• pytorch张量维度变换详解：view、squeeze、transpose

    目录

    1 view函数

    1.1 指定变换后的维度

    1.2 自动推理变换后的维度

    1.3 将tensor展平成一维

    2 reshape函数

    2.1 指定变换后的维度

    2.2 自动推理转换后的维度

    2.3 将tensor展平成一维

    2.4 使用tensor.reshape变换

    3 squeeze函数

    3.1 torch.squeeze去除所有为1的维度

    3.2 torch.squeeze指定dim去除

    3.3 tensor.squeeze去除为1的维度

    4 unsqueeze函数

    4.1 torch.unsqueeze指定dim插入新维度

    4.2 tensor.unsqueeze指定dim插入新维度

    5 transpose函数

    5.1 torch.transpose转置指定维度

    5.2 tensor.transpose转置指定维度

    6 expand函数

    7 repeat函数

    8 permute函数

     

    Pytorch张量维度变化是在构建模型过程中常用且重要的操作，本文从实际应用触发，详细介绍常用的维度变化方法，这些方法包含view、reshap、squeeze、unsqueeze、transpose等。

    1 view函数

    Pytorch中的view函数主要用于Tensor维度的重构，即返回一个有相同数据但不同维度的Tensor。

    view函数的操作对象是Tensor类型，返回的对象类型也为Tensor类型

        def view(self, *size: _int) -> Tensor: ...

    更便于理解的表示形式：

    view(参数a,参数b,…)，其中，总的参数个数表示将张量重构后的维度。

    1.1 指定变换后的维度

    通过手工指定，将一个一维tensor变换为3*8维的tensor

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 
    4.                    13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24])
    5.  
    6. a2 = a1.view(3, 8)
    7. print(a1)
    8. print(a2)
    9. print(a1.shape)
    10. print(a2.shape)

    运行程序显示如下：

    1. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
    2.         19, 20, 21, 22, 23, 24])
    3. tensor([[ 1,  2,  3,  4,  5,  6,  7,  8],
    4.         [ 9, 10, 11, 12, 13, 14, 15, 16],
    5.         [17, 18, 19, 20, 21, 22, 23, 24]])
    6. torch.Size([24])
    7. torch.Size([3, 8])

    1.2 自动推理变换后的维度

    如果某个参数为-1，则表示该维度取决于其它维度，由Pytorch自己补充

    1. import torch
    2.  
    3. a3 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12,
    4.                    13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24])
    5.  
    6. a4 = a3.view(4, -1)
    7. a5 = a3.view(2, 3, -1)
    8. a6 = a3.view(-1, 3, 2)
    9.  
    10. print(a3)
    11. print(a4)
    12. print(a5)
    13. print(a6)
    14. print(a3.shape)
    15. print(a4.shape)
    16. print(a5.shape)
    17. print(a6.shape)

     运行程序显示如下：

    1. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
    2.         19, 20, 21, 22, 23, 24])
    3. tensor([[ 1,  2,  3,  4,  5,  6],
    4.         [ 7,  8,  9, 10, 11, 12],
    5.         [13, 14, 15, 16, 17, 18],
    6.         [19, 20, 21, 22, 23, 24]])
    7. tensor([[[ 1,  2,  3,  4],
    8.          [ 5,  6,  7,  8],
    9.          [ 9, 10, 11, 12]],
    10.         [[13, 14, 15, 16],
    11.          [17, 18, 19, 20],
    12.          [21, 22, 23, 24]]])
    13. tensor([[[ 1,  2],
    14.          [ 3,  4],
    15.          [ 5,  6]],
    16.         [[ 7,  8],
    17.          [ 9, 10],
    18.          [11, 12]],
    19.         [[13, 14],
    20.          [15, 16],
    21.          [17, 18]],
    22.         [[19, 20],
    23.          [21, 22],
    24.          [23, 24]]])
    25. torch.Size([24])
    26. torch.Size([4, 6])
    27. torch.Size([2, 3, 4])
    28. torch.Size([4, 3, 2])

