def _run_multirun_job(cfg: DictConfig) -> float:
"""Runs a maze job which is part of a Hydra --multirun and returns the maximum mean reward of this run.
:param cfg: Hydra configuration for the rollout.
:return: The maximum mean reward achieved throughout the training run.
"""
# required for Hydra --multirun (e.g., sweeper)
clear_global_state()
try:
_run_job(cfg)
# when optimizing hyper parameters a single exception
# in one job should not break the entire experiment
except:
return float(np.finfo(np.float32).min)
# load tensorboard log and return maximum mean reward
# load tensorboard log
tf_summary_files = glob.glob("*events.out.tfevents*")
assert len(
tf_summary_files
) == 1, f"expected exactly 1 tensorflow summary file {tf_summary_files}"
events_df = tensorboard_to_pandas(tf_summary_files[0])
# compute maximum mean reward
max_mean_reward = np.max(
np.asarray(events_df.loc["train_BaseEnvEvents/reward/mean"]))
# Add hparams logging to tensorboard
metrics = [('train_BaseEnvEvents/reward/mean', max_mean_reward, 'max')]
manipulate_hparams_logging_for_exp('.', metrics, clear_hparams=False)
return float(max_mean_reward)
def run(self) -> None:
"""overwrite Process.run()"""
# clean any left-over state from the parent process
clear_global_state()
try:
result = self._target(*self._args, **self._kwargs)
self._cconn.send(result)
except Exception as e:
self._cconn.send(e)
raise e
def perform_algorithm_seeding_test(hydra_overrides: Dict[str, str]):
# Perform base run for comparison ----------------------------------------------------------------------------------
base_dir = os.path.abspath('.')
os.mkdir('./base_exp')
os.chdir('./base_exp')
# run training
with Timeout(seconds=60):
cfg = run_maze_job(hydra_overrides,
config_module="maze.conf",
config_name="conf_train")
# load tensorboard log
tf_summary_files = glob.glob("*events.out.tfevents*")
assert len(
tf_summary_files
) == 1, f"expected exactly 1 tensorflow summary file {tf_summary_files}"
events_df = tensorboard_to_pandas(tf_summary_files[0])
clear_global_state()
# Perform comparison run with same seeds ---------------------------------------------------------------------------
os.chdir(base_dir)
os.mkdir('./exp_pos')
os.chdir('./exp_pos')
hydra_overrides['seeding.agent_base_seed'] = cfg.seeding.agent_base_seed
hydra_overrides['seeding.env_base_seed'] = cfg.seeding.env_base_seed
# run training
with Timeout(seconds=60):
run_maze_job(hydra_overrides,
config_module="maze.conf",
config_name="conf_train")
# load tensorboard log
tf_summary_files = glob.glob("*events.out.tfevents*")
assert len(
tf_summary_files
) == 1, f"expected exactly 1 tensorflow summary file {tf_summary_files}"
events_df_2 = tensorboard_to_pandas(tf_summary_files[0])
clear_global_state()
del hydra_overrides['seeding.agent_base_seed']
del hydra_overrides['seeding.env_base_seed']
assert len(events_df) == len(events_df_2)
for idx, (key, epoch) in enumerate(events_df.index):
if 'time' in key:
continue
assert events_df_2.values[idx] == events_df.values[idx], \
f'Value not equal for key: {key} in epoch: {epoch}'
# Perform second comparison run with different seeds ---------------------------------------------------------------
os.chdir(base_dir)
os.mkdir('./exp_neg')
os.chdir('./exp_neg')
# run training
with Timeout(seconds=60):
run_maze_job(hydra_overrides,
config_module="maze.conf",
config_name="conf_train")
# load tensorboard log
tf_summary_files = glob.glob("*events.out.tfevents*")
assert len(
tf_summary_files
) == 1, f"expected exactly 1 tensorflow summary file {tf_summary_files}"
events_df_2 = tensorboard_to_pandas(tf_summary_files[0])
all_equal = True
for idx, (key, epoch) in enumerate(events_df.index):
if 'time' in key:
continue
all_equal = all_equal and events_df.values[idx] == events_df_2.values[
idx]
assert not all_equal, 'The resulting logs should not be all equal'
def fixture_clear_global_state():
"""clear global state before any test"""
clear_global_state()