def build_components(self,encoder_module=None,decoder_module=None):
"""
Initializes all neural modules
Arguments:
encoder_module: A neural module with the same neural type signature
as the Jasper Encoder
decoder_module: A neural module with the same neural type signature
as the Jasper CTC decoder
"""
self.data_layer = AudioInferDataLayer(sample_rate=self.model_definition['sample_rate'])
self.data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=self.model_definition['sample_rate'],
**self.model_definition["AudioToMelSpectrogramPreprocessor"])
if encoder_module is not None:
# Pass in an already instantiated neural module for the encoder
assert isinstance(encoder_module,NeuralModule), 'encoder is not a neural module'
self.encoder = encoder_module
else:
self.encoder = nemo_asr.JasperEncoder(
feat_in=self.model_definition['AudioToMelSpectrogramPreprocessor']['features'],
**self.model_definition['JasperEncoder'])
if decoder_module is not None:
# Pass in an already instantiated neural module for the decoder
assert isinstance(decoder_module,NeuralModule), 'decoder is not a neural module'
self.decoder = decoder_module
else:
self.decoder = nemo_asr.JasperDecoderForCTC(
feat_in=self.model_definition["JasperEncoder"]["jasper"][-1]["filters"],num_classes=len(self.vocab))
self.greedy_decoder = nemo_asr.GreedyCTCDecoder()
def __init__(self):
"""Loads pre-trained ASR model"""
self.asr_conf = parse_yaml()["asr"]
device = nemo.core.DeviceType.CPU
self.nf = nemo.core.NeuralModuleFactory(placement=device)
# load model configuration
jasper_params = parse_yaml(
os.path.join(self.asr_conf["model_dir"], "quartznet15x5.yaml"))
self.labels = jasper_params["labels"]
self.sample_rate = jasper_params["sample_rate"]
# preprocessor
self.eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"])
self.eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"])
del self.eval_dl_params["train"]
del self.eval_dl_params["eval"]
self.preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate = self.sample_rate,
**jasper_params["AudioPreprocessing"])
# model encoder
feats = jasper_params["AudioPreprocessing"]["features"]
self.jasper_encoder = nemo_asr.JasperEncoder(
feat_in = feats,
**jasper_params["JasperEncoder"])
self.jasper_encoder.restore_from(
os.path.join(self.asr_conf["model_dir"],
"JasperEncoder-STEP-247400.pt"))
# model decoder
filters = jasper_params["JasperEncoder"]["jasper"][-1]["filters"]
self.jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in = filters,
num_classes=len(self.labels))
self.jasper_decoder.restore_from(
os.path.join(self.asr_conf["model_dir"],
"JasperDecoderForCTC-STEP-247400.pt"))
self.nf.logger.info('================================')
self.nf.logger.info(
f"Number of parameters in encoder: {self.jasper_encoder.num_weights}")
self.nf.logger.info(
f"Number of parameters in decoder: {self.jasper_decoder.num_weights}")
self.nf.logger.info(
f"Total number of parameters in model: "
f"{self.jasper_decoder.num_weights + self.jasper_encoder.num_weights}")
self.nf.logger.info('================================')
# CTC decoder
if self.asr_conf["decoder"] == "beam":
self.ctc_decoder = nemo_asr.BeamSearchDecoderWithLM(
vocab = self.labels,
beam_width = self.asr_conf["beam_width"],
alpha = self.asr_conf["alpha"],
beta = self.asr_conf["beta"],
lm_path = self.asr_conf["lm_path"],
num_cpus = max(os.cpu_count(), 1))
else:
self.ctc_decoder = nemo_asr.GreedyCTCDecoder()
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
def main():
parser = argparse.ArgumentParser(description='Jasper')
parser.add_argument("--local_rank", default=None, type=int)
parser.add_argument("--batch_size", default=32, type=int)
parser.add_argument("--model_config", type=str, required=True)
parser.add_argument("--eval_datasets", type=str, required=True)
parser.add_argument("--load_dir", type=str, required=True)
parser.add_argument("--vocab_file", type=str, required=True)
parser.add_argument("--save_logprob", default=None, type=str)
parser.add_argument("--lm_path", default=None, type=str)
parser.add_argument("--beam_width", default=50, type=int)
parser.add_argument("--alpha", default=2.0, type=float)
parser.add_argument("--beta", default=1.0, type=float)
parser.add_argument("--cutoff_prob", default=0.99, type=float)
parser.add_argument("--cutoff_top_n", default=40, type=int)
args = parser.parse_args()
batch_size = args.batch_size
load_dir = args.load_dir
if args.local_rank is not None:
if args.lm_path:
raise NotImplementedError(
"Beam search decoder with LM does not currently support "
"evaluation on multi-gpu.")
