search

Zhendong-Wang / Diffusion-Policies-for-Offline-RL

Apache License 2.0
219 stars 33 forks source link

Compute log_probs with diffusion policy model #1

Closed nuomizai closed 1 year ago

nuomizai commented 1 year ago

Was it possible to compute the action log_probs with the diffusion policy model in your algorithm?

Zhendong-Wang commented 1 year ago

Hi there. the analytical form of action log_probs is intractable, due to the non-linear transformation of neurual networks in each diffusion step. Appoximation for the log prob could be found https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/losses.py#L50.

nuomizai commented 1 year ago

Hi there. the analytical form of action log_probs is intractable, due to the non-linear transformation of neurual networks in each diffusion step. Appoximation for the log prob could be found https://github.com/openai/guided-diffusion/blob/22e0df8183507e13a7813f8d38d51b072ca1e67c/guided_diffusion/losses.py#L50.

Thanks for your reply.