batch_size=128,
lr=1e-3,
# weight_decay=0.01,
),
save_period=5,
),
renderer_kwargs=dict(
# create_image_format='HWC',
# output_image_format='CWH',
output_image_format='CHW',
flatten_image=True,
# normalize_image=False,
),
evaluation_goal_sampling_mode="reset_of_env",
exploration_goal_sampling_mode="vae_prior",
launcher_config=dict(unpack_variant=True, ))
search_space = {
"seed": range(5),
}
sweeper = hyp.DeterministicHyperparameterSweeper(
search_space,
default_parameters=variant,
)
variants = []
for variant in sweeper.iterate_hyperparameters():
variants.append(variant)
run_variants(rig_experiment, variants, process_args)
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
variant['env_kwargs']['env_type'] = env_type
if env_type == 'tray':
variant['env_class'] = SawyerRigMultiobjTrayV0
if env_type == 'pnp':
variant['env_class'] = SawyerRigMultiobjV0
variants.append(variant)
run_variants(awac_rig_experiment,
variants,
run_id=0,
process_args_fn=process_args) #HERE
# 'num_pybullet_objects':[None],
'policy_kwargs.min_log_std': [-6],
'trainer_kwargs.awr_weight': [1.0],
'trainer_kwargs.awr_use_mle_for_vf': [True, ],
'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']
obj = variant['env_kwargs']['object_subset'][0]
eval_goals = 'sasha/presampled_goals/affordances/combined/{0}_{1}_goals.pkl'.format(env_type, obj)
variant['presampled_goal_kwargs']['eval_goals'] = eval_goals
if env_type == 'tray':
variant['env_class'] = SawyerRigMultiobjTrayV0
if env_type == 'pnp':
variant['env_class'] = SawyerRigMultiobjV0
variants.append(variant)
run_variants(awac_rig_experiment, variants, run_id=50) #HERE