    1.3 将tensor展平成一维

    1. import torch
    2.  
    3. a7 = torch.tensor([[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
    4.                    [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]])
    5. a8 = a6.view(-1)
    6. print(a7)
    7. print(a8)
    8. print(a7.shape)
    9. print(a8.shape)

     运行程序显示如下：

    1. tensor([[ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12],
    2.         [13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24]])
    3. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13, 14, 15, 16, 17, 18,
    4.         19, 20, 21, 22, 23, 24])
    5. torch.Size([2, 12])
    6. torch.Size([24])

    2 reshape函数

    返回与 input张量数据大小一样、给定 shape的张量。如果可能，返回的是input 张量的视图，否则返回的是其拷贝。

    torch.reshape(input, shape) → [Tensor]

    也可以直接在Tensor上使用reshape，形式如下：

    tensor.reshape(shape) → [Tensor]

    2.1 指定变换后的维度

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = torch.reshape(a1, (3, 4))
    5. print(a1.shape)
    6. print(a1)
    7. print(a2.shape)
    8. print(a2)

    运行程序显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 4])
    4. tensor([[ 1,  2,  3,  4],
    5.         [ 5,  6,  7,  8],
    6.         [ 9, 10, 11, 12]])

    2.2 自动推理转换后的维度

    1. import torch
    2.  
    3. a3 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a4 = torch.reshape(a1, (-1, 6))
    5. print(a3.shape)
    6. print(a3)
    7. print(a4.shape)
    8. print(a4)

    运行程序显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([2, 6])
    4. tensor([[ 1,  2,  3,  4,  5,  6],
    5.         [ 7,  8,  9, 10, 11, 12]])

    2.3 将tensor展平成一维

    1. import torch
    2.  
    3. a5 = torch.tensor([[1, 2, 3, 4, 5, 6], [7, 8, 9, 10, 11, 12]])
    4. a6 = torch.reshape(a1, (-1,))
    5. print(a5.shape)
    6. print(a5)
    7. print(a6.shape)
    8. print(a6)

    运行程序显示如下：

    1. torch.Size([2, 6])
    2. tensor([[ 1,  2,  3,  4,  5,  6],
    3.         [ 7,  8,  9, 10, 11, 12]])
    4. torch.Size([12])
    5. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])

    2.4 使用tensor.reshape变换

    1. improt torch
    2.  
    3. a7 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a8 = a7.reshape(6, 2)
    5. a9 = a7.reshape(-1, 3)
    6. a10 = a9.reshape(-1)
    7. print(a7.shape)
    8. print(a7)
    9. print(a8.shape)
    10. print(a8)
    11. print(a9.shape)
    12. print(a9)
    13. print(a10.shape)
    14. print(a10)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([6, 2])
    4. tensor([[ 1,  2],
    5.         [ 3,  4],
    6.         [ 5,  6],
    7.         [ 7,  8],
    8.         [ 9, 10],
    9.         [11, 12]])
    10. torch.Size([4, 3])
    11. tensor([[ 1,  2,  3],
    12.         [ 4,  5,  6],
    13.         [ 7,  8,  9],
    14.         [10, 11, 12]])
    15. torch.Size([12])
    16. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])

    3 squeeze函数

    将input张量中所有维度数据为1的维度给移除掉。指定了dim，如果dim对应维度的值不为1 ，则保持不变，为1则移除该维度。

     torch.squeeze(input, dim=None) → [Tensor]

     也可以在tensor上直接使用squeeze，形式如下：

     tensor.squeeze(dim=None) → [Tensor]

    3.1 torch.squeeze去除所有为1的维度

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = a1.reshape(3, 1, 4)
    5. a3 = torch.squeeze(a2)
    6.  
    7. print(a1.shape)
    8. print(a1)
    9. print(a2.shape)
    10. print(a2)
    11. print(a3.shape)
    12. print(a3)

