class PPOTrainer(BaseRLTrainer): r"""Trainer class for PPO algorithm Paper: https://arxiv.org/abs/1707.06347. """ supported_tasks = ["Nav-v0"] SHORT_ROLLOUT_THRESHOLD: float = 0.25 _is_distributed: bool _obs_batching_cache: ObservationBatchingCache envs: VectorEnv agent: PPO actor_critic: Policy def __init__(self, config=None): interrupted_state = load_interrupted_state() if interrupted_state is not None: config = interrupted_state["config"] super().__init__(config) self.actor_critic = None self.agent = None self.envs = None self.obs_transforms = [] self._static_encoder = False self._encoder = None self._obs_space = None # Distirbuted if the world size would be # greater than 1 self._is_distributed = get_distrib_size()[2] > 1 self._obs_batching_cache = ObservationBatchingCache() @property def obs_space(self): if self._obs_space is None and self.envs is not None: self._obs_space = self.envs.observation_spaces[0] return self._obs_space @obs_space.setter def obs_space(self, new_obs_space): self._obs_space = new_obs_space def _all_reduce(self, t: torch.Tensor) -> torch.Tensor: r"""All reduce helper method that moves things to the correct device and only runs if distributed """ if not self._is_distributed: return t orig_device = t.device t = t.to(device=self.device) torch.distributed.all_reduce(t) return t.to(device=orig_device) def _setup_actor_critic_agent(self, ppo_cfg: Config) -> None: r"""Sets up actor critic and agent for PPO. Args: ppo_cfg: config node with relevant params Returns: None """ logger.add_filehandler(self.config.LOG_FILE) policy = baseline_registry.get_policy(self.config.RL.POLICY.name) observation_space = self.obs_space self.obs_transforms = get_active_obs_transforms(self.config) observation_space = apply_obs_transforms_obs_space( observation_space, self.obs_transforms ) self.actor_critic = policy.from_config( self.config, observation_space, self.envs.action_spaces[0] ) self.obs_space = observation_space self.actor_critic.to(self.device) if ( self.config.RL.DDPPO.pretrained_encoder or self.config.RL.DDPPO.pretrained ): pretrained_state = torch.load( self.config.RL.DDPPO.pretrained_weights, map_location="cpu" ) if self.config.RL.DDPPO.pretrained: self.actor_critic.load_state_dict( { k[len("actor_critic.") :]: v for k, v in pretrained_state["state_dict"].items() } ) elif self.config.RL.DDPPO.pretrained_encoder: prefix = "actor_critic.net.visual_encoder." self.actor_critic.net.visual_encoder.load_state_dict( { k[len(prefix) :]: v for k, v in pretrained_state["state_dict"].items() if k.startswith(prefix) } ) if not self.config.RL.DDPPO.train_encoder: self._static_encoder = True for param in self.actor_critic.net.visual_encoder.parameters(): param.requires_grad_(False) if self.config.RL.DDPPO.reset_critic: nn.init.orthogonal_(self.actor_critic.critic.fc.weight) nn.init.constant_(self.actor_critic.critic.fc.bias, 0) self.agent = (DDPPO if self._is_distributed else PPO)( actor_critic=self.actor_critic, clip_param=ppo_cfg.clip_param, ppo_epoch=ppo_cfg.ppo_epoch, num_mini_batch=ppo_cfg.num_mini_batch, value_loss_coef=ppo_cfg.value_loss_coef, entropy_coef=ppo_cfg.entropy_coef, lr=ppo_cfg.lr, eps=ppo_cfg.eps, max_grad_norm=ppo_cfg.max_grad_norm, use_normalized_advantage=ppo_cfg.use_normalized_advantage, ) def _init_envs(self, config=None): if config is None: config = self.config self.envs = construct_envs( config, get_env_class(config.ENV_NAME), workers_ignore_signals=is_slurm_batch_job(), ) def _init_train(self): if self.config.RL.DDPPO.force_distributed: self._is_distributed = True if is_slurm_batch_job(): add_signal_handlers() if self._is_distributed: local_rank, tcp_store = init_distrib_slurm( self.config.RL.DDPPO.distrib_backend ) if rank0_only(): logger.info( "Initialized DD-PPO with {} workers".format( torch.distributed.get_world_size() ) ) self.config.defrost() self.config.TORCH_GPU_ID = local_rank self.config.SIMULATOR_GPU_ID = local_rank # Multiply by the number of simulators to make sure they also get unique seeds self.config.TASK_CONFIG.SEED += ( torch.distributed.get_rank() * self.config.NUM_ENVIRONMENTS ) self.config.freeze() random.seed(self.config.TASK_CONFIG.SEED) np.random.seed(self.config.TASK_CONFIG.SEED) torch.manual_seed(self.config.TASK_CONFIG.SEED) self.num_rollouts_done_store = torch.distributed.PrefixStore( "rollout_tracker", tcp_store ) self.num_rollouts_done_store.set("num_done", "0") if rank0_only() and self.config.VERBOSE: logger.info(f"config: {self.config}") profiling_wrapper.configure( capture_start_step=self.config.PROFILING.CAPTURE_START_STEP, num_steps_to_capture=self.config.PROFILING.NUM_STEPS_TO_CAPTURE, ) self._init_envs() ppo_cfg = self.config.RL.PPO if torch.cuda.is_available(): self.device = torch.device("cuda", self.config.TORCH_GPU_ID) torch.cuda.set_device(self.device) else: self.device = torch.device("cpu") if rank0_only() and not os.path.isdir(self.config.CHECKPOINT_FOLDER): os.makedirs(self.config.CHECKPOINT_FOLDER) self._setup_actor_critic_agent(ppo_cfg) if self._is_distributed: self.agent.init_distributed(find_unused_params=True) logger.info( "agent number of parameters: {}".format( sum(param.numel() for param in self.agent.parameters()) ) ) obs_space = self.obs_space if self._static_encoder: self._encoder = self.actor_critic.net.visual_encoder obs_space = spaces.Dict( { "visual_features": spaces.Box( low=np.finfo(np.float32).min, high=np.finfo(np.float32).max, shape=self._encoder.output_shape, dtype=np.float32, ), **obs_space.spaces, } ) self._nbuffers = 2 if ppo_cfg.use_double_buffered_sampler else 1 self.rollouts = RolloutStorage( ppo_cfg.num_steps, self.envs.num_envs, obs_space, self.envs.action_spaces[0], ppo_cfg.hidden_size, num_recurrent_layers=self.actor_critic.net.num_recurrent_layers, is_double_buffered=ppo_cfg.use_double_buffered_sampler, ) self.rollouts.to(self.device) observations = self.envs.reset() batch = batch_obs( observations, device=self.device, cache=self._obs_batching_cache ) batch = apply_obs_transforms_batch(batch, self.obs_transforms) if self._static_encoder: with torch.no_grad(): batch["visual_features"] = self._encoder(batch) self.rollouts.buffers["observations"][0] = batch self.current_episode_reward = torch.zeros(self.envs.num_envs, 1) self.running_episode_stats = dict( count=torch.zeros(self.envs.num_envs, 1), reward=torch.zeros(self.envs.num_envs, 1), ) self.window_episode_stats = defaultdict( lambda: deque(maxlen=ppo_cfg.reward_window_size) ) self.env_time = 0.0 self.pth_time = 0.0 self.t_start = time.time() @rank0_only @profiling_wrapper.RangeContext("save_checkpoint") def save_checkpoint( self, file_name: str, extra_state: Optional[Dict] = None ) -> None: r"""Save checkpoint with specified name. Args: file_name: file name for checkpoint Returns: None """ checkpoint = { "state_dict": self.agent.state_dict(), "config": self.config, } if extra_state is not None: checkpoint["extra_state"] = extra_state torch.save( checkpoint, os.path.join(self.config.CHECKPOINT_FOLDER, file_name) ) def load_checkpoint(self, checkpoint_path: str, *args, **kwargs) -> Dict: r"""Load checkpoint of specified path as a dict. Args: checkpoint_path: path of target checkpoint *args: additional positional args **kwargs: additional keyword args Returns: dict containing checkpoint info """ return torch.load(checkpoint_path, *args, **kwargs) METRICS_BLACKLIST = {"top_down_map", "collisions.is_collision"} @classmethod def _extract_scalars_from_info( cls, info: Dict[str, Any] ) -> Dict[str, float]: result = {} for k, v in info.items(): if k in cls.METRICS_BLACKLIST: continue if isinstance(v, dict): result.update( { k + "." + subk: subv for subk, subv in cls._extract_scalars_from_info( v ).items() if (k + "." + subk) not in cls.METRICS_BLACKLIST } ) # Things that are scalar-like will have an np.size of 1. # Strings also have an np.size of 1, so explicitly ban those elif np.size(v) == 1 and not isinstance(v, str): result[k] = float(v) return result @classmethod def _extract_scalars_from_infos( cls, infos: List[Dict[str, Any]] ) -> Dict[str, List[float]]: results = defaultdict(list) for i in range(len(infos)): for k, v in cls._extract_scalars_from_info(infos[i]).items(): results[k].append(v) return results def _compute_actions_and_step_envs(self, buffer_index: int = 0): num_envs = self.envs.num_envs env_slice = slice( int(buffer_index * num_envs / self._nbuffers), int((buffer_index + 1) * num_envs / self._nbuffers), ) t_sample_action = time.time() # sample actions with torch.no_grad(): step_batch = self.rollouts.buffers[ self.rollouts.current_rollout_step_idxs[buffer_index], env_slice, ] profiling_wrapper.range_push("compute actions") ( values, actions, actions_log_probs, recurrent_hidden_states, ) = self.actor_critic.act( step_batch["observations"], step_batch["recurrent_hidden_states"], step_batch["prev_actions"], step_batch["masks"], ) # NB: Move actions to CPU. If CUDA tensors are # sent in to env.step(), that will create CUDA contexts # in the subprocesses. # For backwards compatibility, we also call .item() to convert to # an int actions = actions.to(device="cpu") self.pth_time += time.time() - t_sample_action profiling_wrapper.range_pop() # compute actions t_step_env = time.time() for index_env, act in zip( range(env_slice.start, env_slice.stop), actions.unbind(0) ): self.envs.async_step_at(index_env, act.item()) self.env_time += time.time() - t_step_env self.rollouts.insert( next_recurrent_hidden_states=recurrent_hidden_states, actions=actions, action_log_probs=actions_log_probs, value_preds=values, buffer_index=buffer_index, ) def _collect_environment_result(self, buffer_index: int = 0): num_envs = self.envs.num_envs env_slice = slice( int(buffer_index * num_envs / self._nbuffers), int((buffer_index + 1) * num_envs / self._nbuffers), ) t_step_env = time.time() outputs = [ self.envs.wait_step_at(index_env) for index_env in range(env_slice.start, env_slice.stop) ] observations, rewards_l, dones, infos = [ list(x) for x in zip(*outputs) ] self.env_time += time.time() - t_step_env t_update_stats = time.time() batch = batch_obs( observations, device=self.device, cache=self._obs_batching_cache ) batch = apply_obs_transforms_batch(batch, self.obs_transforms) rewards = torch.tensor( rewards_l, dtype=torch.float, device=self.current_episode_reward.device, ) rewards = rewards.unsqueeze(1) not_done_masks = torch.tensor( [[not done] for done in dones], dtype=torch.bool, device=self.current_episode_reward.device, ) done_masks = torch.logical_not(not_done_masks) self.current_episode_reward[env_slice] += rewards current_ep_reward = self.current_episode_reward[env_slice] self.running_episode_stats["reward"][env_slice] += current_ep_reward.where(done_masks, current_ep_reward.new_zeros(())) # type: ignore self.running_episode_stats["count"][env_slice] += done_masks.float() # type: ignore for k, v_k in self._extract_scalars_from_infos(infos).items(): v = torch.tensor( v_k, dtype=torch.float, device=self.current_episode_reward.device, ).unsqueeze(1) if k not in self.running_episode_stats: self.running_episode_stats[k] = torch.zeros_like( self.running_episode_stats["count"] ) self.running_episode_stats[k][env_slice] += v.where(done_masks, v.new_zeros(())) # type: ignore self.current_episode_reward[env_slice].masked_fill_(done_masks, 0.0) if self._static_encoder: with torch.no_grad(): batch["visual_features"] = self._encoder(batch) self.rollouts.insert( next_observations=batch, rewards=rewards, next_masks=not_done_masks, buffer_index=buffer_index, ) self.rollouts.advance_rollout(buffer_index) self.pth_time += time.time() - t_update_stats return