def wrong():
with open("tests/data/jasper_smaller.yaml") as file:
jasper_config = self.yaml.load(file)
labels = jasper_config['labels']
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=labels,
batch_size=4)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**jasper_config['AudioToMelSpectrogramPreprocessor'])
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_config['AudioToMelSpectrogramPreprocessor']
['features'],
**jasper_config['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=1024, num_classes=len(labels))
# DAG definition
audio_signal, audio_signal_len, transcript, transcript_len = \
data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
spec_augment = nemo_asr.SpectrogramAugmentation(rect_masks=5)
aug_signal = spec_augment(input_spec=processed_signal)
encoded, encoded_len = jasper_encoder(audio_signal=aug_signal,
length=processed_signal_len)
log_probs = jasper_decoder(encoder_output=processed_signal)
def create_all_dags(args, neural_factory):
logger = neural_factory.logger
yaml = YAML(typ="safe")
with open(args.model_config) as f:
jasper_params = yaml.load(f)
vocab = jasper_params['labels']
sample_rate = jasper_params['sample_rate']
# Calculate num_workers for dataloader
total_cpus = os.cpu_count()
cpu_per_traindl = max(int(total_cpus / neural_factory.world_size), 1)
# perturb_config = jasper_params.get('perturb', None)
train_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"])
train_dl_params.update(jasper_params["AudioToTextDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
# del train_dl_params["normalize_transcripts"]
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.train_dataset,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
**train_dl_params,
# normalize_transcripts=False
)
N = len(data_layer)
steps_per_epoch = int(
N / (args.batch_size * args.iter_per_step * args.num_gpus))
logger.info('Have {0} examples to train on.'.format(N))
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate,
**jasper_params["AudioToMelSpectrogramPreprocessor"])
multiply_batch_config = jasper_params.get('MultiplyBatch', None)
if multiply_batch_config:
multiply_batch = nemo_asr.MultiplyBatch(**multiply_batch_config)
spectr_augment_config = jasper_params.get('SpectrogramAugmentation', None)
if spectr_augment_config:
data_spectr_augmentation = nemo_asr.SpectrogramAugmentation(
**spectr_augment_config)
eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"])
eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
data_layers_eval = []
if args.eval_datasets:
for eval_datasets in args.eval_datasets:
data_layer_eval = nemo_asr.AudioToTextDataLayer(
manifest_filepath=eval_datasets,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.eval_batch_size,
num_workers=cpu_per_traindl,
**eval_dl_params,
)
data_layers_eval.append(data_layer_eval)
else:
neural_factory.logger.info("There were no val datasets passed")
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_params["AudioToMelSpectrogramPreprocessor"]["features"],
**jasper_params["JasperEncoder"])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(vocab),
factory=neural_factory)
ctc_loss = nemo_asr.CTCLossNM(
num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
logger.info('================================')
logger.info(
f"Number of parameters in encoder: {jasper_encoder.num_weights}")
logger.info(
f"Number of parameters in decoder: {jasper_decoder.num_weights}")
logger.info(
f"Total number of parameters in decoder: "
f"{jasper_decoder.num_weights + jasper_encoder.num_weights}")
logger.info('================================')
# Train DAG
audio_signal_t, a_sig_length_t, \
transcript_t, transcript_len_t = data_layer()
processed_signal_t, p_length_t = data_preprocessor(
input_signal=audio_signal_t,
length=a_sig_length_t)
if multiply_batch_config:
processed_signal_t, p_length_t, transcript_t, transcript_len_t = \
