chensong1995
commented
10 months ago Hi Mona,
Thanks for your question! Can you set a breakpoint and take a look at the value of v?
Closed monajalal closed 9 months ago
chensong1995
commented
10 months ago Hi Mona,
Thanks for your question! Can you set a breakpoint and take a look at the value of v?
monajalal
commented
10 months ago It takes a really long time to get to that point but I will do it shortly. I will also create a separate related issue in the meanwhile.
monajalal
commented
10 months ago I am still working on providing you results but I also noticed this which also looks troublesome:
Epoch: [19][16/19] Time: 0.289 (0.258) Sym: 8.0534 (8.4049) Mask: 0.0785 (0.0957) Pts: 0.0528 (0.0496) Graph: 10.5922 (11.1342) Total: 2.4711 (2.5460)
Epoch: [19][17/19] Time: 0.215 (0.255) Sym: 8.4763 (8.4089) Mask: 0.1002 (0.0959) Pts: 0.0487 (0.0496) Graph: 11.7995 (11.1711) Total: 2.6145 (2.5498)
Epoch: [19][18/19] Time: 0.097 (0.247) Sym: 10.1351 (8.4645) Mask: 0.1134 (0.0965) Pts: 0.0475 (0.0495) Graph: 12.6410 (11.2185) Total: 2.8660 (2.5600)
Testing...
/home/mona/HybridPose/lib/ransac_voting_gpu_layer/ransac_voting_gpu.py:546: UserWarning: indexing with dtype torch.uint8 is now deprecated, please use a dtype torch.bool instead. (Triggered internally at ../aten/src/ATen/native/IndexingUtils.h:27.)
direct = vertex[bi].masked_select(torch.unsqueeze(torch.unsqueeze(cur_mask, 2), 3)) # [tn,vn,2]
/home/mona/HybridPose/./trainers/coretrainer.py:230: RuntimeWarning: Mean of empty slice.
edge_x = graph_pred[i_edge, 0][mask_pred == 1.].mean()
/home/mona/anaconda3/envs/hybridpose/lib/python3.10/site-packages/numpy/core/_methods.py:129: RuntimeWarning: invalid value encountered in divide
ret = ret.dtype.type(ret / rcount)
/home/mona/HybridPose/./trainers/coretrainer.py:231: RuntimeWarning: Mean of empty slice.
edge_y = graph_pred[i_edge, 1][mask_pred == 1.].mean()
/home/mona/HybridPose/./trainers/coretrainer.py:233: RuntimeWarning: invalid value encountered in cast
end = np.int16(np.round(pts2d_gt[start_idx] + edge))Epoch: [199][16/19] Time: 0.451 (0.312) Sym: 4.0442 (3.8591) Mask: 0.0055 (0.0088) Pts: 0.0150 (0.0162) Graph: 1.7577 (2.1512) Total: 0.7361 (0.7721)
Epoch: [199][17/19] Time: 0.275 (0.310) Sym: 4.7510 (3.9087) Mask: 0.0066 (0.0087) Pts: 0.0143 (0.0161) Graph: 2.4858 (2.1698) Total: 0.8730 (0.7777)
Epoch: [199][18/19] Time: 0.171 (0.303) Sym: 2.5082 (3.8635) Mask: 0.0070 (0.0086) Pts: 0.0145 (0.0161) Graph: 1.2050 (2.1387) Total: 0.5235 (0.7695)
Testing...
Loss: 0.6235
Successfully saved model into saved_weights/linemod/ape/checkpoints/0.001/199
value of v is: [6.1583250e+18 7.8760837e+09]
> /home/mona/HybridPose/trainers/coretrainer.py(411)fill_intermediate_predictions()
-> if np.any(v < 0):
(Pdb) v
array([6.1583250e+18, 7.8760837e+09], dtype=float32)
(Pdb) n
> /home/mona/HybridPose/trainers/coretrainer.py(415)fill_intermediate_predictions()
-> if np.any(np.isnan(v)) or np.any(np.isinf(v)):
(Pdb) n
> /home/mona/HybridPose/trainers/coretrainer.py(417)fill_intermediate_predictions()
-> v = np.matrix(np.diag(1. / np.sqrt(v)))
(Pdb) np.sqrt(v)
array([2.4815972e+09, 8.8747305e+04], dtype=float32)
(Pdb) v
array([6.1583250e+18, 7.8760837e+09], dtype=float32)
(Pdb) n
> /home/mona/HybridPose/trainers/coretrainer.py(418)fill_intermediate_predictions()
-> point_inv_half_var[i] = u * v * u.transpose()
(Pdb) v
matrix([[4.0296630e-10, 0.0000000e+00],
[0.0000000e+00, 1.1267948e-05]], dtype=float32)@chensong1995 is this helpful?
