def __init__(
self,
rl_parameters: RLParameters,
use_gpu: bool,
metrics_to_score=None,
actions: Optional[List[str]] = None,
evaluation_parameters: Optional[EvaluationParameters] = None,
loss_reporter=None,
) -> None:
self.minibatch = 0
self.minibatch_size: Optional[int] = None
self.minibatches_per_step: Optional[int] = None
self.rl_parameters = rl_parameters
self.rl_temperature = float(rl_parameters.temperature)
self.maxq_learning = rl_parameters.maxq_learning
self.gamma = rl_parameters.gamma
self.tau = rl_parameters.target_update_rate
self.use_seq_num_diff_as_time_diff = rl_parameters.use_seq_num_diff_as_time_diff
self.time_diff_unit_length = rl_parameters.time_diff_unit_length
self.tensorboard_logging_freq = rl_parameters.tensorboard_logging_freq
self.multi_steps = rl_parameters.multi_steps
self.calc_cpe_in_training = (
evaluation_parameters and evaluation_parameters.calc_cpe_in_training
)
if rl_parameters.q_network_loss == "mse":
self.q_network_loss = F.mse_loss
elif rl_parameters.q_network_loss == "huber":
self.q_network_loss = F.smooth_l1_loss
else:
raise Exception(
"Q-Network loss type {} not valid loss.".format(
rl_parameters.q_network_loss
)
)
if metrics_to_score:
self.metrics_to_score = metrics_to_score + ["reward"]
else:
self.metrics_to_score = ["reward"]
cuda_available = torch.cuda.is_available()
logger.info("CUDA availability: {}".format(cuda_available))
if use_gpu and cuda_available:
logger.info("Using GPU: GPU requested and available.")
self.use_gpu = True
self.device = torch.device("cuda") # type: ignore
else:
logger.info("NOT Using GPU: GPU not requested or not available.")
self.use_gpu = False
self.device = torch.device("cpu") # type: ignore
self.loss_reporter = loss_reporter or LossReporter(actions)
self._actions = actions
def __init__(
self,
parameters,
use_gpu,
additional_feature_types,
metrics_to_score=None,
gradient_handler=None,
actions: Optional[List[str]] = None,
):
self.minibatch = 0
self.parameters = parameters
self.reward_burnin = parameters.rl.reward_burnin
self._additional_feature_types = additional_feature_types
self.rl_temperature = parameters.rl.temperature
self.maxq_learning = parameters.rl.maxq_learning
self.gamma = parameters.rl.gamma
self.tau = parameters.rl.target_update_rate
self.use_seq_num_diff_as_time_diff = parameters.rl.use_seq_num_diff_as_time_diff
self.time_diff_unit_length = parameters.rl.time_diff_unit_length
self.gradient_handler = gradient_handler
self.tensorboard_logging_freq = parameters.rl.tensorboard_logging_freq
self.multi_steps = parameters.rl.multi_steps
if parameters.rl.q_network_loss == "mse":
self.q_network_loss = getattr(F, "mse_loss")
elif parameters.rl.q_network_loss == "huber":
self.q_network_loss = getattr(F, "smooth_l1_loss")
else:
raise Exception(
"Q-Network loss type {} not valid loss.".format(
parameters.rl.q_network_loss
)
)
if metrics_to_score:
self.metrics_to_score = metrics_to_score + ["reward"]
else:
self.metrics_to_score = ["reward"]
cuda_available = torch.cuda.is_available()
logger.info("CUDA availability: {}".format(cuda_available))
if use_gpu and cuda_available:
logger.info("Using GPU: GPU requested and available.")
self.use_gpu = True
self.dtype = torch.cuda.FloatTensor
self.dtypelong = torch.cuda.LongTensor
else:
logger.info("NOT Using GPU: GPU not requested or not available.")
self.use_gpu = False
self.dtype = torch.FloatTensor
self.dtypelong = torch.LongTensor
self.loss_reporter = LossReporter(actions)