group') == -1:
            self.conv = keras.layers.Conv2D(filters, kernel_size, strides, padding)
        else:
            self.conv = GroupConv(filters, kernel_size, strides, 32)
        self.bn = keras.layers.BatchNormalization()
        self.relu = keras.layers.ReLU()

    def call(self, inputs, **kwargs):
        conv = self.conv(inputs)
        bn = self.bn(conv)
        output = self.relu(bn)

        return output


class Conv_Bn(keras.layers.Layer):
    def __init__(self, filters, kernel_size, strides, padding, name):
        super(Conv_Bn, self).__init__()
        self._name = name
        self.conv = keras.layers.Conv2D(filters, kernel_size, strides, padding)
        self.bn = keras.layers.BatchNormalization()

    def call(self, inputs, **kwargs):
        conv = self.conv(inputs)
        output = self.bn(conv)

        return output


def res_block(x, filters, strides, type, name):
    shortcut = x
    x = compose(Conv_Bn_Relu(filters // 2, (1, 1), (1, 1), padding='same', name='{}{}_conv_bn_relu1'.format(type, name)),
                Conv_Bn_Relu(filters // 2, (3, 3), strides, padding='same', name='{}{}_groupconv_bn_relu2'.format(type, name)),
                Conv_Bn(filters, (1, 1), (1, 1), padding='same', name='{}{}_conv_bn3'.format(type, name)))(x)
    if type == 'conv_block':
        shortcut = keras.layers.Conv2D(filters, (1, 1), strides, name='{}{}_shortcut'.format(type, name))(shortcut)
    x = keras.layers.Add(name='{}{}_add'.format(type, name))([x, shortcut])
    x = keras.layers.ReLU(name='{}{}_relu3'.format(type, name))(x)

    return x


def resnext50(input_shape):
    input_tensor = keras.layers.Input(input_shape, name='input')
    x = input_tensor
    x = compose(keras.layers.ZeroPadding2D((3, 3), name='zeropadding'),
                keras.layers.Conv2D(64, (7, 7), (2, 2), name='conv1'),
                keras.layers.BatchNormalization(name='bn'),
                keras.layers.ReLU(name='relu'),
                keras.layers.MaxPool2D((3, 3), (2, 2), padding='same', name='maxpool'))(x)
    filters = [256, 512, 1024, 2048]
    strides = [(1, 1), (2, 2), (2, 2), (2, 2)]
    times = [3, 4, 6, 3]
    for i in range(len(times)):
        x = res_block(x, filters[i], strides[i], 'conv_block', i + 1)
        for j in range(times[i] - 1):
            x = res_block(x, filters[i], (1, 1), 'identity_block{}_'.format(i + 1), j + 1)
    x = compose(keras.layers.GlobalAveragePooling2D(name='global_averagepool'),
                keras.layers.Dense(1000, activation='softmax', name='dense'))(x)
    model = keras.Model(input_tensor, x, name='ResNeXt50')

    return model


if __name__ == '__main__':

    model = resnext50(input_shape=(224, 224, 3))
    model.build(input_shape=(None, 224, 224, 3))
    model.summary()

ResNeXt小结

  ResNeXt是一种非常有效的特征提取网络，ResNeXt参数量为25M，和相同结构的ResNet几乎相同，但是效果优于ResNet，因此是一种实用的特征提取网络。