pytorch中Tensor各种操作

た入场券 2024-04-17 05:22 68阅读 0赞

import torch

### pytorch中Tensor初始化 ###

#返回全零Tensor
    a = torch.zeros(2,3); print(a)
    
    #返回shape和参数一样的全1Tensor, zeros_like类似
    b = torch.ones_like(a); print(b)
    
    #torch.arange(start=0, end, step=1) 
    c = torch.arange(0, 10, 1); print(c)
    
    #初始化指定值的Tensor
    e = torch.full((2,3), 2); print(e)
    
    
    #************随机初始化***********
    #初始化为[0,1)内的均匀分布随机数
    a = torch.rand((2,3)); print(a)
    
    #初始化为[0,1)内的均匀分布随机数，不过shape与参数相同
    b = torch.rand_like(a); print(b)
    
    #返回标准正态分布(0,1)的随机数
    torch.randn(2,3)
    
    #torch.normal(mean, std, out=None)
    c = torch.normal(torch.randn(2,3), torch.randn(2,3)); print(c)

tensor([[0., 0., 0.],
            [0., 0., 0.]])
    tensor([[1., 1., 1.],
            [1., 1., 1.]])
    tensor([0, 1, 2, 3, 4, 5, 6, 7, 8, 9])
    tensor([[2., 2., 2.],
            [2., 2., 2.]])
    tensor([[0.5680, 0.8983, 0.3664],
            [0.1633, 0.9800, 0.8474]])
    tensor([[0.5968, 0.0573, 0.2997],
            [0.7349, 0.0017, 0.4628]])
    tensor([[ 0.5925,  1.8533, -1.8251],
            [-1.9088, -0.1117,  0.6028]])

### pytorch中contiguous ###

调用view之前最好先contiguous  
  
x.contiguous().view()  
  
因为view需要tensor的内存是整块的

x = torch.ones(10, 10)
    print( x.is_contiguous())  # True
    print (x.transpose(0, 1).is_contiguous())  # False
    print (x.transpose(0, 1).contiguous().is_contiguous())

True
    False
    True

### pytorch中transpose ###

transpose就是转置，只能操作两位坐标轴转换

x = torch.randn(2, 3, 4)
    x1 = x.transpose(0, 1)
    x2 = x.transpose(1, 0)
    # x1, x2相同
    x, x1, x2

(tensor([[[ 0.4893, -0.0905, -1.5506,  1.1095],
              [-1.0713, -0.4694, -0.6739, -0.9530],
              [ 1.6341,  0.3761, -1.6181,  0.7335]],
     
             [[-0.4032, -0.3206,  0.1094,  2.7867],
              [-0.1956,  0.3200, -1.3562,  0.7300],
              [-0.3488,  0.5682,  0.8729, -0.4489]]]),
     tensor([[[ 0.4893, -0.0905, -1.5506,  1.1095],
              [-0.4032, -0.3206,  0.1094,  2.7867]],
     
             [[-1.0713, -0.4694, -0.6739, -0.9530],
              [-0.1956,  0.3200, -1.3562,  0.7300]],
     
             [[ 1.6341,  0.3761, -1.6181,  0.7335],
              [-0.3488,  0.5682,  0.8729, -0.4489]]]),
     tensor([[[ 0.4893, -0.0905, -1.5506,  1.1095],
              [-0.4032, -0.3206,  0.1094,  2.7867]],
     
             [[-1.0713, -0.4694, -0.6739, -0.9530],
              [-0.1956,  0.3200, -1.3562,  0.7300]],
     
             [[ 1.6341,  0.3761, -1.6181,  0.7335],
              [-0.3488,  0.5682,  0.8729, -0.4489]]]))

### pytorch中cat ###

cat的第一个参数必须是**tuple of Tensors**, 第二个参数指定连接的方式，通俗来说是**横着**还是**竖着**，默认是横着

a = torch.randn(1,3)
    b = torch.randn(1,3)
    c = torch.cat((a,b))
    print(a.shape); print(b.shape);print(c.shape)
    print('\n')
    a = torch.randn(2,3)
    b = torch.randn(2,1)
    c = torch.cat((a,b), 1)
    print(a.shape); print(b.shape); print(c.shape)

torch.Size([1, 3])
    torch.Size([1, 3])
    torch.Size([2, 3])
    
    
    torch.Size([2, 3])
    torch.Size([2, 1])
    torch.Size([2, 4])

### pytorch中stack ###

与cat操作不同的是stack操作过程中会增加新的维度，具体看例子解释  
  
**All tensors need to be of the same size.**

a = torch.randn(2, 3)
    b = torch.randn(2,3)
    c = torch.stack([a,b], 2)
    print(a.shape); print(b.shape); print(c.shape)
    print(a); print(b); print(c)

torch.Size([2, 3])
    torch.Size([2, 3])
    torch.Size([2, 3, 2])
    tensor([[ 0.1350, -0.6451,  0.4002],
            [ 1.2582,  0.5484,  0.9806]])
    tensor([[-2.3363, -0.8511, -1.5627],
            [-0.3785,  0.2618, -1.0212]])
    tensor([[[ 0.1350, -2.3363],
             [-0.6451, -0.8511],
             [ 0.4002, -1.5627]],
    
            [[ 1.2582, -0.3785],
             [ 0.5484,  0.2618],
             [ 0.9806, -1.0212]]])