pytorch多维张量相乘和广播机制示例

示例：

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

import torch box = torch.tensor([[[0.1000, 0.2000, 0.5000, 0.3000],          [0.6000, 0.6000, 0.9000, 0.9000],          [0.1000, 0.1000, 0.2000, 0.2000]],           [[0.1000, 0.2000, 0.5000, 0.3000],          [0.6000, 0.6000, 0.9000, 0.9000],          [0.1000, 0.1000, 0.2000, 0.2000]]]).to(torch.float32)   wh = torch.tensor([[[200.],          [400.],          [200.],          [400.]],           [[200.],          [400.],          [200.],          [400.]]]).to(torch.float32)   print(box.shape)  # (2, 3 ,4) print(wh.shape)  # (2, 4, 1)   result = box @ wh print(result.shape)  # (2, 3, 1) print(result) # tensor([[[320.], #          [900.], #          [180.]],   #         [[320.], #          [900.], #          [180.]]])

　　

下面这个示例用到了广播机制：

?

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27

import torch box = torch.tensor([[[0.1000, 0.2000, 0.5000, 0.3000],          [0.6000, 0.6000, 0.9000, 0.9000],          [0.1000, 0.1000, 0.2000, 0.2000]],           [[0.1000, 0.2000, 0.5000, 0.3000],          [0.6000, 0.6000, 0.9000, 0.9000],          [0.1000, 0.1000, 0.2000, 0.2000]]]).to(torch.float32)   wh = torch.tensor([[[200.],          [400.],          [200.],          [400.]]]).to(torch.float32)   print(box.shape)  # (2, 3 ,4) print(wh.shape)  # (1, 4, 1)   注意这个wh的第0维度的大小是1   result = box @ wh  # 这里在第0维度会使用广播机制 print(result.shape)  # (2, 3, 1) print(result) # tensor([[[320.], #          [900.], #          [180.]],   #         [[320.], #          [900.], #          [180.]]])