def train_and_eval(trial: optuna.Trial, ex_dir: str, seed: [int, None]):
"""
Objective function for the Optuna `Study` to maximize.
.. note::
Optuna expects only the `trial` argument, thus we use `functools.partial` to sneak in custom arguments.
:param trial: Optuna Trial object for hyper-parameter optimization
:param ex_dir: experiment's directory, i.e. the parent directory for all trials in this study
:param seed: seed value for the random number generators, pass `None` for no seeding
:return: objective function value
"""
# Synchronize seeds between Optuna trials
pyrado.set_seed(seed)
# Environment
env_hparams = dict(dt=1 / 100., max_steps=600)
env = QQubeSim(**env_hparams)
env = ActNormWrapper(env)
# Policy
policy_hparam = dict(
shared_hidden_sizes=trial.suggest_categorical(
'shared_hidden_sizes_policy',
[[16, 16], [32, 32], [64, 64], [16, 16, 16], [32, 32, 32]]),
shared_hidden_nonlin=fcn_from_str(
trial.suggest_categorical('shared_hidden_nonlin_policy',
['to_tanh', 'to_relu'])),
)
policy = TwoHeadedFNNPolicy(spec=env.spec, **policy_hparam)
# Critic
q_fcn_hparam = dict(
hidden_sizes=trial.suggest_categorical(
'hidden_sizes_critic',
[[16, 16], [32, 32], [64, 64], [16, 16, 16], [32, 32, 32]]),
hidden_nonlin=fcn_from_str(
trial.suggest_categorical('hidden_nonlin_critic',
['to_tanh', 'to_relu'])),
)
obsact_space = BoxSpace.cat([env.obs_space, env.act_space])
q_fcn_1 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace),
**q_fcn_hparam)
q_fcn_2 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace),
**q_fcn_hparam)
# Algorithm
algo_hparam = dict(
num_sampler_envs=1, # parallelize via optuna n_jobs
max_iter=100 * env.max_steps,
min_steps=trial.suggest_categorical(
'min_steps_algo', [1]), # , 10, env.max_steps, 10*env.max_steps
memory_size=trial.suggest_loguniform('memory_size_algo',
1e2 * env.max_steps,
1e4 * env.max_steps),
tau=trial.suggest_uniform('tau_algo', 0.99, 1.),
alpha_init=trial.suggest_uniform('alpha_init_algo', 0.1, 0.9),
learn_alpha=trial.suggest_categorical('learn_alpha_algo',
[True, False]),
standardize_rew=trial.suggest_categorical('standardize_rew_algo',
[False]),
gamma=trial.suggest_uniform('gamma_algo', 0.99, 1.),
target_update_intvl=trial.suggest_categorical(
'target_update_intvl_algo', [1, 5]),
num_batch_updates=trial.suggest_categorical('num_batch_updates_algo',
[1, 5]),
batch_size=trial.suggest_categorical('batch_size_algo',
[128, 256, 512]),
lr=trial.suggest_loguniform('lr_algo', 1e-5, 1e-3),
)
csv_logger = create_csv_step_logger(
osp.join(ex_dir, f'trial_{trial.number}'))
algo = SAC(ex_dir,
env,
policy,
q_fcn_1,
q_fcn_2,
**algo_hparam,
logger=csv_logger)
# Train without saving the results
algo.train(snapshot_mode='latest', seed=seed)
# Evaluate
min_rollouts = 1000
sampler = ParallelSampler(
env, policy, num_envs=1,
min_rollouts=min_rollouts) # parallelize via optuna n_jobs
ros = sampler.sample()
mean_ret = sum([r.undiscounted_return() for r in ros]) / min_rollouts
return mean_ret
def train_and_eval(trial: optuna.Trial, study_dir: str, seed: int):
"""
Objective function for the Optuna `Study` to maximize.
.. note::
Optuna expects only the `trial` argument, thus we use `functools.partial` to sneak in custom arguments.
:param trial: Optuna Trial object for hyper-parameter optimization
:param study_dir: the parent directory for all trials in this study
:param seed: seed value for the random number generators, pass `None` for no seeding
:return: objective function value
"""
# Synchronize seeds between Optuna trials
pyrado.set_seed(seed)
# Environment
env_hparams = dict(physicsEngine="Bullet", dt=1 / 100.0, max_steps=500)
env = BallOnPlate2DSim(**env_hparams)
env = ActNormWrapper(env)
# Policy
policy_hparam = dict(
shared_hidden_sizes=trial.suggest_categorical(
"shared_hidden_sizes_policy", [(16, 16), (32, 32), (64, 64),
(16, 16, 16), (32, 32, 32)]),
shared_hidden_nonlin=fcn_from_str(
trial.suggest_categorical("shared_hidden_nonlin_policy",
["to_tanh", "to_relu"])),
)
policy = TwoHeadedFNNPolicy(spec=env.spec, **policy_hparam)
# Critic
qfcn_hparam = dict(
hidden_sizes=trial.suggest_categorical("hidden_sizes_critic",
[(16, 16), (32, 32), (64, 64),
(16, 16, 16), (32, 32, 32)]),
hidden_nonlin=fcn_from_str(
trial.suggest_categorical("hidden_nonlin_critic",
["to_tanh", "to_relu"])),
)
obsact_space = BoxSpace.cat([env.obs_space, env.act_space])
qfcn_1 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace),
**qfcn_hparam)
qfcn_2 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace),
**qfcn_hparam)
# Algorithm
algo_hparam = dict(
num_workers=1, # parallelize via optuna n_jobs
