models/embodied/envs/bsuite.py

import time

import embodied
import numpy as np


class BSuite(embodied.Env):

  def __init__(self, task):
    print(
        'Warning: BSuite result logging is stateful and therefore training ' +
        'runs cannot be interrupted or restarted.')
    np.int = int  # Patch deprecated Numpy alias used inside BSuite.
    from . import from_dm
    if '/' not in task:
      task = f'{task}/0'
    import bsuite
    env = bsuite.from_checkpoint_id(task)
    self.num_episodes = 0
    self.max_episodes = env.bsuite_num_episodes
    self.exit_after = None
    env = from_dm.FromDM(env)
    env = embodied.wrappers.ForceDtypes(env)
    env = embodied.wrappers.FlattenTwoDimObs(env)
    self.env = env

  @property
  def obs_space(self):
    return self.env.obs_space

  @property
  def act_space(self):
    return self.env.act_space

  def step(self, action):
    obs = self.env.step(action)
    if obs['is_last']:
      self.num_episodes += 1
    if self.num_episodes >= self.max_episodes:
      # After reaching the target number of episodes, continue running for 10
      # minutes to make sure logs are flushed and then raise an exception to
      # terminate the program.
      if not self.exit_after:
        self.exit_after = time.time() + 600
      if time.time() > self.exit_after:
        if self.xm:
          wu = self.xm.get_current_work_unit()
          wu.stop(mark_as_completed=True, message='BSuite run complete')
        else:
          raise RuntimeError('BSuite run complete')
    return obs