上一片博客(指路)末尾有提到计算公式,这里就不再赘述了放个截图。
本文将以AlexNet为例,计算AlexNet参数量量和浮点运算次数。
计算参数量
1.首先回顾参数量的计算公式:
如果输入是C channel ×H×W;卷积核是C inchannel × M outchannel ×K ×K;
parameters = [(K ×K)×C inchannel ]×M outchannel+M
2.分析AlexNet网络结构
共进行五次卷积操作,各层之间的kernel大小和输出feature尺寸下图中给出:
各层参数:
parameters = [(K ×K)×C ]×M+M
CONV1:
parameters1 = 11×11×3×96+96=34944
CONV2:
parameters2 = 5×5×96×256+256=307456
CONV3:
parameters3 = 3×3×256×384+384=885120
CONV4:
parameters4 = 3×3×384×384+384=663936
CONV5:
parameters5 = 3×3×384×256+256=442624
FC1:
parameters6 = 6×6×256×4096+4096=37752832
FC2:
parameters7 = 4069×4096+4096=16781312
FC3:
parameters8 = 4069×1000+1000=4097000
计算FLOPs
FLOPs = parameters×H×W
CONV1
FLOPs = parameter1×H×W=34944×55×55
CONV2
FLOPs = parameter2×H×W=307456×27×27
CONV3
FLOPs = parameter3×H×W=34944×13×13
CONV4
FLOPs = parameter4×H×W=34944×13×13
CONV5
FLOPs = parameter5×H×W=34944×13×13
FC1
全连接层不存在权值共享,所以 FLOPs=parameters
FLOPs=parameters=37752832
FC2
FLOPs=parameters=16781312
FC3
FLOPs=parameters=4097000
最后,贴一张百度百科的图:
代码实现
参考GitHub地址FLOPs计算
import torchvision.models as models
import torch
from ptflops import get_model_complexity_info
with torch.cuda.device(0):
#net = models.alexnet()
net = models.resnet50()
macs, params = get_model_complexity_info(net, (3, 224, 224), as_strings=True,
print_per_layer_stat=True, verbose=True)
print('{:<30} {:<8}'.format('Computational complexity: ', macs))
print('{:<30} {:<8}'.format('Number of parameters: ', params))
Alexnet: