metrics.py points cannot contain NaN

[melinawdh]

Hi,

Thanks for the great work!

I set everything up and was also able to run the network training. However, when I try to test the networks and calculate the metrics, I always get the error, that there are points with NaN values - this is the case for inferences on a model I trained myself with the provided code and data, as well as for the model that was provided (PointNet_Model).

Upon inspection it seems like the gt_rest_verts are fine, but something may go wrong in the projection step. Any pointers as to why this happens would be greatly appreciated.

(bodymap-env) ~/codeBodyMap/BodyMAP/scripts$ python metrics.py --files_dir /codeBodyMap/PMM_exps/normal/BodyMAP_PointNet_Model/inferences --save_path /codeBodyMap/PMM_exps/normal/BodyMAP_PointNet_Model/metrics
cover strs:   0%|                                                                                                                                                                                                | 0/3 [00:00<?, ?it/s]
SLP uncover dpx (990, 64, 27, 1) 100.0 ddx (990, 128, 54, 1) dy (990, 162) data_pmap (990, 6890) data_verts (990, 6890, 3) data names (990,)
uncover data:   0%|                                                                                                                                                                                            | 0/990 [00:01<?, ?it/s]
cover strs:   0%|                                                                                                                                                                                                | 0/3 [00:42<?, ?it/s]
Traceback (most recent call last):
  File "/BodyMAP/scripts/metrics.py", line 308, in <module>
    gt_height, gt_chest, gt_waist, gt_hips = compute_anatomy(gt_rest_verts)
                                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/BodyMAP/scripts/metrics.py", line 255, in compute_anatomy
    peri = compute_peripheries(triangles)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/BodyMAP/scripts/metrics.py", line 240, in compute_peripheries
    hull = ConvexHull(points_in_plane)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "scipy/spatial/_qhull.pyx", line 2492, in scipy.spatial._qhull.ConvexHull.__init__
  File "scipy/spatial/_qhull.pyx", line 284, in scipy.spatial._qhull._Qhull.__init__
ValueError: Points cannot contain NaN

Thanks in advance!

[Tandon-A]

Hi @melinawdh ,

Thank you for your interest in our work, and apologies for the delay on my side.

It is a bit weird that it is giving this error with the ground truth vertices. I am not sure about this error.

Since you have already said that the vertices look good, I suspect something is wrong with Shapy. Would recommend you to do a fresh install of Shapy environment outside of the bodymap-env, and then run the metrics file in the new environment.

Please do let me know if that helps you.

[melinawdh]

Hi Abhishek,

Thank you for your reply!

Unfortunetaly, that did not help. I was able to circumvent the error by replacing lines 236-239 with the following:

valid_face_idxs = np.where(collision_faces[ii] >= 0)[0]
if valid_face_idxs.size == 0:
    # plane doesn't intersect
    output[name]['value'].append(torch.tensor(0.0, device=device))
    continue

points_in_plane = np_points[
    ii, valid_face_idxs, :, ][:, :, [0, 2]].reshape(
        -1, 2)
            
finite_mask = np.all(isfinite(points_in_plane), axis=1)
points_in_plane = points_in_plane[finite_mask]
if points_in_plane.shape[0] < 3:
    # Not enough points for a hull
    output[name]['value'].append(torch.tensor(0.0, device=device))
    continue

However, I don't think this is a good fix, as I am simply dropping points, leading to some NaN in the resulting metrics file (specifically for the shape error of chest and waist for all cover modalities and additionally the hips for cover1).

[Tandon-A]

Hi @melinawdh ,

I was able to run the compute anatomy step. It ran perfectly without any NaN errors.

Tested for two participants: p80, p81 for all three cover conditions.

Possible next steps:

1. Double check the data you are using for the metrics.py file. Should be the real_val.txt file for the SLPDataset.
2. Plot the rest verts obtained from get_rest_verts to be sure that they are alright.
3. Would again suggest creating a new environment for Shapy. Can ask Cursor or other such tools to help you set it up. Make sure to use Python 3.8 for the Shapy environemnt. Also maybe check your scipy version with the one Shapy is using.
4. The code for the compute anatomy part (and the functions it is calling internally) is directly borrowed from Shapy without any modifications. Would suggest raising an issue with them if still not fixed.

[Tandon-A]

Please feel free to reopen the issue if needed.

[melinawdh]

Hi @Tandon-A ,

Sorry for the late reply.
Unfortunately, I keep having the same issues, so I will have to do some more investigating for possible issues.

Thank you for checking back and suggesting possible causes!