Pytorch张量操作

Pytorch

1. 数据操作

1. 创建张量

torch.arange()
torch.zeros()
torch.ones()
torch.randn()
torch.tensor()
torch.Tensor()

2. 查看张量属性

x.shape / x.size()   # 张量形状
x.numel()  # 张量元素个数
x.dtype   # 张量类型

3. 改变张量形状

x.view()   # 不生成副本
x.reshape()  # 生成原数据副本

4. 交换维度

# 交换 dim0 和 dim1
torch.transpose(input, dim0, dim1) -> Tensor
x.transpose(dim0, dim1)
# 交换 dims
x.permute(dims) -> Tensor  

5. 压缩维度

x.flatten(dim)        # 从 dim 维度开始压缩 
x.squeeze()          # 压缩维数为1的维度，可指定维数
x.unsqueeze(dim = a)    # 扩充指定维度

6. 降维

A.sum(axis = 0)  # 沿指定维度 0 求和
# 
A.mean() 
A.sum()/A.numel()
# 
A.mean(dim = 0)
A.sum(aixs = 0) / A.shape[0]

7. 非降维求和

sum_A = A.sum(axis = 0, keepdims = True)
A / sum_A
# 沿着某轴计算累积和
A.comsum(axis = 0)

import torch

x = torch.ones(12).reshape(3, -1)
y = torch.zeros(12).reshape(3, -1)

x = x.unsqueeze(0)       # (3,4) -> (1. 3. 4)
x = x.transpose(-1, -2)  # (1, 3, 4) -> (1, 4, 3)
x = x.permute(1,2,0)     #  (1, 4, 3) -> (4, 3, 1)
x = x.flatten(1)         # (4, 3, 1) -> （4，3）
x = x.reshape(3,-1)      # (4, 3\) -> （3, 4）
x.size()

torch.Size([3, 4])

2. 运算符

1. 基本运算

x + y, x - y, x * y, x / y, x ** y

2. 求幂

torch.exp(x)

3. concatenate:

torch.cat((x,y),dim = 0)   # 在现有维度拼接
torch.stack((x,y),dim =  )  # 在一个新维度拼接

4. 逻辑判断

x == y

5. 元素求和：

x.sum()

• $torch.cat()$

z_0 = torch.cat((x, y), dim = 0)
z_1 = torch.cat((x,y), dim = 1)

x.shape, z_0.size(), z_1.size()

(torch.Size([3, 4]), torch.Size([6, 4]), torch.Size([3, 8]))

• $torch.stack()$

z0 = torch.stack((x, y), dim = 0)
z1 = torch.stack((x, y), dim = 1)
z2 = torch.stack((x, y), dim = 2)

z0.size(), z1.size(), z2.size()

(torch.Size([2, 3, 4]), torch.Size([3, 2, 4]), torch.Size([3, 4, 2]))

3. 广播机制

4. 索引 & 切片

5. 转换类型

1. torch.tensor -> numpy.ndarray

a = x.numpy()

2. numpy.ndarray -> torch.tensor

b= torch.from_numpy(a)
b = torch.Tensor(a)
b = torch.tensor(a)

6. 张量运算

1. Hadamard 积

A * B

2. 点积（Dot Product)

x.dot(y)
torch.sum(x * y)

3. 矩阵-向量积

torch.mv(A, x)

4. 矩阵乘法

torch.mm(A, B)
# 常见
torch.mm(A, A.transpose(-1, -2))

5. 范数

torch.norm(input, p, dim, out=None,keepdim=False) → Tensor # 求指定维的L2范数
torch.abs(x)sum() # 求L1范数