Is the path integrator based on autodiff?

[gerwang]

Hi @njroussel

Could you please tell me whether the path integrator is based on autodiff or radiative backpropagation during dr.backward()? Are the forward and backward processes statistically independent? What about the direct_reparam integrator? Thank you!

[njroussel]

Hi @gerwang

Let me point you to a few different sources so you can check this for yourself too:

• The mi.render (source) function creates a CustomOp which defines how dr.forward() and dr.backward() should behave on this operation.
• The CustomOp for the rendering operation is defined here: RenderOp. Basically, it calls integrator.render_backward() for dr.backward(). Reminder: forward here means forward-mode differentiation and not the typical "forward" process which we call "primal" to disambiguate.
• For non-RBIntegrator types (not radiative backropagation style integrators) the integrator.render_backward() method is defined here and basically just used automatic differentiation as you would expect.
• For RBIntegrator integrators, the integrator.render_backward() implementation is located here. As you can see, here we do the two rendering passes: primal and backward mode.

Now for your questions:

1. path is not a RBIntegrator - it uses automatic differentiation.
2. When creating the RenderOp, both processes use different sampler seeds (source).
3. The direct_reparam is in a bit of an odd spot because of its short path length. We handle it differently.

[gerwang]

There remains one thing not intuitive to me: If automatic differentiation is used, is there anything involved with randomness? Does it make a difference whether we set different seeds in primal and backward modes?

[njroussel]

Even with automatic differentiation, there are still practically two renderings:

• The first is typically used to compute the loss (this is the eval() path in the RenderOp).
• The second one happens once you call dr.backward(loss) (it will trigger the backward() path in the RenderOp). This one will use automatic differentiation on its output but it must still be de-correlated from the primal.

Does that make sense?