pytorch替换numpy中的一些组件 //转载请注明来源

numpy可以同时处理两种数据：list和numpy.ndarray。

而torch只能处理张量(Tensor)数据。

---

1.替换 np.asarray()

import torch

data=[[1,2],[3,4]]

data = torch.Tensor(data)

# data = np.asarray(data)

print(data.shape)

2.替换 nn.prod()

#使用torch.prod()替换nn.prod(),其中torch.prod()里面的参数必须得是tensor.

import torch

data=[1,2]

data = torch.Tensor(data)

print(data.shape)

c = torch.prod(data)

c

---

tensor(2.)

Python torch.prod实例讲解 - 码农教程

3.替换 np.max()

#使用torch.max()替换np.max(),其中torch.max()里面的参数必须得是tensor.

import torch

data=[1,2,3,9,4]

data = torch.Tensor(data)

a = torch.max(data)

a

---

tensor(9.)

4.torch.Size转换为np.arrary

import torch

# data = np.asarray(data)

data=[[1,2],[3,4]]

data = torch.Tensor(data)

print(data.shape)

print(np.asarray(data.shape))

c = torch.prod(torch.Tensor(np.asarray(data.shape)))

c

---

torch.Size([2, 2])

[2 2]

tensor(4.)

5.numpy与torch的ravel，它俩用法一样，注意tensor.ravel()后的数据还是Tensor

import torch

# data = np.asarray(data)

data=[[1,2],[3,4]]

data = torch.Tensor(data)

data = data.ravel()

print("ravel:",data)

---

ravel: tensor([1., 2., 3., 4.])

6.numpy与torch的ndim，它俩用法一样

import torch

data=[[1,2],[3,4]]

data_numpy = np.asarray(data).ravel()

data = torch.Tensor(data)

data = data.ravel()

print("ravel:",data)

print(data.shape)

print("tensor_ndim:",data.ndim)

print("data_numpy_ndim:",data_numpy.ndim)

---

ravel: tensor([1., 2., 3., 4.])

torch.Size([4])

tensor_ndim: 1

data_numpy_ndim: 1

 7.numpy与torch.tensor的data.T效果一样

import torch

data=[[1,2],[3,4]]

data_numpy = np.asarray(data)

data = torch.Tensor(data)

# data = data.ravel()

data_trans_tensor=data.T

print("data:",data)

print("data_trans:",data_trans_tensor)

data_trans_numpy=data_numpy.T

print("data_numpy:",data_numpy)

print("data_trans_numpy:",data_trans_numpy)

---

data: tensor([[1., 2.], [3., 4.]])

data_trans: tensor([[1., 3.], [2., 4.]])

data_numpy: [[1 2] [3 4]]

data_trans_numpy: [[1 3] [2 4]]

8.np.arange()与torch.arange()用法一样

timesteps=5

a = np.arange(timesteps)

b = torch.arange(timesteps)

print("a:",a)

print("b:",b)

---

a: [0 1 2 3 4]

b: tensor([0, 1, 2, 3, 4])

Python torch.arange实例讲解 - 码农教程

9.np.ceil()与torch.ceil()用法一样,但torch.ceil()的输入必须是tensor.

a=[0.1,1.1,2.1,3.1]

b=np.ceil(a)

c=torch.ceil(torch.Tensor(a))

print("b:",b)

print("c:",c)

---

b: [1. 2. 3. 4.]

c: tensor([1., 2., 3., 4.])

10.torch.sort()替换np.sort(),torch.sort()还会返回对应原序列的索引

import numpy as np

import torch

a=[5,1.1,2.1,3.1]

b=np.sort(a)

value,indices=torch.sort(torch.Tensor(a))

print("b:",b)

print("c:",value,indices)

---

b: [1.1 2.1 3.1 5. ]

c: tensor([1.1000, 2.1000, 3.1000, 5.0000])  tensor([1, 2, 3, 0])

np.sort()函数的作用_wx5ba0c87f1984b的技术博客_51CTO博客

torch.sort()_James_Bobo的博客-CSDN博客_torch.sort()

11.torch.abs()替换np.abs()

data = [-1 , -2 , 1 , 2]

tensor = torch.FloatTensor(data)

print(

"abs in numpy:\n",np.abs(data),

"\nabs in torch:\n",torch.abs(tensor)

)

---

abs in numpy: [1 2 1 2] abs in torch: tensor([1., 2., 1., 2.])

