【pytorch】1.6 tensor 基本运算

一、逐元素操作

1、 torch.addcdiv(t, v, t1, t2)

   t1 与 t2 按元素除后，乘v 加t

import torch

t = torch.randn(1, 3)
t1 = torch.randn(3, 1)
t2 = torch.randn(1, 3)

a = t + 0.1 *(t1 / t2)
print(a)
# tensor([[-0.2492, -0.6960, 2.3492],
# [-0.1057, -0.3203, 2.2584],
# [-0.0774, -0.2463, 2.2405]])


b = torch.addcdiv(t, 0.1, t1, t2)
print(b)
# tensor([[-0.2492, -0.6960, 2.3492],
# [-0.1057, -0.3203, 2.2584],
# [-0.0774, -0.2463, 2.2405]])

---

2、torch.addcmul(t, v, t1, t2)

   t1 与 t2 按元素乘后，乘v 加t

import torch

t = torch.randn(1, 3)
t1 = torch.randn(3, 1)
t2 = torch.randn(1, 3)

a = t + 0.1 * t1 * t2
print(a)
# tensor([[-0.4994, 1.4826, -0.3377],
# [-0.4893, 1.4880, -0.3338],
# [-0.5957, 1.4314, -0.3756]])


b = torch.addcmul(t, 0.1, t1, t2)
print(b)
# tensor([[-0.4994, 1.4826, -0.3377],
# [-0.4893, 1.4880, -0.3338],
# [-0.5957, 1.4314, -0.3756]])

---

3、torch.clamp(input, min, max)

将张量元素大小限制在指定区间范围内

import torch

x = torch.arange(1, 8)
y = torch.clamp(x, 2, 5)
print(y)
# tensor([2, 2, 3, 4, 5, 5, 5])

---

4、torch.ceil(input) 、torch.floor(input)

torch.ceil(input) ：向上取整
torch.floor(input) ：向下取整

import torch

torch.manual_seed(8)
x = torch.randn(3) * 10
y = torch.ceil(x)
z = torch.floor(x)
print(x)
print(y)
print(z)

---

二、归并操作

---

1、torch.cumprod(t, axis)

在指定维度对 t 进行累积 （cumprod：cumulative product）

import torch

a = torch.arange(1, 10).reshape(3, 3)
print(a)

b_x = torch.cumprod(a, dim=0)  # 沿着y轴累积
print("\ncumulative product:\n", b_x)

b_y = torch.cumprod(a, dim=1)  # 沿着x轴累积
print("\ncumulative product:\n", b_y)

# tensor([[1, 2, 3],
# [4, 5, 6],
# [7, 8, 9]])
# 
# cumulative product:
# tensor([[ 1, 2, 3],
# [ 4, 10, 18],
# [ 28, 80, 162]])
# 
# cumulative product:
# tensor([[ 1, 2, 6],
# [ 4, 20, 120],
# [ 7, 56, 504]])

---

2、torch.cumsum(t, axis)

import torch

a = torch.linspace(0, 10, 6).view(2, 3)

b = a.sum(dim=0)
c = torch.cumsum(a, dim=0)
print(a)
print(b)
print(c)
# tensor([[ 0., 2., 4.],
# [ 6., 8., 10.]])
# 
# tensor([ 6., 10., 14.])
# 
# tensor([[ 0., 2., 4.],
# [ 6., 10., 14.]])


d = a.sum(dim=1)
e = torch.cumsum(a, dim=1)
print(d)
print(e)
# tensor([ 6., 24.])
# 
# tensor([[ 0., 2., 6.],
# [ 6., 14., 24.]])