env_slice.stop - env_slice.start @profiling_wrapper.RangeContext("_collect_rollout_step") def _collect_rollout_step(self): self._compute_actions_and_step_envs() return self._collect_environment_result() @profiling_wrapper.RangeContext("_update_agent") def _update_agent(self): ppo_cfg = self.config.RL.PPO t_update_model = time.time() with torch.no_grad(): step_batch = self.rollouts.buffers[ self.rollouts.current_rollout_step_idx ] next_value = self.actor_critic.get_value( step_batch["observations"], step_batch["recurrent_hidden_states"], step_batch["prev_actions"], step_batch["masks"], ) self.rollouts.compute_returns( next_value, ppo_cfg.use_gae, ppo_cfg.gamma, ppo_cfg.tau ) self.agent.train() value_loss, action_loss, dist_entropy = self.agent.update( self.rollouts ) self.rollouts.after_update() self.pth_time += time.time() - t_update_model return ( value_loss, action_loss, dist_entropy, ) def _coalesce_post_step( self, losses: Dict[str, float], count_steps_delta: int ) -> Dict[str, float]: stats_ordering = sorted(self.running_episode_stats.keys()) stats = torch.stack( [self.running_episode_stats[k] for k in stats_ordering], 0 ) stats = self._all_reduce(stats) for i, k in enumerate(stats_ordering): self.window_episode_stats[k].append(stats[i]) if self._is_distributed: loss_name_ordering = sorted(losses.keys()) stats = torch.tensor( [losses[k] for k in loss_name_ordering] + [count_steps_delta], device="cpu", dtype=torch.float32, ) stats = self._all_reduce(stats) count_steps_delta = int(stats[-1].item()) stats /= torch.distributed.get_world_size() losses = { k: stats[i].item() for i, k in enumerate(loss_name_ordering) } if self._is_distributed and rank0_only(): self.num_rollouts_done_store.set("num_done", "0") self.num_steps_done += count_steps_delta return losses @rank0_only def _training_log( self, writer, losses: Dict[str, float], prev_time: int = 0 ): deltas = { k: ( (v[-1] - v[0]).sum().item() if len(v) > 1 else v[0].sum().item() ) for k, v in self.window_episode_stats.items() } deltas["count"] = max(deltas["count"], 1.0) writer.add_scalar( "reward", deltas["reward"] / deltas["count"], self.num_steps_done, ) # Check to see if there are any metrics # that haven't been logged yet metrics = { k: v / deltas["count"] for k, v in deltas.items() if k not in {"reward", "count"} } if len(metrics) > 0: writer.add_scalars("metrics", metrics, self.num_steps_done) writer.add_scalars( "losses", losses, self.num_steps_done, ) # log stats if self.num_updates_done % self.config.LOG_INTERVAL == 0: logger.info( "update: {}\tfps: {:.3f}\t".format( self.num_updates_done, self.num_steps_done / ((time.time() - self.t_start) + prev_time), ) ) logger.info( "update: {}\tenv-time: {:.3f}s\tpth-time: {:.3f}s\t" "frames: {}".format( self.num_updates_done, self.env_time, self.pth_time, self.num_steps_done, ) ) logger.info( "Average window size: {} {}".format( len(self.window_episode_stats["count"]), " ".join( "{}: {:.3f}".format(k, v / deltas["count"]) for k, v in deltas.items() if k != "count" ), ) ) def should_end_early(self, rollout_step) -> bool: if not self._is_distributed: return False # This is where the preemption of workers happens. If a # worker detects it will be a straggler, it preempts itself! return ( rollout_step >= self.config.RL.PPO.num_steps * self.SHORT_ROLLOUT_THRESHOLD ) and int(self.num_rollouts_done_store.get("num_done")) >= ( self.config.RL.DDPPO.sync_frac * torch.distributed.get_world_size() ) @profiling_wrapper.RangeContext("train") def train(self) -> None: r"""Main method for training DD/PPO. Returns: None """ self._init_train() count_checkpoints = 0 prev_time = 0 lr_scheduler = LambdaLR( optimizer=self.agent.optimizer, lr_lambda=lambda x: 1 - self.percent_done(), ) interrupted_state = load_interrupted_state() if interrupted_state is not None: