Torch. correlation function

torch.nonzero()

Docstring:nonzero(input, *, out=None, as_tuple=False) -> LongTensor or tuple of LongTensors
The parameters involved are mainly input and as_tuple, In fact, it is whether to return to a tuple , Default is not tuple .
Just look at the following example .

"""as_tuple=False By default """
>>> torch.nonzero(torch.tensor([1, 1, 1, 0, 1]))
    tensor([[ 0],
            [ 1],
            [ 2],
            [ 4]])     #  Returns yes or no 0 Index of elements , And is the matrix of the index ,shape= （ Non zero elements ,1）
            
>>> torch.nonzero(torch.tensor([[0.6, 0.0, 0.5, 0.0],
                                [0.0, 0.4, 0.0, 0.0],
                                [0.0, 0.0, 1.2, 0.0],
                                [0.0, 0.0, 0.0,-0.4]]))
    tensor([[ 0,  0],
    	    [ 0,  2],
            [ 1,  1],
            [ 2,  2],
            [ 3,  3]])     # Returns the index of non-zero elements in the matrix , Each element contains a row and column index ,shape= （ Non zero elements ,2）
  

"""as_tuple=True By default """
 >>> torch.nonzero(torch.tensor(5), as_tuple=True)
    (tensor([0]),)        #  Return is also a tensor index , Just put it in a tuple 

    >>> torch.nonzero(torch.tensor([1, 1, 1, 0, 1]), as_tuple=True)
    (tensor([0, 1, 2, 4]),)    # No difference , But to extract an element in the list , Must have two index values , They are a whole , For example, to extract 1,[0][1]
 >>> torch.nonzero(torch.tensor([[0.6, 0.0, 0.5, 0.0],
                                [0.0, 0.4, 0.0, 0.0],
                                [0.0, 0.0, 1.2, 0.0],
                                [0.0, 0.0, 0.0,-0.4]]), as_tuple=True)
                  # What we return are two list tensors   constitute   A tuple of , Tuples [0] Is the row position of non-zero elements in turn , Tuples [1] Is the column position of non-zero elements in turn , 
    (tensor([0, 0, 1, 2, 3]), tensor([0, 2, 1, 2, 3]))  

summary ： Actually as_tuple=True Returning tuples is useless , If you need the row position or column position of non-zero elements, you can set it , In fact, it is as_tuple=False The first and second columns returned .

X.permute(dims)

take tensor Dimensional transposition of
https://zhuanlan.zhihu.com/p/76583143

torch.flatten(t,start_dim= Start dimension ,end_dim= End dimension )

Reduce some dimensions of the matrix , But elements are not lost

t = torch.arange(120,).reshape(2,3,4,5)

t

t.shape

torch.flatten(t,start_dim=1,end_dim=-2)

torch.flatten(t,start_dim=1,end_dim=-2).shape