12.torch.ones_like()替换np.ones_like(),但torch.ones_like参数必须是Tensor

import torch

import numpy as np

a=[[1,2],[3,4]]

input = torch.empty(2, 3)

print(np.ones_like(a))

print(torch.ones_like(input))

---

[[1 1] [1 1]]

tensor([[1., 1., 1.], [1., 1., 1.]])

13. torch.split替换np.split//未做比较solid的实验，因为np.split可以接受tensor，而且np.split和torch.split的用法不太一样，但能达到相同的效果，未深入研究

import torch

import numpy as np

a=[[1,2],[3,4]]

input = torch.tensor(a)

input_1 = torch.tensor(input)

print(input)

print(input_1)

a = np.split(input,1)

print(a)

b = torch.split(input,1)

print(b)

---

tensor([[1, 2], [3, 4]])

tensor([[1, 2], [3, 4]])

[tensor([[1, 2], [3, 4]])]

(tensor([[1, 2]]), tensor([[3, 4]]))
 

np.split 用法详解_Wanderer001的博客-CSDN博客_np.split函数

numpy的ones_like函数_God_s_apple的博客-CSDN博客_numpy ones_like

pytorch之torch.ones_like_刘嘉晨Amber的博客-CSDN博客

14.torch.empty_like()替换np.empty_like()//torch.empty_like()创建一个使用未初始化值填满的tensor

import torch

import numpy as np

a=[[1,2],[3,4]]

input = torch.tensor(a)

print(np.empty_like(a))

print(torch.empty_like(input))

---

[[1 2] [3 4]]

tensor([[0, 2], [3, 4]])

Python numpy.empty_like用法及代码示例 - 纯净天空

pytorch每日一学22(torch.empty()、torch.empty_like()、torch.empty_strided())创建未初始化数据的tensor_Fluid_ray的博客-CSDN博客_torch.empty

15.tensor.sum()替换np.sum()

import torch

import numpy as np

a=[[1,2],[3,4]]

input = torch.tensor(a)

 

d = np.array(input)

d1=d.sum(axis=0)

print(d1)

e = torch.tensor(d)

# print(e)

d2=e.sum(axis=0)

print(d2)

---

[4 6]

tensor([4, 6])

16.np.transpose(可处理高维)与torch.transpose不一样，torch.transpose只能处理2D情况，torch.permute可以处理高维情况

待补

python transpose与permute函数详解_ice123muxin的博客-CSDN博客_transpose和permute

np.transpose_cool whidpers的博客-CSDN博客_np transpose

17. torch.squeeze替换np.squeeze

写到这里突然发现，大部分np函数可以处理tensor，处理完后一般结果是np,再转换为tensor就好了

18. tensor.T与np.arrary.T的效果一样

19.A@B//矩阵乘法

import torch

import numpy as np

a=[[1,2],[3,4]]

input = torch.tensor(a)

 

d = np.array(input)

print(d)

e = torch.tensor(d)

print(e)

f = torch.tensor(np.squeeze(e))

print(type(f))

g=f.T

print(g)

k = e @ g

print(k)

print(k.type())

---

[[1 2] [3 4]]

tensor([[1, 2], [3, 4]])

tensor([[1, 3], [2, 4]])

tensor([[ 5, 11], [11, 25]]) torch.LongTensor

/tmp/ipykernel_7717/3140886201.py:

11: UserWarning: To copy construct from a tensor, it is recommended to use sourceTensor.clone().detach() or sourceTensor.clone().detach().requires_grad_(True), rather than torch.tensor(sourceTensor). f = torch.tensor(np.squeeze(e))

Python中@符号是什么意思 - 知乎