动机
一些以往不清楚的深度学习方面的基本概念。学习后做下记录。
分类1
-
网络的depth、width、cardinality
depth
:整个网络的层数width
:每一层的通道(channel)数cardinality
:首次出现在ResNeXt中, 原文解释为‘the size of the set of transformations’,可理解为同一层网络中的支路数。
The architecture of neural networks often specified by the width and the depth of the networks. The depth of a network is defined as its number of layers (including output layer but excluding input layer); while the width of a network is defined to be the maximal number of nodes in a layer. The number of input nodes, i.e. the input dimension, is denoted as n n n.
参考-The Expressive Power of Neural Networks - NIPS Proceedings -
FLOPs
- floating point operations,意指浮点运算数,理解为计算量。可以用来衡量算法/模型的复杂度。
- 与FLOPS(注意全大写)区分开来,FLOPS是floating point operations per second的缩写,意指每秒浮点运算次数,理解为计算速度。是一个衡量硬件性能的指标。