Is the (un)biased AD handled already?

[saeedhd96]

Hi,

I recall in Mitsuba2, for optimizing using pure AD, we needed to do the primal mode twice, and replace the AD graph in the second pirmal mode before doing backward mode.

Is this already handled in this version?

The use case is in integrator 'path' and calling dr.backward(loss).

Thanks

[ziyi-zhang]

Hi Saeed,
This is handled by mi.render() customOP defined here. An example of render_for/backward() can be found here.
But we can still run into that issue if we manually call render_backward without specifying a different seed.