PyTorch基础知识1.0

原创已于 2022-01-20 14:44:06 修改 · 1.4k 阅读

3 ·

CC 4.0 BY-SA版权

文章标签：

#pytorch #深度学习 #python

于 2022-01-20 14:43:03 首次发布

深度学习之美同时被 3 个专栏收录

209 篇文章

订阅专栏

Python

131 篇文章

订阅专栏

Code

110 篇文章

订阅专栏

本文介绍了PyTorch中张量的基本概念与操作，包括张量的创建、基本运算及高级操作等。通过实例展示了不同类型的张量创建方法，如标量、向量、矩阵等，并演示了加减乘除等运算。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

PyTorch基础知识1.0

1.Tensor的基本概念和操作

import torch
import numpy as np

# 1. 标量
a = torch.tensor(4.0)
print('标量： ', a)

标量：  tensor(4.)

# 2. 向量
b = torch.tensor([1, 5, 2])
print('向量： ', b)

向量：  tensor([1, 5, 2])

# 3. 矩阵（张量）
c = torch.tensor([[1, 2, 5],
                  [5, 6, 9]])
print('矩阵： ', c)

矩阵：  tensor([[1, 2, 5],
        [5, 6, 9]])

# 4. 多维矩阵
d = torch.tensor([[1,2,5], [5,6,1], [9,7,5]])
print("多维矩阵 ： ", d)

多维矩阵 ：  tensor([[1, 2, 5],
        [5, 6, 1],
        [9, 7, 5]])

# 5. 创建一个空的张量
e = torch.empty(5, 2)
print(e)

tensor([[0., 0.],
        [0., 0.],
        [0., 0.],
        [0., 0.],
        [0., 0.]])

# 6. 创建一个张量，元素随机产生
f = torch.rand(5, 2)
print(f)

tensor([[0.4557, 0.6264],
        [0.0040, 0.6414],
        [0.1270, 0.4115],
        [0.2843, 0.4744],
        [0.5275, 0.2181]])

# 7. 创建一个全是0的张量
g = torch.zeros(5, 2, dtype = torch.long)
print(g)

g1 = torch.zeros(5, 4)
print(g1)

tensor([[0, 0],
        [0, 0],
        [0, 0],
        [0, 0],
        [0, 0]])
tensor([[0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.]])

# 8. 创建一个全是0的张量
h = torch.ones(5, 2, dtype = torch.int)
print(h)

h1 = torch.ones(6, 2, dtype = torch.long)
print(h1)

tensor([[1, 1],
        [1, 1],
        [1, 1],
        [1, 1],
        [1, 1]], dtype=torch.int32)
tensor([[1, 1],
        [1, 1],
        [1, 1],
        [1, 1],
        [1, 1],
        [1, 1]])

# 9. 创建一个产生随机元素的张量，其shape和另一个张量的shape大小一致
i = torch.rand_like(h1, dtype = torch.float)
print(i)
print(i.shape)

tensor([[0.3572, 0.3757],
        [0.6741, 0.3193],
        [0.0902, 0.0936],
        [0.0152, 0.1191],
        [0.2416, 0.5334],
        [0.1352, 0.3116]])
torch.Size([6, 2])

2. Tensor 的基本运算

# 10. 加法
# 第一种
j = torch.tensor([2, 5, 6])
k = torch.tensor([8, 9, 6])
j_k = j + k
print(j_k)

tensor([10, 14, 12])

# 第二种
j = torch.tensor([4, 5, 8])
k = torch.tensor([8, 9, 51])
j_k = torch.add(j, k)
print(j_k)

tensor([12, 14, 59])

# 第三种
j = torch.tensor([4, 5, 8])
k = torch.tensor([8, 9, 51])
j.add_(k)
print(j)
print(k)

tensor([12, 14, 59])
tensor([ 8,  9, 51])

# 11. 除法
# 第一种
l = torch.tensor([4, 8, 9])
m = torch.tensor([6, 7, 85])
print(l // m)

tensor([0, 1, 0])


D:\anaconda\envs\mytorch\lib\site-packages\torch\_tensor.py:575: UserWarning: floor_divide is deprecated, and will be removed in a future version of pytorch. It currently rounds toward 0 (like the 'trunc' function NOT 'floor'). This results in incorrect rounding for negative values.
To keep the current behavior, use torch.div(a, b, rounding_mode='trunc'), or for actual floor division, use torch.div(a, b, rounding_mode='floor'). (Triggered internally at  ..\aten\src\ATen\native\BinaryOps.cpp:467.)
  return torch.floor_divide(self, other)

# 第二种
print(torch.true_divide(l, m))

tensor([0.6667, 1.1429, 0.1059])

