如何理解 "torch.randn()" size* 参数参数？

Question

据我了解，torch.randn(layers/depth, rows, columns)，执行时可以看到：torch.randn(2, 3, 3) ==> 2层（3x3）矩阵：

tensor([[[ 1.4838,  1.2926,  1.6147],
     [ 0.7923,  0.6414, -0.2676],
     [-0.1949,  0.3859, -0.6940]],

    [[ 0.2454, -1.9215, -0.3078],
     [ 0.8544,  0.9726,  0.0330],
     [ 0.3579,  0.8247,  2.1288]]])

但是在 size* 参数中添加一个额外的项意味着什么？如：torch.randn(2, 1, 3, 3)

tensor([[[[ 0.6206, -1.3697, -0.2267],
      [ 1.0511,  2.3375, -0.9598],
      [-0.8148, -0.0911, -2.1211]]],


    [[[ 0.0659,  1.0764,  0.6150],
      [-1.7226,  0.5038, -0.9544],
      [-0.6447, -0.3325,  0.2048]]]])

“1”添加到创建的张量中是什么？

Answer 1

简而言之，您只是添加了一个包含单个元素的维度。注意额外的大括号。

也许重写输出会帮助您理解发生了什么：

torch.randn(2, 3, 3)
> tensor(
tensor(
[    # 1st dim - 1st element
  [  # 2nd dim - 1st element
    [ 1.4838,     # 3rd dim - 1st element
      1.2926,     # 3rd dim - 2nd element
      1.6147],    # 3rd dim - 3rd element
    [ 0.7923,  0.6414, -0.2676], # 2nd dim - 2nd element ( 3rd dim ... )
    [-0.1949,  0.3859, -0.6940]  # 2nd dim - 3rd element ( 3rd dim ... )
  ], #1st dim - 2nd element
  [  ## .. same as above
    [ 0.2454, -1.9215, -0.3078],
    [ 0.8544,  0.9726,  0.0330],
    [ 0.3579,  0.8247,  2.1288]
  ]
])

torch.randn(2, 1, 3, 3)
> 
tensor(
[     # 1st dim - 1st element
  [   # 2nd dim - 1st element
    [ # 3rd dim - 1st element
      [ 0.6206,  # 4th dim - 1st element
        -1.3697, # 4th dim - 2nd element
        -0.2267],# 4th dim - 3rd element
      [ 1.0511,  2.3375, -0.9598], # 3rd dim - 2nd element
      [-0.8148, -0.0911, -2.1211]  # 3rd dim - 3rd element
    ]
     # 2nd dim - nothing more ...
  ], # 1st dim - 2nd element
  [  # .. same as above
    [  
      [ 0.0659,  1.0764,  0.6150],
      [-1.7226,  0.5038, -0.9544],
      [-0.6447, -0.3325,  0.2048]
    ]
  ]
])

Answer 2

你引入的每个数字都是指矩阵的一个维度。人类很难想象超过 3 个维度，但计算机可以。

在这种特殊情况下，您可以将额外维度视为类似于批量大小的东西。

如何理解 "torch.randn()" size* 参数参数？

How to understand "torch.randn()" size* parameter arguments?

python

python-3.x

torch

pytorch