max_iter=100 * env.max_steps,
min_steps=trial.suggest_categorical(
"min_steps_algo", [1]), # 10, env.max_steps, 10*env.max_steps
memory_size=trial.suggest_loguniform("memory_size_algo",
1e2 * env.max_steps,
1e4 * env.max_steps),
tau=trial.suggest_uniform("tau_algo", 0.99, 1.0),
ent_coeff_init=trial.suggest_uniform("ent_coeff_init_algo", 0.1, 0.9),
learn_ent_coeff=trial.suggest_categorical("learn_ent_coeff_algo",
[True, False]),
standardize_rew=trial.suggest_categorical("standardize_rew_algo",
[False]),
gamma=trial.suggest_uniform("gamma_algo", 0.99, 1.0),
target_update_intvl=trial.suggest_categorical(
"target_update_intvl_algo", [1, 5]),
num_updates_per_step=trial.suggest_categorical(
"num_batch_updates_algo", [1, 5]),
batch_size=trial.suggest_categorical("batch_size_algo",
[128, 256, 512]),
lr=trial.suggest_loguniform("lr_algo", 1e-5, 1e-3),
)
csv_logger = create_csv_step_logger(
osp.join(study_dir, f"trial_{trial.number}"))
algo = SAC(study_dir,
env,
policy,
qfcn_1,
qfcn_2,
**algo_hparam,
logger=csv_logger)
# Train without saving the results
algo.train(snapshot_mode="latest", seed=seed)
# Evaluate
min_rollouts = 1000
sampler = ParallelRolloutSampler(
env, policy, num_workers=1,
min_rollouts=min_rollouts) # parallelize via optuna n_jobs
ros = sampler.sample()
mean_ret = sum([r.undiscounted_return() for r in ros]) / min_rollouts
return mean_ret
env = OneMassOscillatorSim(**env_hparams, task_args=dict(task_args=dict(state_des=np.array([0.5, 0]))))
env = ActNormWrapper(env)
# Policy
policy_hparam = dict(
shared_hidden_sizes=[32, 32],
shared_hidden_nonlin=to.relu,
)
policy = TwoHeadedFNNPolicy(spec=env.spec, **policy_hparam)
# Critic
qfcn_hparam = dict(
hidden_sizes=[32, 32],
hidden_nonlin=to.relu
)
obsact_space = BoxSpace.cat([env.obs_space, env.act_space])
qfcn_1 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace), **qfcn_hparam)
qfcn_2 = FNNPolicy(spec=EnvSpec(obsact_space, ValueFunctionSpace), **qfcn_hparam)
# Algorithm
algo_hparam = dict(
max_iter=1000*env.max_steps,
memory_size=100*env.max_steps,
gamma=0.995,
num_batch_updates=1,
tau=0.995,
ent_coeff_init=0.2,
learn_ent_coeff=True,
target_update_intvl=5,
standardize_rew=False,
min_steps=1,
def test_snapshots_notmeta(ex_dir, env: SimEnv, policy, algo_class,
algo_hparam):
# Collect hyper-parameters, create algorithm, and train
common_hparam = dict(max_iter=1, num_workers=1)
common_hparam.update(algo_hparam)
if issubclass(algo_class, ActorCritic):
common_hparam.update(
min_rollouts=3,
critic=GAE(
vfcn=FNNPolicy(spec=EnvSpec(env.obs_space, ValueFunctionSpace),
hidden_sizes=[16, 16],
hidden_nonlin=to.tanh)),
)
elif issubclass(algo_class, ParameterExploring):
common_hparam.update(num_init_states_per_domain=1)
elif issubclass(algo_class, (DQL, SAC)):
common_hparam.update(memory_size=1000,
num_updates_per_step=2,
gamma=0.99,
min_rollouts=1)
fnn_hparam = dict(hidden_sizes=[8, 8], hidden_nonlin=to.tanh)
if issubclass(algo_class, DQL):
# Override the setting
env = BallOnBeamDiscSim(env.dt, env.max_steps)
net = FNN(
input_size=DiscreteActQValPolicy.get_qfcn_input_size(env.spec),
output_size=DiscreteActQValPolicy.get_qfcn_output_size(),
**fnn_hparam,
)
policy = DiscreteActQValPolicy(spec=env.spec, net=net)
else:
# Override the setting
env = ActNormWrapper(env)
policy = TwoHeadedGRUPolicy(env.spec,
shared_hidden_size=8,
shared_num_recurrent_layers=1)
obsact_space = BoxSpace.cat([env.obs_space, env.act_space])
common_hparam.update(qfcn_1=FNNPolicy(
spec=EnvSpec(obsact_space, ValueFunctionSpace), **fnn_hparam))
common_hparam.update(qfcn_2=FNNPolicy(
spec=EnvSpec(obsact_space, ValueFunctionSpace), **fnn_hparam))
else:
raise NotImplementedError
# Simulate training
algo = algo_class(ex_dir, env, policy, **common_hparam)
algo.policy.param_values += to.tensor([42.0])
if isinstance(algo, ActorCritic):
algo.critic.vfcn.param_values += to.tensor([42.0])
# Save and load
algo.save_snapshot(meta_info=None)
algo_loaded = Algorithm.load_snapshot(load_dir=ex_dir)
assert isinstance(algo_loaded, Algorithm)
policy_loaded = algo_loaded.policy
if isinstance(algo, ActorCritic):
critic_loaded = algo_loaded.critic
# Check
assert all(algo.policy.param_values == policy_loaded.param_values)
if isinstance(algo, ActorCritic):
assert all(
algo.critic.vfcn.param_values == critic_loaded.vfcn.param_values)
# Load the experiment. Since we did not save any hyper-parameters, we ignore the errors when loading.
env, policy, extra = load_experiment(ex_dir)
assert isinstance(env, Env)
assert isinstance(policy, Policy)
assert isinstance(extra, dict)