monajalal
commented
10 months ago my bad, here is the pdb.set_trace() for when if np.any(v < 0):
Epoch: [199][15/19] Time: 0.289 (0.295) Sym: 4.3407 (3.7749) Mask: 0.0074 (0.0089) Pts: 0.0170 (0.0160) Graph: 1.9972 (2.0208) Total: 0.8114 (0.7480)
Epoch: [199][16/19] Time: 0.448 (0.304) Sym: 3.9754 (3.7867) Mask: 0.0057 (0.0087) Pts: 0.0160 (0.0160) Graph: 1.5828 (1.9950) Total: 0.7215 (0.7465)
Epoch: [199][17/19] Time: 0.260 (0.302) Sym: 4.7488 (3.8401) Mask: 0.0072 (0.0086) Pts: 0.0143 (0.0159) Graph: 2.4331 (2.0194) Total: 0.8682 (0.7532)
Epoch: [199][18/19] Time: 0.172 (0.295) Sym: 2.4919 (3.7966) Mask: 0.0077 (0.0086) Pts: 0.0130 (0.0158) Graph: 1.3690 (1.9984) Total: 0.5236 (0.7458)
Testing...
Loss: 0.6320
Successfully saved model into saved_weights/linemod/ape/checkpoints/0.001/199
value of v is: [ 7.4060767e+18 -1.6713240e+11]
> /home/mona/HybridPose/trainers/coretrainer.py(413)fill_intermediate_predictions()
-> print("Covariance matrix has negative eigenvalues")
(Pdb) l
408 print('value of v is: ', v)
409
410 # Check eigenvalues
411 if np.any(v < 0):
412 pdb.set_trace()
413 -> print("Covariance matrix has negative eigenvalues")
414
415 # Check eigenvectors
416 if np.any(np.isnan(v)) or np.any(np.isinf(v)):
417 print("Eigenvectors have NaNs or Infs")
418 v = np.matrix(np.diag(1. / np.sqrt(v)))
(Pdb) n
Covariance matrix has negative eigenvalues
> /home/mona/HybridPose/trainers/coretrainer.py(416)fill_intermediate_predictions()
-> if np.any(np.isnan(v)) or np.any(np.isinf(v)):
(Pdb) n
> /home/mona/HybridPose/trainers/coretrainer.py(418)fill_intermediate_predictions()
-> v = np.matrix(np.diag(1. / np.sqrt(v)))
(Pdb) n
/home/mona/HybridPose/./trainers/coretrainer.py:418: RuntimeWarning: invalid value encountered in sqrt
v = np.matrix(np.diag(1. / np.sqrt(v)))
> /home/mona/HybridPose/trainers/coretrainer.py(419)fill_intermediate_predictions()
-> point_inv_half_var[i] = u * v * u.transpose()
(Pdb) v
matrix([[3.6745645e-10, 0.0000000e+00],
[0.0000000e+00, nan]], dtype=float32)as you see we end up with a nan since v has negative values.
When this happens:
(Pdb) cov
matrix([[6.5462130e+18, 2.3725201e+18],
[2.3725201e+18, 8.5986365e+17]], dtype=float32)and as you see there is no inf or nan values inside cov matrix.
With this information, would you please be able to provide a solution? Here the covariance matrix is not positive definite hence we have negative eigenvalues. Given the situation, I am not sure what can be done here?
chensong1995
commented
10 months ago Hi Mona,
Can you confirm you have downloaded the pre-trained weights as instructed by README and set the load_dir (here) to the downloaded path?
monajalal
commented
10 months ago This problem happens even with the newly and checked with md5sum weights unfortunately.
chensong1995
commented
10 months ago Hi Mona,
Thanks for your question! Can you run three iterations of trainer.test (here) and see if the visualizations look good to you?
Is this ok or should it be resolved?