def __init__(
self,
seq2slate_net: Seq2SlateTransformerNet,
reward_net_path: str,
minibatch_size: int,
parameters: Seq2SlateParameters,
baseline_net: Optional[BaselineNet] = None,
baseline_warmup_num_batches: int = 0,
use_gpu: bool = False,
policy_optimizer: Optimizer__Union = field( # noqa: B008
default_factory=Optimizer__Union.default
),
baseline_optimizer: Optimizer__Union = field( # noqa: B008
default_factory=Optimizer__Union.default
),
) -> None:
self.reward_net_path = reward_net_path
# loaded when used
self.reward_net = None
self.parameters = parameters
self.minibatch_size = minibatch_size
self.use_gpu = use_gpu
self.device = torch.device("cuda") if use_gpu else torch.device("cpu")
self.permutation_index = torch.tensor(
list(
permutations(
# pyre-fixme[6]: Expected `Iterable[Variable[itertools._T]]` for
# 1st param but got `Tensor`.
torch.arange(seq2slate_net.max_src_seq_len),
seq2slate_net.max_tgt_seq_len,
)
),
device=self.device,
).long()
if self.parameters.simulation_distance_penalty is not None:
# pyre-fixme[16]: `Optional` has no attribute `__gt__`.
assert self.parameters.simulation_distance_penalty > 0
self.permutation_distance = (
torch.tensor(
[swap_dist(x.tolist()) for x in self.permutation_index],
device=self.device,
)
.unsqueeze(1)
.float()
)
self.MAX_DISTANCE = torch.max(self.permutation_distance)
self.trainer = Seq2SlateTrainer(
seq2slate_net,
minibatch_size,
self.parameters,
baseline_net=baseline_net,
baseline_warmup_num_batches=baseline_warmup_num_batches,
use_gpu=use_gpu,
policy_optimizer=policy_optimizer,
baseline_optimizer=baseline_optimizer,
)
self.seq2slate_net = self.trainer.seq2slate_net
self.baseline_net = self.trainer.baseline_net
def __init__(
self,
seq2slate_net: Seq2SlateTransformerNet,
parameters: Seq2SlateTransformerParameters,
minibatch_size: int,
reward_net_path: str,
baseline_net: Optional[BaselineNet] = None,
use_gpu: bool = False,
) -> None:
self.reward_net_path = reward_net_path
# loaded when used
self.reward_net = None
self.parameters = parameters
self.minibatch_size = minibatch_size
self.use_gpu = use_gpu
self.device = torch.device("cuda") if use_gpu else torch.device("cpu")
self.permutation_index = torch.tensor(
list(
permutations(
torch.arange(seq2slate_net.max_src_seq_len),
seq2slate_net.max_tgt_seq_len,
)),
device=self.device,
).long()
self.trainer = Seq2SlateTrainer(seq2slate_net, parameters,
minibatch_size, baseline_net, use_gpu)
self.seq2slate_net = self.trainer.seq2slate_net
self.baseline_net = self.trainer.baseline_net
def create_trainer(seq2slate_net, batch_size, learning_rate, device,
on_policy):
use_gpu = False if device == torch.device("cpu") else True
return Seq2SlateTrainer(
seq2slate_net=seq2slate_net,
minibatch_size=batch_size,
parameters=Seq2SlateParameters(on_policy=on_policy),
policy_optimizer=Optimizer__Union.default(lr=learning_rate),
use_gpu=use_gpu,
print_interval=100,
)
def create_trainer(
seq2slate_net,
learning_rate,
seq2slate_params,
policy_gradient_interval,
):
return Seq2SlateTrainer(
seq2slate_net=seq2slate_net,
params=seq2slate_params,
policy_optimizer=Optimizer__Union(SGD=classes["SGD"](
lr=learning_rate)),
policy_gradient_interval=policy_gradient_interval,
print_interval=1,
)
def create_trainer(
seq2slate_net,
batch_size,
learning_rate,
device,
seq2slate_params,
policy_gradient_interval,
):
use_gpu = False if device == torch.device("cpu") else True
return Seq2SlateTrainer(
seq2slate_net=seq2slate_net,
minibatch_size=batch_size,
parameters=seq2slate_params,
policy_optimizer=Optimizer__Union(SGD=classes["SGD"](lr=learning_rate)),
use_gpu=use_gpu,
policy_gradient_interval=policy_gradient_interval,
print_interval=1,
)
def __init__(
self,
seq2slate_net: Seq2SlateTransformerNet,
minibatch_size: int,
parameters: Seq2SlateParameters,
baseline_net: Optional[BaselineNet] = None,
baseline_warmup_num_batches: int = 0,
use_gpu: bool = False,
policy_optimizer: Optimizer__Union = field( # noqa: B008
default_factory=Optimizer__Union.default
),
baseline_optimizer: Optimizer__Union = field( # noqa: B008
default_factory=Optimizer__Union.default
),
policy_gradient_interval: int = 1,
print_interval: int = 100,
) -> None:
self.sim_param = parameters.simulation
assert self.sim_param is not None
# loaded when used
self.reward_name_and_net = {}
self.parameters = parameters
self.minibatch_size = minibatch_size
self.use_gpu = use_gpu
self.policy_gradient_interval = policy_gradient_interval
self.print_interval = print_interval
self.device = torch.device("cuda") if use_gpu else torch.device("cpu")
self.MAX_DISTANCE = (
seq2slate_net.max_src_seq_len * (seq2slate_net.max_src_seq_len - 1) / 2
)
self.trainer = Seq2SlateTrainer(
seq2slate_net,
minibatch_size,
self.parameters,
baseline_net=baseline_net,
baseline_warmup_num_batches=baseline_warmup_num_batches,
use_gpu=use_gpu,
policy_optimizer=policy_optimizer,
baseline_optimizer=baseline_optimizer,
policy_gradient_interval=policy_gradient_interval,
print_interval=print_interval,
)
self.seq2slate_net = self.trainer.seq2slate_net
self.baseline_net = self.trainer.baseline_net
def __init__(
self,
seq2slate_net: Seq2SlateTransformerNet,
parameters: Seq2SlateTransformerParameters,
minibatch_size: int,
reward_net_path: str,
baseline_net: Optional[BaselineNet] = None,
use_gpu: bool = False,
) -> None:
self.reward_net_path = reward_net_path
# loaded when used
self.reward_net = None
self.parameters = parameters
self.minibatch_size = minibatch_size
self.use_gpu = use_gpu
self.device = torch.device("cuda") if use_gpu else torch.device("cpu")
self.permutation_index = torch.tensor(
list(
permutations(
torch.arange(seq2slate_net.max_src_seq_len),
seq2slate_net.max_tgt_seq_len,
)),
device=self.device,
).long()
if self.parameters.simulation_distance_penalty is not None:
assert self.parameters.simulation_distance_penalty > 0
self.permutation_distance = torch.tensor(
[swap_dist(x.tolist()) for x in self.permutation_index],
device=self.device,
).float()
self.MAX_DISTANCE = torch.max(self.permutation_distance)
self.trainer = Seq2SlateTrainer(seq2slate_net, parameters,
minibatch_size, baseline_net, use_gpu)
self.seq2slate_net = self.trainer.seq2slate_net
self.baseline_net = self.trainer.baseline_net