def pl_train_with_tune(cfg, pl_module_cls, checkpoint_dir=None):
cfg = munch_to_dictconfig(Munch(cfg))
checkpoint_path = (None if not checkpoint_dir else os.path.join(
checkpoint_dir, f"{pl_module_cls.__name__}.ckpt"))
trainer_extra_args = dict(gpus=1 if cfg.gpu else None,
progress_bar_refresh_rate=0,
resume_from_checkpoint=checkpoint_path,
callbacks=[TuneReportCheckpointCallback()])
pl_train(cfg, pl_module_cls, **trainer_extra_args)
def main(cfg: OmegaConf):
with StdoutTee('train.stdout'), StderrTee('train.stderr'):
print(OmegaConf.to_yaml(cfg))
pl_module_cls = DistillLightningModel if cfg.train.get(
'crossfit_size', 1) == 1 else DistillCrossfitLightningModel
if cfg.runner.name == 'pl':
pl_train(cfg, pl_module_cls)
else:
assert cfg.runner.name == 'ray', 'Only pl and ray runners are supported'
# Shouldn't need to install ray unless doing distributed training
from ray_runner import ray_train
ray_train(cfg, pl_module_cls)
def main(cfg: OmegaConf):
with StdoutTee('train.stdout'), StderrTee('train.stderr'):
print(OmegaConf.to_yaml(cfg))
if cfg.runner.name == 'pl':
trainer, model = pl_train(cfg, LightningModel)
else:
assert cfg.runner.name == 'ray', 'Only pl and ray runners are supported'
# Shouldn't need to install ray unless doing distributed training
from ray_runner import ray_train
ray_train(cfg, LightningModel)
def main(cfg: OmegaConf):
print(cfg.pretty())
if cfg.runner.name == 'pl':
from pl_runner import pl_train
trainer, model = pl_train(cfg, RNNTraining)
elif cfg.runner.name == 'ray':
# Shouldn't need to install ray unless doing distributed training
from ray_runner import ray_train
ray_train(cfg, RNNTraining)
else:
assert False, 'Only pl and ray runners are supported'
def main(cfg: OmegaConf):
# We want to add fields to cfg so need to call OmegaConf.set_struct
OmegaConf.set_struct(cfg, False)
print(OmegaConf.to_yaml(cfg))
if cfg.runner.name == 'pl':
from pl_runner import pl_train
trainer, model = pl_train(cfg, RNNTraining)
elif cfg.runner.name == 'ray':
# Shouldn't need to install ray unless doing distributed training
from ray_runner import ray_train
ray_train(cfg, RNNTraining)
else:
assert False, 'Only pl and ray runners are supported'