def train_eval_lightning(
train_dataset,
eval_dataset,
trainer_module,
num_epochs,
use_gpu,
batch_preprocessor=None,
reader_options: Optional[ReaderOptions] = None,
checkpoint_path: Optional[str] = None,
) -> pl.Trainer:
reader_options = reader_options or ReaderOptions()
datamodule = PetastormLightningDataModule(train_dataset, eval_dataset,
batch_preprocessor,
reader_options)
# pyre-fixme[16]: Module `pl` has no attribute `Trainer`.
# pyre-fixme[16]: Module `pl` has no attribute `Trainer`.
trainer = pl.Trainer(
max_epochs=num_epochs * 1000,
gpus=int(use_gpu),
reload_dataloaders_every_epoch=True,
resume_from_checkpoint=checkpoint_path,
callbacks=[StoppingEpochCallback(num_epochs)],
)
trainer.fit(trainer_module, datamodule=datamodule)
# TODO: evaluate
return trainer
def train_eval_lightning(
train_dataset,
eval_dataset,
test_dataset,
trainer_module,
data_module,
num_epochs,
logger_name: str,
batch_preprocessor=None,
reader_options: Optional[ReaderOptions] = None,
checkpoint_path: Optional[str] = None,
resource_options: Optional[ResourceOptions] = None,
) -> pl.Trainer:
resource_options = resource_options or ResourceOptions()
use_gpu = resource_options.use_gpu
reader_options = reader_options or ReaderOptions()
datamodule = data_module or PetastormLightningDataModule(
train_dataset, eval_dataset, batch_preprocessor, reader_options)
trainer = pl.Trainer(
logger=OssTensorboardLogger(save_dir="pl_log_tensorboard",
name=logger_name),
max_epochs=num_epochs * 1000,
gpus=int(use_gpu),
reload_dataloaders_every_epoch=True,
resume_from_checkpoint=checkpoint_path,
callbacks=[StoppingEpochCallback(num_epochs)],
)
trainer.fit(trainer_module, datamodule=datamodule)
trainer.test()
if checkpoint_path is not None:
# Overwrite the warmstart path with the new model
trainer_module.set_clean_stop(True)
trainer.save_checkpoint(checkpoint_path)
return trainer
def train_eval_lightning(
train_dataset,
eval_dataset,
test_dataset,
trainer_module,
data_module,
num_epochs,
use_gpu,
logger_name: str,
batch_preprocessor=None,
reader_options: Optional[ReaderOptions] = None,
checkpoint_path: Optional[str] = None,
resource_options: Optional[ResourceOptions] = None,
) -> pl.Trainer:
reader_options = reader_options or ReaderOptions()
datamodule = data_module or PetastormLightningDataModule(
train_dataset, eval_dataset, batch_preprocessor, reader_options)
# pyre-fixme[16]: Module `pl` has no attribute `Trainer`.
# pyre-fixme[16]: Module `pl` has no attribute `Trainer`.
trainer = pl.Trainer(
logger=TensorBoardLogger(save_dir="pl_log_tensorboard",
name=logger_name),
max_epochs=num_epochs * 1000,
gpus=int(use_gpu),
reload_dataloaders_every_epoch=True,
resume_from_checkpoint=checkpoint_path,
callbacks=[StoppingEpochCallback(num_epochs)],
)
trainer.fit(trainer_module, datamodule=datamodule)
trainer.test()
return trainer