device = nemo.core.DeviceType.AllGpu
else:
device = nemo.core.DeviceType.GPU
# Instantiate Neural Factory with supported backend
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=args.local_rank,
optimization_level=nemo.core.Optimization.mxprO1,
placement=device)
logger = neural_factory.logger
if args.local_rank is not None:
logger.info('Doing ALL GPU')
yaml = YAML(typ="safe")
with open(args.model_config) as f:
jasper_params = yaml.load(f)
vocab = load_vocab(args.vocab_file)
sample_rate = jasper_params['sample_rate']
eval_datasets = args.eval_datasets
eval_dl_params = copy.deepcopy(jasper_params["AudioToTextDataLayer"])
eval_dl_params.update(jasper_params["AudioToTextDataLayer"]["eval"])
eval_dl_params["normalize_transcripts"] = False
del eval_dl_params["train"]
del eval_dl_params["eval"]
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=eval_datasets,
sample_rate=sample_rate,
labels=vocab,
batch_size=batch_size,
**eval_dl_params)
n = len(data_layer)
logger.info('Evaluating {0} examples'.format(n))
data_preprocessor = nemo_asr.AudioPreprocessing(
sample_rate=sample_rate,
**jasper_params["AudioPreprocessing"])
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_params["AudioPreprocessing"]["features"],
**jasper_params["JasperEncoder"])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
if args.lm_path:
beam_width = args.beam_width
alpha = args.alpha
beta = args.beta
cutoff_prob = args.cutoff_prob
cutoff_top_n = args.cutoff_top_n
beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM(
vocab=vocab,
beam_width=beam_width,
alpha=alpha,
beta=beta,
cutoff_prob=cutoff_prob,
cutoff_top_n=cutoff_top_n,
lm_path=args.lm_path,
num_cpus=max(os.cpu_count(), 1))
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('================================')
audio_signal_e1, a_sig_length_e1, transcript_e1, transcript_len_e1 = \
data_layer()
processed_signal_e1, p_length_e1 = data_preprocessor(
input_signal=audio_signal_e1,
length=a_sig_length_e1)
encoded_e1, encoded_len_e1 = jasper_encoder(
audio_signal=processed_signal_e1,
length=p_length_e1)
log_probs_e1 = jasper_decoder(encoder_output=encoded_e1)
predictions_e1 = greedy_decoder(log_probs=log_probs_e1)
eval_tensors = [log_probs_e1, predictions_e1,
transcript_e1, transcript_len_e1, encoded_len_e1]
if args.lm_path:
beam_predictions_e1 = beam_search_with_lm(
log_probs=log_probs_e1, log_probs_length=encoded_len_e1)
eval_tensors.append(beam_predictions_e1)
evaluated_tensors = neural_factory.infer(
tensors=eval_tensors,
checkpoint_dir=load_dir,
)
greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab)
references = post_process_transcripts(
evaluated_tensors[2], evaluated_tensors[3], vocab)
cer = word_error_rate(hypotheses=greedy_hypotheses,
references=references,
use_cer=True)
logger.info("Greedy CER {:.2f}%".format(cer * 100))
if args.lm_path:
beam_hypotheses = []
# Over mini-batch
for i in evaluated_tensors[-1]:
# Over samples
for j in i:
beam_hypotheses.append(j[0][1])
cer = word_error_rate(
hypotheses=beam_hypotheses, references=references, use_cer=True)
logger.info("Beam CER {:.2f}".format(cer * 100))
if args.save_logprob:
# Convert logits to list of numpy arrays
logprob = []
for i, batch in enumerate(evaluated_tensors[0]):
for j in range(batch.shape[0]):
logprob.append(
batch[j][:evaluated_tensors[4][i][j], :].cpu().numpy())
with open(args.save_logprob, 'wb') as f:
pickle.dump(logprob, f, protocol=pickle.HIGHEST_PROTOCOL)
def main():
parser = argparse.ArgumentParser(parents=[nm_argparse.NemoArgParser()],
description='AN4 ASR',
conflict_handler='resolve')
# Overwrite default args
parser.add_argument("--train_dataset",
type=str,
help="training dataset path")
parser.add_argument("--eval_datasets",
type=str,
nargs=1,
help="validation dataset path")
# Create new args
parser.add_argument("--lm", default="./an4-lm.3gram.binary", type=str)
parser.add_argument("--test_after_training", action='store_true')
parser.add_argument("--momentum", type=float)
parser.add_argument("--beta1", default=0.95, type=float)
parser.add_argument("--beta2", default=0.25, type=float)
parser.set_defaults(
model_config="./configs/jasper_an4.yaml",
train_dataset="/home/mrjenkins/TestData/an4_dataset/an4_train.json",
eval_datasets="/home/mrjenkins/TestData/an4_dataset/an4_val.json",
work_dir="./tmp",
checkpoint_dir="./tmp",
optimizer="novograd",
num_epochs=50,
batch_size=32,
eval_batch_size=16,
lr=0.02,
weight_decay=0.005,
checkpoint_save_freq=1000,
eval_freq=100,
amp_opt_level="O1")
args = parser.parse_args()
betas = (args.beta1, args.beta2)
wer_thr = 0.20
beam_wer_thr = 0.15