multiply_batch(
in_x=processed_signal_t, in_x_len=p_length_t,
in_y=transcript_t,
in_y_len=transcript_len_t)
if spectr_augment_config:
processed_signal_t = data_spectr_augmentation(
input_spec=processed_signal_t)
encoded_t, encoded_len_t = jasper_encoder(
audio_signal=processed_signal_t,
length=p_length_t)
log_probs_t = jasper_decoder(encoder_output=encoded_t)
predictions_t = greedy_decoder(log_probs=log_probs_t)
loss_t = ctc_loss(
log_probs=log_probs_t,
targets=transcript_t,
input_length=encoded_len_t,
target_length=transcript_len_t)
# Callbacks needed to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss_t, predictions_t, transcript_t, transcript_len_t],
print_func=partial(
monitor_asr_train_progress,
labels=vocab,
logger=logger),
get_tb_values=lambda x: [("loss", x[0])],
tb_writer=neural_factory.tb_writer,
)
chpt_callback = nemo.core.CheckpointCallback(
folder=neural_factory.checkpoint_dir,
step_freq=args.checkpoint_save_freq)
callbacks = [train_callback, chpt_callback]
# assemble eval DAGs
for i, eval_dl in enumerate(data_layers_eval):
audio_signal_e, a_sig_length_e, transcript_e, transcript_len_e = \
eval_dl()
processed_signal_e, p_length_e = data_preprocessor(
input_signal=audio_signal_e,
length=a_sig_length_e)
encoded_e, encoded_len_e = jasper_encoder(
audio_signal=processed_signal_e,
length=p_length_e)
log_probs_e = jasper_decoder(encoder_output=encoded_e)
predictions_e = greedy_decoder(log_probs=log_probs_e)
loss_e = ctc_loss(
log_probs=log_probs_e,
targets=transcript_e,
input_length=encoded_len_e,
target_length=transcript_len_e)
# create corresponding eval callback
tagname = os.path.basename(args.eval_datasets[i]).split(".")[0]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[loss_e, predictions_e,
transcript_e, transcript_len_e],
user_iter_callback=partial(
process_evaluation_batch,
labels=vocab),
user_epochs_done_callback=partial(
process_evaluation_epoch,
tag=tagname,
logger=logger),
eval_step=args.eval_freq,
tb_writer=neural_factory.tb_writer)
callbacks.append(eval_callback)
return loss_t, callbacks, steps_per_epoch
yaml = YAML(typ="safe")
with open("./NeMo/examples/asr/configs/jasper12x1SEP.yaml") as f:
jasper_model_definition = yaml.load(f)
labels = jasper_model_definition['labels']
# Instantiate neural modules
data_layer = nemo_asr.AudioToTextDataLayer(manifest_filepath=train_dataset,
labels=labels,
batch_size=32)
data_layer_val = nemo_asr.AudioToTextDataLayer(manifest_filepath=eval_datasets,
labels=labels,
batch_size=32,
shuffle=False)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor()
spec_augment = nemo_asr.SpectrogramAugmentation(rect_masks=5)
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=64, **jasper_model_definition['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForCTC(feat_in=1024,
num_classes=len(labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(labels))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# Training DAG (Model)
audio_signal, audio_signal_len, transcript, transcript_len = data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
aug_signal = spec_augment(input_spec=processed_signal)
encoded, encoded_len = jasper_encoder(audio_signal=aug_signal,
length=processed_signal_len)
def create_all_dags(args, neural_factory):
'''
creates train and eval dags as well as their callbacks
returns train loss tensor and callbacks'''
# parse the config files
yaml = YAML(typ="safe")
with open(args.model_config) as f:
quartz_params = yaml.load(f)
vocab = quartz_params['labels']
sample_rate = quartz_params['sample_rate']
# Calculate num_workers for dataloader
total_cpus = os.cpu_count()
cpu_per_traindl = max(int(total_cpus / neural_factory.world_size), 1)
# create data layer for training
train_dl_params = copy.deepcopy(quartz_params["AudioToTextDataLayer"])
train_dl_params.update(quartz_params["AudioToTextDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
