FLOPS: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。衡量硬件性能的指标。
FLOPs:floating point operations的缩写(s表复数),意指浮点运算数,理解为计算量。衡量算法/模型的复杂度。
假设采用滑动窗实现卷积且忽略非线性计算开销,则卷积核的FLOPs为,其中、和分别为输入特征图(就是输入图片了)的高度、宽度和通道数,为核宽度,为输出通道数。
全连接层网络FLOPs为,其中,为输入维数,为输出维数。
Ref: Molchanov P, Tyree S, Karras T, et al. Pruning Convolutional Neural Networks for Resource Efficient Inference[J]. 2016.
https://blog.csdn.net/zhaoyin214/article/details/83616164
https://blog.csdn.net/wangdongwei0/article/details/88389775
https://zhuanlan.zhihu.com/p/33992733
https://zhuanlan.zhihu.com/p/34475420
https://zhuanlan.zhihu.com/p/29024978
https://blog.csdn.net/peaceinmind/article/details/78079263
https://blog.csdn.net/warm_in_spring/article/details/79133459
https://blog.csdn.net/hzhj2007/article/details/80164909
https://blog.csdn.net/hzhj2007/article/details/80164909
https://blog.csdn.net/qian99/article/details/79008053
https://blog.csdn.net/wgx571859177/article/details/80983043
https://fomoro.com/projects/project/receptive-field-calculator