`for cam in cameras` raises `ValueError`

[Mason-McGough]

🐛 Bugs / Unexpected behaviors

Using for to loop over child classes of CamerasBase attempts to index beyond len(cameras) and raises a ValueError.

Instructions To Reproduce the Issue:

With the following script I am able to reproduce the issue.

from pytorch3d.renderer import (
    look_at_view_transform,
    PerspectiveCameras
)

n_iters = 10
R, T = look_at_view_transform(2.7, 0, 180)
R, T = R.repeat(n_iters, 1, 1), T.repeat(n_iters, 1)
cameras = PerspectiveCameras(R=R, T=T)

for i, cam in enumerate(cameras):
    print(f'{i}: {cam}')

This produces the following output:

0: PerspectiveCameras()
1: PerspectiveCameras()
2: PerspectiveCameras()
3: PerspectiveCameras()
4: PerspectiveCameras()
5: PerspectiveCameras()
6: PerspectiveCameras()
7: PerspectiveCameras()
8: PerspectiveCameras()
9: PerspectiveCameras()
Traceback (most recent call last):
  File "/home/mmcgoug/Workspace/NeRF/pytorch3d/_test_cameras_iterator.py", line 11, in <module>
    for i, cam in enumerate(cameras):
  File "/home/mmcgoug/miniconda3/envs/pytorch3d/lib/python3.9/site-packages/pytorch3d/renderer/cameras.py", line 413, in __getitem__
    raise ValueError(f"Index {max(index)} is out of bounds for select cameras")
ValueError: Index 10 is out of bounds for select cameras

Seems that the iterator eventually produces an index>len(cameras), which causes it to raise a ValueError at this line. Since CamerasBase does not define an __iter__ method, should __getitem__ instead raise StopIteration somewhere?

The issue shows up for any value n_iters>0 I have tried. I have also tried catching StopIteration explicitly but it does not seem to be raised.

Versions:
pytorch==1.13.0
pytorch-cuda==11.6
pytorch3d==0.7.2

[bottler]

Thanks for raising this. If __getitem__ is changed to raising IndexError the behaviour will be as expected. I'll make the change soon.

[bottler]

Fixed in 84851c8

[dav-ell]

Oh great, I just ran into this and was looking for a solution. For those who haven't updated, this is a workaround:

do_thing(cameras[i]) for i in range(len(cameras))