pytorch基础操作

1 机器学习基本构成元素

分类是离散，回归是连续

2 pytorch基本概念

tensor张量 任意维度

3 tensor与机器学习的关系

4 tensor创建编程实例

import torch

a = torch.Tensor([[1,2],[3,4]])
print(a)
print(a.type())

a=torch.Tensor(2,3)
print(a)
print(a.type())

'''几种特殊的tensor'''
a=torch.ones(2,2)
print(a)
print(a.type())

a=torch.eye(2,2)
print(a)
print(a.type())

a=torch.zeros(2,2)
print(a)
print(a.type())

b = torch.Tensor(2,3)
b = torch.ones_like(b)
b = torch.zeros_like(b)
print(b)
print(b.type())

'''随机'''
a= torch.rand(2,2)
print(a)
print(a.type())

a = torch.normal(mean=0.0, std=torch.rand(5)) #正态分布
print(a)
print(a.type())

a = torch.normal(mean=torch.rand(5), std=torch.rand(5)) #正态分布
print(a)
print(a.type())

'''a = torch.Tensor(2,2).uniform(-1,1) #均匀分布要指定Tensor的大小
print(a)
print(a.type())'''

a = torch.arange(0,10,1)
print(a)
print(a.type())

a = torch.linspace(2,10,4) #拿到等间隔的n个数字
print(a)
print(a.type())

a = torch.randperm(10)
print(a)
print(a.type())

################
import numpy as np
a = np.array([[1,2],[3,4]])
print(a)
b = np.eye(3)
print(b)

5 tensor的属性

稀疏（零元素的个数越多越稀疏）
①稀疏使模型更简单
②减少内存的开销

6 稀疏张量的编程实践

import torch

dev = torch.device("cpu")
dev = torch.device("cuda")
a = torch.tensor([2,2],dtype=torch.float32,device=dev)
print(a)

i = torch.tensor([[0,1,2],[0,1,2]])
v = torch.tensor([1,2,3])
a= torch.sparse_coo_tensor(i,v,(4,4))
print(a)   #稀疏的张量

i = torch.tensor([[0,1,2],[0,1,2]])
v = torch.tensor([1,2,3])
a= torch.sparse_coo_tensor(i,v,(4,4),
                           dtype=torch.float32,
                           device=dev).to_dense() #device是对资源的分配cpu,gpu
print(a)  # 转换成稠密的张量

7 tensor的算数运算

带下划线运算之后，a的值被修改为相加的结果

8 tensor算术运算编程实例

import torch

#add
a=torch.rand(2,3)
b=torch.rand(2,3)
print(a)
print(b)

print(a+b)
print(a.add(b))
print(torch.add(a,b))
print(a)
print(a.add_(b))
print(a)

#sub
print("==== sub ====")
print(a-b)
print(a.sub(b))
print(torch.sub(a,b))
print(a)
print(a.sub_(b))
print(a)

#mul
print("==== mul ====")
print(a*b)
print(a.mul(b))
print(torch.mul(a,b))
print(a)
print(a.mul_(b))
print(a)

print("==== div ====")
print(a/b)
print(a.div(b))
print(torch.div(a,b))
print(a.div_(b))
print(a)

###matmul
a = torch.ones(2,1)
b = torch.ones(1,2)
print(a @ b)
print(a.matmul(b))
print(torch.matmul(a,b))
print(torch.mm(a,b))
print(a.mm(b))

###高维tensor
a = torch.ones(1,2,3,4)
b = torch.ones(1,2,4,3)
print(a.matmul(b).shape)

##pow
a = torch.tensor([1,2])
print(a)
print(torch.pow(a,3))
print(a.pow(3))
print(a**3)
print(a.pow_(3))
print(a)

#exp
a = torch.tensor([1,2],dtype=torch.float32)
print(torch.exp(a))
print(torch.exp_(a))
print(a.exp())
print(a.exp_())

###log
a = torch.tensor([10,2],dtype=torch.float32)
print(torch.log(a))
print(torch.log_(a))
print(a.log())
print(a.log_())

