如何数四维矩阵

 

如何数四维矩阵，如下是一个四位矩阵，形状为（4,3,2,5）

tensor([[[[-0.3037,  0.5017,  0.2698,  0.7181,  0.7317],
          [ 0.2041, -2.1096,  2.0244, -2.0167, -0.4281]],

         [[ 0.0483,  0.0486, -0.9498, -0.1509, -0.9814],
          [-0.0850, -0.8587,  1.4044,  0.6037, -0.9757]],

         [[ 0.8890, -0.4697, -0.8317, -0.8372,  1.5004],
          [-0.4455,  0.9579, -2.2947,  1.9205,  0.3391]]],

        [[[-0.0613, -0.3637,  2.9507,  1.0946,  1.9070],
          [ 1.9137, -0.8911, -0.3479,  0.9942, -1.7937]],

         [[-1.1416, -1.5116, -0.6588,  1.0701, -0.3426],
          [-1.3269,  0.5529, -0.6450,  0.0393,  1.5098]],

         [[ 1.4184, -0.7669, -1.0354, -0.7632,  0.4466],
          [-1.9081, -0.1087,  0.0065, -0.3233, -0.7373]]],

        [[[-0.8933, -1.6219,  1.0330, -1.1950, -1.0639],
          [ 0.7363,  0.4656, -0.8343,  1.3202,  0.7524]],

         [[-0.4619, -0.4536,  0.6632,  0.2840,  1.1670],
          [-0.4915, -0.8160, -1.3393, -1.4767, -0.1640]],

         [[-1.4513, -0.3191,  0.2184,  1.7790, -1.0141],
          [-1.8744, -1.0015, -0.8923,  1.4281, -0.3708]]],

        [[[-0.4469,  0.7219, -0.7110,  0.3740,  0.0478],
          [-0.4859,  0.3819, -0.4086,  1.0739,  0.0245]],

         [[ 0.7192,  2.0502,  0.0091, -1.4356,  1.1417],
          [ 0.5979,  0.4916,  0.3360,  0.3793, -0.5015]],

         [[-0.5521,  0.0183, -0.0687, -0.5918,  0.1760],
          [-1.7947, -0.9572, -0.3511, -0.0038,  1.1138]]]])

最里面的是2*5的矩阵，形状为（2,5）

[[-0.3037,  0.5017,  0.2698,  0.7181,  0.7317],
 [ 0.2041, -2.1096,  2.0244, -2.0167, -0.4281]]

在其次是有3个2*5的矩阵，形状为（3,2,5）

[[[-0.3037,  0.5017,  0.2698,  0.7181,  0.7317],
  [ 0.2041, -2.1096,  2.0244, -2.0167, -0.4281]],

  [[ 0.0483,  0.0486, -0.9498, -0.1509, -0.9814],
   [-0.0850, -0.8587,  1.4044,  0.6037, -0.9757]],

  [[ 0.8890, -0.4697, -0.8317, -0.8372,  1.5004],
   [-0.4455,  0.9579, -2.2947,  1.9205,  0.3391]]]

最后有4个（3,2,5），最终形状为（4,3,2,5）

为了更清晰地知道如何应用四维矩阵，每个矩阵的含义，以音频为例，假设输入的是经过预处理后的音频形状为（B,C,T,F）。

索引DC = d[:,0,:,:]

结果：

tensor([[[ 0.1216,  0.7147, -0.5788,  0.2586, -0.6516],
         [-2.0123, -0.5601,  2.0672,  0.7074, -1.2746]],

        [[-0.1756,  0.1880, -1.5323, -0.0562, -2.0827],
         [ 0.4644, -0.5828,  0.0791,  0.2199,  0.9643]],

        [[-0.3476,  1.2914, -1.0125, -1.8291, -2.1911],
         [ 0.2787,  1.5103,  0.9466,  1.1055,  0.7242]],

        [[ 1.5504, -0.1351, -0.2468, -0.7308,  1.2519],
         [ 0.6720, -1.0799,  0.2046,  0.3113, -1.6933]]])

含义是取第一个通道即音频的实部，得到的是三维的数据，规律是被索引的那一维变成具体的数据，其他维数不变。