     运行结果显示如下：（a2的第二个维度被移除）

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 1, 4])
    4. tensor([[[ 1,  2,  3,  4]],
    5.  
    6.         [[ 5,  6,  7,  8]],
    7.  
    8.         [[ 9, 10, 11, 12]]])
    9. torch.Size([3, 4])
    10. tensor([[ 1,  2,  3,  4],
    11.         [ 5,  6,  7,  8],
    12.         [ 9, 10, 11, 12]])

    3.2 torch.squeeze指定dim去除为1的维度

    1. import torch
    2.  
    3. a4 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a5 = a1.reshape(3, 1, 4)
    5. a6 = torch.squeeze(a5, 0)
    6. a7 = torch.squeeze(a5, 1)
    7.  
    8. print(a4.shape)
    9. print(a4)
    10. print(a5.shape)
    11. print(a5)
    12. print(a6.shape)
    13. print(a6)
    14. print(a7.shape)
    15. print(a7)

    运行结果显示如下：（a5的第一个维度不为1，所以保持不变；a5的第二个维度为1，所以被移除）

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 1, 4])
    4. tensor([[[ 1,  2,  3,  4]],
    5.  
    6.         [[ 5,  6,  7,  8]],
    7.  
    8.         [[ 9, 10, 11, 12]]])
    9. torch.Size([3, 1, 4])
    10. tensor([[[ 1,  2,  3,  4]],
    11.  
    12.         [[ 5,  6,  7,  8]],
    13.  
    14.         [[ 9, 10, 11, 12]]])
    15. torch.Size([3, 4])
    16. tensor([[ 1,  2,  3,  4],
    17.         [ 5,  6,  7,  8],
    18.         [ 9, 10, 11, 12]])

    3.3 tensor.squeeze指定dim去除为1的维度

    1. import torch
    2.  
    3. a8 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a9 = a8.reshape(3, 1, 4)
    5. a10 = a9.squeeze()
    6. a11 = a9.squeeze(0)
    7. a12 = a9.squeeze(1)
    8.  
    9. print(a8.shape)
    10. print(a8)
    11. print(a9.shape)
    12. print(a9)
    13. print(a10.shape)
    14. print(a10)
    15. print(a11.shape)
    16. print(a11)
    17. print(a12.shape)
    18. print(a12)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 1, 4])
    4. tensor([[[ 1,  2,  3,  4]],
    5.  
    6.         [[ 5,  6,  7,  8]],
    7.  
    8.         [[ 9, 10, 11, 12]]])
    9. torch.Size([3, 4])
    10. tensor([[ 1,  2,  3,  4],
    11.         [ 5,  6,  7,  8],
    12.         [ 9, 10, 11, 12]])
    13. torch.Size([3, 1, 4])
    14. tensor([[[ 1,  2,  3,  4]],
    15.  
    16.         [[ 5,  6,  7,  8]],
    17.  
    18.         [[ 9, 10, 11, 12]]])
    19. torch.Size([3, 4])
    20. tensor([[ 1,  2,  3,  4],
    21.         [ 5,  6,  7,  8],
    22.         [ 9, 10, 11, 12]])

    4 unsqueeze函数

    在给定的 dim 维度位置插入一个新的维度，维度数值为 1，dim 的范围在 [-dim()-1, dim()+1)，包首不包尾

    torch.unsqueeze(input, dim) → [Tensor]

     也可以在tensor上直接使用unsqueeze，形式如下：

    torch.unsqueeze(dim) → [Tensor]