self.agent.load_state_dict(interrupted_state["state_dict"]) self.agent.optimizer.load_state_dict( interrupted_state["optim_state"] ) lr_scheduler.load_state_dict(interrupted_state["lr_sched_state"]) requeue_stats = interrupted_state["requeue_stats"] self.env_time = requeue_stats["env_time"] self.pth_time = requeue_stats["pth_time"] self.num_steps_done = requeue_stats["num_steps_done"] self.num_updates_done = requeue_stats["num_updates_done"] self._last_checkpoint_percent = requeue_stats[ "_last_checkpoint_percent" ] count_checkpoints = requeue_stats["count_checkpoints"] prev_time = requeue_stats["prev_time"] self._last_checkpoint_percent = requeue_stats[ "_last_checkpoint_percent" ] ppo_cfg = self.config.RL.PPO with ( TensorboardWriter( self.config.TENSORBOARD_DIR, flush_secs=self.flush_secs ) if rank0_only() else contextlib.suppress() ) as writer: while not self.is_done(): profiling_wrapper.on_start_step() profiling_wrapper.range_push("train update") if ppo_cfg.use_linear_clip_decay: self.agent.clip_param = ppo_cfg.clip_param * ( 1 - self.percent_done() ) if EXIT.is_set(): profiling_wrapper.range_pop() # train update self.envs.close() if REQUEUE.is_set() and rank0_only(): requeue_stats = dict( env_time=self.env_time, pth_time=self.pth_time, count_checkpoints=count_checkpoints, num_steps_done=self.num_steps_done, num_updates_done=self.num_updates_done, _last_checkpoint_percent=self._last_checkpoint_percent, prev_time=(time.time() - self.t_start) + prev_time, ) save_interrupted_state( dict( state_dict=self.agent.state_dict(), optim_state=self.agent.optimizer.state_dict(), lr_sched_state=lr_scheduler.state_dict(), config=self.config, requeue_stats=requeue_stats, ) ) requeue_job() return self.agent.eval() count_steps_delta = 0 profiling_wrapper.range_push("rollouts loop") profiling_wrapper.range_push("_collect_rollout_step") for buffer_index in range(self._nbuffers): self._compute_actions_and_step_envs(buffer_index) for step in range(ppo_cfg.num_steps): is_last_step = ( self.should_end_early(step + 1) or (step + 1) == ppo_cfg.num_steps ) for buffer_index in range(self._nbuffers): count_steps_delta += self._collect_environment_result( buffer_index ) if (buffer_index + 1) == self._nbuffers: profiling_wrapper.range_pop() # _collect_rollout_step if not is_last_step: if (buffer_index + 1) == self._nbuffers: profiling_wrapper.range_push( "_collect_rollout_step" ) self._compute_actions_and_step_envs(buffer_index) if is_last_step: break profiling_wrapper.range_pop() # rollouts loop if self._is_distributed: self.num_rollouts_done_store.add("num_done", 1) ( value_loss, action_loss, dist_entropy, ) = self._update_agent() if ppo_cfg.use_linear_lr_decay: lr_scheduler.step() # type: ignore self.num_updates_done += 1 losses = self._coalesce_post_step( dict(value_loss=value_loss, action_loss=action_loss), count_steps_delta, ) self._training_log(writer, losses, prev_time) # checkpoint model if rank0_only() and self.should_checkpoint(): self.save_checkpoint( f"ckpt.{count_checkpoints}.pth", dict( step=self.num_steps_done, wall_time=(time.time() - self.t_start) + prev_time, ), ) count_checkpoints += 1 profiling_wrapper.range_pop() # train update self.envs.close() def _eval_checkpoint( self, checkpoint_path: str, writer: TensorboardWriter, checkpoint_index: int = 0, ) -> None: r"""Evaluates a single checkpoint. Args: checkpoint_path: path of checkpoint writer: tensorboard writer object for logging to tensorboard checkpoint_index: index of cur checkpoint for logging Returns: None """ if self._is_distributed: raise RuntimeError("Evaluation does not support distributed mode") # Map location CPU is almost always better than mapping to a CUDA device. ckpt_dict = self.load_checkpoint(checkpoint_path, map_location="cpu") if self.config.EVAL.USE_CKPT_CONFIG: config = self._setup_eval_config(ckpt_dict["config"]) else: config = self.config.clone() ppo_cfg = config.RL.PPO