def __init__(
self,
model: models.ConditionedModel = Ref("model"),
src_file: Union[None, str, Sequence[str]] = None,
trg_file: Optional[str] = None,
dev_every: numbers.Integral = 0,
dev_zero: bool = False,
batcher: batchers.Batcher = bare(batchers.SrcBatcher, batch_size=32),
loss_calculator: loss_calculators.LossCalculator = bare(
loss_calculators.MLELoss),
trainer: optimizers.XnmtOptimizer = bare(optimizers.SimpleSGDTrainer,
e0=0.1),
run_for_epochs: Optional[numbers.Integral] = None,
lr_decay: numbers.Real = 1.0,
lr_decay_times: numbers.Integral = 3,
patience: numbers.Integral = 1,
initial_patience: Optional[numbers.Integral] = None,
dev_tasks: Sequence[eval_tasks.EvalTask] = None,
dev_combinator: Optional[str] = None,
restart_trainer: bool = False,
reload_command: Optional[str] = None,
name: str = "{EXP}",
sample_train_sents: Optional[numbers.Integral] = None,
max_num_train_sents: Optional[numbers.Integral] = None,
max_src_len: Optional[numbers.Integral] = None,
max_trg_len: Optional[numbers.Integral] = None,
loss_comb_method: str = Ref("exp_global.loss_comb_method",
default="sum"),
update_every: numbers.Integral = 1,
commandline_args: dict = Ref("exp_global.commandline_args", default={})
) -> None:
super().__init__(model=model,
src_file=src_file,
trg_file=trg_file,
dev_every=dev_every,
batcher=batcher,
loss_calculator=loss_calculator,
run_for_epochs=run_for_epochs,
lr_decay=lr_decay,
lr_decay_times=lr_decay_times,
patience=patience,
initial_patience=initial_patience,
dev_tasks=dev_tasks,
dev_combinator=dev_combinator,
restart_trainer=restart_trainer,
reload_command=reload_command,
name=name,
sample_train_sents=sample_train_sents,
max_num_train_sents=max_num_train_sents,
max_src_len=max_src_len,
max_trg_len=max_trg_len)
self.dev_zero = dev_zero
self.trainer = trainer or optimizers.SimpleSGDTrainer(e0=0.1)
self.dynet_profiling = commandline_args.get(
"dynet_profiling", 0) if commandline_args else 0
self.train_loss_tracker = loss_trackers.TrainLossTracker(self)
self.loss_comb_method = loss_comb_method
self.update_every = update_every
self.num_updates_skipped = 0
def __init__(
self,
tasks: Sequence[train_tasks.TrainingTask],
task_weights: Optional[Sequence[numbers.Real]] = None,
trainer: optimizers.XnmtOptimizer = bare(optimizers.SimpleSGDTrainer,
e0=0.1),
dev_zero: bool = False,
loss_comb_method: str = Ref("exp_global.loss_comb_method",
default="sum"),
update_every_within: numbers.Integral = 1,
update_every_across: numbers.Integral = 1,
commandline_args=Ref("exp_global.commandline_args", default=None)
) -> None:
super().__init__(tasks=tasks,
trainer=trainer,
dev_zero=dev_zero,
update_every=update_every_across,
commandline_args=commandline_args)
if update_every_within != 1 and update_every_across != 1:
raise ValueError(
"update_every_within and update_every_across cannot be mixed.")
self.update_every_within = update_every_within
self.task_weights = task_weights or [1. / len(tasks)] * len(tasks)
if len(self.task_weights) != len(self.tasks):
raise ValueError(
f"number of tasks must match number of task weights; "
f"found: {len(self.task_weights)} != {len(self.tasks)}")
self.train_loss_trackers = {
task: loss_trackers.TrainLossTracker(task)
for task in tasks
}
self.loss_comb_method = loss_comb_method
def __init__(
self,
tasks: Sequence[train_tasks.TrainingTask],
trainer: optimizers.XnmtOptimizer = bare(optimizers.SimpleSGDTrainer,
e0=0.1),
dev_zero: bool = False,
per_task_backward: bool = True,
loss_comb_method: str = Ref("exp_global.loss_comb_method",
default="sum"),
update_every: numbers.Integral = 1,
n_task_steps: Optional[Sequence[numbers.Integral]] = None,
commandline_args: dict = Ref("exp_global.commandline_args",
default=None)
) -> None:
super().__init__(tasks=tasks,
trainer=trainer,
dev_zero=dev_zero,
update_every=update_every,
commandline_args=commandline_args)
self.train_loss_trackers = {
task: loss_trackers.TrainLossTracker(task)
for task in tasks
}
self.per_task_backward = per_task_backward
self.loss_comb_method = loss_comb_method
self.n_task_steps = n_task_steps or [1] * len(tasks)
if len(self.n_task_steps) != len(tasks):
raise ValueError(
f"number of tasks and steps per task do not match: {len(tasks)} != {len(self.n_task_steps)}"
)
def __init__(
self,
tasks: Sequence[train_tasks.TrainingTask],
trainer: optimizers.XnmtOptimizer = bare(optimizers.SimpleSGDTrainer,
e0=0.1),
dev_zero: bool = False,
loss_comb_method: str = Ref("exp_global.loss_comb_method",
default="sum"),
update_every: numbers.Integral = 1,
commandline_args: dict = Ref("exp_global.commandline_args",
default=None)
) -> None:
super().__init__(tasks=tasks,
trainer=trainer,
dev_zero=dev_zero,
commandline_args=commandline_args,
update_every=update_every)
self.train_loss_trackers = {
task: loss_trackers.TrainLossTracker(task)
for task in tasks
}
self.loss_comb_method = loss_comb_method