网络模型的参数量和FLOPs的计算 Pytorch

FLOPS和FLOPs的区别：

• FLOPS：注意全大写，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。
• FLOPs：注意s小写，是floating point operations的缩写（s表复数），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。

在介绍torchstat包和thop包之前，先总结一下：

• torchstat包可以统计卷积神经网络和全连接神经网络的参数和计算量。
• thop包可以统计统计卷积神经网络、全连接神经网络以及循环神经网络的参数和计算量，程序示例等详见下文。

1、torchstat

pip install torchstat -i https://pypi.tuna.tsinghua.edu.cn/simple

在实际操作中，我们可以调用torchstat包，帮助我们统计模型的parameters和FLOPs。如果不修改这个包里面的一些代码，那么这个包只适用于输入为3通道的图像的模型。

import torch
import torch.nn as nn
from torchstat import stat
 
 
class Simple(nn.Module):
    def __init__(self):
        super().__init__()
        self.conv1 = nn.Conv2d(3, 16, 3, 1, padding=1, bias=False)
        self.conv2 = nn.Conv2d(16, 32, 3, 1, padding=1, bias=False)
 
    def forward(self, x):
        x = self.conv1(x)
        x = self.conv2(x)
        return x
 
 
model = Simple()
stat(model, (3, 244, 244))   # 统计模型的参数量和FLOPs，（3,244,244）是输入图像的size

如果把torchstat包中的一行程序进行一点点改动，那么这个包可以用来统计全连接神经网络的参数量和计算量。当然手动计算全连接神经网络的参数量和计算量也很快 =_= 。进入torchstat源代码之后，如下图所示，注释掉圈红的地方，就可以用torchstat包统计全连接神经网络的参数量和计算量了。

2、thop

pip install thop -i https://pypi.tuna.tsinghua.edu.cn/simple

import torch
import torch.nn as nn
from thop import profile
 
class Simple(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc1 = nn.Linear(10, 10)
 
    def forward(self, x):
        x = self.fc1(x)
        return x
 
net = Simple()
input = torch.randn(1, 10)  # batchsize=1, 输入向量长度为10
macs, params = profile(net, inputs=(input, ))
print(' FLOPs: ', macs*2)   # 一般来讲，FLOPs是macs的两倍
print('params: ', params)

3、fvcore

pip install fvcore -i https://pypi.tuna.tsinghua.edu.cn/simple

用它比较好

import torch
from torchvision.models import resnet50
from fvcore.nn import FlopCountAnalysis, parameter_count_table

# 创建resnet50网络
model = resnet50(num_classes=1000)

# 创建输入网络的tensor
tensor = (torch.rand(1, 3, 224, 224),)

# 分析FLOPs
flops = FlopCountAnalysis(model, tensor)
print("FLOPs: ", flops.total())

# 分析parameters
print(parameter_count_table(model))

终端输出结果如下，FLOPs为4089184256，模型参数数量约为25.6M（这里的参数数量和我自己计算的有些出入，主要是在BN模块中，这里只计算了beta和gamma两个训练参数，没有统计moving_mean和moving_var两个参数），具体可以看下我在官方提的issue。
通过终端打印的信息我们可以发现在计算FLOPs时并没有包含BN层，池化层还有普通的add操作（我发现计算FLOPs时并没有统一的规定，在github上看的计算FLOPs项目基本每个都不同，但计算出来的结果大同小异）。

注意：在使用fvcore模块计算模型的flops时，遇到了问题，记录一下解决方案。首先是在jit_analysis.py的589行出错。经过调试发现，op_counts.values()的类型是int32，但是计算要求的类型只能是int、float、np.float64和np.int64，因此需要手动进行强制转换。修改如下：

4、flops_counter

pip install ptflops -i https://pypi.tuna.tsinghua.edu.cn/simple

用它也很好，结果和fvcore一样

from ptflops import get_model_complexity_info

macs, params = get_model_complexity_info(model, (112, 9, 9), 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))

5、自定义统计函数

import torch
import numpy as np

def calc_flops(model, input):
    def conv_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = self.kernel_size[0] * self.kernel_size[1] * (self.in_channels / self.groups) * (
            2 if multiply_adds else 1)
        bias_ops = 1 if self.bias is not None else 0

        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_conv.append(flops)

    def linear_hook(self, input, output):
        batch_size = input[0].size(0) if input[0].dim() == 2 else 1
        num_steps = input[0].size(0)
        weight_ops = self.weight.nelement() * (2 if multiply_adds else 1)
        bias_ops = self.bias.nelement() if self.bias is not None else 0

        flops = batch_size * (weight_ops + bias_ops)
        flops *= num_steps
        list_linear.append(flops)

    def fsmn_hook(self, input, output):
        batch_size = input[0].size(0) if input[0].dim() == 2 else 1

        weight_ops = self.filter.nelement() * (2 if multiply_adds else 1)
        num_steps = input[0].size(0)
        flops = num_steps * weight_ops
        flops *= batch_size
        list_fsmn.append(flops)

    def gru_cell(input_size, hidden_size, bias=True):
        total_ops = 0
        # r = \sigma(W_{ir} x + b_{ir} + W_{hr} h + b_{hr}) \\
        # z = \sigma(W_{iz} x + b_{iz} + W_{hz} h + b_{hz}) \\
        state_ops = (hidden_size + input_size) * hidden_size + hidden_size
        if bias:
            state_ops += hidden_size * 2
        total_ops += state_ops * 2

