---

一、tensor加减乘除

加法操作

import torch

x = torch.randn(2, 3)
y = torch.randn(2, 3)

z = x + y
print(z)

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

y.add_(x)
print(y)

其他操作类似：减法：sub(-)， 乘法：mul(*)， 除法：div(/)

二、tensor矩阵运算

# 二维矩阵相乘
a = torch.full([2, 2], 3, dtype=torch.long)
b = torch.ones(2, 2, dtype=torch.long)

print(a)
print(b)
print(torch.mm(a, b))

#  matmul 和 @ 可以用于二维矩阵计算，也可以是多维
# 四维，计算的时候就是前两维不变，后两维进行计算。
a = torch.rand(1, 1, 3, 2)
b = torch.rand(1, 1, 2, 4)

c = torch.matmul(a, b)
print(a)
print(b)
print(c)

pow

a = torch.full([2, 2], 6)
print(a.pow(3))

a = torch.full([2, 2], 6)
print(a**2)

sqrt: 平方根
rsqrt: 平方根倒数

a = torch.full([2, 2], 1024)

print(a.sqrt())
print(a.rsqrt())
print(a**0.5)

exp log

a = torch.ones(2, 2)

print(torch.exp(a))
print(torch.log(a))
print(torch.log2(a))

.floor()——往下近似
.ceil()——往上近似
.trunc()——裁剪为整数部分
.frac()——裁剪成小数部分

a = torch.tensor(3.1415926)

print(a.floor())
print(a.ceil())
print(a.trunc())
print(a.frac())

torch.round()——四舍五入

a = torch.tensor(3.1415926)

print(a.round())

.item() 转化为python number

x = torch.randn(1)

print(x)
print(x.item())

四、tensor切片操作

a = torch.randn(4, 3)
print(a)

# 取第二列
print(a[:, 1])

# 取前两列
print(a[:, :2])

五、tensor改变形状

x = torch.randn(4, 4)
y = x.view(16)

# -1, 自动匹配个数
z = x.view(-1, 8)

print(x)
print(y)
print(z)

六、tensor 和 numpy.array相互转换

# 底层内存共享
x = torch.ones(5)
print(x)

y = x.numpy()
print(y)

x.add_(1)
print(y)

import numpy as np
x = np.ones(5)
y = torch.from_numpy(x)
print(y)

七、tensor 转到GPU上

if torch.cuda.is_available():
    
    device = torch.device("cuda")
    x = torch.randn(2, 3)
    print(x)

    y = x.to(device)
    print(y)

    z = torch.randn(2, 3, device="cuda")
    print(z)
    
    # 同时在GPU上才能相加
    print(y + z)

    # 转换会cpu
    print(z.to("cpu"))