config.defrost() config.TASK_CONFIG.DATASET.SPLIT = config.EVAL.SPLIT config.freeze() if len(self.config.VIDEO_OPTION) > 0: config.defrost() config.TASK_CONFIG.TASK.MEASUREMENTS.append("TOP_DOWN_MAP") config.TASK_CONFIG.TASK.MEASUREMENTS.append("COLLISIONS") config.freeze() if config.VERBOSE: logger.info(f"env config: {config}") self._init_envs(config) self._setup_actor_critic_agent(ppo_cfg) self.agent.load_state_dict(ckpt_dict["state_dict"]) self.actor_critic = self.agent.actor_critic observations = self.envs.reset() batch = batch_obs( observations, device=self.device, cache=self._obs_batching_cache ) batch = apply_obs_transforms_batch(batch, self.obs_transforms) current_episode_reward = torch.zeros( self.envs.num_envs, 1, device="cpu" ) test_recurrent_hidden_states = torch.zeros( self.config.NUM_ENVIRONMENTS, self.actor_critic.net.num_recurrent_layers, ppo_cfg.hidden_size, device=self.device, ) prev_actions = torch.zeros( self.config.NUM_ENVIRONMENTS, 1, device=self.device, dtype=torch.long, ) not_done_masks = torch.zeros( self.config.NUM_ENVIRONMENTS, 1, device=self.device, dtype=torch.bool, ) stats_episodes: Dict[ Any, Any ] = {} # dict of dicts that stores stats per episode rgb_frames = [ [] for _ in range(self.config.NUM_ENVIRONMENTS) ] # type: List[List[np.ndarray]] if len(self.config.VIDEO_OPTION) > 0: os.makedirs(self.config.VIDEO_DIR, exist_ok=True) number_of_eval_episodes = self.config.TEST_EPISODE_COUNT if number_of_eval_episodes == -1: number_of_eval_episodes = sum(self.envs.number_of_episodes) else: total_num_eps = sum(self.envs.number_of_episodes) if total_num_eps < number_of_eval_episodes: logger.warn( f"Config specified {number_of_eval_episodes} eval episodes" ", dataset only has {total_num_eps}." ) logger.warn(f"Evaluating with {total_num_eps} instead.") number_of_eval_episodes = total_num_eps pbar = tqdm.tqdm(total=number_of_eval_episodes) self.actor_critic.eval() while ( len(stats_episodes) < number_of_eval_episodes and self.envs.num_envs > 0 ): current_episodes = self.envs.current_episodes() with torch.no_grad(): ( _, actions, _, test_recurrent_hidden_states, ) = self.actor_critic.act( batch, test_recurrent_hidden_states, prev_actions, not_done_masks, deterministic=False, ) prev_actions.copy_(actions) # type: ignore # NB: Move actions to CPU. If CUDA tensors are # sent in to env.step(), that will create CUDA contexts # in the subprocesses. # For backwards compatibility, we also call .item() to convert to # an int step_data = [a.item() for a in actions.to(device="cpu")] outputs = self.envs.step(step_data) observations, rewards_l, dones, infos = [ list(x) for x in zip(*outputs) ] batch = batch_obs( observations, device=self.device, cache=self._obs_batching_cache, ) batch = apply_obs_transforms_batch(batch, self.obs_transforms) not_done_masks = torch.tensor( [[not done] for done in dones], dtype=torch.bool, device="cpu", ) rewards = torch.tensor( rewards_l, dtype=torch.float, device="cpu" ).unsqueeze(1) current_episode_reward += rewards next_episodes = self.envs.current_episodes() envs_to_pause = [] n_envs = self.envs.num_envs for i in range(n_envs): if ( next_episodes[i].scene_id, next_episodes[i].episode_id, ) in stats_episodes: envs_to_pause.append(i) # episode ended if not not_done_masks[i].item(): pbar.update() episode_stats = {} episode_stats["reward"] = current_episode_reward[i].item() episode_stats.update( self._extract_scalars_from_info(infos[i]) ) current_episode_reward[i] = 0 # use scene_id + episode_id as unique id for storing stats stats_episodes[ ( current_episodes[i].scene_id, current_episodes[i].episode_id, ) ] = episode_stats if len(self.config.VIDEO_OPTION) > 0: generate_video( video_option=self.config.VIDEO_OPTION, video_dir=self.config.VIDEO_DIR, images=rgb_frames[i], episode_id=current_episodes[i].episode_id, checkpoint_idx=checkpoint_index, metrics=self._extract_scalars_from_info(infos[i]), tb_writer=writer, ) rgb_frames[i] = [] # episode