    4.1 torch.unsqueeze指定dim插入新维度

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = a1.reshape(3, 4)
    5. a3 = torch.unsqueeze(a2, 0)
    6. a4 = torch.unsqueeze(a2, 2)
    7.  
    8. print(a1.shape)
    9. print(a1)
    10. print(a2.shape)
    11. print(a2)
    12. print(a3.shape)
    13. print(a3)
    14. print(a4.shape)
    15. print(a4)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 4])
    4. tensor([[ 1,  2,  3,  4],
    5.         [ 5,  6,  7,  8],
    6.         [ 9, 10, 11, 12]])
    7. torch.Size([1, 3, 4])
    8. tensor([[[ 1,  2,  3,  4],
    9.          [ 5,  6,  7,  8],
    10.          [ 9, 10, 11, 12]]])
    11. torch.Size([3, 4, 1])
    12. tensor([[[ 1],
    13.          [ 2],
    14.          [ 3],
    15.          [ 4]],
    16.  
    17.         [[ 5],
    18.          [ 6],
    19.          [ 7],
    20.          [ 8]],
    21.  
    22.         [[ 9],
    23.          [10],
    24.          [11],
    25.          [12]]])

    4.2 tensor.unsqueeze指定dim插入新维度

    1. import torch
    2.  
    3. a5 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a6 = a5.reshape(3, 4)
    5. a7 = a6.unsqueeze(0)
    6. a8 = a6.unsqueeze(1)
    7.  
    8. print(a5.shape)
    9. print(a5)
    10. print(a6.shape)
    11. print(a6)
    12. print(a7.shape)
    13. print(a7)
    14. print(a8.shape)
    15. print(a8)

    运行结果显示如下：

    1. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    2. torch.Size([3, 4])
    3. tensor([[ 1,  2,  3,  4],
    4.         [ 5,  6,  7,  8],
    5.         [ 9, 10, 11, 12]])
    6. torch.Size([1, 3, 4])
    7. tensor([[[ 1,  2,  3,  4],
    8.          [ 5,  6,  7,  8],
    9.          [ 9, 10, 11, 12]]])
    10. torch.Size([3, 1, 4])
    11. tensor([[[ 1,  2,  3,  4]],
    12.  
    13.         [[ 5,  6,  7,  8]],
    14.  
    15.         [[ 9, 10, 11, 12]]])

    5 transpose函数

    返回 input 张量的转置，dim0与dim1交换位置

    torch.transpose(input, dim0, dim1) → [Tensor]

      也可以在tensor上直接使用unsqueeze，形式如下：

    tensor.transpose(dim0, dim1) → [Tensor]

    参数：

    • input ([Tensor] 输入的张量
    • dim0 ([int] 第一个要转置的维度
    • dim1 ([int] 第二个要转置的维度

    5.1 torch.transpose转置指定维度

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = a1.reshape(4, 3, 1)
    5. a3 = torch.transpose(a2, 0, 1)
    6. a4 = torch.transpose(a2, 1, 2)
    7.  
    8. print(a1.shape)
    9. print(a1)
    10. print(a2.shape)
    11. print(a2)
    12. print(a3.shape)
    13. print(a3)
    14. print(a4.shape)
    15. print(a4)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([4, 3, 1])
    4. tensor([[[ 1],
    5.          [ 2],
    6.          [ 3]],
    7.         [[ 4],
    8.          [ 5],
    9.          [ 6]],
    10.         [[ 7],
    11.          [ 8],
    12.          [ 9]],
    13.         [[10],
    14.          [11],
    15.          [12]]])
    16. torch.Size([3, 4, 1])
    17. tensor([[[ 1],
    18.          [ 4],
    19.          [ 7],
    20.          [10]],
    21.         [[ 2],
    22.          [ 5],
    23.          [ 8],
    24.          [11]],
    25.         [[ 3],
    26.          [ 6],
    27.          [ 9],
    28.          [12]]])
    29. torch.Size([4, 1, 3])
    30. tensor([[[ 1,  2,  3]],
    31.         [[ 4,  5,  6]],
    32.         [[ 7,  8,  9]],
    33.         [[10, 11, 12]]])

