Python torch.inverse返回单位矩阵。通过计算前打印输入解决错误（但为什么？）

inverse（）仅返回标识矩阵。（参见下面的非正常输出）。它在第一次迭代后重复出现

如果我尝试先打印任何

pose\u pre

，问题就会迎刃而解。（见下面的正常输出）

以下是代码的一部分：

---

#print（pose_pre）#我用from替换了
手电筒。这个问题被忽略了，但仍然想知道为什么会发生

tensor([[[ 8.7334e-01, -4.8659e-01,  2.2528e-02,  6.4885e+03],                                          | 0/90 [00:00<?, ?it/s]
         [ 4.8635e-01,  8.7363e-01,  1.5527e-02, -4.6202e+03],
         [-2.7237e-02, -2.6037e-03,  9.9963e-01, -6.3445e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]],

        [[ 8.7334e-01, -4.8659e-01,  2.2528e-02,  6.4885e+03],
         [ 4.8635e-01,  8.7363e-01,  1.5527e-02, -4.6202e+03],
         [-2.7237e-02, -2.6037e-03,  9.9963e-01, -6.3445e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]]],
       device='cuda:0'),)
saving hidden-state image hc_state_14_31_18_734196.png to ../../visual_result/lstm_hidden_states
epochs:   0%|                                                                      | 0/1 [00:01<?, ?it/s, loss=4.14, lr=0.0003](tensor([[[1., 0., 0., 0.],                                                         | 1/90 [00:01<01:49,  1.23s/it, total_it=1]
         [0., 1., 0., 0.],
         [0., 0., 1., 0.],
         [0., 0., 0., 1.]],

        [[1., 0., 0., 0.],
         [0., 1., 0., 0.],
         [0., 0., 1., 0.],
         [0., 0., 0., 1.]]], device='cuda:0'),)
saving hidden-state image hc_state_14_31_19_915028.png to ../../visual_result/lstm_hidden_states
epochs:   0%|                                                                    | 0/1 [00:02<?, ?it/s, loss=3.45, lr=0.000305](tensor([[[1., 0., 0., 0.],                                                         | 2/90 [00:02<01:34,  1.07s/it, total_it=2]
         [0., 1., 0., 0.],
         [0., 0., 1., 0.],
         [0., 0., 0., 1.]],

        [[1., 0., 0., 0.],
         [0., 1., 0., 0.],
         [0., 0., 1., 0.],
         [0., 0., 0., 1.]]], device='cuda:0'),)


(tensor([[[ 8.7334e-01, -4.8659e-01,  2.2528e-02,  6.4885e+03],
         [ 4.8635e-01,  8.7363e-01,  1.5527e-02, -4.6202e+03],
         [-2.7237e-02, -2.6037e-03,  9.9963e-01, -6.3445e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]],

        [[ 8.7334e-01, -4.8659e-01,  2.2528e-02,  6.4885e+03],
         [ 4.8635e-01,  8.7363e-01,  1.5527e-02, -4.6202e+03],
         [-2.7237e-02, -2.6037e-03,  9.9963e-01, -6.3445e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]]],
       device='cuda:0'),)
saving hidden-state image hc_state_14_32_08_581169.png to ../../visual_result/lstm_hidden_states
epochs:   0%|                                                                      | 0/1 [00:01<?, ?it/s, loss=3.94, lr=0.0003]tensor([ 8.7283e-01,  4.8730e-01, -2.6772e-02, -3.4203e+03], device='cuda:0')       | 1/90 [00:01<01:47,  1.20s/it, total_it=1]
(tensor([[[ 8.7283e-01, -4.8751e-01,  2.2528e-02,  6.4922e+03],
         [ 4.8730e-01,  8.7312e-01,  1.4569e-02, -4.6133e+03],
         [-2.6772e-02, -1.7379e-03,  9.9964e-01, -6.8084e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]],

        [[ 8.7235e-01, -4.8834e-01,  2.3104e-02,  6.4954e+03],
         [ 4.8815e-01,  8.7265e-01,  1.3706e-02, -4.6071e+03],
         [-2.6854e-02, -6.7833e-04,  9.9964e-01, -7.5985e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]]],
       device='cuda:0'),)
saving hidden-state image hc_state_14_32_09_748923.png to ../../visual_result/lstm_hidden_states
epochs:   0%|                                                                    | 0/1 [00:02<?, ?it/s, loss=3.63, lr=0.000305]tensor([ 8.7182e-01,  4.8909e-01, -2.6878e-02, -3.4184e+03], device='cuda:0')       | 2/90 [00:02<01:32,  1.05s/it, total_it=2]
(tensor([[[ 8.7182e-01, -4.8925e-01,  2.3774e-02,  6.4990e+03],
         [ 4.8909e-01,  8.7214e-01,  1.2556e-02, -4.6002e+03],
         [-2.6878e-02,  6.8170e-04,  9.9964e-01, -8.5845e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]],

        [[ 8.7129e-01, -4.9016e-01,  2.4445e-02,  6.5026e+03],
         [ 4.9002e-01,  8.7163e-01,  1.1885e-02, -4.5933e+03],
         [-2.7133e-02,  1.6239e-03,  9.9963e-01, -9.3492e+01],
         [ 0.0000e+00,  0.0000e+00,  0.0000e+00,  1.0000e+00]]],
       device='cuda:0'),)