HS-ResNet，模型涨点的新方案

前言

近几年的模型结构的改进，越来越偏向于宽度的改进，如SKnet以及ResNeSt，都获得了不错的结果，本文讨论的是百度所提出的HS-ResNet，通过对网络的复现以及论文的阅读，感觉本片paper是把Res2Net思想和ResNeSt结合到一起了。

模型介绍

网络模型如图所示，先像ResNeSt一样，在channel这个层级上把结构split为5分，每一份都和前面的一份concat（类似Res2Net）操作，后进行卷积，最后再把得到的特征进行concat。

代码实现

import tensorflow as tf

x = tf.placeholder(tf.float32,[None, 500, 500, 3])#输入图片大小

def Hierarchical_split(x, channel):
    channel_begin=x.get_shape().as_list()[-1]
    ##第一阶段
    layer1=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(x,channel,1,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))
    layer1_0,layer1_1,layer1_2,layer1_3,layer1_4=tf.split(layer1,5,-1)
    ##第二阶段
    channel1=layer1_1.get_shape().as_list()[-1]*2
    layer1_1=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(layer1_1,channel1,3,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))
    layer1_1_0,layer1_1_1=tf.split(layer1_1,2,-1)    
    ##第三阶段
    channel2=layer1_2.get_shape().as_list()[-1]*2

    layer1_2=tf.concat([layer1_2,layer1_1_1],-1)
    layer1_2=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(layer1_2,channel2,3,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))
    layer1_2_0,layer1_2_1=tf.split(layer1_2,2,-1)          
    ##第四阶段
    channel3=layer1_3.get_shape().as_list()[-1]*2
    layer1_3=tf.concat([layer1_3,layer1_2_1],-1)
    layer1_3=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(layer1_3,channel3,3,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))
    layer1_3_0,layer1_3_1=tf.split(layer1_3,2,-1)             
    ##第五阶段
    channel4=layer1_4.get_shape().as_list()[-1]*2
    layer1_4=tf.concat([layer1_4,layer1_3_1],-1)
    layer1_4=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(layer1_4,channel4,3,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))      
    ##第六阶段
    layer_all=tf.concat([layer1_0, layer1_1_0, layer1_2_0, layer1_3_0, layer1_4],-1)  
    layer_all=tf.nn.relu(tf.layers.batch_normalization(tf.layers.conv2d(layer_all,channel_begin,1,strides=1, padding='same'),axis=-1,momentum=0.99,epsilon=0.001, center=True, scale=True,))
    return x+layer_all
layer1=tf.nn.relu(tf.layers.conv2d(x,256,3,strides=1, padding='same'))
layer2=Hierarchical_split(layer1,250)
print(layer2)