# del train_dl_params["normalize_transcripts"]
data_layer_train = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.train_dataset,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
**train_dl_params,
# normalize_transcripts=False
)
N = len(data_layer_train)
steps_per_epoch = int(
N / (args.batch_size * args.iter_per_step * args.num_gpus))
# create separate data layers for eval
# we need separate eval dags for separate eval datasets
# but all other modules in these dags will be shared
eval_dl_params = copy.deepcopy(quartz_params["AudioToTextDataLayer"])
eval_dl_params.update(quartz_params["AudioToTextDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
data_layers_eval = []
if args.eval_datasets:
for eval_dataset in args.eval_datasets:
data_layer_eval = nemo_asr.AudioToTextDataLayer(
manifest_filepath=eval_dataset,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.eval_batch_size,
num_workers=cpu_per_traindl,
**eval_dl_params,
)
data_layers_eval.append(data_layer_eval)
else:
nemo.logging.warning("There were no val datasets passed")
# create shared modules
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate,
**quartz_params["AudioToMelSpectrogramPreprocessor"])
# (QuartzNet uses the Jasper baseline encoder and decoder)
encoder = nemo_asr.JasperEncoder(
feat_in=quartz_params["AudioToMelSpectrogramPreprocessor"]["features"],
**quartz_params["JasperEncoder"])
decoder = nemo_asr.JasperDecoderForCTC(
feat_in=quartz_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(vocab))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# create augmentation modules (only used for training) if their configs
# are present
multiply_batch_config = quartz_params.get('MultiplyBatch', None)
if multiply_batch_config:
multiply_batch = nemo_asr.MultiplyBatch(**multiply_batch_config)
spectr_augment_config = quartz_params.get('SpectrogramAugmentation', None)
if spectr_augment_config:
data_spectr_augmentation = nemo_asr.SpectrogramAugmentation(
**spectr_augment_config)
# assemble train DAG
audio_signal_t, a_sig_length_t, \
transcript_t, transcript_len_t = data_layer_train()
processed_signal_t, p_length_t = data_preprocessor(
input_signal=audio_signal_t, length=a_sig_length_t)
if multiply_batch_config:
processed_signal_t, p_length_t, transcript_t, transcript_len_t = \
multiply_batch(
in_x=processed_signal_t, in_x_len=p_length_t,
in_y=transcript_t,
in_y_len=transcript_len_t)
if spectr_augment_config:
processed_signal_t = data_spectr_augmentation(
input_spec=processed_signal_t)
encoded_t, encoded_len_t = encoder(audio_signal=processed_signal_t,
length=p_length_t)
log_probs_t = decoder(encoder_output=encoded_t)
predictions_t = greedy_decoder(log_probs=log_probs_t)
loss_t = ctc_loss(log_probs=log_probs_t,
targets=transcript_t,
input_length=encoded_len_t,
target_length=transcript_len_t)
# create train callbacks
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss_t, predictions_t, transcript_t, transcript_len_t],
print_func=partial(monitor_asr_train_progress, labels=vocab),
get_tb_values=lambda x: [["loss", x[0]]],
tb_writer=neural_factory.tb_writer)
callbacks = [train_callback]
if args.checkpoint_dir or args.load_dir:
chpt_callback = nemo.core.CheckpointCallback(
folder=args.checkpoint_dir,
load_from_folder=args.load_dir,
step_freq=args.checkpoint_save_freq)
callbacks.append(chpt_callback)
# assemble eval DAGs
for i, eval_dl in enumerate(data_layers_eval):
audio_signal_e, a_sig_length_e, transcript_e, transcript_len_e = \
eval_dl()
processed_signal_e, p_length_e = data_preprocessor(
input_signal=audio_signal_e, length=a_sig_length_e)
encoded_e, encoded_len_e = encoder(audio_signal=processed_signal_e,
length=p_length_e)
log_probs_e = decoder(encoder_output=encoded_e)
predictions_e = greedy_decoder(log_probs=log_probs_e)
loss_e = ctc_loss(log_probs=log_probs_e,
