pytorch学习笔记（二）

import torch
import numpy as np

一、合并与切割

1.concat

a1 = torch.rand(4, 3, 32, 32)
a2 = torch.rand(5, 3, 32, 32)

# 将a1与a2的第0个维度相加
torch.cat([a1, a2], dim=0).shape

out:torch.Size([9, 3, 32, 32])

a2 = torch.rand(4, 1, 32, 32)
# 将a1与a2的第1个维度相加
torch.cat([a1, a2], dim=1).shape

out:torch.Size([4, 4, 32, 32])

# 如果维度不匹配，则会报错
torch.cat([a1, a2], dim=0).shape

---------------------------------------------------------------------------

RuntimeError                              Traceback (most recent call last)

<ipython-input-7-c6c20381a457> in <module>
      1 # 如果维度不匹配，则会报错
----> 2 torch.cat([a1, a2], dim=0).shape

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 3 and 1 in dimension 1 at ..\aten\src\TH/generic/THTensor.cpp:711

2.stack

a = torch.rand(32, 8)
b = torch.rand(32, 8)

# stack的作用就是将两个指定tensor合并成一个新的tensor，在指定维度扩展一个新的维度，其他地方保持不变
# 比如将第0个维度合并得到一个新的(2,32,8)的tensor
c = torch.stack([a, b], dim = 0)
c.shape

out:torch.Size([2, 32, 8])

# 合并后的(2,32,8)的tensor的第0的维度既是a， b两个tensor
print(torch.equal(torch.stack([a, b], dim = 0)[0], a))

out:True

torch.stack([a, b], dim = 1).shape

torch.Size([32, 2, 8])

# stack对维度和每一维度的通道数有严格的要求,若不合适则会报错
c = torch.rand(30, 8)
torch.stack([a, c], dim=0)

---------------------------------------------------------------------------

RuntimeError                              Traceback (most recent call last)

<ipython-input-12-2f6aa1a168d8> in <module>
      1 # stack对维度和每一维度的通道数有严格的要求,若不合适则会报错
      2 c = torch.rand(30, 8)
----> 3 torch.stack([a, c], dim=0)

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 0. Got 32 and 30 in dimension 1 at ..\aten\src\TH/generic/THTensor.cpp:711

3.split

# 之前一步将a，b合并，若想将其拆开，则使用split
# 因为合并是在第0维度的，拆开则在第0维度
aa, bb = c.split([1,1], dim = 0)
aa.shape, bb.shape

out:(torch.Size([1, 32, 8]), torch.Size([1, 32, 8]))

# 验证一下拆开后的确实和原本的一致
print(torch.equal(aa[0], a))

out:True

4.chunk

# 与上一步的需求一样，使用chunk的拆分，比如将C拆成两份
aa, bb = c.chunk(2, dim = 0)
aa.shape, bb.shape

out:(torch.Size([1, 32, 8]), torch.Size([1, 32, 8]))

d = torch.rand(8, 32, 8)
aa, bb, cc = d.chunk(3, dim = 0)
aa.shape, bb.shape, cc.shape

out:(torch.Size([3, 32, 8]), torch.Size([3, 32, 8]), torch.Size([2, 32, 8]))

# 使用split则可以指定维度的通道数大小
aa, bb, cc = d.split([4, 3, 1], dim = 0)
aa.shape, bb.shape, cc.shape

out:(torch.Size([4, 32, 8]), torch.Size([3, 32, 8]), torch.Size([1, 32, 8]))

二、运算

1.乘法

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

# a * b 是简单的对应位置的元素相乘
a * b

out:tensor([[3, 3],
        [3, 3]])

# 矩阵相乘(不推荐，只适合2d的)
torch.mm(a, b)

out:tensor([[6, 6],
        [6, 6]])

# 矩阵相乘
torch.matmul(a, b)

out:tensor([[6, 6],
        [6, 6]])

# 矩阵相乘
a @ b

out:tensor([[6, 6],
        [6, 6]])

# 多维乘法还是对后两维进行矩阵运算，相当于多个矩阵进行并行乘法计算
a = torch.rand(4, 3, 28, 64)
b = torch.rand(4, 3, 64, 28)
c = torch.matmul(a, b)
c.shape

out:torch.Size([4, 3, 28, 28])

2.其他

# 建立一个（2,2）的，全部都是3的tensor
a = torch.full([2,2], 3)
a.pow(2)

out:tensor([[9., 9.],
        [9., 9.]])

aa = a ** 2
aa

out:tensor([[9., 9.],
        [9., 9.]])

a = torch.tensor(3.14)
# 向下取整， 向上取整， 整数部分， 小数部分, 四舍五入
a, a.floor(),a.ceil(), a.trunc(), a.frac(), a.round()

out:(tensor(3.1400),
 tensor(3.),
 tensor(4.),
 tensor(3.),
 tensor(0.1400),
 tensor(3.))

grad = torch.rand(2, 3) * 15
grad

out:tensor([[ 5.0516, 12.1264,  3.0216],
        [12.0383,  7.7159, 13.8156]])

grad.max()

out:tensor(13.8156)

grad.median()

out:tensor(7.7159)

# 将小于10的全部命为10
grad.clamp(10)

out:tensor([[10.0000, 12.1264, 10.0000],
        [12.0383, 10.0000, 13.8156]])

l1范数：
$$\Vert x{\Vert }_1=\sum _{i=1}^n\left|x_i\right|$$
l2范数：
$$\Vert x{\Vert }_2=\sqrt{\sum _{i=1}^N{x}_i^2}$$

a = torch.full([8], 1)
b = a.view(2, 4)
c = a.view(2, 2, 2)

out:tensor([[[1., 1.],
         [1., 1.]],

        [[1., 1.],
         [1., 1.]]])

a

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

b

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

c

out:tensor([[[1., 1.],
         [1., 1.]],

        [[1., 1.],
         [1., 1.]]])

# 求l1范数
a.norm(1), b.norm(1), c.norm(1)

out:(tensor(8.), tensor(8.), tensor(8.))

# 求指定维度的l1范数
b.norm(1, dim = 1)

out:torch.Size([2])

# 求指定维度的l1范数
b.norm(1, dim = 0)

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

# l2范数
a.norm(2), b.norm(2), c.norm(2)

out:(tensor(2.8284), tensor(2.8284), tensor(2.8284))