You can not select more than 25 topics Topics must start with a chinese character,a letter or number, can include dashes ('-') and can be up to 35 characters long.
 
 
 
 
huicui 20c6023900 Single_Path_NAS Ascend910 2 years ago
..
lr_scheduler Single_Path_NAS Ascend910 2 years ago
CrossEntropySmooth.py Single_Path_NAS Ascend910 2 years ago
__init__.py Single_Path_NAS Ascend910 2 years ago
config.py Single_Path_NAS Ascend910 2 years ago
dataset.py Single_Path_NAS Ascend910 2 years ago
spnasnet.py Single_Path_NAS Ascend910 2 years ago
utils.py Single_Path_NAS Ascend910 2 years ago

Single_Path_NAS通过一个7x7的卷积来替换3x3、5x5和7x7的三种卷积(外边一圈mask清零可以变成3x3或5x5),这个大的卷积成为superkernel,于是整个网络只有一种卷积,看起来是一个直筒结构。搜索空间是基于block的直筒结构,跟ProxylessNAS和FBNet一样,都采用Inverted Bottleneck 作为cell, 层数跟MobileNetV2一样都是22层,每层只有两个参数 expansion rate, kernel size需要搜索,其他都已固定,比如22层中每层的filter number已固定死,跟FBNet一样,跟MobileNetV2比略有变化。论文中的kernel size和FBNet、 ProxylessNAS一样只有3x3和5x5两种,没有用上7x7。论文中的expansion ratio也只有3和6两种选择。kernel size 和 expansion ratio都只有2种选择,论文选择用Lightnn这篇论文中的手法,把离散选择用连续的光滑函数来表示,阈值用group Lasso term。论文用了跟ProxylessNAS一样的手法来表达skip connection, 用一个zero layer表示。

Python C++ Shell Text

Contributors (2)