def test_discrete_action(self):
state_dim = 8
action_dim = 4
model = DuelingQNetwork(
layers=[state_dim, 8, 4, action_dim],
activations=["relu", "relu", "linear"],
use_batch_norm=True,
)
input = model.input_prototype()
self.assertEqual((1, state_dim), input.state.float_features.shape)
# Using batch norm requires more than 1 example in training, avoid that
model.eval()
q_values = model(input)
self.assertEqual((1, action_dim), q_values.q_values.shape)
def test_save_load_discrete_action_batch_norm(self):
state_dim = 8
action_dim = 4
model = DuelingQNetwork(
layers=[state_dim, 8, 4, action_dim],
activations=["relu", "relu", "linear"],
use_batch_norm=False,
)
# Freezing batch_norm
model.eval()
# Number of expected params is the same because DuelingQNetwork always
# initialize batch norm layer even if it doesn't use it.
expected_num_params, expected_num_inputs, expected_num_outputs = 22, 1, 1
check_save_load(self, model, expected_num_params, expected_num_inputs,
expected_num_outputs)
def _get_model(self, training_parameters, dueling_architecture=False):
if dueling_architecture:
return DuelingQNetwork(
training_parameters.layers,
training_parameters.activations,
action_dim=self.num_action_features,
)
elif training_parameters.factorization_parameters is None:
return FullyConnectedNetwork(
training_parameters.layers,
training_parameters.activations,
use_noisy_linear_layers=training_parameters.
use_noisy_linear_layers,
)
else:
return ParametricInnerProduct(
FullyConnectedNetwork(
training_parameters.factorization_parameters.state.layers,
training_parameters.factorization_parameters.state.
activations,
),
FullyConnectedNetwork(
training_parameters.factorization_parameters.action.layers,
training_parameters.factorization_parameters.action.
activations,
),
self.num_state_features,
self.num_action_features,
)
def build_q_network(
self,
state_normalization_parameters: Dict[int, NormalizationParameters],
output_dim: int,
) -> ModelBase:
state_dim = self._get_input_dim(state_normalization_parameters)
return DuelingQNetwork(
layers=[state_dim] + self.config.sizes + [output_dim],
activations=self.config.activations + ["linear"],
)
def test_save_load_discrete_action(self):
state_dim = 8
action_dim = 4
model = DuelingQNetwork(
layers=[state_dim, 8, 4, action_dim],
activations=["relu", "relu", "linear"],
use_batch_norm=False,
)
expected_num_params, expected_num_inputs, expected_num_outputs = 22, 1, 1
check_save_load(self, model, expected_num_params, expected_num_inputs,
expected_num_outputs)
def create_dqn_trainer_from_params(
model: DiscreteActionModelParameters,
normalization_parameters: Dict[int, NormalizationParameters],
use_gpu: bool = False,
use_all_avail_gpus: bool = False,
metrics_to_score=None,
):
metrics_to_score = metrics_to_score or []
if model.rainbow.quantile:
q_network = QuantileDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=len(model.actions),
num_atoms=model.rainbow.num_atoms,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
elif model.rainbow.categorical:
q_network = CategoricalDQN( # type: ignore
state_dim=get_num_output_features(normalization_parameters),
action_dim=len(model.actions),
num_atoms=model.rainbow.num_atoms,
qmin=model.rainbow.qmin,
qmax=model.rainbow.qmax,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
use_gpu=use_gpu,
)
elif model.rainbow.dueling_architecture:
q_network = DuelingQNetwork( # type: ignore
layers=[get_num_output_features(normalization_parameters)] +
model.training.layers[1:-1] + [len(model.actions)],
activations=model.training.activations,
)
else:
q_network = FullyConnectedDQN( # type: ignore
state_dim=get_num_output_features(normalization_parameters),
action_dim=len(model.actions),
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
if use_gpu and torch.cuda.is_available():
q_network = q_network.cuda()
q_network_target = q_network.get_target_network()
reward_network, q_network_cpe, q_network_cpe_target = None, None, None
if model.evaluation.calc_cpe_in_training:
# Metrics + reward
num_output_nodes = (len(metrics_to_score) + 1) * len(model.actions)
reward_network = FullyConnectedDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=num_output_nodes,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
q_network_cpe = FullyConnectedDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=num_output_nodes,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
if use_gpu and torch.cuda.is_available():
reward_network.cuda()
q_network_cpe.cuda()
