Convert pytorch tensor to mi.Vector3f

[Hochheilige]

Hello, I'm trying to write my integrator using mitsuba and I want to spawn new rays based on some calculations that were done using pytorch. So I have:
sampled_dir - torch.Tensor with size [N, 3]
and I want to use it to spawn ray like this:
ray = si.spawn_ray(si.to_world(sampled_dir))

What is the right way to convert this tensor to Vector3f?

[njroussel]

Hi @Hochheilige

The dimensions need to to be re-ordered prior to conversion, like this:

import torch
import mitsuba as mi

mi.set_variant('cuda_ad_rgb')

torch_dirs = torch.arange(30).view(10, 3).cuda() # [10, 3]
mitsuba_dirs = mi.Vector3f(torch_dirs.permute(1, 0)) # re-order before cast

[Hochheilige]

Thanks, @njroussel, that make sense and I also tried to transpose tensor earlier, but unfortunately it doesn't work. And even your code gives me this error:

TypeError: Vector3f constructor expects: 1 or 3 values of type "Float", a matching list/tuple, or a NumPy/PyTorch/TF/Jax array.

I also tried mi.cuda_ad_rgb.Vector3f instead of just mi.Vector3f, but same error. Maybe I have no something installed or need to add some decorators before function I use?

[njroussel]

That's surprising, I just tried that small script I had sent on different operating systems and it worked fine.
Here's the versions I'm using:

mitsuba==3.6.4
drjit==1.0.5
torch==2.6.0
torchaudio==2.6.0
torchvision==0.21.0

Do you also have these ?

[Hochheilige]

Oh, really. I checked the versions and I have really old one drjit for some reason. With it updated everything works fine. Thank you!