# 12. 乘法
# 第一种
n = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(torch.mm(n, p))

tensor([[184, 207, 138],
        [184, 207, 138],
        [184, 207, 138]])

n = torch.tensor([[8, 9, 6]])
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(torch.mm(n, p))

tensor([[184, 207, 138]])

n = torch.tensor([8, 9, 6])# n是向量不是矩阵，torch.mm:是两个矩阵之间的乘法函数
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(torch.mm(n, p))

---------------------------------------------------------------------------

RuntimeError                              Traceback (most recent call last)

C:\Users\Public\Documents\Wondershare\CreatorTemp/ipykernel_12572/1070864940.py in <module>
      1 n = torch.tensor([8, 9, 6])
      2 p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
----> 3 print(torch.mm(n, p))


RuntimeError: self must be a matrix

# 第二种
n = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(torch.matmul(n, p))

tensor([[184, 207, 138],
        [184, 207, 138],
        [184, 207, 138]])

# 第三种
n = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(n.mm(p))

tensor([[184, 207, 138],
        [184, 207, 138],
        [184, 207, 138]])

# 第四种
n = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
p = torch.tensor([[8, 9, 6],[8, 9, 6],[8, 9, 6]])
print(n@p)

tensor([[184, 207, 138],
        [184, 207, 138],
        [184, 207, 138]])

# 13. 幂运算
# 第一种
q = torch.tensor([2, 3])
print(torch.pow(q, 2))

tensor([4, 9])

# 第二种
q = torch.tensor([2, 3])
print(q.pow(2))

tensor([4, 9])

# 第三种
q = torch.tensor([2, 3])
print(q**2)

tensor([4, 9])

# 14. 开根号运算（开方运算）
# 第一种
r = torch.tensor(4)
r1 = torch.tensor([5, 9])
r2 = torch.tensor([1,5,6])
r3 = torch.tensor([[7, 8, 9], [4, 5, 6]])
print(torch.sqrt(r))
print(torch.sqrt(r1))
print(torch.sqrt(r2))
print(torch.sqrt(r3))
print(r.sqrt())
print(r1.sqrt())
print(r2.sqrt())
print(r3.sqrt())

tensor(2.)
tensor([2.2361, 3.0000])
tensor([1.0000, 2.2361, 2.4495])
tensor([[2.6458, 2.8284, 3.0000],
        [2.0000, 2.2361, 2.4495]])
tensor(2.)
tensor([2.2361, 3.0000])
tensor([1.0000, 2.2361, 2.4495])
tensor([[2.6458, 2.8284, 3.0000],
        [2.0000, 2.2361, 2.4495]])

# 第二种
r = torch.tensor(4)
r1 = torch.tensor([5, 9])
r2 = torch.tensor([1,5,6])
r3 = torch.tensor([[7, 8, 9], [4, 5, 6]])
print(r.sqrt())
print(r1.sqrt())
print(r2.sqrt())
print(r3.sqrt())

tensor(2.)
tensor([2.2361, 3.0000])
tensor([1.0000, 2.2361, 2.4495])
tensor([[2.6458, 2.8284, 3.0000],
        [2.0000, 2.2361, 2.4495]])

# 15. 对数运算
# 第一种
s1 = torch.tensor(4)
s2 = torch.tensor([5, 6])
s3 = torch.tensor([[1,5], [6,8]])
print(torch.log2(s1))
print(torch.log2(s2))
print(torch.log2(s3))
print(torch.log(s1))
print(torch.log(s2))
print(torch.log(s3))
print(torch.log10(s1))
print(torch.log10(s2))
print(torch.log10(s3))

tensor(2.)
tensor([2.3219, 2.5850])
tensor([[0.0000, 2.3219],
        [2.5850, 3.0000]])
tensor(1.3863)
tensor([1.6094, 1.7918])
tensor([[0.0000, 1.6094],
        [1.7918, 2.0794]])
tensor(0.6021)
tensor([0.6990, 0.7782])
tensor([[0.0000, 0.6990],
        [0.7782, 0.9031]])

# 第二种
s1 = torch.tensor(4)
s2 = torch.tensor([5, 6])
s3 = torch.tensor([[1,5], [6,8]])
print(s1.log2())
print(s2.log2())
print(s3.log2())
print(s1.log())
print(s2.log())
print(s3.log())
print(s1.log10())
print(s2.log10())
print(s3.log10())

tensor(2.)
tensor([2.3219, 2.5850])
tensor([[0.0000, 2.3219],
        [2.5850, 3.0000]])
tensor(1.3863)
tensor([1.6094, 1.7918])
tensor([[0.0000, 1.6094],
        [1.7918, 2.0794]])
tensor(0.6021)
tensor([0.6990, 0.7782])
tensor([[0.0000, 0.6990],
        [0.7782, 0.9031]])

# 16. 取整/取余运算
t = torch.tensor([5.26161, 9.16651, 8.51665])

# 第一种 向上取整
print(t.floor())

tensor([5., 9., 8.])