q_network_cpe_target = q_network_cpe.get_target_network()
if (use_all_avail_gpus and not model.rainbow.categorical
and not model.rainbow.quantile):
q_network = q_network.get_distributed_data_parallel_model()
reward_network = (reward_network.get_distributed_data_parallel_model()
if reward_network else None)
q_network_cpe = (q_network_cpe.get_distributed_data_parallel_model()
if q_network_cpe else None)
if model.rainbow.quantile:
assert (not use_all_avail_gpus
), "use_all_avail_gpus not implemented for distributional RL"
return QRDQNTrainer(
q_network,
q_network_target,
model,
use_gpu,
metrics_to_score=metrics_to_score,
)
elif model.rainbow.categorical:
assert (not use_all_avail_gpus
), "use_all_avail_gpus not implemented for distributional RL"
return C51Trainer(
q_network,
q_network_target,
model,
use_gpu,
metrics_to_score=metrics_to_score,
)
else:
return DQNTrainer(
q_network,
q_network_target,
reward_network,
model,
use_gpu,
q_network_cpe=q_network_cpe,
q_network_cpe_target=q_network_cpe_target,
metrics_to_score=metrics_to_score,
)
def get_modular_sarsa_trainer_reward_boost(
self,
environment,
reward_shape,
dueling,
categorical,
quantile,
use_gpu=False,
use_all_avail_gpus=False,
clip_grad_norm=None,
):
assert not quantile or not categorical
parameters = self.get_sarsa_parameters(environment, reward_shape,
dueling, categorical, quantile,
clip_grad_norm)
if quantile:
if dueling:
q_network = DuelingQuantileDQN(
layers=[
get_num_output_features(environment.normalization)
] + parameters.training.layers[1:-1] +
[len(environment.ACTIONS)],
activations=parameters.training.activations,
num_atoms=parameters.rainbow.num_atoms,
)
else:
q_network = QuantileDQN(
state_dim=get_num_output_features(
environment.normalization),
action_dim=len(environment.ACTIONS),
num_atoms=parameters.rainbow.num_atoms,
sizes=parameters.training.layers[1:-1],
activations=parameters.training.activations[:-1],
)
elif categorical:
assert not dueling
q_network = CategoricalDQN(
state_dim=get_num_output_features(environment.normalization),
action_dim=len(environment.ACTIONS),
num_atoms=parameters.rainbow.num_atoms,
qmin=-100,
qmax=200,
sizes=parameters.training.layers[1:-1],
activations=parameters.training.activations[:-1],
)
else:
if dueling:
q_network = DuelingQNetwork(
layers=[
get_num_output_features(environment.normalization)
] + parameters.training.layers[1:-1] +
[len(environment.ACTIONS)],
activations=parameters.training.activations,
)
else:
q_network = FullyConnectedDQN(
state_dim=get_num_output_features(
environment.normalization),
action_dim=len(environment.ACTIONS),
sizes=parameters.training.layers[1:-1],
activations=parameters.training.activations[:-1],
)
q_network_cpe, q_network_cpe_target, reward_network = None, None, None
if parameters.evaluation and parameters.evaluation.calc_cpe_in_training:
q_network_cpe = FullyConnectedDQN(
state_dim=get_num_output_features(environment.normalization),
action_dim=len(environment.ACTIONS),
sizes=parameters.training.layers[1:-1],
activations=parameters.training.activations[:-1],
)
q_network_cpe_target = q_network_cpe.get_target_network()
reward_network = FullyConnectedDQN(
state_dim=get_num_output_features(environment.normalization),
action_dim=len(environment.ACTIONS),
sizes=parameters.training.layers[1:-1],
activations=parameters.training.activations[:-1],
)
if use_gpu:
q_network = q_network.cuda()
if parameters.evaluation.calc_cpe_in_training:
reward_network = reward_network.cuda()
q_network_cpe = q_network_cpe.cuda()
q_network_cpe_target = q_network_cpe_target.cuda()
if use_all_avail_gpus and not categorical:
q_network = q_network.get_distributed_data_parallel_model()
reward_network = reward_network.get_distributed_data_parallel_model(
)
q_network_cpe = q_network_cpe.get_distributed_data_parallel_model(
)
q_network_cpe_target = (
q_network_cpe_target.get_distributed_data_parallel_model())
if quantile:
trainer = QRDQNTrainer(
q_network,
q_network.get_target_network(),
parameters,
use_gpu,
reward_network=reward_network,
q_network_cpe=q_network_cpe,
q_network_cpe_target=q_network_cpe_target,
)
elif categorical:
trainer = C51Trainer(q_network, q_network.get_target_network(),
parameters, use_gpu)
else:
parameters = DQNTrainerParameters.from_discrete_action_model_parameters(
parameters)
trainer = DQNTrainer(
q_network,
q_network.get_target_network(),
reward_network,
parameters,
use_gpu,
q_network_cpe=q_network_cpe,
q_network_cpe_target=q_network_cpe_target,
)
return trainer
def __init__(
self,
parameters: DiscreteActionModelParameters,
