def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
if torch.cuda.is_available():
torch.cuda.manual_seed_all(args.seed)
if args.from_pretrained is not None:
model = Summarizer.load_from_checkpoint(args.from_pretrained, args)
else:
model = Summarizer(args)
model.hf_datasets = nlp.load_dataset('scientific_papers', 'arxiv')
logger = pl_loggers.CometLogger(save_dir='logs/')
checkpoint_callback = ModelCheckpoint(filepath=os.path.join(
args.save_dir, args.save_prefix, "checkpoints"),
save_top_k=5,
verbose=True,
monitor='avg_val_loss',
mode='min',
period=-1,
prefix='')
print(args)
args.dataset_size = 203037 # hardcode dataset size. Needed to compute number of steps for the lr scheduler
trainer = pl.Trainer(
gpus=args.gpus,
distributed_backend='ddp' if torch.cuda.is_available() else None,
track_grad_norm=-1,
max_epochs=args.epochs if not args.debug else 100,
max_steps=None if not args.debug else 1,
replace_sampler_ddp=False,
accumulate_grad_batches=args.grad_accum,
val_check_interval=args.val_every if not args.debug else 1,
num_sanity_val_steps=2 if not args.debug else 0,
check_val_every_n_epoch=1 if not args.debug else 1,
val_percent_check=args.val_percent_check,
test_percent_check=args.val_percent_check,
logger=logger,
checkpoint_callback=checkpoint_callback
if not args.disable_checkpointing else False,
show_progress_bar=not args.no_progress_bar,
use_amp=not args.fp32,
amp_level='O2',
resume_from_checkpoint=args.resume_ckpt,
)
if not args.test:
trainer.fit(model)
trainer.test(model)
def main(args, model_name: str, reproducible: bool, comet: bool, wandb: bool):
if reproducible:
seed_everything(42)
args.deterministic = True
args.benchmark = True
if args.default_root_dir is None:
args.default_root_dir = 'results'
if comet:
from pytorch_lightning.loggers import CometLogger
comet_logger = loggers.CometLogger(
api_key=os.environ.get('COMET_API_KEY'),
workspace=os.environ.get('COMET_WORKSPACE'), # Optional
project_name=os.environ.get('COMET_PROJECT_NAME'), # Optional
experiment_name=model_name # Optional
)
args.logger = comet_logger
if wandb:
from pytorch_lightning.loggers import WandbLogger
wandb_logger = WandbLogger(project=os.environ.get('WANDB_PROJECT_NAME'), log_model=True, sync_step=True)
args.logger = wandb_logger
# Save best model
model_checkpoint = ModelCheckpoint(
filename=model_name + '_{epoch}',
save_top_k=1,
monitor='val_iou',
mode='max',
)
args.checkpoint_callback = model_checkpoint
data = TwoDomainDM(dataPath=args.dataPath, augment=args.augment, batch_size=args.batch_size, num_workers=8)
model = RightLaneSTModule(lr=args.learningRate, lrRatio=args.lrRatio, decay=args.decay, num_cls=4)
# Parse all trainer options available from the command line
trainer = Trainer.from_argparse_args(args)
trainer.fit(model, datamodule=data)
# Reload best model
model = RightLaneSTModule.load_from_checkpoint(model_checkpoint.best_model_path, dataPath=args.dataPath, num_cls=4)
# Upload weights
if comet:
comet_logger.experiment.log_model(model_name + '_weights', model_checkpoint.best_model_path)
# Perform testing
trainer.test(model, datamodule=data)
def main(args=None):
pl.seed_everything(52)
parser = ArgumentParser()
dm_cls = NeutronDataLoader
checkpoint_callback = ModelCheckpoint(monitor="Valid Loss", save_last=True, save_top_k=3, mode='min')
script_args, _ = parser.parse_known_args(args)
parser = dm_cls.add_argparse_args(parser)
parser = MyModel.add_model_specific_args(parser)
parser = pl.Trainer.add_argparse_args(parser)
args = parser.parse_args(args)
comet_logger = pl_loggers.CometLogger(save_dir="comet_logs", experiment_name=EXPERIMENT_NAME,
project_name="HIDA", offline=True)
dm = dm_cls.from_argparse_args(args)
model = MyModel(**vars(args))
trainer = pl.Trainer.from_argparse_args(args, logger=comet_logger, checkpoint_callback=checkpoint_callback)
trainer.fit(model, dm)
def main(args=None):
pl.seed_everything(52)
parser = ArgumentParser()
dm_cls = NeutronDataLoader
script_args, _ = parser.parse_known_args(args)
parser = dm_cls.add_argparse_args(parser)
parser = MyModel.add_model_specific_args(parser)
parser = pl.Trainer.add_argparse_args(parser)
args = parser.parse_args(args)
comet_logger = pl_loggers.CometLogger(save_dir="comet_logs",
experiment_name=EXPERIMENT_NAME,
project_name="HIDA",
offline=True)
dm = dm_cls.from_argparse_args(args)
dm.setup()
model = MyModel.load_from_checkpoint(CHECKPOINT, **vars(args))
trainer = pl.Trainer.from_argparse_args(args, logger=comet_logger)
trainer.test(model, test_dataloaders=dm.test_dataloader())
'lr': 1e-3,
'num_warmup_steps': 0,
'batch_size': 64,
'accumulate_grad_batches': 1,
'bptt': 140
})
if path.exists('.comet.config'):
import configparser
comet_config = configparser.ConfigParser()
comet_config.read('.comet.config')
logger = loggers.CometLogger(
api_key=comet_config['comet']['api_key'],
project_name="lstm-chatbot-test",
workspace="luungoc2005"
)
for k, v in config.items():
logger.experiment.log_parameter(k, v)
else:
logger = loggers.TensorBoardLogger()
checkpoint_callback = ModelCheckpoint(filepath='./checkpoints/')
model = LanguageModelTrainer(config)
trainer = Trainer(
gradient_clip_val=.5,
gpus=1,