class PunctuationCapitalizationConfig(NemoConfig):
"""
A config for punctuation model training and testing.
See an example of full config in
`nemo/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml
<https://github.com/NVIDIA/NeMo/blob/main/examples/nlp/token_classification/conf/punctuation_capitalization_config.yaml>`_
"""
pretrained_model: Optional[str] = None
"""Can be an NVIDIA's NGC cloud model or a path to a .nemo checkpoint. You can get list of possible cloud options
by calling method
:func:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel.list_available_models`.
"""
name: Optional[str] = 'Punctuation_and_Capitalization'
"""A name of the model. Used for naming output directories and ``.nemo`` checkpoints."""
do_training: bool = True
"""Whether to perform training of the model."""
do_testing: bool = False
"""Whether ot perform testing of the model."""
model: PunctuationCapitalizationModelConfig = PunctuationCapitalizationModelConfig()
"""A configuration for the
:class:`~nemo.collections.nlp.models.token_classification.punctuation_capitalization_model.PunctuationCapitalizationModel`
model."""
trainer: Optional[TrainerConfig] = TrainerConfig()
"""Contains ``Trainer`` Lightning class constructor parameters."""
exp_manager: Optional[ExpManagerConfig] = ExpManagerConfig(name=name, files_to_copy=[])
"""A configuration with various NeMo training options such as output directories, resuming from checkpoint,
class ParallelAlignmentConfig:
model: Optional[str] = None # name
predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True, num_workers=4)
aligner_args: K2AlignerWrapperModelConfig = K2AlignerWrapperModelConfig()
output_path: str = MISSING
model_stride: int = 8
trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp")
# there arguments will be ignored
return_predictions: bool = False
use_cer: bool = False
class AppConfig(Config):
"""
This is structured config for this application.
Args:
name: Description of the application.
trainer: configuration of the trainer.
model: configuation of the model.
"""
name: str = "Training of a LeNet-5 Model using a pure PyTorchLightning approach - using DDP on 2 GPUs."
trainer: TrainerConfig = TrainerConfig(gpus=2, accelerator="ddp")
model: MNISTLeNet5Config = MNISTLeNet5Config()
class ParallelTranscriptionConfig:
model: Optional[str] = None # name
predict_ds: ASRDatasetConfig = ASRDatasetConfig(return_sample_id=True,
num_workers=4)
output_path: str = MISSING
# when return_predictions is enabled, the prediction call would keep all the predictions in memory and return them when prediction is done
return_predictions: bool = False
use_cer: bool = False
# decoding strategy for RNNT models
rnnt_decoding: RNNTDecodingConfig = RNNTDecodingConfig()
trainer: TrainerConfig = TrainerConfig(gpus=-1, accelerator="ddp")
def nemo_convert(argv):
args = get_args(argv)
loglevel = logging.INFO
# assuming loglevel is bound to the string value obtained from the
# command line argument. Convert to upper case to allow the user to
# specify --log=DEBUG or --log=debug
if args.verbose is not None:
numeric_level = getattr(logging, args.verbose.upper(), None)
if not isinstance(numeric_level, int):
raise ValueError('Invalid log level: %s' % numeric_level)
loglevel = numeric_level
logger = logging.getLogger(__name__)
if logger.handlers:
for handler in logger.handlers:
logger.removeHandler(handler)
logging.basicConfig(level=loglevel,
format='%(asctime)s [%(levelname)s] %(message)s')
logging.info("Logging level set to {}".format(loglevel))
"""Convert a .nemo saved model trained on previous versions of nemo into a nemo fie with current version."""
nemo_in = args.source
out = args.out
# Create a PL trainer object which is required for restoring Megatron models
cfg_trainer = TrainerConfig(
gpus=1,
accelerator="ddp",
num_nodes=1,
# Need to set the following two to False as ExpManager will take care of them differently.
logger=False,
checkpoint_callback=False,
)
trainer = pl.Trainer(cfg_trainer)
logging.info("Restoring NeMo model from '{}'".format(nemo_in))
try:
# If the megatron based NLP model was trained on NeMo < 1.5, then we need to update the lm_checkpoint on the model config
if args.megatron_legacy:
if args.megatron_checkpoint:
connector = NLPSaveRestoreConnector()
model_cfg = ModelPT.restore_from(
restore_path=nemo_in,
save_restore_connector=connector,
trainer=trainer,
return_config=True)
OmegaConf.set_struct(model_cfg, True)
with open_dict(model_cfg):
model_cfg.language_model.lm_checkpoint = args.megatron_checkpoint
model_cfg['megatron_legacy'] = True
model_cfg['masked_softmax_fusion'] = False
model_cfg['bias_gelu_fusion'] = False
model = ModelPT.restore_from(
restore_path=nemo_in,
save_restore_connector=connector,
trainer=trainer,
override_config_path=model_cfg,
)
else:
logging.error(
"Megatron Checkpoint must be provided if Megatron legacy is chosen"
)
else:
model = ModelPT.restore_from(restore_path=nemo_in, trainer=trainer)
logging.info("Model {} restored from '{}'".format(
model.cfg.target, nemo_in))
# Save the model
model.save_to(out)
logging.info("Successfully converted to {}".format(out))
del model
except Exception as e:
logging.error(
"Failed to restore model from NeMo file : {}. Please make sure you have the latest NeMo package installed with [all] dependencies."
.format(nemo_in))
raise e