state_normalization_parameters: Dict[int, NormalizationParameters],
use_gpu: bool = False,
additional_feature_types:
AdditionalFeatureTypes = DEFAULT_ADDITIONAL_FEATURE_TYPES,
metrics_to_score=None,
gradient_handler=None,
use_all_avail_gpus: bool = False,
) -> None:
self.double_q_learning = parameters.rainbow.double_q_learning
self.warm_start_model_path = parameters.training.warm_start_model_path
self.minibatch_size = parameters.training.minibatch_size
self._actions = parameters.actions if parameters.actions is not None else []
if parameters.training.cnn_parameters is None:
self.state_normalization_parameters: Optional[Dict[
int, NormalizationParameters]] = state_normalization_parameters
self.num_features = get_num_output_features(
state_normalization_parameters)
logger.info("Number of state features: " + str(self.num_features))
parameters.training.layers[0] = self.num_features
else:
self.state_normalization_parameters = None
parameters.training.layers[-1] = self.num_actions
RLTrainer.__init__(
self,
parameters,
use_gpu,
additional_feature_types,
metrics_to_score,
gradient_handler,
actions=self._actions,
)
self.reward_boosts = torch.zeros([1,
len(self._actions)]).type(self.dtype)
if parameters.rl.reward_boost is not None:
for k in parameters.rl.reward_boost.keys():
i = self._actions.index(k)
self.reward_boosts[0, i] = parameters.rl.reward_boost[k]
if parameters.rainbow.dueling_architecture:
self.q_network = DuelingQNetwork(
parameters.training.layers,
parameters.training.activations,
use_batch_norm=parameters.training.use_batch_norm,
)
else:
if parameters.training.cnn_parameters is None:
self.q_network = FullyConnectedNetwork(
parameters.training.layers,
parameters.training.activations,
use_noisy_linear_layers=parameters.training.
use_noisy_linear_layers,
min_std=parameters.training.weight_init_min_std,
use_batch_norm=parameters.training.use_batch_norm,
)
else:
self.q_network = ConvolutionalNetwork(
parameters.training.cnn_parameters,
parameters.training.layers,
parameters.training.activations,
use_noisy_linear_layers=parameters.training.
use_noisy_linear_layers,
min_std=parameters.training.weight_init_min_std,
use_batch_norm=parameters.training.use_batch_norm,
)
self.q_network_target = deepcopy(self.q_network)
self.q_network._name = "training"
self.q_network_target._name = "target"
self._set_optimizer(parameters.training.optimizer)
self.q_network_optimizer = self.optimizer_func(
self.q_network.parameters(),
lr=parameters.training.learning_rate,
weight_decay=parameters.training.l2_decay,
)
self._init_cpe_networks(parameters, use_all_avail_gpus)
if self.use_gpu:
self.q_network.cuda()
self.q_network_target.cuda()
if use_all_avail_gpus:
self.q_network = torch.nn.DataParallel(self.q_network)
self.q_network_target = torch.nn.DataParallel(
self.q_network_target)
def create_dqn_trainer_from_params(
model: DiscreteActionModelParameters,
normalization_parameters: Dict[int, NormalizationParameters],
use_gpu: bool = False,
metrics_to_score=None,
):
metrics_to_score = metrics_to_score or []
if model.rainbow.dueling_architecture:
q_network = DuelingQNetwork(
layers=[get_num_output_features(normalization_parameters)] +
model.training.layers[1:-1] + [len(model.actions)],
activations=model.training.activations,
)
else:
q_network = FullyConnectedDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=len(model.actions),
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
if use_gpu and torch.cuda.is_available():
q_network = q_network.cuda()
q_network_target = q_network.get_target_network()
reward_network, q_network_cpe, q_network_cpe_target = None, None, None
if model.evaluation.calc_cpe_in_training:
# Metrics + reward
num_output_nodes = (len(metrics_to_score) + 1) * len(model.actions)
reward_network = FullyConnectedDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=num_output_nodes,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
q_network_cpe = FullyConnectedDQN(
state_dim=get_num_output_features(normalization_parameters),
action_dim=num_output_nodes,
sizes=model.training.layers[1:-1],
activations=model.training.activations[:-1],
dropout_ratio=model.training.dropout_ratio,
)
if use_gpu and torch.cuda.is_available():
reward_network.cuda()
q_network_cpe.cuda()
q_network_cpe_target = q_network_cpe.get_target_network()
return DQNTrainer(
q_network,
q_network_target,
reward_network,
model,
use_gpu,
q_network_cpe=q_network_cpe,
q_network_cpe_target=q_network_cpe_target,
metrics_to_score=metrics_to_score,
)