exploring_exploration/utils/eval.py [667:787]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - obs_im, obs_sm, obs_lm = get_obs(obs) obs_collns = obs["collisions"] if actor_type == "frontier": delta_ego = torch.zeros((num_processes, 3)).to(device) frontier_agent.reset() if use_policy: recurrent_hidden_states = torch.zeros(num_processes, feat_shape_sim[0]).to( device ) masks = torch.zeros(num_processes, 1).to(device) nav_policy.reset() prev_action = torch.zeros(num_processes, 1).long().to(device) prev_collision = obs_collns obs_odometer = torch.zeros(num_processes, 4).to(device) per_proc_collisions = [0.0 for _ in range(num_processes)] # ================================================================= # ==================== Perform exploration ======================== # ================================================================= for step in range(num_steps_exp): if use_policy: encoder_inputs = [obs_im] if encoder_type == "rgb+map": encoder_inputs += [obs_sm, obs_lm] with torch.no_grad(): policy_feats = encoder(*encoder_inputs) policy_inputs = {"features": policy_feats} if use_action_embedding: policy_inputs["actions"] = prev_action if use_collision_embedding: policy_inputs["collisions"] = prev_collision.long() policy_outputs = actor_critic.act( policy_inputs, recurrent_hidden_states, masks, deterministic=False, ) _, action, _, recurrent_hidden_states = policy_outputs elif actor_type == "oracle": action = obs["oracle_action"].long() elif actor_type == "random": action = torch.randint( 0, envs.action_space.n, (num_processes, 1) ).long() elif actor_type == "forward": action = torch.Tensor(np.ones((num_processes, 1)) * forward_action_id) action = action.long() elif actor_type == "forward-plus": action = torch.Tensor(np.ones((num_processes, 1)) * forward_action_id) collision_mask = prev_collision > 0 action[collision_mask] = turn_action_id action = action.long() elif actor_type == "frontier": # This assumes that num_processes = 1 occ_map = obs["highres_coarse_occupancy"][0].cpu().numpy() occ_map = occ_map.transpose(1, 2, 0) occ_map = np.ascontiguousarray(occ_map) occ_map = occ_map.astype(np.uint8) action = frontier_agent.act( occ_map, delta_ego[0].cpu().numpy(), prev_collision[0].item() ) action = torch.Tensor([[action]]).long() obs, reward, done, infos = envs.step(action) # Processing environment inputs obs_im, obs_sm, obs_lm = get_obs(obs) obs_collns = obs["collisions"] obs_odometer_curr = process_odometer(obs["delta"]) if actor_type == "frontier": delta_ego = compute_egocentric_coors( obs_odometer_curr, obs_odometer, occ_map_scale, ) # (N, 3) --- (dx_ego, dy_ego, dt_ego) # Always set masks to 1 (does not matter for now) masks = torch.FloatTensor([[1.0] for _ in range(num_processes)]).to(device) obs_odometer = obs_odometer + obs_odometer_curr # This must not reach done = True assert done[0] == False # Update collisions metric for pr in range(num_processes): per_proc_collisions[pr] += obs_collns[pr, 0].item() prev_collision = obs_collns prev_action = action # Verifying correctness if step == num_steps_exp - 1: assert infos[0]["finished_exploration"] elif step < num_steps_exp - 1: assert not infos[0]["finished_exploration"] exploration_topdown_map = infos[0]["topdown_map"] # Update Exploration statistics for pr in range(num_processes): episode_environment_statistics.append(infos[pr]["environment_statistics"]) exp_area_covered.append(infos[pr]["seen_area"]) exp_collisions.append(per_proc_collisions[pr]) # ================================================================= # ===================== Navigation evaluation ===================== # ================================================================= # gather statistics for visualization per_proc_rgb = [[] for pr in range(num_processes)] per_proc_depth = [[] for pr in range(num_processes)] per_proc_fine_occ = [[] for pr in range(num_processes)] per_proc_coarse_occ = [[] for pr in range(num_processes)] per_proc_topdown_map = [[] for pr in range(num_processes)] per_proc_planner_vis = [[] for pr in range(num_processes)] per_proc_gt_topdown_map = [[] for pr in range(num_processes)] per_proc_initial_planner_vis = [[] for pr in range(num_processes)] per_proc_exploration_topdown_map = [[] for pr in range(num_processes)] WIDTH, HEIGHT = 300, 300 