targets=transcript_e,
input_length=encoded_len_e,
target_length=transcript_len_e)
# create corresponding eval callback
tagname = os.path.basename(args.eval_datasets[i]).split(".")[0]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[
loss_e, predictions_e, transcript_e, transcript_len_e
],
user_iter_callback=partial(process_evaluation_batch, labels=vocab),
user_epochs_done_callback=partial(process_evaluation_epoch,
tag=tagname),
eval_step=args.eval_freq,
tb_writer=neural_factory.tb_writer)
callbacks.append(eval_callback)
return loss_t, callbacks, steps_per_epoch
def test_simple_dags(self):
# module instantiation
with open("tests/data/jasper_smaller.yaml") as file:
jasper_model_definition = self.yaml.load(file)
labels = jasper_model_definition['labels']
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=labels,
batch_size=4)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**jasper_model_definition['AudioToMelSpectrogramPreprocessor'])
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_model_definition[
'AudioToMelSpectrogramPreprocessor']['features'],
**jasper_model_definition['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForCTC(feat_in=1024,
num_classes=len(labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(labels))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# DAG definition
audio_signal, audio_signal_len, transcript, transcript_len = \
data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
spec_augment = nemo_asr.SpectrogramAugmentation(rect_masks=5)
aug_signal = spec_augment(input_spec=processed_signal)
encoded, encoded_len = jasper_encoder(audio_signal=aug_signal,
length=processed_signal_len)
log_probs = jasper_decoder(encoder_output=encoded)
predictions = greedy_decoder(log_probs=log_probs)
loss = ctc_loss(log_probs=log_probs,
targets=transcript,
input_length=encoded_len,
target_length=transcript_len)
def wrong():
with open("tests/data/jasper_smaller.yaml") as file:
jasper_config = self.yaml.load(file)
labels = jasper_config['labels']
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=labels,
batch_size=4)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**jasper_config['AudioToMelSpectrogramPreprocessor'])
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_config['AudioToMelSpectrogramPreprocessor']
['features'],
**jasper_config['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=1024, num_classes=len(labels))
# DAG definition
audio_signal, audio_signal_len, transcript, transcript_len = \
data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
spec_augment = nemo_asr.SpectrogramAugmentation(rect_masks=5)
aug_signal = spec_augment(input_spec=processed_signal)
encoded, encoded_len = jasper_encoder(audio_signal=aug_signal,
length=processed_signal_len)
log_probs = jasper_decoder(encoder_output=processed_signal)
self.assertRaises(NeuralPortNmTensorMismatchError, wrong)
def create_dag(args, cfg, logger, num_gpus):
# Defining nodes
data = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.train_dataset,
labels=cfg['target']['labels'],
batch_size=cfg['optimization']['batch_size'],
eos_id=cfg['target']['eos_id'],
**cfg['AudioToTextDataLayer']['train'])
data_evals = []
if args.eval_datasets:
for val_path in args.eval_datasets:
data_evals.append(
nemo_asr.AudioToTextDataLayer(
manifest_filepath=val_path,
labels=cfg['target']['labels'],
batch_size=cfg['inference']['batch_size'],
eos_id=cfg['target']['eos_id'],
**cfg['AudioToTextDataLayer']['eval']))
else:
logger.info("There were no val datasets passed")
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**cfg['AudioToMelSpectrogramPreprocessor'])
data_augmentation = nemo_asr.SpectrogramAugmentation(
**cfg['SpectrogramAugmentation'])
encoder = nemo_asr.JasperEncoder(
feat_in=cfg["AudioToMelSpectrogramPreprocessor"]["features"],
**cfg['JasperEncoder'])
if args.encoder_checkpoint is not None \