continues elif len(self.config.VIDEO_OPTION) > 0: # TODO move normalization / channel changing out of the policy and undo it here frame = observations_to_image( {k: v[i] for k, v in batch.items()}, infos[i] ) rgb_frames[i].append(frame) not_done_masks = not_done_masks.to(device=self.device) ( self.envs, test_recurrent_hidden_states, not_done_masks, current_episode_reward, prev_actions, batch, rgb_frames, ) = self._pause_envs( envs_to_pause, self.envs, test_recurrent_hidden_states, not_done_masks, current_episode_reward, prev_actions, batch, rgb_frames, ) num_episodes = len(stats_episodes) aggregated_stats = {} for stat_key in next(iter(stats_episodes.values())).keys(): aggregated_stats[stat_key] = ( sum(v[stat_key] for v in stats_episodes.values()) / num_episodes ) for k, v in aggregated_stats.items(): logger.info(f"Average episode {k}: {v:.4f}") step_id = checkpoint_index if "extra_state" in ckpt_dict and "step" in ckpt_dict["extra_state"]: step_id = ckpt_dict["extra_state"]["step"] writer.add_scalars( "eval_reward", {"average reward": aggregated_stats["reward"]}, step_id, ) metrics = {k: v for k, v in aggregated_stats.items() if k != "reward"} if len(metrics) > 0: writer.add_scalars("eval_metrics", metrics, step_id) self.envs.close()
def _worker_fn(world_rank: int, world_size: int, port: int, unused_params: bool): device = (torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")) tcp_store = distrib.TCPStore( # type: ignore "127.0.0.1", port, world_size, world_rank == 0) distrib.init_process_group("gloo", store=tcp_store, rank=world_rank, world_size=world_size) config = get_config("habitat_baselines/config/test/ppo_pointnav_test.yaml") obs_space = gym.spaces.Dict({ IntegratedPointGoalGPSAndCompassSensor.cls_uuid: gym.spaces.Box( low=np.finfo(np.float32).min, high=np.finfo(np.float32).max, shape=(2, ), dtype=np.float32, ) }) action_space = ActionSpace({"move": EmptySpace()}) actor_critic = PointNavBaselinePolicy.from_config(config, obs_space, action_space) # This use adds some arbitrary parameters that aren't part of the computation # graph, so they will mess up DDP if they aren't correctly ignored by it if unused_params: actor_critic.unused = nn.Linear(64, 64) actor_critic.to(device=device) ppo_cfg = config.RL.PPO agent = DDPPO( actor_critic=actor_critic, clip_param=ppo_cfg.clip_param, ppo_epoch=ppo_cfg.ppo_epoch, num_mini_batch=ppo_cfg.num_mini_batch, value_loss_coef=ppo_cfg.value_loss_coef, entropy_coef=ppo_cfg.entropy_coef, lr=ppo_cfg.lr, eps=ppo_cfg.eps, max_grad_norm=ppo_cfg.max_grad_norm, use_normalized_advantage=ppo_cfg.use_normalized_advantage, ) agent.init_distributed() rollouts = RolloutStorage( ppo_cfg.num_steps, 2, obs_space, action_space, ppo_cfg.hidden_size, num_recurrent_layers=actor_critic.net.num_recurrent_layers, is_double_buffered=False, ) rollouts.to(device) for k, v in rollouts.buffers["observations"].items(): rollouts.buffers["observations"][k] = torch.randn_like(v) # Add two steps so batching works rollouts.advance_rollout() rollouts.advance_rollout() # Get a single batch batch = next(rollouts.recurrent_generator(rollouts.buffers["returns"], 1)) # Call eval actions through the internal wrapper that is used in # agent.update value, action_log_probs, dist_entropy, _ = agent._evaluate_actions( batch["observations"], batch["recurrent_hidden_states"], batch["prev_actions"], batch["masks"], batch["actions"], ) # Backprop on things (value.mean() + action_log_probs.mean() + dist_entropy.mean()).backward() # Make sure all ranks have very similar parameters for param in actor_critic.parameters(): if param.grad is not None: grads = [param.grad.detach().clone() for _ in range(world_size)] distrib.all_gather(grads, grads[world_rank]) for i in range(world_size): assert torch.isclose(grads[i], grads[world_rank]).all()