nav_policy.reset() - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - exploring_exploration/utils/eval.py [1012:1132]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - obs_im, obs_sm, obs_lm = get_obs(obs) obs_collns = obs["collisions"] if actor_type == "frontier": delta_ego = torch.zeros((num_processes, 3)).to(device) frontier_agent.reset() if use_policy: recurrent_hidden_states = torch.zeros(num_processes, feat_shape_sim[0]).to( device ) masks = torch.zeros(num_processes, 1).to(device) nav_policy.reset() prev_action = torch.zeros(num_processes, 1).long().to(device) prev_collision = obs_collns obs_odometer = torch.zeros(num_processes, 4).to(device) per_proc_collisions = [0.0 for _ in range(num_processes)] # =================== Perform exploration ======================== for step in range(num_steps_exp): if use_policy: encoder_inputs = [obs_im] if encoder_type == "rgb+map": encoder_inputs += [obs_sm, obs_lm] with torch.no_grad(): policy_feats = encoder(*encoder_inputs) policy_inputs = {"features": policy_feats} if use_action_embedding: policy_inputs["actions"] = prev_action if use_collision_embedding: policy_inputs["collisions"] = prev_collision.long() policy_outputs = actor_critic.act( policy_inputs, recurrent_hidden_states, masks, deterministic=False, ) _, action, _, recurrent_hidden_states = policy_outputs elif actor_type == "oracle": action = obs["oracle_action"].long() elif actor_type == "random": action = torch.randint( 0, envs.action_space.n, (num_processes, 1) ).long() elif actor_type == "forward": action = torch.Tensor(np.ones((num_processes, 1)) * forward_action_id) action = action.long() elif actor_type == "forward-plus": action = torch.Tensor(np.ones((num_processes, 1)) * forward_action_id) collision_mask = prev_collision > 0 action[collision_mask] = turn_action_id action = action.long() elif actor_type == "frontier": # This assumes that num_processes = 1 occ_map = obs["highres_coarse_occupancy"][0].cpu().numpy() occ_map = occ_map.transpose(1, 2, 0) occ_map = np.ascontiguousarray(occ_map) occ_map = occ_map.astype(np.uint8) action = frontier_agent.act( occ_map, delta_ego[0].cpu().numpy(), prev_collision[0].item() ) action = torch.Tensor([[action]]).long() obs, reward, done, infos = envs.step(action) # Processing environment inputs obs_im, obs_sm, obs_lm = get_obs(obs) obs_collns = obs["collisions"] obs_odometer_curr = process_odometer(obs["delta"]) if actor_type == "frontier": delta_ego = compute_egocentric_coors( obs_odometer_curr, obs_odometer, occ_map_scale, ) # (N, 3) --- (dx_ego, dy_ego, dt_ego) # Always set masks to 1 (does not matter for now) masks = torch.FloatTensor([[1.0] for _ in range(num_processes)]).to(device) obs_odometer = obs_odometer + obs_odometer_curr # This must not reach done = True assert done[0] == False # Update collisions metric for pr in range(num_processes): per_proc_collisions[pr] += obs_collns[pr, 0].item() prev_collision = obs_collns prev_action = action # Debug stuff if step == num_steps_exp - 1: assert infos[0]["finished_exploration"] elif step < num_steps_exp - 1: assert not infos[0]["finished_exploration"] exploration_topdown_map = infos[0]["topdown_map"] # ================================================================= # ===================== Navigation evaluation ===================== # ================================================================= # Exploration statistics for pr in range(num_processes): episode_environment_statistics.append(infos[pr]["environment_statistics"]) exp_area_covered.append(infos[pr]["seen_area"]) exp_collisions.append(per_proc_collisions[pr]) # gather statistics for visualization per_proc_rgb = [[] for pr in range(num_processes)] per_proc_depth = [[] for pr in range(num_processes)] per_proc_fine_occ = [[] for pr in range(num_processes)] per_proc_coarse_occ = [[] for pr in range(num_processes)] per_proc_topdown_map = [[] for pr in range(num_processes)] per_proc_planner_vis = [[] for pr in range(num_processes)] per_proc_gt_topdown_map = [[] for pr in range(num_processes)] per_proc_initial_planner_vis = [[] for pr in range(num_processes)] per_proc_exploration_topdown_map = [[] for pr in range(num_processes)] WIDTH, HEIGHT = 300, 300 nav_policy.reset() - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -