35 lines
675 B
Python
35 lines
675 B
Python
"""
|
|
Advantage-weighted regression (AWR) for diffusion policy.
|
|
|
|
"""
|
|
|
|
import logging
|
|
import torch
|
|
|
|
log = logging.getLogger(__name__)
|
|
|
|
from model.diffusion.diffusion_rwr import RWRDiffusion
|
|
|
|
|
|
class AWRDiffusion(RWRDiffusion):
|
|
|
|
def __init__(
|
|
self,
|
|
actor,
|
|
critic,
|
|
**kwargs,
|
|
):
|
|
super().__init__(network=actor, **kwargs)
|
|
self.critic = critic.to(self.device)
|
|
|
|
# assign actor
|
|
self.actor = self.network
|
|
|
|
def loss_critic(self, obs, advantages):
|
|
# get advantage
|
|
adv = self.critic(obs)
|
|
|
|
# Update critic
|
|
loss_critic = torch.mean((adv - advantages) ** 2)
|
|
return loss_critic
|