    5.2 tensor.transpose转置指定维度

    1. import torch
    2.  
    3. a5 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a6 = a1.reshape(4, 3, 1)
    5. a7 = a6.transpose(0, 1)
    6. a8 = a6.transpose(1, 2)
    7.  
    8. print(a5.shape)
    9. print(a5)
    10. print(a6.shape)
    11. print(a6)
    12. print(a7.shape)
    13. print(a7)
    14. print(a8.shape)
    15. print(a8)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([4, 3, 1])
    4. tensor([[[ 1],
    5.          [ 2],
    6.          [ 3]],
    7.         [[ 4],
    8.          [ 5],
    9.          [ 6]],
    10.         [[ 7],
    11.          [ 8],
    12.          [ 9]],
    13.         [[10],
    14.          [11],
    15.          [12]]])
    16. torch.Size([3, 4, 1])
    17. tensor([[[ 1],
    18.          [ 4],
    19.          [ 7],
    20.          [10]],
    21.         [[ 2],
    22.          [ 5],
    23.          [ 8],
    24.          [11]],
    25.         [[ 3],
    26.          [ 6],
    27.          [ 9],
    28.          [12]]])
    29. torch.Size([4, 1, 3])
    30. tensor([[[ 1,  2,  3]],
    31.         [[ 4,  5,  6]],
    32.         [[ 7,  8,  9]],
    33.         [[10, 11, 12]]])

    6 expand函数

    返回张量的新视图，其某个维度 size 扩展到更大的 size，如果当前维度 size 为 -1 ，表示当前维度 size 保持不变。

    Tensor也可以扩展到更多的维度，新的会追加在最前面。对于新维度，大小不能设置为 -1；

    扩展张量不会分配新内存，而只会在现有张量上创建一个新视图。任何大小为1的维度都可以扩展为任意值，而无需分配新内存。

    Tensor.expand( *sizes) → [Tensor]

    参数：

    • sizes (torch.Size or [int] – 指定维度复制的次数

     

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = a1.reshape(3, 1, 4, 1)
    5.  
    6. # 维度为 1 的 size 可以扩展成什么任意的 size
    7. a3 = a2.expand(3, 5, 4, 2)
    8.  
    9. # -1 表示对应的维度size不变，但如果第一个维度3扩展成6则会报错，维度不为1不能扩展
    10. a4 = a2.expand(-1, 5, -1, -1)
    11.  
    12. # 可以扩展新的维度，但只会放到最前面，不能放到后面（会报错）且不能设置为-1
    13. a5 = a2.expand(2, -1, 5, -1, -1)
    14.  
    15. print(a1.shape)
    16. print(a1)
    17. print(a2.shape)
    18. print(a2)
    19. print(a3.shape)
    20. print(a3)
    21. print(a4.shape)
    22. print(a4)
    23. print(a5.shape)
    24. print(a5)

