| import torch | |
| def batch_simulate_on_environment(policy, env, verbose = True): | |
| if verbose: | |
| print("*** In batch_simulate_on_environment ***") | |
| from Dataset import Trajectory, TrajectoryDataset | |
| from math import ceil | |
| dataset = TrajectoryDataset() | |
| trajectories = [Trajectory() for _ in range(env.bsize)] | |
| batch_obs = env.reset() | |
| batch_done = [False,]*env.bsize | |
| while not all(batch_done): | |
| with torch.no_grad(): | |
| actions = policy(batch_obs) | |
| batch_feedback = env.step(actions) | |
| for i, feedback in zip(range(env.bsize), batch_feedback): | |
| if feedback is None: | |
| continue | |
| next_obs, r, done = feedback | |
| trajectories[i].append({"observation": batch_obs[i], | |
| "action": actions[i], | |
| "reward": r, | |
| "next_observation": next_obs, | |
| "done": done, | |
| }) | |
| batch_obs[i] = next_obs | |
| batch_done[i] = done | |
| for trajectory in trajectories: | |
| dataset.append_trajectory(trajectory) | |
| print(trajectory.transitions[-1].next_observation) | |
| dataset.check_consistency() | |
| if verbose: | |
| print("Data Coollection is Complete. Returns: \n", dataset.get_all_trajectory_returns(), "\n with mean: ",dataset.mean_trajectory_return(), "\n" ) | |
| return dataset | |