and os.path.exists(args.encoder_checkpoint):
if cfg['JasperEncoder']['load']:
encoder.restore_from(args.encoder_checkpoint, args.local_rank)
logger.info(f'Loaded weights for encoder'
f' from {args.encoder_checkpoint}')
if cfg['JasperEncoder']['freeze']:
encoder.freeze()
logger.info(f'Freeze encoder weights')
connector = nemo_asr.JasperRNNConnector(
in_channels=cfg['JasperEncoder']['jasper'][-1]['filters'],
out_channels=cfg['DecoderRNN']['hidden_size'])
decoder = nemo.backends.pytorch.DecoderRNN(voc_size=len(
cfg['target']['labels']),
bos_id=cfg['target']['bos_id'],
**cfg['DecoderRNN'])
if args.decoder_checkpoint is not None \
and os.path.exists(args.decoder_checkpoint):
if cfg['DecoderRNN']['load']:
decoder.restore_from(args.decoder_checkpoint, args.local_rank)
logger.info(f'Loaded weights for decoder'
f' from {args.decoder_checkpoint}')
if cfg['DecoderRNN']['freeze']:
decoder.freeze()
logger.info(f'Freeze decoder weights')
if cfg['decoder']['unfreeze_attn']:
for name, param in decoder.attention.named_parameters():
param.requires_grad = True
logger.info(f'Unfreeze decoder attn weights')
num_data = len(data)
batch_size = cfg['optimization']['batch_size']
num_epochs = cfg['optimization']['params']['num_epochs']
steps_per_epoch = int(num_data / (batch_size * num_gpus))
total_steps = num_epochs * steps_per_epoch
vsc = ValueSetterCallback
tf_callback = ValueSetterCallback(
decoder,
'teacher_forcing',
policies=[vsc.Policy(vsc.Method.Const(1.0), start=0.0, end=1.0)],
total_steps=total_steps)
seq_loss = nemo.backends.pytorch.SequenceLoss(
pad_id=cfg['target']['pad_id'],
smoothing_coef=cfg['optimization']['smoothing_coef'],
sample_wise=cfg['optimization']['sample_wise'])
se_callback = ValueSetterCallback(seq_loss,
'smoothing_coef',
policies=[
vsc.Policy(vsc.Method.Const(
seq_loss.smoothing_coef),
start=0.0,
end=1.0),
],
total_steps=total_steps)
beam_search = nemo.backends.pytorch.BeamSearch(
decoder=decoder,
pad_id=cfg['target']['pad_id'],
bos_id=cfg['target']['bos_id'],
eos_id=cfg['target']['eos_id'],
max_len=cfg['target']['max_len'],
beam_size=cfg['inference']['beam_size'])
uf_callback = UnfreezeCallback(
[encoder, decoder], start_epoch=cfg['optimization']['start_unfreeze'])
saver_callback = nemo.core.ModuleSaverCallback(
save_modules_list=[encoder, connector, decoder],
folder=args.checkpoint_dir,
step_freq=args.eval_freq)
# Creating DAG
audios, audio_lens, transcripts, _ = data()
processed_audios, processed_audio_lens = data_preprocessor(
input_signal=audios, length=audio_lens)
augmented_spec = data_augmentation(input_spec=processed_audios)
encoded, _ = encoder(audio_signal=augmented_spec,
length=processed_audio_lens)
encoded = connector(tensor=encoded)
log_probs, _ = decoder(targets=transcripts, encoder_outputs=encoded)
train_loss = seq_loss(log_probs=log_probs, targets=transcripts)
evals = []
for i, data_eval in enumerate(data_evals):
audios, audio_lens, transcripts, _ = data_eval()
processed_audios, processed_audio_lens = data_preprocessor(
input_signal=audios, length=audio_lens)
encoded, _ = encoder(audio_signal=processed_audios,
length=processed_audio_lens)
encoded = connector(tensor=encoded)
log_probs, _ = decoder(targets=transcripts, encoder_outputs=encoded)
loss = seq_loss(log_probs=log_probs, targets=transcripts)
predictions, aw = beam_search(encoder_outputs=encoded)
evals.append((args.eval_datasets[i], (loss, log_probs, transcripts,
predictions, aw)))
# Update config
cfg['num_params'] = {
'encoder': encoder.num_weights,
'connector': connector.num_weights,
'decoder': decoder.num_weights
}
cfg['num_params']['total'] = sum(cfg['num_params'].values())
cfg['input']['train'] = {'num_data': num_data}
cfg['optimization']['steps_per_epoch'] = steps_per_epoch
cfg['optimization']['total_steps'] = total_steps
return (train_loss, evals), cfg, [
tf_callback, se_callback, uf_callback, saver_callback
]