    运行结果显示如下 ：（维度不为1则不能扩展）

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 1, 4, 1])
    4. tensor([[[[ 1],
    5.           [ 2],
    6.           [ 3],
    7.           [ 4]]],
    8.         [[[ 5],
    9.           [ 6],
    10.           [ 7],
    11.           [ 8]]],
    12.         [[[ 9],
    13.           [10],
    14.           [11],
    15.           [12]]]])
    16. torch.Size([3, 5, 4, 2])
    17. tensor([[[[ 1,  1],
    18.           [ 2,  2],
    19.           [ 3,  3],
    20.           [ 4,  4]],
    21.          [[ 1,  1],
    22.           [ 2,  2],
    23.           [ 3,  3],
    24.           [ 4,  4]],
    25.          [[ 1,  1],
    26.           [ 2,  2],
    27.           [ 3,  3],
    28.           [ 4,  4]],
    29.          [[ 1,  1],
    30.           [ 2,  2],
    31.           [ 3,  3],
    32.           [ 4,  4]],
    33.          [[ 1,  1],
    34.           [ 2,  2],
    35.           [ 3,  3],
    36.           [ 4,  4]]],
    37.         [[[ 5,  5],
    38.           [ 6,  6],
    39.           [ 7,  7],
    40.           [ 8,  8]],
    41.          [[ 5,  5],
    42.           [ 6,  6],
    43.           [ 7,  7],
    44.           [ 8,  8]],
    45.          [[ 5,  5],
    46.           [ 6,  6],
    47.           [ 7,  7],
    48.           [ 8,  8]],
    49.          [[ 5,  5],
    50.           [ 6,  6],
    51.           [ 7,  7],
    52.           [ 8,  8]],
    53.          [[ 5,  5],
    54.           [ 6,  6],
    55.           [ 7,  7],
    56.           [ 8,  8]]],
    57.         [[[ 9,  9],
    58.           [10, 10],
    59.           [11, 11],
    60.           [12, 12]],
    61.          [[ 9,  9],
    62.           [10, 10],
    63.           [11, 11],
    64.           [12, 12]],
    65.          [[ 9,  9],
    66.           [10, 10],
    67.           [11, 11],
    68.           [12, 12]],
    69.          [[ 9,  9],
    70.           [10, 10],
    71.           [11, 11],
    72.           [12, 12]],
    73.          [[ 9,  9],
    74.           [10, 10],
    75.           [11, 11],
    76.           [12, 12]]]])
    77. torch.Size([3, 5, 4, 1])
    78. tensor([[[[ 1],
    79.           [ 2],
    80.           [ 3],
    81.           [ 4]],
    82.          [[ 1],
    83.           [ 2],
    84.           [ 3],
    85.           [ 4]],
    86.          [[ 1],
    87.           [ 2],
    88.           [ 3],
    89.           [ 4]],
    90.          [[ 1],
    91.           [ 2],
    92.           [ 3],
    93.           [ 4]],
    94.          [[ 1],
    95.           [ 2],
    96.           [ 3],
    97.           [ 4]]],
    98.         [[[ 5],
    99.           [ 6],
    100.           [ 7],
    101.           [ 8]],
    102.          [[ 5],
    103.           [ 6],
    104.           [ 7],
    105.           [ 8]],
    106.          [[ 5],
    107.           [ 6],
    108.           [ 7],
    109.           [ 8]],
    110.          [[ 5],
    111.           [ 6],
    112.           [ 7],
    113.           [ 8]],
    114.          [[ 5],
    115.           [ 6],
    116.           [ 7],
    117.           [ 8]]],
    118.         [[[ 9],
    119.           [10],
    120.           [11],
    121.           [12]],
    122.          [[ 9],
    123.           [10],
    124.           [11],
    125.           [12]],
    126.          [[ 9],
    127.           [10],
    128.           [11],
