Cute image classification -- a general explanation of the article "what are the common flops in CNN model?"

《CNN Common in the model flops What is it 》

Let me write it out front

One 、 So what is FLOPs? 

Two 、 How to calculate FLOPs？

---

Let me write it out front

Reading recently CNN Various papers of , We found an index called flops It can be said that every paper must see , Even when comparing models , This flops It is also almost used . therefore , I can't help being right flops There was a strong interest in , After learning , I can't help but think of sharing what I have learned 、

Like this picture ： 

One 、 So what is FLOPs? 

• FLOPS： Pay attention to all capitals , yes floating point operations per second Abbreviation , Floating point operations per second , It means computing speed . Is a measure of hardware performance .
• FLOPs： Be careful s A lowercase letter , yes floating point operations Abbreviation （s Plural ）, It means floating point arithmetic , It means the amount of calculation . It can be used to measure the algorithm / The complexity of the model .

        We all know ： Let's go to the Evaluate a model When , The first element is, of course, to see this The accuracy of the model Acc.

         But second , Different models achieve the same accuracy Acc The computational power required for forward propagation calculation may vary greatly , and flops It is this calculating power that may be very different , This reflects the hardware GPU Performance requirements .

         such as ： Arrive at the same 87% about Of Accuracy rate ,ConvNeXt Of FLOPs by 34.4G and Swin Transformer It is 103.9G！

         besides ：1） Parameter quantity                                 （ It reflects the amount of memory occupied

                          2）throughout（image/s)      （ It reflects the speed of model training

         And so on are also very important indicators of the evaluation model

  Two 、 How to calculate FLOPs？

          Use FLOPs Open source library for Computing ：  

import torchvision.models as models
from flops counter import get model_complexity_info

net = models.vgg16

flops, params = get model complexity info(net, (224,224),as_strings=True, print per layer_stat=True)

print( FLops:0'.format(flops)

print('Params: '+params)

         The picture above —— Example code calculation VGG16 Of FLOPs And parameter quantities . 

         The following pictures —— Show the running results :

        Several common models FLOPs And parameter quantity ：