深度学习中模型计算量(FLOPs)和参数量(Params)的理解以及四种计算方法总结

接下来要分别概述以下内容：

１　首先什么是参数量，什么是计算量

２　如何计算　参数量，如何统计　计算量

３　换算参数量，把他换算成我们常用的单位，比如:mb

４　对于各个经典网络，论述他们是计算量大还是参数量，有什么好处

1 首先什么是计算量，什么是参数量

计算量对应我们之前的时间复杂度，参数量对应于我们之前的空间复杂度，这么说就很明显了

也就是计算量要看网络执行时间的长短，参数量要看占用显存的量

2 如何计算：参数量，计算量

对于卷积层：

参数量就是

(kernel*kernel) *channel_input*channel_output kernel*kernel 就是 weight * weight 其中kernel*kernel ＝ 1个feature的参数量 

计算量就是

(kernel*kernel*map*map) *channel_input*channel_output kernel*kernel 就是weight*weight map*map是下个featuremap的大小，也就是上个weight*weight到底做了多少次运算 其中kernel*kernel*map*map＝　1个feature的计算量 

（２）针对于池化层：

无参数

（３）针对于全连接层：

参数量＝计算量＝weight_in*weight_out 

3 对于换算计算量

• 一般一个参数是值一个float，也就是４个字节
• 1kb=1024字节

4 对于各个经典网络：

(1)换算

以alexnet为例：

参数量：6000万

设每个参数都是float,也就是一个参数是4字节，

总的字节数是24000万字节

24000万字节= 24000万/1024/1024=228mb

(2)为什么模型之间差距这么大

这个关乎于模型的设计了，其中模型里面最费参数的就是全连接层，这个可以看alex和vgg,

alex,vgg有很多fc(全连接层)

resnet就一个fc

inceptionv1(googlenet)也是就一个fc

(3)计算量

densenet其实这个模型不大，也就是参数量不大，因为就1个fc

但是他的计算量确实很大，因为每一次都把上一个feature加进来，所以计算量真的很大

5 计算量与参数量对于硬件要求

计算量，参数量对于硬件的要求是不同的

计算量的要求是在于芯片的floaps（指的是gpu的运算能力）

参数量取决于显存大小

6 计算量(FLOPs)和参数量(Params)

6.1 第一种方法：thop

第一步：安装模块

pip install thop 

第二步：计算

# -- coding: utf-8 -- import torch import torchvision from thop import profile # Model print('==> Building model..') model = torchvision.models.alexnet(pretrained=False) dummy_input = torch.randn(1, 3, 224, 224) flops, params = profile(model, (dummy_input,)) print('flops: ', flops, 'params: ', params) print('flops: %.2f M, params: %.2f M' % (flops / .0, params / .0)) 

结果

==> Building model.. [INFO] Register count_convNd() for <class 'torch.nn.modules.conv.Conv2d'>. [INFO] Register zero_ops() for <class 'torch.nn.modules.activation.ReLU'>. [INFO] Register zero_ops() for <class 'torch.nn.modules.pooling.MaxPool2d'>. [WARN] Cannot find rule for <class 'torch.nn.modules.container.Sequential'>. Treat it as zero Macs and zero Params. [INFO] Register count_adap_avgpool() for <class 'torch.nn.modules.pooling.AdaptiveAvgPool2d'>. [INFO] Register zero_ops() for <class 'torch.nn.modules.dropout.Dropout'>. [INFO] Register count_linear() for <class 'torch.nn.modules.linear.Linear'>. [WARN] Cannot find rule for <class 'torchvision.models.alexnet.AlexNet'>. Treat it as zero Macs and zero Params. flops: .0 params: .0 flops: 714.69 M, params: 61.10 M 

注意：

• 输入input的第一维度是批量(batch size)，批量的大小不回影响参数量， 计算量是batch_size=1的倍数
• profile(net, (inputs,))的 (inputs,)中必须加上逗号，否者会报错

6.2 第二种方法：ptflops

# -- coding: utf-8 -- import torchvision from ptflops import get_model_complexity_info model = torchvision.models.alexnet(pretrained=False) flops, params = get_model_complexity_info(model, (3, 224, 224), as_strings=True, print_per_layer_stat=True) print('flops: ', flops, 'params: ', params) 