nf = nemo.core.NeuralModuleFactory(local_rank=args.local_rank,
optimization_level=args.amp_opt_level,
random_seed=0,
log_dir=args.work_dir,
checkpoint_dir=args.checkpoint_dir,
create_tb_writer=True,
cudnn_benchmark=args.cudnn_benchmark)
tb_writer = nf.tb_writer
checkpoint_dir = nf.checkpoint_dir
args.checkpoint_dir = nf.checkpoint_dir
# Load model definition
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']
# build train and eval model
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"]
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.train_dataset,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.batch_size,
**train_dl_params)
num_samples = len(data_layer)
total_steps = int(num_samples * args.num_epochs / args.batch_size)
print("Train samples=", num_samples, "num_steps=", total_steps)
data_preprocessor = nemo_asr.AudioPreprocessing(
sample_rate=sample_rate, **jasper_params["AudioPreprocessing"])
# data_augmentation = nemo_asr.SpectrogramAugmentation(
# **jasper_params['SpectrogramAugmentation']
# )
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_layer_eval = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.eval_datasets,
sample_rate=sample_rate,
labels=vocab,
batch_size=args.eval_batch_size,
**eval_dl_params)
num_samples = len(data_layer_eval)
nf.logger.info(f"Eval samples={num_samples}")
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_params["AudioPreprocessing"]["features"],
**jasper_params["JasperEncoder"])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(vocab))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# Training model
audio, audio_len, transcript, transcript_len = data_layer()
processed, processed_len = data_preprocessor(input_signal=audio,
length=audio_len)
encoded, encoded_len = jasper_encoder(audio_signal=processed,
length=processed_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)
# Evaluation model
audio_e, audio_len_e, transcript_e, transcript_len_e = data_layer_eval()
processed_e, processed_len_e = data_preprocessor(input_signal=audio_e,
length=audio_len_e)
encoded_e, encoded_len_e = jasper_encoder(audio_signal=processed_e,
length=processed_len_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)
nf.logger.info("Num of params in encoder: {0}".format(
jasper_encoder.num_weights))
# Callbacks to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss, predictions, transcript, transcript_len],
print_func=lambda x: monitor_asr_train_progress(x, labels=vocab),
get_tb_values=lambda x: [["loss", x[0]]],
tb_writer=tb_writer,
)
checkpointer_callback = nemo.core.CheckpointCallback(
folder=checkpoint_dir, step_freq=args.checkpoint_save_freq)
eval_tensors = [loss_e, predictions_e, transcript_e, transcript_len_e]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=eval_tensors,
user_iter_callback=lambda x, y: process_evaluation_batch(
x, y, labels=vocab),
user_epochs_done_callback=process_evaluation_epoch,
eval_step=args.eval_freq,
tb_writer=tb_writer)
nf.train(tensors_to_optimize=[loss],
callbacks=[train_callback, eval_callback, checkpointer_callback],
optimizer=args.optimizer,
lr_policy=CosineAnnealing(total_steps=total_steps),
optimization_params={
"num_epochs": args.num_epochs,
"max_steps": args.max_steps,
"lr": args.lr,
"momentum": args.momentum,
"betas": betas,
"weight_decay": args.weight_decay,
"grad_norm_clip": None
},
batches_per_step=args.iter_per_step)
if args.test_after_training:
# Create BeamSearch NM
beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM(
vocab=vocab,
beam_width=64,
alpha=2.,
beta=1.5,
lm_path=args.lm,
num_cpus=max(os.cpu_count(), 1))
beam_predictions = beam_search_with_lm(log_probs=log_probs_e,
log_probs_length=encoded_len_e)
eval_tensors.append(beam_predictions)
evaluated_tensors = nf.infer(eval_tensors)
greedy_hypotheses = post_process_predictions(evaluated_tensors[1],
vocab)
references = post_process_transcripts(evaluated_tensors[2],
evaluated_tensors[3], vocab)
wer = word_error_rate(hypotheses=greedy_hypotheses,
references=references)
nf.logger.info("Greedy WER: {:.2f}".format(wer * 100))
assert wer <= wer_thr, (
"Final eval greedy WER {:.2f}% > than {:.2f}%".format(
wer * 100, wer_thr * 100))
beam_hypotheses = []
# Over mini-batch
for i in evaluated_tensors[-1]:
# Over samples
for j in i:
beam_hypotheses.append(j[0][1])
beam_wer = word_error_rate(hypotheses=beam_hypotheses,
references=references)
nf.logger.info("Beam WER {:.2f}%".format(beam_wer * 100))
assert beam_wer <= beam_wer_thr, (
"Final eval beam WER {:.2f}% > than {:.2f}%".format(
beam_wer * 100, beam_wer_thr * 100))
assert beam_wer <= wer, ("Final eval beam WER > than the greedy WER.")
