def experiment(variant):
env_params = variant['env_params']
env = MultiTaskSawyerXYZReachingEnv(env_params)
tdm_normalizer = TdmNormalizer(
env,
vectorized=True,
max_tau=variant['ddpg_tdm_kwargs']['tdm_kwargs']['max_tau'],
)
qf = TdmQf(
env=env,
vectorized=True,
hidden_sizes=[variant['hidden_sizes'], variant['hidden_sizes']],
structure='norm_difference',
tdm_normalizer=tdm_normalizer,
)
policy = TdmPolicy(
env=env,
hidden_sizes=[variant['hidden_sizes'], variant['hidden_sizes']],
tdm_normalizer=tdm_normalizer,
)
es = OUStrategy(
action_space=env.action_space,
**variant['es_kwargs']
)
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = HerReplayBuffer(
env=env,
**variant['her_replay_buffer_kwargs']
)
qf_criterion = variant['qf_criterion_class']()
ddpg_tdm_kwargs = copy.deepcopy(variant['ddpg_tdm_kwargs'])
ddpg_tdm_kwargs['ddpg_kwargs']['qf_criterion'] = qf_criterion
algorithm = TdmDdpg(
env,
qf=qf,
replay_buffer=replay_buffer,
policy=policy,
exploration_policy=exploration_policy,
**variant['ddpg_tdm_kwargs']
)
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
def experiment(variant):
vectorized = variant['sac_tdm_kwargs']['tdm_kwargs']['vectorized']
# env = NormalizedBoxEnv(variant['env_class'](**variant['env_kwargs']))
env = NormalizedBoxEnv(variant['env_class'](**variant['env_kwargs']))
max_tau = variant['sac_tdm_kwargs']['tdm_kwargs']['max_tau']
tdm_normalizer = TdmNormalizer(
env,
vectorized,
max_tau=max_tau,
**variant['tdm_normalizer_kwargs']
)
qf = TdmQf(
env=env,
vectorized=vectorized,
tdm_normalizer=tdm_normalizer,
**variant['qf_kwargs']
)
vf = TdmVf(
env=env,
vectorized=vectorized,
tdm_normalizer=tdm_normalizer,
**variant['vf_kwargs']
)
policy = StochasticTdmPolicy(
env=env,
tdm_normalizer=tdm_normalizer,
**variant['policy_kwargs']
)
replay_buffer = HerReplayBuffer(
env=env,
**variant['her_replay_buffer_kwargs']
)
algorithm = TdmSac(
env=env,
policy=policy,
qf=qf,
vf=vf,
replay_buffer=replay_buffer,
**variant['sac_tdm_kwargs']
)
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
vectorized = variant['vectorized']
norm_order = variant['norm_order']
variant['ddpg_tdm_kwargs']['tdm_kwargs']['vectorized'] = vectorized
variant['ddpg_tdm_kwargs']['tdm_kwargs']['norm_order'] = norm_order
env = NormalizedBoxEnv(variant['env_class'](**variant['env_kwargs']))
max_tau = variant['ddpg_tdm_kwargs']['tdm_kwargs']['max_tau']
tdm_normalizer = TdmNormalizer(env,
vectorized,
max_tau=max_tau,
**variant['tdm_normalizer_kwargs'])
qf = TdmQf(env=env,
vectorized=vectorized,
norm_order=norm_order,
tdm_normalizer=tdm_normalizer,
**variant['qf_kwargs'])
policy = TdmPolicy(env=env,
tdm_normalizer=tdm_normalizer,
**variant['policy_kwargs'])
es = OUStrategy(action_space=env.action_space, **variant['es_kwargs'])
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = HerReplayBuffer(env=env,
**variant['her_replay_buffer_kwargs'])
qf_criterion = variant['qf_criterion_class'](
**variant['qf_criterion_kwargs'])
ddpg_tdm_kwargs = variant['ddpg_tdm_kwargs']
ddpg_tdm_kwargs['ddpg_kwargs']['qf_criterion'] = qf_criterion
ddpg_tdm_kwargs['tdm_kwargs']['tdm_normalizer'] = tdm_normalizer
algorithm = TdmDdpg(env,
qf=qf,
replay_buffer=replay_buffer,
policy=policy,
exploration_policy=exploration_policy,
**variant['ddpg_tdm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env = NormalizedBoxEnv(variant['env_class'](**variant['env_kwargs']))
observation_dim = int(np.prod(env.observation_space.low.shape))
action_dim = int(np.prod(env.action_space.low.shape))
obs_normalizer = TorchFixedNormalizer(observation_dim)
goal_normalizer = TorchFixedNormalizer(env.goal_dim)
action_normalizer = TorchFixedNormalizer(action_dim)
distance_normalizer = TorchFixedNormalizer(env.goal_dim)
tdm_normalizer = TdmNormalizer(env,
obs_normalizer=obs_normalizer,
goal_normalizer=goal_normalizer,
action_normalizer=action_normalizer,
distance_normalizer=distance_normalizer,