        # n = \tanh(W_{in} x + b_{in} + r * (W_{hn} h + b_{hn})) \\
        total_ops += (hidden_size + input_size) * hidden_size + hidden_size
        if bias:
            total_ops += hidden_size * 2
        # r hadamard : r * (~)
        total_ops += hidden_size

        # h' = (1 - z) * n + z * h
        # hadamard hadamard add
        total_ops += hidden_size * 3

        return total_ops

    def gru_hook(self, input, output):

        batch_size = input[0].size(0) if input[0].dim() == 2 else 1
        if self.batch_first:
            batch_size = input[0].size(0)
            num_steps = input[0].size(1)
        else:
            batch_size = input[0].size(1)
            num_steps = input[0].size(0)
        total_ops = 0
        bias = self.bias
        input_size = self.input_size
        hidden_size = self.hidden_size
        num_layers = self.num_layers
        total_ops = 0
        total_ops += gru_cell(input_size, hidden_size, bias)
        for i in range(num_layers - 1):
            total_ops += gru_cell(hidden_size, hidden_size, bias)
        total_ops *= batch_size
        total_ops *= num_steps

        list_lstm.append(total_ops)

    def lstm_cell(input_size, hidden_size, bias):
        total_ops = 0
        state_ops = (input_size + hidden_size) * hidden_size + hidden_size
        if bias:
            state_ops += hidden_size * 2
        total_ops += state_ops * 4
        total_ops += hidden_size * 3
        total_ops += hidden_size
        return total_ops

    def lstm_hook(self, input, output):

        batch_size = input[0].size(0) if input[0].dim() == 2 else 1
        if self.batch_first:
            batch_size = input[0].size(0)
            num_steps = input[0].size(1)
        else:
            batch_size = input[0].size(1)
            num_steps = input[0].size(0)
        total_ops = 0
        bias = self.bias
        input_size = self.input_size
        hidden_size = self.hidden_size
        num_layers = self.num_layers
        total_ops = 0
        total_ops += lstm_cell(input_size, hidden_size, bias)
        for i in range(num_layers - 1):
            total_ops += lstm_cell(hidden_size, hidden_size, bias)
        total_ops *= batch_size
        total_ops *= num_steps

        list_lstm.append(total_ops)

    def bn_hook(self, input, output):
        list_bn.append(input[0].nelement())

    def relu_hook(self, input, output):
        list_relu.append(input[0].nelement())

    def pooling_hook(self, input, output):
        batch_size, input_channels, input_height, input_width = input[0].size()
        output_channels, output_height, output_width = output[0].size()

        kernel_ops = self.kernel_size * self.kernel_size
        bias_ops = 0
        params = output_channels * (kernel_ops + bias_ops)
        flops = batch_size * params * output_height * output_width

        list_pooling.append(flops)

    def foo(net):
        childrens = list(net.children())
        if not childrens:
            print(net)
            if isinstance(net, torch.nn.Conv2d) or isinstance(net, torch.nn.ConvTranspose2d):
                net.register_forward_hook(conv_hook)
                # print('conv_hook_ready')
            if isinstance(net, torch.nn.Linear):
                net.register_forward_hook(linear_hook)
                # print('linear_hook_ready')
            if isinstance(net, torch.nn.BatchNorm2d):
                net.register_forward_hook(bn_hook)
                # print('batch_norm_hook_ready')
            if isinstance(net, torch.nn.ReLU) or isinstance(net, torch.nn.PReLU):
                net.register_forward_hook(relu_hook)
                # print('relu_hook_ready')
            if isinstance(net, torch.nn.MaxPool2d) or isinstance(net, torch.nn.AvgPool2d):
                net.register_forward_hook(pooling_hook)
                # print('pooling_hook_ready')
            if isinstance(net, torch.nn.LSTM):
                net.register_forward_hook(lstm_hook)
                # print('lstm_hook_ready')
            if isinstance(net, torch.nn.GRU):
                net.register_forward_hook(gru_hook)

            # if isinstance(net, FSMNZQ):
            #     net.register_forward_hook(fsmn_hook)
                # print('fsmn_hook_ready')
            return
        for c in childrens:
            foo(c)

    multiply_adds = False
    list_conv, list_bn, list_relu, list_linear, list_pooling, list_lstm, list_fsmn = [], [], [], [], [], [], []
    foo(model)

    _ = model(input)

    total_flops = (sum(list_conv) + sum(list_linear) + sum(list_bn) + sum(list_relu) + sum(list_pooling) + sum(
        list_lstm) + sum(list_fsmn))
    fsmn_flops = (sum(list_fsmn) + sum(list_linear))
    lstm_flops = sum(list_lstm)

    model_parameters = filter(lambda p: p.requires_grad, model.parameters())
    params = sum([np.prod(p.size()) for p in model_parameters])
    print('The network has {} params.'.format(params))

    print(total_flops, fsmn_flops, lstm_flops)
    print('  + Number of FLOPs: %.2f M' % (total_flops / 1000 ** 2))
    return total_flops

if __name__ == '__main__':
    from torchvision.models import resnet18

    model = resnet18(num_classes=1000)
    imput_size = torch.rand((1,3,224,224))
    calc_flops(model, imput_size)