In early 2019, ApacheCN organized volunteers to translate PyTorch1.0 Chinese documents (Making the address), as well as being officially licensed by PyTorch, I’m sure there are already a lot of people out thereChinese Official websiteYes. But so farproofreadingWe are short of staff, so I hope everyone will participate. We’ve been in an email exchange with PyTorch’s Bruce Lin for a while. At an appropriate time, we will organize volunteers to work on other PyTorch projects. Please join us and follow us. More hope our series of work can be helpful to you.
Translator: bat67
Proofread by FontTian
By Soumith Chintala
PyTorch is a Python-based scientific computing package aimed at two groups of people:
- As an alternative to NumPy, computations can take advantage of the GPU’s performance
- As a high flexibility, fast deep learning platform
An introduction to
tensor
Tensor is similar to NumPy’s DARray, but you can also use it on your GPU to speed up calculations.
from __future__ import print_function
import torch
Copy the code
Create an uninitialized 5*3 matrix:
x = torch.empty(5.3)
print(x)
Copy the code
Output:
tensor([[2.2391 e-19.4.5869 e-41.1.4191 e-17],
[4.5869 e-41.0.0000 e+00.0.0000 e+00],
[0.0000 e+00.0.0000 e+00.0.0000 e+00],
[0.0000 e+00.0.0000 e+00.0.0000 e+00],
[0.0000 e+00.0.0000 e+00.0.0000 e+00]])
Copy the code
Create a random initialization matrix:
x = torch.rand(5.3)
print(x)
Copy the code
Output:
tensor([[0.5307.0.9752.0.5376],
[0.2789.0.7219.0.1254],
[0.6700.0.6100.0.3484],
[0.0922.0.0779.0.2446],
[0.2967.0.9481.0.1311]])
Copy the code
Construct a matrix filled with zeros and data type long:
x = torch.zeros(5.3, dtype=torch.long)
print(x)
Copy the code
Output:
tensor([[0.0.0],
[0.0.0],
[0.0.0],
[0.0.0],
[0.0.0]])
Copy the code
Construct tensors directly from data:
x = torch.tensor([5.5.3])
print(x)
Copy the code
Output:
tensor([5.5000.3.0000])
Copy the code
Or build a new tensor from an existing tensor. Unless the user provides a new value, these methods will reuse the properties of the input tensor, such as dtype, etc. :
x = x.new_ones(5.3, dtype=torch.double) # new_* methods take in sizes
print(x)
x = torch.randn_like(x, dtype=torch.float) # overloaded dtype!
print(x) # result size consistent
Copy the code
Output:
tensor([[1..1..1.],
[1..1..1.],
[1..1..1.],
[1..1..1.],
[1..1..1.]], dtype=torch.float64)
tensor([[ 1.6040, -0.6769.0.0555],
[ 0.6273.0.7683, -0.2838], [...0.7159, -0.5566, -0.2020],
[ 0.6266.0.3566.1.4497], [...0.8092, -0.6741.0.0406]])
Copy the code
Get the shape of the tensor:
print(x.size())
Copy the code
Output:
torch.Size([5.3])
Copy the code
Note:
Torch.Size is essentially a tuple, so all operations on tuples are supported.
operation
An operation has multiple syntax. In the following example, we will examine addition.
Addition: Form one
y = torch.rand(5.3)
print(x + y)
Copy the code
Output:
tensor([[ 2.5541.0.0943.0.9835],
[ 1.4911.1.3117.0.5220], [...0.0078, -0.1161.0.6687],
[ 0.8176.1.1179.1.9194], [...0.3251, -0.2236.0.7653]])
Copy the code
Addition: Form two
print(torch.add(x, y))
Copy the code
Output:
tensor([[ 2.5541.0.0943.0.9835],
[ 1.4911.1.3117.0.5220], [...0.0078, -0.1161.0.6687],
[ 0.8176.1.1179.1.9194], [...0.3251, -0.2236.0.7653]])
Copy the code
Addition: Given an output tensor as a parameter
result = torch.empty(5.3)
torch.add(x, y, out=result)
print(result)
Copy the code
Output:
tensor([[ 2.5541.0.0943.0.9835],
[ 1.4911.1.3117.0.5220], [...0.0078, -0.1161.0.6687],
[ 0.8176.1.1179.1.9194], [...0.3251, -0.2236.0.7653]])
Copy the code
Addition: in-place operation
# adds x to y
y.add_(x)
print(y)
Copy the code
Output:
tensor([[ 2.5541.0.0943.0.9835],
[ 1.4911.1.3117.0.5220], [...0.0078, -0.1161.0.6687],
[ 0.8176.1.1179.1.9194], [...0.3251, -0.2236.0.7653]])
Copy the code
Note:
Any in-place operation that changes a tensor is followed by a fixed _. For example, x. popy _(y) and x.t_() will change x
You can also use various index operations like standard NumPy:
print(x[:, 1])
Copy the code
Output:
tensor([-0.6769.0.7683, -0.5566.0.3566, -0.6741])
Copy the code
Change shape: If you want to change shape, use torch. View
x = torch.randn(4.4)
y = x.view(16)
z = x.view(-1.8) # the size -1 is inferred from other dimensions
print(x.size(), y.size(), z.size())
Copy the code
Output:
torch.Size([4.4]) torch.Size([16]) torch.Size([2.8])
Copy the code
If you have a tensor with just one element, you can use.item() to get python values
x = torch.randn(1)
print(x)
print(x.item())
Copy the code
Output:
tensor([0.0445])
0.0445479191839695
Copy the code
Follow-up reading:
More than 100 tensor operations, including transposes, indexes, slicing, math, linear algebra, random numbers, go here
Bridge NumPy
Converting a Torch tensor to a NumPy array is a snap, and vice versa.
The Torch tensor and the NumPy array share their underlying memory locations, so when one changes, the other changes.
Translate Torch’s Tensor into a NumPy array
Input:
a = torch.ones(5)
print(a)
Copy the code
Output:
tensor([1..1..1..1..1.])
Copy the code
Input:
b = a.numpy()
print(b)
Copy the code
Output:
[1. 1. 1. 1. 1.]
Copy the code
See how the NumPy array changes the values inside:
a.add_(1)
print(a)
print(b)
Copy the code
Output:
tensor([2..2..2..2..2.[])2. 2. 2. 2. 2.]
Copy the code
Convert the NumPy array to the Torch tensor
See how changing the NumPy array automatically changes the Torch tensor:
import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)
print(a)
print(b)
Copy the code
Output:
[2. 2. 2. 2. 2.]
tensor([2..2..2..2..2.], dtype=torch.float64)
Copy the code
All tensors on the CPU (except CharTensor) support conversions to and from Numpy.
Tensor on CUDA
Tensors can be moved to any device using the. To method:
# We can run the following code when the GPU is available
We're going to use 'torch. Device' to put the tensor in and out of the GPU
if torch.cuda.is_available():
device = torch.device("cuda") # a CUDA device object
y = torch.ones_like(x, device=device) Create tensor directly on the GPU
x = x.to(device) # or use the. To ("cuda") method
z = x + y
print(z)
print(z.to("cpu", torch.double)) # '. To 'can also change the dtype while moving
Copy the code
Output:
tensor([1.0445], device='cuda:0')
tensor([1.0445], dtype=torch.float64)
Copy the code