Questions about NLL computation in the "maze2d" branch

[Jang-seunghwan]

While attempting to compute the negative log-likelihood (NLL) for the Maze2D task,
I ran into a few points that I’m hoping you can clarify:

1. L_0 term in vb_terms_bpd
   The current implementation comments out the L_0 calculation

• Is it correct to simply uncomment this block when we want to include L_0 in the NLL?
• I assume the L_T term was omitted in function because its contribution is ~0; is it right?

2. Potential mismatch in the ELBO call
   In diffuser/guides/policies.py the ELBO is computed with
   elbo_i = self.diffusion_model._vb_terms_bpd( diffuser_state, conditions, diffusion[:, step, :, :], timesteps )
   Here, diffusion[:, step, :, :] seems to be an x_t sampled from an independent diffusion trajectory,
   so it may not align with diffuser_state. Is this intentional, or should both tensors come from the same trajectory?

3. Division NLL by 255
   sum_elbo = sum_elbo.detach().cpu().numpy()[0] / 255 # ave
   It makes sense for image data, where pixel values lie in 0…255. But the Maze2D trajectories are already scaled to the range −1,1,
   so this extra division seems unnecessary.
   Could you share the reasoning behind this 255-scaling step?

[Weixy21]

1. Yes, you can uncomment this block if you want to include L0

2. The diffuser is comparing with itself to get NLL, so we used a trajectory from diffuser as a reference.

3. Here 255 is the diffusion timestep. Dividing by 255 is not necessary. We just get the average NLL in the diffusion process in this case.