def _set_env_attr(self, env, player_id, attr_chain, value):
"""Allows us to set an arbitrary attribute of the environment, e.g. attr_chain can be unwrapped.foo.bar"""
assert player_id == self.player_id
attrs = attr_chain.split('.')
curr_attr = env
try:
for attr_name in attrs[:-1]:
curr_attr = getattr(curr_attr, attr_name)
except AttributeError:
log.error('Env does not have an attribute %s', attr_chain)
attr_to_set = attrs[-1]
setattr(curr_attr, attr_to_set, value)
def reset(self):
self._ensure_initialized()
if self.record_to is not None and not self.is_multiplayer:
# does not work in multiplayer (uses different mechanism)
if not os.path.exists(self.record_to):
os.makedirs(self.record_to)
demo_path = self.demo_path(self._num_episodes)
log.warning('Recording episode demo to %s', demo_path)
self.game.new_episode(demo_path)
else:
if self._num_episodes > 0:
# no demo recording (default)
self.game.new_episode()
self.state = self.game.get_state()
img = None
try:
img = self.state.screen_buffer
except AttributeError:
# sometimes Doom does not return screen buffer at all??? Rare bug
pass
if img is None:
log.error(
'Game returned None screen buffer! This is not supposed to happen!'
)
img = self._black_screen()
# Swap current and previous histogram
if self.current_histogram is not None and self.previous_histogram is not None:
swap = self.current_histogram
self.current_histogram = self.previous_histogram
self.previous_histogram = swap
self.current_histogram.fill(0)
self._actions_flattened = None
self._last_episode_info = copy.deepcopy(self._prev_info)
self._prev_info = None
self._num_episodes += 1
return np.transpose(img, (1, 2, 0))
def _selected_weapon_rewards(self, selected_weapon, selected_weapon_ammo,
deltas):
# we must keep the weapon ready for a certain number of frames to get rewards
unholstered = len(self.selected_weapon) > 4 and all(
sw == selected_weapon for sw in self.selected_weapon)
reward = 0.0
if selected_weapon_ammo > 0 and unholstered:
try:
reward = self.reward_shaping_scheme['selected_weapon'][
f'SELECTED{weapon}']
except KeyError:
log.error('%r', self.reward_shaping_scheme)
log.error('%r', selected_weapon)
weapon_key = f'weapon{selected_weapon}'
deltas.append((weapon_key, reward))
self.reward_structure[weapon_key] = self.reward_structure.get(
weapon_key, 0.0) + reward
return reward
def read_seeds_file(filename, has_keys):
seeds = []
with open(filename, 'r') as seed_file:
lines = seed_file.readlines()
for line in lines:
try:
if has_keys:
seed, cache_key = line.split(' ')
else:
seed = line
seed = int(seed)
seeds.append(seed)
except Exception:
log.error(
'Could not read seed value from the file! File potentially corrupted'
)
log.exception('Exception when reading seeds file')
return seeds
def sample(self, proc_idx):
# workers should ignore Ctrl+C because the termination is handled in the event loop by a special msg
signal.signal(signal.SIGINT, signal.SIG_IGN)
timing = Timing()
psutil.Process().nice(10)
num_envs = len(DMLAB30_LEVELS_THAT_USE_LEVEL_CACHE)
assert self.cfg.num_workers % num_envs == 0, f'should have an integer number of workers per env, e.g. {1 * num_envs}, {2 * num_envs}, etc...'
assert self.cfg.num_envs_per_worker == 1, 'use populate_cache with 1 env per worker'
with timing.timeit('env_init'):
env_key = 'env'
env_desired_num_levels = 0
global_env_id = proc_idx * self.cfg.num_envs_per_worker
env_config = AttrDict(worker_index=proc_idx,
vector_index=0,
env_id=global_env_id)
env = create_env(self.cfg.env, cfg=self.cfg, env_config=env_config)
env.seed(global_env_id)
# this is to track the performance for individual DMLab levels
if hasattr(env.unwrapped, 'level_name'):
env_key = env.unwrapped.level_name
env_level = env.unwrapped.level
approx_num_episodes_per_1b_frames = DMLAB30_APPROX_NUM_EPISODES_PER_BILLION_FRAMES[
env_key]
num_billions = DESIRED_TRAINING_LENGTH / int(1e9)
num_workers_for_env = self.cfg.num_workers // num_envs
env_desired_num_levels = int(
(approx_num_episodes_per_1b_frames * num_billions) /
num_workers_for_env)
env_num_levels_generated = len(
dmlab_level_cache.DMLAB_GLOBAL_LEVEL_CACHE[0].
all_seeds[env_level]) // num_workers_for_env
log.warning('Worker %d (env %s) generated %d/%d levels!',
proc_idx, env_key, env_num_levels_generated,
env_desired_num_levels)
time.sleep(4)
env.reset()
env_uses_level_cache = env.unwrapped.env_uses_level_cache
self.report_queue.put(dict(proc_idx=proc_idx, finished_reset=True))
self.start_event.wait()
try:
with timing.timeit('work'):
last_report = last_report_frames = total_env_frames = 0
while not self.terminate.value and total_env_frames < self.cfg.sample_env_frames_per_worker:
action = env.action_space.sample()
with timing.add_time(f'{env_key}.step'):
env.step(action)
total_env_frames += 1
with timing.add_time(f'{env_key}.reset'):
env.reset()
env_num_levels_generated += 1
log.debug('Env %s done %d/%d resets', env_key,
env_num_levels_generated,
env_desired_num_levels)
if env_num_levels_generated >= env_desired_num_levels:
log.debug('%s finished %d/%d resets, sleeping...',
env_key, env_num_levels_generated,
env_desired_num_levels)
time.sleep(30) # free up CPU time for other envs
# if env does not use level cache, there is no need to run it
# let other workers proceed
if not env_uses_level_cache:
log.debug('Env %s does not require cache, sleeping...',
env_key)
time.sleep(200)
with timing.add_time('report'):
now = time.time()
if now - last_report > self.report_every_sec:
last_report = now
frames_since_last_report = total_env_frames - last_report_frames
last_report_frames = total_env_frames
self.report_queue.put(
dict(proc_idx=proc_idx,
env_frames=frames_since_last_report))
if get_free_disk_space_mb() < 3 * 1024:
log.error('Not enough disk space! %d',
get_free_disk_space_mb())
time.sleep(200)
except:
log.exception('Unknown exception')
log.error('Unknown exception in worker %d, terminating...',
proc_idx)
self.report_queue.put(dict(proc_idx=proc_idx, crash=True))
time.sleep(proc_idx * 0.1 + 0.1)
log.info('Process %d finished sampling. Timing: %s', proc_idx, timing)
env.close()