欢迎您访问程序员文章站本站旨在为大家提供分享程序员计算机编程知识!
您现在的位置是: 首页  >  IT编程

pytorch中的numel函数用法说明

程序员文章站 2022-06-22 23:02:37
获取tensor中一共包含多少个元素import torchx = torch.randn(3,3)print("number elements of x is ",x.numel())y = tor...

获取tensor中一共包含多少个元素

import torch
x = torch.randn(3,3)
print("number elements of x is ",x.numel())
y = torch.randn(3,10,5)
print("number elements of y is ",y.numel())

输出:

number elements of x is 9

number elements of y is 150

27和150分别位x和y中各有多少个元素或变量

补充:pytorch获取张量元素个数numel()的用法

numel就是"number of elements"的简写。

numel()可以直接返回int类型的元素个数

import torch 
a = torch.randn(1, 2, 3, 4)
b = a.numel()
print(type(b)) # int
print(b) # 24

通过numel()函数,我们可以迅速查看一个张量到底又多少元素。

补充:pytorch 卷积结构和numel()函数

看代码吧~

from torch import nn 
class cnn(nn.module):
    def __init__(self, num_channels=1, d=56, s=12, m=4):
        super(cnn, self).__init__()
        self.first_part = nn.sequential(
            nn.conv2d(num_channels, d, kernel_size=3, padding=5//2),
            nn.conv2d(num_channels, d, kernel_size=(1,3), padding=5//2),
            nn.conv2d(num_channels, d, kernel_size=(3,1), padding=5//2),
            nn.prelu(d)
        )
 
    def forward(self, x):
        x = self.first_part(x)
        return x
 
model = cnn()
for m in model.first_part:
    if isinstance(m, nn.conv2d):
        # print('m:',m.weight.data)
        print('m:',m.weight.data[0])
        print('m:',m.weight.data[0][0])
        print('m:',m.weight.data.numel()) #numel() 计算矩阵中元素的个数
 
结果:
m: tensor([[[-0.2822,  0.0128, -0.0244],
         [-0.2329,  0.1037,  0.2262],
         [ 0.2845, -0.3094,  0.1443]]]) #卷积核大小为3x3
m: tensor([[-0.2822,  0.0128, -0.0244],
        [-0.2329,  0.1037,  0.2262],
        [ 0.2845, -0.3094,  0.1443]]) #卷积核大小为3x3
m: 504   # = 56 x (3 x 3)  输出通道数为56,卷积核大小为3x3
m: tensor([-0.0335,  0.2945,  0.2512,  0.2770,  0.2071,  0.1133, -0.1883,  0.2738,
         0.0805,  0.1339, -0.3000, -0.1911, -0.1760,  0.2855, -0.0234, -0.0843,
         0.1815,  0.2357,  0.2758,  0.2689, -0.2477, -0.2528, -0.1447, -0.0903,
         0.1870,  0.0945, -0.2786, -0.0419,  0.1577, -0.3100, -0.1335, -0.3162,
        -0.1570,  0.3080,  0.0951,  0.1953,  0.1814, -0.1936,  0.1466, -0.2911,
        -0.1286,  0.3024,  0.1143, -0.0726, -0.2694, -0.3230,  0.2031, -0.2963,
         0.2965,  0.2525, -0.2674,  0.0564, -0.3277,  0.2185, -0.0476,  0.0558]) bias偏置的值
m: tensor([[[ 0.5747, -0.3421,  0.2847]]]) 卷积核大小为1x3
m: tensor([[ 0.5747, -0.3421,  0.2847]]) 卷积核大小为1x3
m: 168 # = 56 x (1 x 3) 输出通道数为56,卷积核大小为1x3
m: tensor([ 0.5328, -0.5711, -0.1945,  0.2844,  0.2012, -0.0084,  0.4834, -0.2020,
        -0.0941,  0.4683, -0.2386,  0.2781, -0.1812, -0.2990, -0.4652,  0.1228,
        -0.0627,  0.3112, -0.2700,  0.0825,  0.4345, -0.0373, -0.3220, -0.5038,
        -0.3166, -0.3823,  0.3947, -0.3232,  0.1028,  0.2378,  0.4589,  0.1675,
        -0.3112, -0.0905, -0.0705,  0.2763,  0.5433,  0.2768, -0.3804,  0.4855,
        -0.4880, -0.4555,  0.4143,  0.5474,  0.3305, -0.0381,  0.2483,  0.5133,
        -0.3978,  0.0407,  0.2351,  0.1910, -0.5385,  0.1340,  0.1811, -0.3008]) bias偏置的值
m: tensor([[[0.0184],
         [0.0981],
         [0.1894]]]) 卷积核大小为3x1
m: tensor([[0.0184],
        [0.0981],
        [0.1894]]) 卷积核大小为3x1
m: 168 # = 56 x (3 x 1) 输出通道数为56,卷积核大小为3x1
m: tensor([-0.2951, -0.4475,  0.1301,  0.4747, -0.0512,  0.2190,  0.3533, -0.1158,
         0.2237, -0.1407, -0.4756,  0.1637, -0.4555, -0.2157,  0.0577, -0.3366,
        -0.3252,  0.2807,  0.1660,  0.2949, -0.2886, -0.5216,  0.1665,  0.2193,
         0.2038, -0.1357,  0.2626,  0.2036,  0.3255,  0.2756,  0.1283, -0.4909,
         0.5737, -0.4322, -0.4930, -0.0846,  0.2158,  0.5565,  0.3751, -0.3775,
        -0.5096, -0.4520,  0.2246, -0.5367,  0.5531,  0.3372, -0.5593, -0.2780,
        -0.5453, -0.2863,  0.5712, -0.2882,  0.4788,  0.3222, -0.4846,  0.2170]) bias偏置的值
  