max_tau=1,
**variant['tdm_normalizer_kwargs'])
qf = HerQFunction(env=env, **variant['qf_kwargs'])
policy = HerPolicy(env=env,
tdm_normalizer=tdm_normalizer,
**variant['policy_kwargs'])
es = OUStrategy(action_space=env.action_space, **variant['es_kwargs'])
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = HerReplayBuffer(env=env,
**variant['her_replay_buffer_kwargs'])
qf_criterion = variant['qf_criterion_class'](
**variant['qf_criterion_kwargs'])
ddpg_tdm_kwargs = variant['ddpg_tdm_kwargs']
ddpg_tdm_kwargs['ddpg_kwargs']['qf_criterion'] = qf_criterion
algorithm = HER(env,
qf=qf,
replay_buffer=replay_buffer,
policy=policy,
exploration_policy=exploration_policy,
**variant['ddpg_tdm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()
def experiment(variant):
env_params = variant['env_params']
env = MultiTaskSawyerXYZReachingEnv(**env_params)
max_tau = variant['ddpg_tdm_kwargs']['tdm_kwargs']['max_tau']
tdm_normalizer = TdmNormalizer(
env,
vectorized=True,
max_tau=max_tau,
)
qf = TdmQf(env=env,
vectorized=True,
norm_order=2,
tdm_normalizer=tdm_normalizer,
**variant['qf_kwargs'])
policy = TdmPolicy(env=env,
tdm_normalizer=tdm_normalizer,
**variant['policy_kwargs'])
es = OUStrategy(action_space=env.action_space, **variant['es_kwargs'])
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=policy,
)
replay_buffer = HerReplayBuffer(env=env,
**variant['her_replay_buffer_kwargs'])
qf_criterion = variant['qf_criterion_class']()
ddpg_tdm_kwargs = variant['ddpg_tdm_kwargs']
ddpg_tdm_kwargs['ddpg_kwargs']['qf_criterion'] = qf_criterion
ddpg_tdm_kwargs['tdm_kwargs']['tdm_normalizer'] = tdm_normalizer
algorithm = TdmDdpg(env,
qf=qf,
replay_buffer=replay_buffer,
policy=policy,
exploration_policy=exploration_policy,
**variant['ddpg_tdm_kwargs'])
if ptu.gpu_enabled():
algorithm.cuda()
algorithm.train()
def experiment(variant):
vectorized = variant['sac_tdm_kwargs']['tdm_kwargs']['vectorized']
env = NormalizedBoxEnv(variant['env_class'](**variant['env_kwargs']))
max_tau = variant['sac_tdm_kwargs']['tdm_kwargs']['max_tau']
qf = TdmQf(env, vectorized=vectorized, **variant['qf_kwargs'])
tdm_normalizer = TdmNormalizer(env,
vectorized,
max_tau=max_tau,
**variant['tdm_normalizer_kwargs'])
implicit_model = TdmToImplicitModel(
env,
qf,
tau=0,
)
vf = TdmVf(env=env,
vectorized=vectorized,
tdm_normalizer=tdm_normalizer,
**variant['vf_kwargs'])
policy = StochasticTdmPolicy(env=env,
tdm_normalizer=tdm_normalizer,
**variant['policy_kwargs'])
replay_buffer = HerReplayBuffer(env=env,
**variant['her_replay_buffer_kwargs'])
goal_slice = env.ob_to_goal_slice
lbfgs_mpc_controller = TdmLBfgsBCMC(implicit_model,
env,
goal_slice=goal_slice,
multitask_goal_slice=goal_slice,
**variant['mpc_controller_kwargs'])
state_only_mpc_controller = TdmLBfgsBStateOnlyCMC(
vf,
policy,
env,
goal_slice=goal_slice,
multitask_goal_slice=goal_slice,
**variant['state_only_mpc_controller_kwargs'])
es = GaussianStrategy(action_space=env.action_space,
**variant['es_kwargs'])
if variant['explore_with'] == 'TdmLBfgsBCMC':
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=lbfgs_mpc_controller,
)
variant['sac_tdm_kwargs']['base_kwargs']['exploration_policy'] = (
exploration_policy)
elif variant['explore_with'] == 'TdmLBfgsBStateOnlyCMC':
exploration_policy = PolicyWrappedWithExplorationStrategy(
exploration_strategy=es,
policy=state_only_mpc_controller,
)
variant['sac_tdm_kwargs']['base_kwargs']['exploration_policy'] = (
exploration_policy)
if variant['eval_with'] == 'TdmLBfgsBCMC':
variant['sac_tdm_kwargs']['base_kwargs']['eval_policy'] = (
lbfgs_mpc_controller)
elif variant['eval_with'] == 'TdmLBfgsBStateOnlyCMC':
variant['sac_tdm_kwargs']['base_kwargs']['eval_policy'] = (
state_only_mpc_controller)
algorithm = TdmSac(env=env,
policy=policy,
qf=qf,
vf=vf,
replay_buffer=replay_buffer,
**variant['sac_tdm_kwargs'])
algorithm.to(ptu.device)
algorithm.train()