""" Advantage-weighted regression (AWR) for diffusion policy. Advantage = discounted-reward-to-go - V(s) Do not support pixel input right now. """ import os import pickle import einops import numpy as np import torch import logging import wandb from copy import deepcopy log = logging.getLogger(__name__) from util.timer import Timer from collections import deque from agent.finetune.train_agent import TrainAgent from util.scheduler import CosineAnnealingWarmupRestarts def td_values( states, rewards, dones, state_values, gamma=0.99, alpha=0.95, lam=0.95, ): """ Gives a list of TD estimates for a given list of samples from an RL environment. The TD(λ) estimator is used for this computation. :param replay_buffers: The replay buffers filled by exploring the RL environment. Includes: states, rewards, "final state?"s. :param state_values: The currently estimated state values. :return: The TD estimates. """ sample_count = len(states) tds = np.zeros_like(state_values, dtype=np.float32) dones[-1] = 1 next_value = 1 - dones[-1] val = 0.0 for i in range(sample_count - 1, -1, -1): # next_value = 0.0 if dones[i] else state_values[i + 1] # get next_value for vectorized if i < sample_count - 1: next_value = state_values[i + 1] next_value = next_value * (1 - dones[i]) state_value = state_values[i] error = rewards[i] + gamma * next_value - state_value val = alpha * error + gamma * lam * (1 - dones[i]) * val tds[i] = val + state_value return tds class TrainAWRDiffusionAgent(TrainAgent): def __init__(self, cfg): super().__init__(cfg) self.logprob_batch_size = cfg.train.get("logprob_batch_size", 10000) # note the discount factor gamma here is applied to reward every act_steps, instead of every env step self.gamma = cfg.train.gamma # Wwarm up period for critic before actor updates self.n_critic_warmup_itr = cfg.train.n_critic_warmup_itr # Optimizer self.actor_optimizer = torch.optim.AdamW( self.model.actor.parameters(), lr=cfg.train.actor_lr, weight_decay=cfg.train.actor_weight_decay, ) self.actor_lr_scheduler = CosineAnnealingWarmupRestarts( self.actor_optimizer, first_cycle_steps=cfg.train.actor_lr_scheduler.first_cycle_steps, cycle_mult=1.0, max_lr=cfg.train.actor_lr, min_lr=cfg.train.actor_lr_scheduler.min_lr, warmup_steps=cfg.train.actor_lr_scheduler.warmup_steps, gamma=1.0, ) self.critic_optimizer = torch.optim.AdamW( self.model.critic.parameters(), lr=cfg.train.critic_lr, weight_decay=cfg.train.critic_weight_decay, ) self.critic_lr_scheduler = CosineAnnealingWarmupRestarts( self.critic_optimizer, first_cycle_steps=cfg.train.critic_lr_scheduler.first_cycle_steps, cycle_mult=1.0, max_lr=cfg.train.critic_lr, min_lr=cfg.train.critic_lr_scheduler.min_lr, warmup_steps=cfg.train.critic_lr_scheduler.warmup_steps, gamma=1.0, ) # Buffer size self.buffer_size = cfg.train.buffer_size # Reward exponential self.beta = cfg.train.beta # Max weight for AWR self.max_adv_weight = cfg.train.max_adv_weight # Scaling reward self.scale_reward_factor = cfg.train.scale_reward_factor # Updates self.replay_ratio = cfg.train.replay_ratio self.critic_update_ratio = cfg.train.critic_update_ratio def run(self): # make a FIFO replay buffer for obs, action, and reward obs_buffer = deque(maxlen=self.buffer_size) action_buffer = deque(maxlen=self.buffer_size) reward_buffer = deque(maxlen=self.buffer_size) done_buffer = deque(maxlen=self.buffer_size) first_buffer = deque(maxlen=self.buffer_size) # Start training loop timer = Timer() run_results = [] last_itr_eval = False done_venv = np.zeros((1, self.n_envs)) while self.itr < self.n_train_itr: # Prepare video paths for each envs --- only applies for the first set of episodes if allowing reset within iteration and each iteration has multiple episodes from one env options_venv = [{} for _ in range(self.n_envs)] if self.itr % self.render_freq == 0 and self.render_video: for env_ind in range(self.n_render): options_venv[env_ind]["video_path"] = os.path.join( self.render_dir, f"itr-{self.itr}_trial-{env_ind}.mp4" ) # Define train or eval - all envs restart eval_mode = self.itr % self.val_freq == 0 and not self.force_train self.model.eval() if eval_mode else self.model.train() last_itr_eval = eval_mode # Reset env before iteration starts (1) if specified, (2) at eval mode, or (3) right after eval mode firsts_trajs = np.zeros((self.n_steps + 1, self.n_envs)) if self.reset_at_iteration or eval_mode or last_itr_eval: prev_obs_venv = self.reset_env_all(options_venv=options_venv) firsts_trajs[0] = 1 else: firsts_trajs[0] = ( done_venv # if done at the end of last iteration, then the envs are just reset ) reward_trajs = np.empty((0, self.n_envs)) # Collect a set of trajectories from env for step in range(self.n_steps): if step % 10 == 0: print(f"Processed step {step} of {self.n_steps}") # Select action with torch.no_grad(): cond = { "state": torch.from_numpy(prev_obs_venv["state"]) .float() .to(self.device) } samples = ( self.model( cond=cond, deterministic=eval_mode, ) .cpu() .numpy() ) action_venv = samples[:, : self.act_steps] # Apply multi-step action obs_venv, reward_venv, done_venv, info_venv = self.venv.step( action_venv ) reward_trajs = np.vstack((reward_trajs, reward_venv[None])) # add to buffer obs_buffer.append(prev_obs_venv["state"]) action_buffer.append(action_venv) reward_buffer.append(reward_venv * self.scale_reward_factor) done_buffer.append(done_venv) first_buffer.append(firsts_trajs[step]) firsts_trajs[step + 1] = done_venv prev_obs_venv = obs_venv # Summarize episode reward --- this needs to be handled differently depending on whether the environment is reset after each iteration. Only count episodes that finish within the iteration. episodes_start_end = [] for env_ind in range(self.n_envs): env_steps = np.where(firsts_trajs[:, env_ind] == 1)[0] for i in range(len(env_steps) - 1): start = env_steps[i] end = env_steps[i + 1] if end - start > 1: episodes_start_end.append((env_ind, start, end - 1)) if len(episodes_start_end) > 0: reward_trajs_split = [ reward_trajs[start : end + 1, env_ind] for env_ind, start, end in episodes_start_end ] num_episode_finished = len(reward_trajs_split) episode_reward = np.array( [np.sum(reward_traj) for reward_traj in reward_trajs_split] ) episode_best_reward = np.array( [ np.max(reward_traj) / self.act_steps for reward_traj in reward_trajs_split ] ) avg_episode_reward = np.mean(episode_reward) avg_best_reward = np.mean(episode_best_reward) success_rate = np.mean( episode_best_reward >= self.best_reward_threshold_for_success ) else: episode_reward = np.array([]) num_episode_finished = 0 avg_episode_reward = 0 avg_best_reward = 0 success_rate = 0 log.info("[WARNING] No episode completed within the iteration!") # Update models if not eval_mode: obs_trajs = np.array(deepcopy(obs_buffer)) # assume only state reward_trajs = np.array(deepcopy(reward_buffer)) dones_trajs = np.array(deepcopy(done_buffer)) obs_t = einops.rearrange( torch.from_numpy(obs_trajs).float().to(self.device), "s e h d -> (s e) h d", ) values_trajs = np.array( self.model.critic({"state": obs_t}).detach().cpu().numpy() ).reshape(-1, self.n_envs) td_trajs = td_values(obs_trajs, reward_trajs, dones_trajs, values_trajs) td_t = torch.from_numpy(td_trajs.flatten()).float().to(self.device) # Update critic num_batch = int( self.n_steps * self.n_envs / self.batch_size * self.replay_ratio ) for _ in range(num_batch // self.critic_update_ratio): inds = np.random.choice(len(obs_trajs), self.batch_size) loss_critic = self.model.loss_critic( {"state": obs_t[inds]}, td_t[inds] ) self.critic_optimizer.zero_grad() loss_critic.backward() self.critic_optimizer.step() # Update policy - use a new copy of data obs_trajs = np.array(deepcopy(obs_buffer)) samples_trajs = np.array(deepcopy(action_buffer)) reward_trajs = np.array(deepcopy(reward_buffer)) dones_trajs = np.array(deepcopy(done_buffer)) obs_t = einops.rearrange( torch.from_numpy(obs_trajs).float().to(self.device), "s e h d -> (s e) h d", ) values_trajs = np.array( self.model.critic({"state": obs_t}).detach().cpu().numpy() ).reshape(-1, self.n_envs) td_trajs = td_values(obs_trajs, reward_trajs, dones_trajs, values_trajs) advantages_trajs = td_trajs - values_trajs # flatten obs_trajs = einops.rearrange( obs_trajs, "s e h d -> (s e) h d", ) samples_trajs = einops.rearrange( samples_trajs, "s e h d -> (s e) h d", ) advantages_trajs = einops.rearrange( advantages_trajs, "s e -> (s e)", ) for _ in range(num_batch): # Sample batch inds = np.random.choice(len(obs_trajs), self.batch_size) obs_b = { "state": torch.from_numpy(obs_trajs[inds]) .float() .to(self.device) } actions_b = ( torch.from_numpy(samples_trajs[inds]).float().to(self.device) ) advantages_b = ( torch.from_numpy(advantages_trajs[inds]).float().to(self.device) ) advantages_b = (advantages_b - advantages_b.mean()) / ( advantages_b.std() + 1e-6 ) advantages_b_scaled = torch.exp(self.beta * advantages_b) advantages_b_scaled.clamp_(max=self.max_adv_weight) # Update policy with collected trajectories loss = self.model.loss( actions_b, obs_b, advantages_b_scaled.detach(), ) self.actor_optimizer.zero_grad() loss.backward() if self.itr >= self.n_critic_warmup_itr: if self.max_grad_norm is not None: torch.nn.utils.clip_grad_norm_( self.model.actor.parameters(), self.max_grad_norm ) self.actor_optimizer.step() # Update lr self.actor_lr_scheduler.step() self.critic_lr_scheduler.step() # Save model if self.itr % self.save_model_freq == 0 or self.itr == self.n_train_itr - 1: self.save_model() # Log loss and save metrics run_results.append( { "itr": self.itr, } ) if self.itr % self.log_freq == 0: time = timer() if eval_mode: log.info( f"eval: success rate {success_rate:8.4f} | avg episode reward {avg_episode_reward:8.4f} | avg best reward {avg_best_reward:8.4f}" ) if self.use_wandb: wandb.log( { "success rate - eval": success_rate, "avg episode reward - eval": avg_episode_reward, "avg best reward - eval": avg_best_reward, "num episode - eval": num_episode_finished, }, step=self.itr, commit=False, ) run_results[-1]["eval_success_rate"] = success_rate run_results[-1]["eval_episode_reward"] = avg_episode_reward run_results[-1]["eval_best_reward"] = avg_best_reward else: log.info( f"{self.itr}: loss {loss:8.4f} | reward {avg_episode_reward:8.4f} |t:{time:8.4f}" ) if self.use_wandb: wandb.log( { "loss": loss, "loss - critic": loss_critic, "avg episode reward - train": avg_episode_reward, "num episode - train": num_episode_finished, }, step=self.itr, commit=True, ) run_results[-1]["loss"] = loss run_results[-1]["loss_critic"] = loss_critic run_results[-1]["train_episode_reward"] = avg_episode_reward run_results[-1]["time"] = time with open(self.result_path, "wb") as f: pickle.dump(run_results, f) self.itr += 1