'''初始化后'''
class cnn(nn.module):
    def __init__(self, num_channels=1, d=56, s=12, m=4):
        super(cnn, self).__init__()
        self.first_part = nn.sequential(
            nn.conv2d(num_channels, d, kernel_size=3, padding=5//2),
            nn.conv2d(num_channels, d, kernel_size=(1,3), padding=5//2),
            nn.conv2d(num_channels, d, kernel_size=(3,1), padding=5//2),
            nn.prelu(d)
        )
        self._initialize_weights()
    def _initialize_weights(self):
        for m in self.first_part:
            if isinstance(m, nn.conv2d):
                nn.init.normal_(m.weight.data, mean=0.0, std=math.sqrt(2/(m.out_channels*m.weight.data[0][0].numel())))
                nn.init.zeros_(m.bias.data)
 
    def forward(self, x):
        x = self.first_part(x)
        return x
 
model = cnn()
for m in model.first_part:
    if isinstance(m, nn.conv2d):
        # print('m:',m.weight.data)
        print('m:',m.weight.data[0])
        print('m:',m.weight.data[0][0])
        print('m:',m.weight.data.numel()) #numel() 计算矩阵中元素的个数
 
结果:
m: tensor([[[-0.0284, -0.0585,  0.0271],
         [ 0.0125,  0.0554,  0.0511],
         [-0.0106,  0.0574, -0.0053]]])
m: tensor([[-0.0284, -0.0585,  0.0271],
        [ 0.0125,  0.0554,  0.0511],
        [-0.0106,  0.0574, -0.0053]])
m: 504
m: tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0.])
m: tensor([[[ 0.0059,  0.0465, -0.0725]]])
m: tensor([[ 0.0059,  0.0465, -0.0725]])
m: 168
m: tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0.])
m: tensor([[[ 0.0599],
         [-0.1330],
         [ 0.2456]]])
m: tensor([[ 0.0599],
        [-0.1330],
        [ 0.2456]])
m: 168
m: tensor([0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0.,
        0., 0., 0., 0., 0., 0., 0., 0.])
 

以上为个人经验,希望能给大家一个参考,也希望大家多多支持。如有错误或未考虑完全的地方,望不吝赐教。