    129.           [12]],
    130.          [[ 9],
    131.           [10],
    132.           [11],
    133.           [12]],
    134.          [[ 9],
    135.           [10],
    136.           [11],
    137.           [12]]]])
    138. torch.Size([2, 3, 5, 4, 1])
    139. tensor([[[[[ 1],
    140.            [ 2],
    141.            [ 3],
    142.            [ 4]],
    143.           [[ 1],
    144.            [ 2],
    145.            [ 3],
    146.            [ 4]],
    147.           [[ 1],
    148.            [ 2],
    149.            [ 3],
    150.            [ 4]],
    151.           [[ 1],
    152.            [ 2],
    153.            [ 3],
    154.            [ 4]],
    155.           [[ 1],
    156.            [ 2],
    157.            [ 3],
    158.            [ 4]]],
    159.          [[[ 5],
    160.            [ 6],
    161.            [ 7],
    162.            [ 8]],
    163.           [[ 5],
    164.            [ 6],
    165.            [ 7],
    166.            [ 8]],
    167.           [[ 5],
    168.            [ 6],
    169.            [ 7],
    170.            [ 8]],
    171.           [[ 5],
    172.            [ 6],
    173.            [ 7],
    174.            [ 8]],
    175.           [[ 5],
    176.            [ 6],
    177.            [ 7],
    178.            [ 8]]],
    179.          [[[ 9],
    180.            [10],
    181.            [11],
    182.            [12]],
    183.           [[ 9],
    184.            [10],
    185.            [11],
    186.            [12]],
    187.           [[ 9],
    188.            [10],
    189.            [11],
    190.            [12]],
    191.           [[ 9],
    192.            [10],
    193.            [11],
    194.            [12]],
    195.           [[ 9],
    196.            [10],
    197.            [11],
    198.            [12]]]],
    199.         [[[[ 1],
    200.            [ 2],
    201.            [ 3],
    202.            [ 4]],
    203.           [[ 1],
    204.            [ 2],
    205.            [ 3],
    206.            [ 4]],
    207.           [[ 1],
    208.            [ 2],
    209.            [ 3],
    210.            [ 4]],
    211.           [[ 1],
    212.            [ 2],
    213.            [ 3],
    214.            [ 4]],
    215.           [[ 1],
    216.            [ 2],
    217.            [ 3],
    218.            [ 4]]],
    219.          [[[ 5],
    220.            [ 6],
    221.            [ 7],
    222.            [ 8]],
    223.           [[ 5],
    224.            [ 6],
    225.            [ 7],
    226.            [ 8]],
    227.           [[ 5],
    228.            [ 6],
    229.            [ 7],
    230.            [ 8]],
    231.           [[ 5],
    232.            [ 6],
    233.            [ 7],
    234.            [ 8]],
    235.           [[ 5],
    236.            [ 6],
    237.            [ 7],
    238.            [ 8]]],
    239.          [[[ 9],
    240.            [10],
    241.            [11],
    242.            [12]],
    243.           [[ 9],
    244.            [10],
    245.            [11],
    246.            [12]],
    247.           [[ 9],
    248.            [10],
    249.            [11],
    250.            [12]],
    251.           [[ 9],
    252.            [10],
    253.            [11],
    254.            [12]],
    255.           [[ 9],
    256.            [10],
    257.            [11],
    258.            [12]]]]])

