def get_sac_trainer(
env: OpenAIGymEnvironment,
rl_parameters: RLParameters,
trainer_parameters: SACTrainerParameters,
critic_training: FeedForwardParameters,
actor_training: FeedForwardParameters,
sac_value_training: Optional[FeedForwardParameters],
use_gpu: bool,
) -> SACTrainer:
assert rl_parameters == trainer_parameters.rl
state_dim = get_num_output_features(env.normalization)
action_dim = get_num_output_features(env.normalization_action)
q1_network = FullyConnectedParametricDQN(state_dim, action_dim,
critic_training.layers,
critic_training.activations)
q2_network = None
# TODO:
# if trainer_parameters.use_2_q_functions:
# q2_network = FullyConnectedParametricDQN(
# state_dim,
# action_dim,
# critic_training.layers,
# critic_training.activations,
# )
value_network = None
if sac_value_training:
value_network = FullyConnectedNetwork(
[state_dim] + sac_value_training.layers + [1],
sac_value_training.activations + ["linear"],
)
actor_network = GaussianFullyConnectedActor(state_dim, action_dim,
actor_training.layers,
actor_training.activations)
min_action_range_tensor_training = torch.full((1, action_dim), -1 + 1e-6)
max_action_range_tensor_training = torch.full((1, action_dim), 1 - 1e-6)
min_action_range_tensor_serving = (
torch.from_numpy(env.action_space.low).float().unsqueeze(
dim=0) # type: ignore
)
max_action_range_tensor_serving = (
torch.from_numpy(env.action_space.high).float().unsqueeze(
dim=0) # type: ignore
)
if use_gpu:
q1_network.cuda()
if q2_network:
q2_network.cuda()
if value_network:
value_network.cuda()
actor_network.cuda()
min_action_range_tensor_training = min_action_range_tensor_training.cuda(
)
max_action_range_tensor_training = max_action_range_tensor_training.cuda(
)
min_action_range_tensor_serving = min_action_range_tensor_serving.cuda(
)
max_action_range_tensor_serving = max_action_range_tensor_serving.cuda(
)
return SACTrainer(
q1_network,
actor_network,
trainer_parameters,
use_gpu=use_gpu,
value_network=value_network,
q2_network=q2_network,
min_action_range_tensor_training=min_action_range_tensor_training,
max_action_range_tensor_training=max_action_range_tensor_training,
min_action_range_tensor_serving=min_action_range_tensor_serving,
max_action_range_tensor_serving=max_action_range_tensor_serving,
)
def get_sac_trainer(
self,
env,
use_gpu,
use_2_q_functions=False,
logged_action_uniform_prior=True,
constrain_action_sum=False,
use_value_network=True,
use_alpha_optimizer=True,
entropy_temperature=None,
):
q_network_params = FeedForwardParameters(layers=[128, 64],
activations=["relu", "relu"])
value_network_params = FeedForwardParameters(
layers=[128, 64], activations=["relu", "relu"])
actor_network_params = FeedForwardParameters(
layers=[128, 64], activations=["relu", "relu"])
state_dim = get_num_output_features(env.normalization)
action_dim = get_num_output_features(
env.normalization_continuous_action)
q1_network = FullyConnectedParametricDQN(state_dim, action_dim,
q_network_params.layers,
q_network_params.activations)
q2_network = None
if use_2_q_functions:
q2_network = FullyConnectedParametricDQN(
state_dim,
action_dim,
q_network_params.layers,
q_network_params.activations,
)
if constrain_action_sum:
actor_network = DirichletFullyConnectedActor(
state_dim,
action_dim,
actor_network_params.layers,
actor_network_params.activations,
)
else:
actor_network = GaussianFullyConnectedActor(
state_dim,
action_dim,
actor_network_params.layers,
actor_network_params.activations,
)
value_network = None
if use_value_network:
value_network = FullyConnectedNetwork(
[state_dim] + value_network_params.layers + [1],
value_network_params.activations + ["linear"],
)
if use_gpu:
q1_network.cuda()
if q2_network:
q2_network.cuda()
if value_network:
value_network.cuda()
actor_network.cuda()
parameters = SACTrainerParameters(
rl=RLParameters(gamma=DISCOUNT, target_update_rate=0.5),
minibatch_size=self.minibatch_size,
q_network_optimizer=OptimizerParameters(),
value_network_optimizer=OptimizerParameters(),
actor_network_optimizer=OptimizerParameters(),
alpha_optimizer=OptimizerParameters()
if use_alpha_optimizer else None,
entropy_temperature=entropy_temperature,
logged_action_uniform_prior=logged_action_uniform_prior,
)
return SACTrainer(
q1_network,
actor_network,
parameters,
use_gpu=use_gpu,
value_network=value_network,
q2_network=q2_network,
)