结果

AlexNet( 61.101 M, 100.000% Params, 0.716 GMac, 100.000% MACs, (features): Sequential( 2.47 M, 4.042% Params, 0.657 GMac, 91.804% MACs, (0): Conv2d(0.023 M, 0.038% Params, 0.07 GMac, 9.848% MACs, 3, 64, kernel_size=(11, 11), stride=(4, 4), padding=(2, 2)) (1): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.027% MACs, inplace=True) (2): MaxPool2d(0.0 M, 0.000% Params, 0.0 GMac, 0.027% MACs, kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) (3): Conv2d(0.307 M, 0.503% Params, 0.224 GMac, 31.316% MACs, 64, 192, kernel_size=(5, 5), stride=(1, 1), padding=(2, 2)) (4): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.020% MACs, inplace=True) (5): MaxPool2d(0.0 M, 0.000% Params, 0.0 GMac, 0.020% MACs, kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) (6): Conv2d(0.664 M, 1.087% Params, 0.112 GMac, 15.681% MACs, 192, 384, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (7): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.009% MACs, inplace=True) (8): Conv2d(0.885 M, 1.448% Params, 0.15 GMac, 20.902% MACs, 384, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (9): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.006% MACs, inplace=True) (10): Conv2d(0.59 M, 0.966% Params, 0.1 GMac, 13.936% MACs, 256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1)) (11): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.006% MACs, inplace=True) (12): MaxPool2d(0.0 M, 0.000% Params, 0.0 GMac, 0.006% MACs, kernel_size=3, stride=2, padding=0, dilation=1, ceil_mode=False) ) (avgpool): AdaptiveAvgPool2d(0.0 M, 0.000% Params, 0.0 GMac, 0.001% MACs, output_size=(6, 6)) (classifier): Sequential( 58.631 M, 95.958% Params, 0.059 GMac, 8.195% MACs, (0): Dropout(0.0 M, 0.000% Params, 0.0 GMac, 0.000% MACs, p=0.5, inplace=False) (1): Linear(37.753 M, 61.788% Params, 0.038 GMac, 5.276% MACs, in_features=9216, out_features=4096, bias=True) (2): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.001% MACs, inplace=True) (3): Dropout(0.0 M, 0.000% Params, 0.0 GMac, 0.000% MACs, p=0.5, inplace=False) (4): Linear(16.781 M, 27.465% Params, 0.017 GMac, 2.345% MACs, in_features=4096, out_features=4096, bias=True) (5): ReLU(0.0 M, 0.000% Params, 0.0 GMac, 0.001% MACs, inplace=True) (6): Linear(4.097 M, 6.705% Params, 0.004 GMac, 0.573% MACs, in_features=4096, out_features=1000, bias=True) ) ) flops: 0.72 GMac params: 61.1 M 

6.3 第三种方法：pytorch_model_summary

import torch import torchvision from pytorch_model_summary import summary # Model print('==> Building model..') model = torchvision.models.alexnet(pretrained=False) dummy_input = torch.randn(1, 3, 224, 224) print(summary(model, dummy_input, show_input=False, show_hierarchical=False)) 

结果

==> Building model.. ----------------------------------------------------------------------------- Layer (type) Output Shape Param # Tr. Param # ============================================================================= Conv2d-1 [1, 64, 55, 55] 23,296 23,296 ReLU-2 [1, 64, 55, 55] 0 0 MaxPool2d-3 [1, 64, 27, 27] 0 0 Conv2d-4 [1, 192, 27, 27] 307,392 307,392 ReLU-5 [1, 192, 27, 27] 0 0 MaxPool2d-6 [1, 192, 13, 13] 0 0 Conv2d-7 [1, 384, 13, 13] 663,936 663,936 ReLU-8 [1, 384, 13, 13] 0 0 Conv2d-9 [1, 256, 13, 13] 884,992 884,992 ReLU-10 [1, 256, 13, 13] 0 0 Conv2d-11 [1, 256, 13, 13] 590,080 590,080 ReLU-12 [1, 256, 13, 13] 0 0 MaxPool2d-13 [1, 256, 6, 6] 0 0 AdaptiveAvgPool2d-14 [1, 256, 6, 6] 0 0 Dropout-15 [1, 9216] 0 0 Linear-16 [1, 4096] 37,752,832 37,752,832 ReLU-17 [1, 4096] 0 0 Dropout-18 [1, 4096] 0 0 Linear-19 [1, 4096] 16,781,312 16,781,312 ReLU-20 [1, 4096] 0 0 Linear-21 [1, 1000] 4,097,000 4,097,000 ============================================================================= Total params: 61,100,840 Trainable params: 61,100,840 Non-trainable params: 0 ----------------------------------------------------------------------------- 

6.4 第四种方法：参数总量和可训练参数总量

import torch import torchvision from pytorch_model_summary import summary # Model print('==> Building model..') model = torchvision.models.alexnet(pretrained=False) pytorch_total_params = sum(p.numel() for p in model.parameters()) trainable_pytorch_total_params = sum(p.numel() for p in model.parameters() if p.requires_grad) print('Total - ', pytorch_total_params) print('Trainable - ', trainable_pytorch_total_params) 

结果

==> Building model.. Total -  Trainable -  

7 输入数据对模型的参数量和计算量的影响

# -- coding: utf-8 -- import torch import torchvision from thop import profile # Model print('==> Building model..') model = torchvision.models.alexnet(pretrained=False) dummy_input = torch.randn(1, 3, 224, 224) flops, params = profile(model, (dummy_input,)) print('flops: ', flops, 'params: ', params) print('flops: %.2f M, params: %.2f M' % (flops / .0, params / .0)) 

• 输入数据：(1, 3, 224, 224)，一张224*224的RGB图像

flops: .0 params: .0 flops: 714.69 M, params: 61.10 M 

• 输入数据：(1, 3, 512, 512)，一张512*512的RGB图像

flops: .0 params: .0 flops: 3710.03 M params: 61.10 M 

• 输入数据：(8, 3, 224, 224)，八张224*224的RGB图像

flops: .0 params: .0 flops: 5717.54 M params: 61.10 M 

参考资料

1. https://www.cnblogs.com/lllcccddd/p/10671879.html
2. https://blog.csdn.net/Caesar6666/article/details/