'trainer_kwargs.awr_sample_actions': [
False,
],
'trainer_kwargs.clip_score': [
2,
],
'trainer_kwargs.awr_min_q': [
True,
],
'trainer_kwargs.reward_transform_kwargs': [
None,
],
'trainer_kwargs.terminal_transform_kwargs': [
dict(m=0, b=0),
],
'qf_kwargs.output_activation': [Clamp(max=0)],
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
env_type = variant['env_type']
eval_goals = 'sasha/presampled_goals/affordances/combined/{0}_goals.pkl'.format(
env_type)
variant['presampled_goal_kwargs']['eval_goals'] = eval_goals
if env_type in ['top_drawer', 'bottom_drawer']:
variant['env_class'] = SawyerRigAffordancesV0
policy_class=GaussianPolicy,
policy_kwargs=dict(
hidden_sizes=[
256,
256,
],
max_log_std=0,
min_log_std=-6,
std_architecture="values",
),
qf_kwargs=dict(
hidden_sizes=[
256,
256,
],
output_activation=Clamp(max=0), # rewards are <= 0
),
version="normal",
collection_mode="batch",
trainer_kwargs=dict(
discount=0.99,
soft_target_tau=5e-3,
target_update_period=1,
policy_lr=3e-4,
qf_lr=3e-4,
reward_scale=1,
beta=1,
use_automatic_entropy_tuning=False,
alpha=0,
compute_bc=False,
awr_min_q=True,