---

三、比较操作

比较操作一般是进行逐元素比较，有些是按指定方向比较

1、torch.eq(a, b) 元素是否相等

a = torch.tensor([[1, 2, 3]])
b = torch.tensor([[1, 2, 4]])

print(a.eq(b))
# tensor([[ True, True, False]])

print(torch.eq(a, b))
# tensor([[ True, True, False]])

2、torch.equal(a, b) tensor是否完全相等

比较两个张量的形状和各个元素是否都相等

a = torch.tensor([[1, 2, 3]])
b = torch.tensor([[1, 2, 4]])
print(a.equal(b))  # False
print(torch.equal(a, b))  # False

c = torch.tensor([[1, 2, 3]]) 
d = torch.tensor([[1, 2, 3]]) 
print(c.equal(d))  # True
print(torch.equal(c, d))  # True

3、torch.ge 、torch.le 、torch.gt、torch.lt 大于等于/小于等于/大于/小于

a = torch.tensor([[1, 2, 5]])
b = torch.tensor([[1, 3, 3]])

# 大于等于
print(a.ge(b))  # tensor([[ True, False, True]])
print(torch.ge(a, b))  # tensor([[ True, False, True]])

# 小于等于
print(a.le(b))  # tensor([[ True, True, False]])
print(torch.le(a, b))  # tensor([[ True, True, False]])

# 大于
print(a.gt(b))  # tensor([[False, False, True]])
print(torch.gt(a, b))  # tensor([[False, False, True]])

# 小于
print(a.lt(b))  # tensor([[False, True, False]])
print(torch.lt(a, b))  # tensor([[False, True, False]])

4、torch.max、torch.min 最大值、最小值

若指定axis， 则额外返回索引下标

a = torch.tensor([[1, 8, 3],
                  [2, 5, 3]])

print(a.max())  # tensor(8)

print(torch.min(a))  # tensor(1)

print(a.max(0))
# torch.return_types.max(
# values=tensor([2, 8, 3]),
# indices=tensor([1, 0, 0]))

print(torch.min(a, 0))
# torch.return_types.min(
# values=tensor([1, 5, 3]),
# indices=tensor([0, 1, 0]))

5、torch.topk 指定维度上最高的k个值

a = torch.tensor([[1, 8, 3],
                  [2, 5, 3]])

# 维度 1 上的最大的 2 个值
print(a.topk(2, 1))
# torch.return_types.topk(
# values=tensor([[8, 3],
# [5, 3]]),
# indices=tensor([[1, 2],
# [1, 2]]))

---

四、矩阵操作

1、torch.dot(a, b) 点积

a = torch.tensor([2, 3])
b = torch.tensor([3, 4])

print(torch.dot(a, b))
# tensor(18)

Torch的 dot 只能对两个为1D 的张量进行点积运算，否则会报错；Numpy中的dot无此限制。

a = torch.tensor([[2, 3],
                  [3, 4]])

b = torch.tensor([[3, 4],
                  [1, 2]])

print(torch.dot(a, b))

2、torch.mm(a, b) 矩阵乘法

3、torch.mul(a, b)、 a * b 逐元素相乘

a = torch.tensor([[2, 3],
                  [3, 4]])

b = torch.tensor([[3, 4],
                  [1, 2]])

print(torch.mm(a, b))
# tensor([[ 9, 14],
# [13, 20]])

print(torch.mul(a, b))
# tensor([[ 6, 12],
# [ 3, 8]])

print(a * b)
# tensor([[ 6, 12],
# [ 3, 8]])

4、torch.mv(a, b) 矩阵与向量乘法

torch.mv(a, b), 矩阵a为第一个参数，向量b为第二个参数，位置不能换，否则会报错

a = torch.tensor([[1, 2, 3],
                [2, 3, 4]])

b = torch.tensor([1, 2, 3])

print(torch.mv(a, b))
# tensor([14, 20])

5、tensor.T 转置

a = torch.randint(10, (2, 3))

print(a)
# tensor([[6, 9, 6],
# [0, 4, 8]])

print(a.T)
# tensor([[6, 0],
# [9, 4],
# [6, 8]])

6、 torch.svd(a) 矩阵的SVD分解

a = torch.randn(2, 3)
print(torch.svd(a))

# torch.return_types.svd(
# U=tensor([[-0.5960, 0.8030],
# [-0.8030, -0.5960]]),
# S=tensor([3.3907, 1.0873]),
# V=tensor([[ 0.2531, 0.8347],
# [-0.7722, 0.4789],
# [-0.5828, -0.2721]]))