Got NAN value in training process by the same command in this code

Sorry to bother. I download Undistorted_megadepth datasets, and used this command "python -m pixloc.pixlib.train pixloc_megadepth_reproduce --conf pixloc/pixlib/configs/train_pixloc_megadepth.yaml". When in process, I got NAN as follows:

Evaluation: 97%|#########7| 99/102 [00:21<00:00, 5.18it/s]

Evaluation: 98%|#########8| 100/102 [00:21<00:00, 5.18it/s]

Evaluation: 99%|#########9| 101/102 [00:22<00:00, 4.92it/s]

Evaluation: 100%|##########| 102/102 [00:22<00:00, 4.60it/s]

[11/26/2021 01:22:48 pixloc INFO] [Validation] {R_error/0 3.900E+00, t_error/0 1.549E-01, R_error/1 4.264E+00, t_error/1 1.560E-01, R_error/2 3.748E+00, t_error/2 1.468E-01, R_error 3.748E+00, R_error_median 4.480E+00, t_error 1.468E-01, t_error_median 1.756E-01, R_error/init 1.718E+00, t_error/init 1.331E-01, loss/total 1.939E+01, loss/reprojection_error/0 3.426E+01, loss/reprojection_error/1 3.524E+01, loss/reprojection_error/2 2.786E+01, loss/reprojection_error 2.786E+01, loss/reprojection_error_median 3.192E+01, loss/reprojection_error/init 1.634E+01, loss/reprojection_error/init_median 1.069E+01}

[11/26/2021 01:23:24 pixloc.pixlib.models.two_view_refiner WARNING] Few points in batch [('0035', 472, 741)]

[11/26/2021 01:23:31 pixloc INFO] [E 3 | it 550] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 1.343E+01}

[11/26/2021 01:24:15 pixloc INFO] [E 3 | it 600] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 1.996E+01}

[11/26/2021 01:24:28 pixloc.pixlib.models.two_view_refiner WARNING] Few points in batch [('1017', 962, 1173)]

[11/26/2021 01:24:58 pixloc INFO] [E 3 | it 650] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 2.845E+01}

[11/26/2021 01:25:41 pixloc INFO] [E 3 | it 700] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 1.092E+01}

[11/26/2021 01:26:13 pixloc.pixlib.models.two_view_refiner WARNING] Few points in batch [('0348', 167, 168)]

[11/26/2021 01:26:24 pixloc INFO] [E 3 | it 750] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 8.301E+00}

[11/26/2021 01:27:08 pixloc INFO] [E 3 | it 800] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 2.569E+01}

[11/26/2021 01:27:19 pixloc.pixlib.models.two_view_refiner WARNING] Few points in batch [('5009', 168, 165)]

[11/26/2021 01:27:51 pixloc INFO] [E 3 | it 850] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 2.809E+01}

[11/26/2021 01:28:23 pixloc.pixlib.models.two_view_refiner WARNING] Few points in batch [('5005', 140, 136)]

[11/26/2021 01:28:34 pixloc INFO] [E 3 | it 900] loss {total NAN, reprojection_error/0 NAN, reprojection_error/1 NAN, reprojection_error/2 NAN, reprojection_error NAN, reprojection_error/init 8.891E+00}

It occurred in epoch 3, but when I try another the same command, it occurred in epoch 1 iteration 1180.

I just want to reproduce your results and haven't change any parameters, just the same as you post. I wondered is there something I ignored or something going wrong? My pytorch version is 1.7.1, Numpy 1.21.2.

Thank you so much.

cvg / pixloc

Got NAN value in training process by the same command in this code #21