def objective(self, trial: optuna.Trial) -> float:
kwargs = self._hyperparams.copy()
# Hack to use DDPG/TD3 noise sampler
trial.n_actions = self.n_actions
# Hack when using HerReplayBuffer
trial.using_her_replay_buffer = kwargs.get(
"replay_buffer_class") == HerReplayBuffer
if trial.using_her_replay_buffer:
trial.her_kwargs = kwargs.get("replay_buffer_kwargs", {})
# Sample candidate hyperparameters
sampled_hyperparams = HYPERPARAMS_SAMPLER[self.algo](trial,
self.n_envs)
kwargs.update(sampled_hyperparams)
n_envs = 1 if self.algo == "ars" else self.n_envs
env = self.create_envs(n_envs, no_log=True)
# By default, do not activate verbose output to keep
# stdout clean with only the trials results
trial_verbosity = 0
# Activate verbose mode for the trial in debug mode
# See PR #214
if self.verbose >= 2:
trial_verbosity = self.verbose
model = ALGOS[self.algo](
env=env,
tensorboard_log=None,
# We do not seed the trial
seed=None,
verbose=trial_verbosity,
device=self.device,
**kwargs,
)
eval_env = self.create_envs(n_envs=self.n_eval_envs, eval_env=True)
optuna_eval_freq = int(self.n_timesteps / self.n_evaluations)
# Account for parallel envs
optuna_eval_freq = max(optuna_eval_freq // self.n_envs, 1)
# Use non-deterministic eval for Atari
path = None
if self.optimization_log_path is not None:
path = os.path.join(self.optimization_log_path,
f"trial_{str(trial.number)}")
callbacks = get_callback_list({"callback": self.specified_callbacks})
eval_callback = TrialEvalCallback(
eval_env,
trial,
best_model_save_path=path,
log_path=path,
n_eval_episodes=self.n_eval_episodes,
eval_freq=optuna_eval_freq,
deterministic=self.deterministic_eval,
)
callbacks.append(eval_callback)
learn_kwargs = {}
# Special case for ARS
if self.algo == "ars" and self.n_envs > 1:
learn_kwargs["async_eval"] = AsyncEval([
lambda: self.create_envs(n_envs=1, no_log=True)
for _ in range(self.n_envs)
], model.policy)
try:
model.learn(self.n_timesteps, callback=callbacks, **learn_kwargs)
# Free memory
model.env.close()
eval_env.close()
except (AssertionError, ValueError, EOFError, BrokenPipeError) as e:
# Sometimes, random hyperparams can generate NaN
# Free memory
# if using subprocvecenv, the following calls can produce errors themselves
traceback.print_exc()
try:
model.env.close()
except:
pass
try:
eval_env.close()
except:
pass
# Prune hyperparams that generate NaNs
print(e)
print("============")
print("Sampled hyperparams:")
pprint(sampled_hyperparams)
return 0
raise optuna.exceptions.TrialPruned()
is_pruned = eval_callback.is_pruned
reward = eval_callback.last_mean_reward
del model.env, eval_env
del model
if is_pruned:
raise optuna.exceptions.TrialPruned()
return reward
def objective(self, trial: optuna.Trial) -> float:
kwargs = self._hyperparams.copy()
# Hack to use DDPG/TD3 noise sampler
trial.n_actions = self.n_actions
# Hack when using HerReplayBuffer
trial.using_her_replay_buffer = kwargs.get(
"replay_buffer_class") == HerReplayBuffer
if trial.using_her_replay_buffer:
trial.her_kwargs = kwargs.get("replay_buffer_kwargs", {})
# Sample candidate hyperparameters
sampled_hyperparams = HYPERPARAMS_SAMPLER[self.algo](trial)
kwargs.update(sampled_hyperparams)
model = ALGOS[self.algo](
env=self.create_envs(self.n_envs, no_log=True),
tensorboard_log=None,
# We do not seed the trial
seed=None,
verbose=0,
**kwargs,
)
model.trial = trial
eval_env = self.create_envs(n_envs=self.n_eval_envs, eval_env=True)
optuna_eval_freq = int(self.n_timesteps / self.n_evaluations)
# Account for parallel envs
optuna_eval_freq = max(optuna_eval_freq // model.get_env().num_envs, 1)
# Use non-deterministic eval for Atari
path = None
if self.optimization_log_path is not None:
path = os.path.join(self.optimization_log_path,
f"trial_{str(trial.number)}")
callbacks = get_callback_list({"callback": self.specified_callbacks})
eval_callback = TrialEvalCallback(
eval_env,
trial,
best_model_save_path=path,
log_path=path,
n_eval_episodes=self.n_eval_episodes,
eval_freq=optuna_eval_freq,
deterministic=self.deterministic_eval,
)
callbacks.append(eval_callback)
try:
model.learn(self.n_timesteps, callback=callbacks)
# Free memory
model.env.close()
eval_env.close()
except (AssertionError, ValueError) as e:
# Sometimes, random hyperparams can generate NaN
# Free memory
model.env.close()
eval_env.close()
# Prune hyperparams that generate NaNs
print(e)
print("============")
print("Sampled hyperparams:")
pprint(sampled_hyperparams)
raise optuna.exceptions.TrialPruned()
is_pruned = eval_callback.is_pruned
reward = eval_callback.last_mean_reward
del model.env, eval_env
del model
if is_pruned:
raise optuna.exceptions.TrialPruned()
return reward
def objective(self, trial: optuna.Trial) -> float:
kwargs = self._hyperparams.copy()
# Hack to use DDPG/TD3 noise sampler
trial.n_actions = self.n_actions
# Hack when using HerReplayBuffer
trial.using_her_replay_buffer = kwargs.get(
"replay_buffer_class") == HerReplayBuffer
if trial.using_her_replay_buffer:
trial.her_kwargs = kwargs.get("replay_buffer_kwargs", {})
# Sample candidate hyperparameters
kwargs.update(HYPERPARAMS_SAMPLER[self.algo](trial))
model = ALGOS[self.algo](
env=self.create_envs(self.n_envs, no_log=True),
tensorboard_log=None,
# We do not seed the trial
seed=None,
verbose=0,
**kwargs,
)
model.trial = trial
eval_env = self.create_envs(n_envs=1, eval_env=True)
eval_freq = int(self.n_timesteps / self.n_evaluations)
# Account for parallel envs
eval_freq_ = max(eval_freq // model.get_env().num_envs, 1)
# Use non-deterministic eval for Atari
eval_callback = TrialEvalCallback(
eval_env,
trial,
n_eval_episodes=self.n_eval_episodes,
eval_freq=eval_freq_,
deterministic=self.deterministic_eval,
)
try:
model.learn(self.n_timesteps, callback=eval_callback)
# Free memory
model.env.close()
eval_env.close()
except AssertionError as e:
# Sometimes, random hyperparams can generate NaN
# Free memory
model.env.close()
eval_env.close()
# Prune hyperparams that generate NaNs
print(e)
raise optuna.exceptions.TrialPruned()
is_pruned = eval_callback.is_pruned
reward = eval_callback.last_mean_reward
del model.env, eval_env
del model
if is_pruned:
raise optuna.exceptions.TrialPruned()
return reward