    7 repeat函数

    根据指定维度复制张量，与 expand 不同的是，该方法会拷贝原张量的数据

    Tensor.repeat( *sizes) → [Tensor]

    参数：

    • sizes (torch.Size or [int] – 指定维度复制的次数 
    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. print(a1.storage().data_ptr())
    5. a2 = a1.reshape(3, 1, 4)
    6. print(a2.storage().data_ptr())
    7. a3 = a2.expand(3, 3, -1)
    8.  
    9. # expand 操作后，张量的内存地址没变
    10. print(a3.storage().data_ptr())
    11.  
    12. a4 = a2.repeat(2, 4, 1)
    13.  
    14. # repeat 操作后，张量的内存地址会改变
    15. print(a4.storage().data_ptr())
    16.  
    17. print(a1.shape)
    18. print(a1)
    19. print(a2.shape)
    20. print(a2)
    21. print(a3.shape)
    22. print(a3)
    23. print(a4.shape)

     运行结果显示如下：

    1. 1974461518528
    2. 1974461518528
    3. 1974461518528
    4. 1974462302208
    5. torch.Size([12])
    6. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    7. torch.Size([3, 1, 4])
    8. tensor([[[ 1,  2,  3,  4]],
    9.         [[ 5,  6,  7,  8]],
    10.         [[ 9, 10, 11, 12]]])
    11. torch.Size([3, 3, 4])
    12. tensor([[[ 1,  2,  3,  4],
    13.          [ 1,  2,  3,  4],
    14.          [ 1,  2,  3,  4]],
    15.         [[ 5,  6,  7,  8],
    16.          [ 5,  6,  7,  8],
    17.          [ 5,  6,  7,  8]],
    18.         [[ 9, 10, 11, 12],
    19.          [ 9, 10, 11, 12],
    20.          [ 9, 10, 11, 12]]])
    21. torch.Size([6, 4, 4])
    22. tensor([[[ 1,  2,  3,  4],
    23.          [ 1,  2,  3,  4],
    24.          [ 1,  2,  3,  4],
    25.          [ 1,  2,  3,  4]],
    26.         [[ 5,  6,  7,  8],
    27.          [ 5,  6,  7,  8],
    28.          [ 5,  6,  7,  8],
    29.          [ 5,  6,  7,  8]],
    30.         [[ 9, 10, 11, 12],
    31.          [ 9, 10, 11, 12],
    32.          [ 9, 10, 11, 12],
    33.          [ 9, 10, 11, 12]],
    34.         [[ 1,  2,  3,  4],
    35.          [ 1,  2,  3,  4],
    36.          [ 1,  2,  3,  4],
    37.          [ 1,  2,  3,  4]],
    38.         [[ 5,  6,  7,  8],
    39.          [ 5,  6,  7,  8],
    40.          [ 5,  6,  7,  8],
    41.          [ 5,  6,  7,  8]],
    42.         [[ 9, 10, 11, 12],
    43.          [ 9, 10, 11, 12],
    44.          [ 9, 10, 11, 12],
    45.          [ 9, 10, 11, 12]]])

    8 permute函数

    返回重新排列的张量

    torch.permute(input, dims) → [Tensor]

     也可以在tensor上直接使用permute，形式如下： 

    tensor.permute(dims) → [Tensor]

    参数：

    • input ([Tensor] 要重新排列的张量
    • dims (tuple of python:int) 需要重排的维度索引数组

     

    1. import torch
    2.  
    3. a1 = torch.tensor([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12])
    4. a2 = a1.reshape(3, 1, 4)
    5. a3 = torch.permute(a2, (2, 0, 1))
    6. a4 = torch.permute(a2, (1, 0, 2))
    7. a5 = a2.permute(1, 2, 0)
    8.  
    9. print(a1.shape)
    10. print(a1)
    11. print(a2.shape)
    12. print(a2)
    13. print(a3.shape)
    14. print(a3)
    15. print(a4.shape)
    16. print(a4)
    17. print(a5.shape)
    18. print(a5)

    运行结果显示如下：

    1. torch.Size([12])
    2. tensor([ 1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12])
    3. torch.Size([3, 1, 4])
    4. tensor([[[ 1,  2,  3,  4]],
    5.         [[ 5,  6,  7,  8]],
    6.         [[ 9, 10, 11, 12]]])
    7. torch.Size([4, 3, 1])
    8. tensor([[[ 1],
    9.          [ 5],
    10.          [ 9]],
    11.         [[ 2],
    12.          [ 6],
    13.          [10]],
    14.         [[ 3],
    15.          [ 7],
    16.          [11]],
    17.         [[ 4],
    18.          [ 8],
    19.          [12]]])
    20. torch.Size([1, 3, 4])
    21. tensor([[[ 1,  2,  3,  4],
    22.          [ 5,  6,  7,  8],
    23.          [ 9, 10, 11, 12]]])
    24. torch.Size([1, 4, 3])
    25. tensor([[[ 1,  5,  9],
    26.          [ 2,  6, 10],
    27.          [ 3,  7, 11],
    28.          [ 4,  8, 12]]])

     

     

  • 相关阅读:
    发布订阅模式
    教你1分钟搞定2小时字幕
    解决：使用WileyNJDv5_Template模板时，无法生成pdf文件。
    UOS服务器操作系统搭建离线yum仓库
    迈向数字化发展新阶段，某商业银行数据存储创新方案及实践经验
    Java随笔-反射
    Selenium 模拟浏览器操作案例
    多线程的创建及状态描述
    pytorch的searchsorted解释
    【一行记录】达梦timestamp转yyyy-mm-dd
  • 原文地址：https://blog.csdn.net/lsb2002/article/details/132905346