# 第二种 向下取整
print(t.ceil())

tensor([ 6., 10.,  9.])

# 第三种 四舍五入
print(t.round())

tensor([5., 9., 9.])

# 第四种 裁剪，只取整数部分
print(t.trunc())

tensor([5., 9., 8.])

# 第五种 只取小数部分
print(t.frac())

tensor([0.2616, 0.1665, 0.5167])

# 第六种 取余
print(torch.tensor([8, 9]) % 7)

tensor([1, 2])

3.其他Tensor操作

# 17. shape， reshape, resize, view
# shape 显示矩阵的, 形状，各个维度上的大小
u = torch.tensor([[5, 8, 9], [8, 1, 8]])
print(u.shape)

torch.Size([2, 3])

# reshape 改变张量的形状
print(u.reshape(3,2))

tensor([[5, 8],
        [9, 8],
        [1, 8]])

# resize 改变张量的形状
print(u.resize(2, 3))

tensor([[5, 8, 9],
        [8, 1, 8]])


D:\anaconda\envs\mytorch\lib\site-packages\torch\_tensor.py:490: UserWarning: non-inplace resize is deprecated
  warnings.warn("non-inplace resize is deprecated")

# view 产生一个境像，改变张量形状，但不改变原张量的形状
print(u.view(3, 2))
print(u)

tensor([[5, 8],
        [9, 8],
        [1, 8]])
tensor([[5, 8, 9],
        [8, 1, 8]])

# 18.获取一个tensor的值（注意：只针对一个值的tensor），获取它的标量值
v = torch.tensor(3.3)
print(v.item())

3.299999952316284

# 19. Tensor排序（会显示原始位置的下标）
# 升序
w = torch.tensor([8,5,6,1,161,16,15198,119800])
print(w.sort())
print(torch.sort(w))

torch.return_types.sort(
values=tensor([     1,      5,      6,      8,     16,    161,  15198, 119800]),
indices=tensor([3, 1, 2, 0, 5, 4, 6, 7]))
torch.return_types.sort(
values=tensor([     1,      5,      6,      8,     16,    161,  15198, 119800]),
indices=tensor([3, 1, 2, 0, 5, 4, 6, 7]))

# 降序
print(torch.sort(w, descending = True))

torch.return_types.sort(
values=tensor([119800,  15198,    161,     16,      8,      6,      5,      1]),
indices=tensor([7, 6, 4, 5, 0, 2, 1, 3]))

# 20. 沿着指定维度返回最大K个数值及其索引值
x = torch.tensor([8,5,6,1,161,16,15198,119800] )
# dim:按照那个维度进行排序，k：最大的k个数，largest：是否沿着从大到小这样排列
print(torch.topk(x, k = 2, dim = 0, largest = True, out = None))

torch.return_types.topk(
values=tensor([119800,  15198]),
indices=tensor([7, 6]))

# 沿着指定维度返回第最小值及其索引值
print(torch.kthvalue(x, 5, dim = 0))

torch.return_types.kthvalue(
values=tensor(16),
indices=tensor(5))

# 21. 判断是否为nan， inf, finite
# nan；not a number
# inf: infinity:无穷大
# finite:有限
# 判断张量中是否存在不是数字的元素
y = torch.tensor([2,3,4])
print(torch.isnan(y))

tensor([False, False, False])

# 检查其参数是否是无穷大，如果number是有效数字（可以转换为有限数字），返回True
print(torch.isfinite(y))

tensor([True, True, True])

# 检查是否为无穷大
print(torch.isinf(y))

tensor([False, False, False])

# 22. tensor 转化成numpy    numpy转化成tensor
# tensor 转化成numpy 
z = torch.tensor([7, 2, 8])
print(z.numpy())

[7 2 8]

# numpy转化成tensor
import numpy as np
z1 = np.ones(6)
print(z1)
print(torch.from_numpy(z1))

[1. 1. 1. 1. 1. 1.]
tensor([1., 1., 1., 1., 1., 1.], dtype=torch.float64)

# 23. 判断是否有GPU,如果有，则部署到GPU上运行
if torch.cuda.is_available():
    device = torch.device('gpu')
else:
    device = torch.device('cpu')
a1 = torch.tensor([4, 5, 6], device = device)
a2 = torch.tensor([5, 15, 151], device = device)
print(a1)
print(a2)
a3 = a1 + a2
print(a3)

tensor([4, 5, 6])
tensor([  5,  15, 151])
tensor([  9,  20, 157])