简单数值计算时是否需要把数据转化成tensor格式在gpu上跑呢

原创已于 2022-11-30 22:37:00 修改 · 372 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#python #pytorch

于 2022-11-30 22:36:41 首次发布

本文对比了使用Tensor（CUDA启用和禁用）、math模块以及直接函数调用进行数值计算的效率。结果显示，在不使用GPU的情况下，直接使用math模块和函数调用的效率更高，且差距极小。对于简单计算，没有必要将数据转换为Tensor格式。

部署运行你感兴趣的模型镜像

1，使用tensor，USE_CUDA=True

import torch,time,math
USE_CUDA=torch.cuda.is_available()
USE_CUDA=True

def get_location(elevation,azimuth):
    elevation_tensor=torch.tensor(elevation)
    azimuth_tensor=torch.tensor(azimuth)
    if USE_CUDA:
        elevation_tensor.cuda()
        azimuth_tensor.cuda()

    x_tensor = torch.sin(torch.pi * elevation_tensor / 180) *  torch.cos( torch.pi * azimuth_tensor / 180)
    y_tensor = torch.sin(torch.pi * elevation_tensor / 180) *  torch.sin( torch.pi * azimuth_tensor / 180)
    z_tensor = torch.cos(torch.pi * elevation_tensor / 180)
    x=x_tensor.item()
    y=y_tensor.item()
    z=z_tensor.item()
    
    print(x,y,z)
    return x,y,z
    
start = time.time()
get_location(0,90)
end = time.time()
print(end-start)

结果
-0.0 0.0 1.0
3.648632049560547

2,使用tensor，USE_CUDA=False

import torch,time,math
USE_CUDA=torch.cuda.is_available()
USE_CUDA=False

def get_location(elevation,azimuth):
    elevation_tensor=torch.tensor(elevation)
    azimuth_tensor=torch.tensor(azimuth)
    if USE_CUDA:
        elevation_tensor.cuda()
        azimuth_tensor.cuda()

    x_tensor = torch.sin(torch.pi * elevation_tensor / 180) *  torch.cos( torch.pi * azimuth_tensor / 180)
    y_tensor = torch.sin(torch.pi * elevation_tensor / 180) *  torch.sin( torch.pi * azimuth_tensor / 180)
    z_tensor = torch.cos(torch.pi * elevation_tensor / 180)
    x=x_tensor.item()
    y=y_tensor.item()
    z=z_tensor.item()
    
    print(x,y,z)
    return x,y,z
    
start = time.time()
get_location(0,90)
end = time.time()
print(end-start)

结果
-0.0 0.0 1.0
0.00550079345703125

3，math模块完成运算

import torch,time,math
elevation=0
azimuth=90
start = time.time()
x1 = math.sin(math.pi * elevation / 180) * math.cos(math.pi * azimuth / 180)
y1 = math.sin(math.pi * elevation / 180) * math.sin(math.pi * azimuth / 180)
z1 = math.cos(math.pi * elevation / 180)
print(x1,y1,z1)
end = time.time()
print(end-start)

结果
0.0 0.0 1.0
4.124641418457031e-05

4，使用函数调用

import torch,time,math
def get_location1(elevation,azimuth):
    x1 = math.sin(math.pi * elevation / 180) * math.cos(math.pi * azimuth / 180)
    y1 = math.sin(math.pi * elevation / 180) * math.sin(math.pi * azimuth / 180)
    z1 = math.cos(math.pi * elevation / 180)
    print(x1,y1,z1)
    return x1,y1,z1

start = time.time()
get_location1(0,90)
end = time.time()
print(end-start)