每天讲解一点PyTorch 17Spatial Affinity代码实现分析

Posted 2023-02-27 cv.exp

>>> import torch
>>> import torch.nn as nn
>>> fm_in = torch.randn(1,3,2,3)
>>> fm_in
tensor([[[[-0.1291, -0.0966,  0.0632],
          [-0.1640, -0.2832,  1.0553]],

         [[ 1.2854,  0.3400,  1.6823],
          [ 0.1555, -1.2052,  0.1288]],

         [[ 0.5609,  0.3766,  0.7720],
          [-2.0410,  0.2177,  1.4301]]]])
>>> 
>>> fm_in.shape
torch.Size([1, 3, 2, 3])
>>> 
>>> fm_in = fm_in.view(fm_in.size(0), fm_in.size(1), -1)
>>> fm_in
tensor([[[-0.1291, -0.0966,  0.0632, -0.1640, -0.2832,  1.0553],
         [ 1.2854,  0.3400,  1.6823,  0.1555, -1.2052,  0.1288],
         [ 0.5609,  0.3766,  0.7720, -2.0410,  0.2177,  1.4301]]])
>>> fm_in.shape
torch.Size([1, 3, 6])
>>> 
>>> pow_out = torch.pow(fm_in,2)
>>> pow_out
tensor([[[1.6674e-02, 9.3340e-03, 3.9945e-03, 2.6887e-02, 8.0227e-02,
          1.1137e+00],
         [1.6523e+00, 1.1559e-01, 2.8300e+00, 2.4166e-02, 1.4526e+00,
          1.6588e-02],
         [3.1456e-01, 1.4179e-01, 5.9594e-01, 4.1658e+00, 4.7385e-02,
          2.0452e+00]]])
>>> pow_out.shape
torch.Size([1, 3, 6])
>>> 
>>> sum_out = torch.sum(pow_out, 1)
>>> sum_out
tensor([[1.9836, 0.2667, 3.4300, 4.2168, 1.5802, 3.1755]])
>>> sum_out.shape
torch.Size([1, 6])
>>> 
>>> sqrt_out = torch.sqrt(sum_out)
>>> sqrt_out
tensor([[1.4084, 0.5164, 1.8520, 2.0535, 1.2571, 1.7820]])
>>> sqrt_out.shape
torch.Size([1, 6])
>>> 
>>> unsqueeze_out = sqrt_out.unsqueeze(1)
>>> unsqueeze_out
tensor([[[1.4084, 0.5164, 1.8520, 2.0535, 1.2571, 1.7820]]])
>>> unsqueeze_out.shape
torch.Size([1, 1, 6])
>>> 
>>> expand_out = unsqueeze_out.expand(fm_in.shape)
>>> expand_out
tensor([[[1.4084, 0.5164, 1.8520, 2.0535, 1.2571, 1.7820],
         [1.4084, 0.5164, 1.8520, 2.0535, 1.2571, 1.7820],
         [1.4084, 0.5164, 1.8520, 2.0535, 1.2571, 1.7820]]])
>>> 
>>> fm_in.shape
torch.Size([1, 3, 6])
>>> expand_out.shape
torch.Size([1, 3, 6])
>>> 
>>> norm_fm_in = fm_in / (expand_out + 0.0000001)
>>> norm_fm_in
tensor([[[-0.0917, -0.1871,  0.0341, -0.0799, -0.2253,  0.5922],
         [ 0.9127,  0.6583,  0.9083,  0.0757, -0.9588,  0.0723],
         [ 0.3982,  0.7291,  0.4168, -0.9939,  0.1732,  0.8025]]])
>>> fm_in.shape
torch.Size([1, 3, 6])
>>> norm_fm_in.shape
torch.Size([1, 3, 6])
>>> 
>>> transpose_out = norm_fm_in.transpose(1,2)
>>> transpose_out
tensor([[[-0.0917,  0.9127,  0.3982],
         [-0.1871,  0.6583,  0.7291],
         [ 0.0341,  0.9083,  0.4168],
         [-0.0799,  0.0757, -0.9939],
         [-0.2253, -0.9588,  0.1732],
         [ 0.5922,  0.0723,  0.8025]]])
>>> transpose_out.shape
torch.Size([1, 6, 3])
>>> 
>>> fm_out = transpose_out.bmm(norm_fm_in)
>>> 
>>> fm_out
tensor([[[ 1.0000,  0.9083,  0.9919, -0.3194, -0.7854,  0.3313],
         [ 0.9083,  1.0000,  0.8955, -0.6599, -0.4628,  0.5219],
         [ 0.9919,  0.8955,  1.0000, -0.3483, -0.8064,  0.4204],
         [-0.3194, -0.6599, -0.3483,  1.0000, -0.2267, -0.8395],
         [-0.7854, -0.4628, -0.8064, -0.2267,  1.0000, -0.0638],
         [ 0.3313,  0.5219,  0.4204, -0.8395, -0.0638,  1.0000]]])
>>> fm_out.shape
torch.Size([1, 6, 6])
>>> 
>>> fm_out = fm_out.unsqueeze(1)
>>> fm_out.shape
torch.Size([1, 1, 6, 6])
>>> 
>>> fm_out
tensor([[[[ 1.0000,  0.9083,  0.9919, -0.3194, -0.7854,  0.3313],
          [ 0.9083,  1.0000,  0.8955, -0.6599, -0.4628,  0.5219],
          [ 0.9919,  0.8955,  1.0000, -0.3483, -0.8064,  0.4204],
          [-0.3194, -0.6599, -0.3483,  1.0000, -0.2267, -0.8395],
          [-0.7854, -0.4628, -0.8064, -0.2267,  1.0000, -0.0638],
          [ 0.3313,  0.5219,  0.4204, -0.8395, -0.0638,  1.0000]]]])
>>>