Camera to_world differentiability

[satyajitghana]

Hi, is there any example to get camera to_world to be differentiable?

I've been trying, but didn't work: #251

[njroussel]

Hi @satyajitghana,

The PR you linked adds a test case in src/integrators/tests/test_ad_integrators.py which is a setup where the sensor is being translated.
You can use that as an example.

The output you posted in the PR is rather surprising.

(...)/pytorch3d/lib/python3.9/site-packages/mitsuba/python/ad/optimizers.py:51, in Optimizer.setitem(self, key, value)

Did you manually install your own build? If not, I think your current setup is picking up an old mitsuba install which doesn't have the PR yet.

Also, in the future, could you please just stick to a single thread - spreading this conversation makes it hard to follow for us or anyone that would have the same issue :)

[satyajitghana]

Yes, I've built the latest build.

I'll look into the test case. I tried to understand that, but it was a little difficult. Thanks though.