PyTorch的学习笔记

PyTorch的Tensor

官网上是类比Numpy的Ndarrays，相比于Ndarrays多了GPU相关的加速运算。关于Ndarrays其实就是N-dimension Arrays，是一个N维的数组对象。Pytorch里提供的许多方法也与Numpy很像。
下面提供了关于创建矩阵的方法，包括empty，rand，zeros，tensor等。
torch创建的对象是tensor。

torch.empty()

empty方法创建的矩阵不是空矩阵，而是未初始化的矩阵，所以里面的值不一定是0。

import torch
x = torch.empty(5,3)
print(x)

tensor([[ 1.1018e-08,  4.5818e-41, -1.2501e+11],
        [ 4.5916e-41,  4.9724e-17,  4.5818e-41],
        [ 4.9696e-17,  4.5818e-41,  1.3498e-08],
        [ 4.5818e-41,  1.3416e-08,  4.5818e-41],
        [ 4.9691e-17,  4.5818e-41,  4.9691e-17]])

torch.rand()

rand方法生成的是一个用随机数初始化的矩阵，里面的数值为限定时是在[0,1]之间随机生成

x = torch.rand(3,5)
print(x)

tensor([[ 0.9101,  0.5218,  0.6148,  0.2900,  0.3983],
        [ 0.8021,  0.3228,  0.2530,  0.3052,  0.6225],
        [ 0.0999,  0.2756,  0.8993,  0.1512,  0.0917]])

torch.zeros()

zeros方法顾名思义是生成0矩阵，detype参数指定了生成的数据的类型

x = torch.zeros(5, 3, dtype=torch.long)
print(x)

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

torch.tensor()

tensor 方法能直接生成tensor构成的数据（这里有点迷），不过生成的数据自动保留四位小数

x = torch.tensor([5.5, 3,9.99999,8.16])
print(x)

tensor([ 5.5000,  3.0000, 10.0000,  8.1600])

从现成的tensor来生成新的tensor

x = x.new_ones(5, 3, dtype=torch.double)      # new_* methods take in sizes
print(x)

x = torch.randn_like(x, dtype=torch.float)    # override dtype!
print(x)                                      # result has the same size

tensor([[ 1.,  1.,  1.],
        [ 1.,  1.,  1.],
        [ 1.,  1.,  1.],
        [ 1.,  1.,  1.],
        [ 1.,  1.,  1.]], dtype=torch.float64)
tensor([[ 0.2710,  1.6084, -0.8171],
        [ 0.8977,  1.4150,  0.2287],
        [ 0.2518, -0.0245, -0.5036],
        [-0.5529, -0.1147, -1.3930],
        [-0.7114,  0.2698,  2.3081]])

Torch里面重载的一些基本运算操作

矩阵的+法运算

加法运算要求矩阵的大小相同，dtype也需要相同才能进行运算,add方法与所重载的+号类似。
add方法的接受参数

• (Tensor input, Tensor other, float alpha, Tensor out)
• (Tensor input, float other, float alpha, Tensor out)

y = torch.rand(5,3,dtype=torch.float)
print(y)
z = x + y
print(z)

tensor([[ 0.1392,  0.6817,  0.1232],
        [ 0.6517,  0.5228,  0.1931],
        [ 0.8584,  0.7995,  0.4617],
        [ 0.1270,  0.0686,  0.8771],
        [ 0.1968,  0.9849,  0.8087]])
tensor([[ 0.4103,  2.2900, -0.6938],
        [ 1.5495,  1.9379,  0.4218],
        [ 1.1102,  0.7750, -0.0419],
        [-0.4260, -0.0461, -0.5159],
        [-0.5146,  1.2546,  3.1168]])

z = torch.add(x , y)
print(z)

tensor([[ 0.4103,  2.2900, -0.6938],
        [ 1.5495,  1.9379,  0.4218],
        [ 1.1102,  0.7750, -0.0419],
        [-0.4260, -0.0461, -0.5159],
        [-0.5146,  1.2546,  3.1168]])

output = torch.empty(5,3)
print(output)
torch.add(x,y,out=output)
print(output)

tensor(1.00000e-08 *
       [[ 1.1018,  0.0000,  1.1018],
        [ 0.0000,  0.0000,  0.0000],
        [ 0.0000,  0.0000,  0.0000],
        [ 0.0000,  0.0000,  0.0000],
        [ 0.0000,  0.0000,  0.0000]])
tensor([[ 0.4103,  2.2900, -0.6938],
        [ 1.5495,  1.9379,  0.4218],
        [ 1.1102,  0.7750, -0.0419],
        [-0.4260, -0.0461, -0.5159],
        [-0.5146,  1.2546,  3.1168]])

替换性操作

在提供的方法中，如果是(operation)_ (arg)格式的方法的，操作后会替换到调用这个方法的对象

y.add_(x)
print(y)

tensor([[ 0.4103,  2.2900, -0.6938],
        [ 1.5495,  1.9379,  0.4218],
        [ 1.1102,  0.7750, -0.0419],
        [-0.4260, -0.0461, -0.5159],
        [-0.5146,  1.2546,  3.1168]])

x.copy_(y)
print(x)

tensor([[ 0.4103,  2.2900, -0.6938],
        [ 1.5495,  1.9379,  0.4218],
        [ 1.1102,  0.7750, -0.0419],
        [-0.4260, -0.0461, -0.5159],
        [-0.5146,  1.2546,  3.1168]])

类Numpy的索引方式

print(x[:,1])

tensor([ 2.2900,  1.9379,  0.7750, -0.0461,  1.2546])

torch.view()

改变矩阵的大小，类似resize，reshape

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())

torch.Size([4, 4]) torch.Size([16]) torch.Size([2, 8])

print(x,y,z)

tensor([[-0.3858, -0.6874,  1.0538, -1.2053],
        [-0.2992,  0.7963,  0.5221, -1.0758],
        [-1.1194,  0.1516,  2.0523,  0.4788],
        [ 0.3233, -0.0533,  0.3937, -2.0091]]) tensor([-0.3858, -0.6874,  1.0538, -1.2053, -0.2992,  0.7963,  0.5221,
        -1.0758, -1.1194,  0.1516,  2.0523,  0.4788,  0.3233, -0.0533,
         0.3937, -2.0091]) tensor([[-0.3858, -0.6874,  1.0538, -1.2053, -0.2992,  0.7963,  0.5221,
         -1.0758],
        [-1.1194,  0.1516,  2.0523,  0.4788,  0.3233, -0.0533,  0.3937,
         -2.0091]])

对于单值的tensor可用.item()方法获取其详细的值。

x = torch.rand(1)
print(x)
print(x.item())

tensor([ 0.5580])
0.5579903721809387

从Torch的Tensor到Numpy的转换

numpy方法

x = torch.ones(5)
print(x)
y = x.numpy()
print(y)

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

x.add_(1)
print(x)
print(y)

tensor([ 2.,  2.,  2.,  2.,  2.])
[2. 2. 2. 2. 2.]

可以发现对x操作会影响到y

从Numpy到Tensor的转换

import numpy as np
a = np.ones(5)
b = torch.from_numpy(a)
np.add(a, 1, out=a)
print(a)
print(b)

[2. 2. 2. 2. 2.]
tensor([ 2.,  2.,  2.,  2.,  2.], dtype=torch.float64)

可以看到从numpy生成的tensor也会共同影响

Reference:

1. What is PyTorch
2. 关于tensor的操作