###sqrt
a = torch.tensor([10,2],dtype=torch.float32)
print(torch.sqrt(a))
print(torch.sqrt_(a))
print(a.sqrt())
print(a.sqrt_())

9 in-place的概念和广播机制

import torch
a = torch.rand(2,1,1,3)
b = torch.rand(4,2,3)
# a, 2*1
# b, 1*2
# c, 2*2
# c, 2*4*2*3
c = a+b
print(a)
print(b)
print(c)
print(c.shape)

10 取整-余

11 比较运算-排序

import torch

a = torch.rand(2,3)
b = torch.rand(2,3)
print(a)
print(b)

print(torch.eq(a,b))  #等于
print(torch.equal(a,b))

print(torch.ge(a,b))  #大于等于
print(torch.gt(a,b))  #大于
print(torch.le(a,b))  #小于等于
print(torch.lt(a,b))  #小于
print(torch.ne(a,b))  #不等于

####
a = torch.tensor([1,4,4,3,5])
print(torch.sort(a))
print(torch.sort(a,descending=True))

b = torch.tensor([[1,4,4,3,5],
                  [2,3,1,3,5]])
print(b.shape)
print(torch.sort(b,dim=1,descending=False))  #dim=0对2这个维度排序，dim=1对5这个维度排序

#####topk
a= torch.tensor([[2,4,3,1,5],
                 [2,3,5,1,4]])
print(a.shape)
print(torch.topk(a,k=2,dim=1))
print(torch.topk(a,k=1,dim=0))

print(torch.kthvalue(a,k=2,dim=0))
print(torch.kthvalue(a,k=2,dim=1)) #取第二小的数

a = torch.rand(2,3)
print(a)
print(a/0)
print(torch.isfinite(a)) #看看是否有界
print(torch.isfinite(a/0))
print(torch.isinf(a/0))  #看看是否无界
print(torch.isnan(a))

import numpy as np
b = torch.tensor([1,2,np.nan])
print(torch.isnan(b))

12 三角函数

import torch
a = torch.zeros(2,3)
b = torch.cos(a)
print(a)
print(b)

13 其他数学函数

14 pytorch与统计学方法

import torch
a = torch.rand(2,2)
print(a)
print(torch.mean(a))
print(torch.sum(a))
print(torch.prod(a))

b = torch.rand(2,2)
print(b)
print(torch.mean(b,dim=0))
print(torch.sum(b,dim=0))
print(torch.prod(b,dim=0))

print(torch.argmax(a,dim=0))
print(torch.argmin(a,dim=0))
print(torch.std(a))  #标准差
print(torch.var(a))  #方差
print(torch.median(a)) #中值
print(torch.mode(a)) #众数

a = torch.rand(2,2)*10
print(a)
print(torch.histc(a,6,0,0))

a= torch.randint(0,10,[10])
print(a)
print(torch.bincount(a)) #bincount只能处理1维的
#统计某一类别样本的个数

15 pytorch与分布函数

16 pytorch与随机抽样

import torch

torch.manual_seed(1)
mean = torch.rand(1,2)
std = torch.rand(1,2)
print(torch.normal(mean,std)) #正态分布

17 pytorch与线性代数

0范数是指向量中非零元素的个数
1范数是指元素的绝对值的和
通过范数约束可以完成对模型参数的约束

18 pytorch与矩阵分解-PCA

矩阵分解：解决和分析最优化问题时会用到

LU、QR用的不多
EVD、SVD用的较多，需要掌握。与EVD分解相对应的是PCA算法（无监督的算法），与SVD分解相对应的是LDA算法（可以完成特征工程的降维，引入了学习的因素，有监督的算法）。

满秩的矩阵称为奇异矩阵（用EVD），不满秩的矩阵称为非奇异矩阵（用SVD）。

方差越大，说明特征值越丰富，相关性越低，信息之间的冗余度越低
把n维中能量低的去掉，保留其中k维能量高的。

19 SVD分解-LDA

S_B是同类物体之间的协方差矩阵
S_w是不同类物体之间的协方差矩阵
类内要小，类间要大

20 pytorch与张量裁剪

import torch
a = torch.rand(2,2)*10
print(a)
a = a.clamp(2,5)
print(a)

运行结果：