# Reload model weights and train for extra 10 epochs
checkpointer_callback = nemo.core.CheckpointCallback(
folder=checkpoint_dir,
step_freq=args.checkpoint_save_freq,
force_load=True)
nf.reset_trainer()
nf.train(tensors_to_optimize=[loss],
callbacks=[train_callback, checkpointer_callback],
optimizer=args.optimizer,
optimization_params={
"num_epochs": args.num_epochs + 10,
"lr": args.lr,
"momentum": args.momentum,
"betas": betas,
"weight_decay": args.weight_decay,
"grad_norm_clip": None
},
reset=True)
evaluated_tensors = nf.infer(eval_tensors[:-1])
greedy_hypotheses = post_process_predictions(evaluated_tensors[1],
vocab)
references = post_process_transcripts(evaluated_tensors[2],
evaluated_tensors[3], vocab)
wer_new = word_error_rate(hypotheses=greedy_hypotheses,
references=references)
nf.logger.info("New greedy WER: {:.2f}%".format(wer_new * 100))
assert wer_new <= wer * 1.1, (
f"Fine tuning: new WER {wer * 100:.2f}% > than the previous WER "
f"{wer_new * 100:.2f}%")
def main():
parser = argparse.ArgumentParser(description='Jasper')
parser.add_argument("--local_rank", default=None, type=int)
parser.add_argument("--batch_size", default=32, type=int)
parser.add_argument("--model_config", type=str, required=True)
parser.add_argument("--eval_datasets", type=str, required=True)
parser.add_argument("--load_dir", type=str, required=True)
parser.add_argument("--save_logprob", default=None, type=str)
parser.add_argument("--lm_path", default=None, type=str)
parser.add_argument('--alpha',
default=2.,
type=float,
help='value of LM weight',
required=False)
parser.add_argument(
'--alpha_max',
type=float,
help='maximum value of LM weight (for a grid search in \'eval\' mode)',
required=False)
parser.add_argument('--alpha_step',
type=float,
help='step for LM weight\'s tuning in \'eval\' mode',
required=False,
default=0.1)
parser.add_argument('--beta',
default=1.5,
type=float,
help='value of word count weight',
required=False)
parser.add_argument(
'--beta_max',
type=float,
help='maximum value of word count weight (for a grid search in \
\'eval\' mode',
required=False)
parser.add_argument(
'--beta_step',
type=float,
help='step for word count weight\'s tuning in \'eval\' mode',
required=False,
default=0.1)
parser.add_argument("--beam_width", default=128, type=int)
args = parser.parse_args()
batch_size = args.batch_size
load_dir = args.load_dir
if args.local_rank is not None:
if args.lm_path:
raise NotImplementedError(
"Beam search decoder with LM does not currently support "
"evaluation on multi-gpu.")
device = nemo.core.DeviceType.AllGpu
else:
device = nemo.core.DeviceType.GPU
# Instantiate Neural Factory with supported backend
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=args.local_rank,
optimization_level=nemo.core.Optimization.mxprO1,
placement=device)
logger = neural_factory.logger
if args.local_rank is not None:
logger.info('Doing ALL GPU')
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']
eval_datasets = args.eval_datasets
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_layer = nemo_asr.AudioToTextDataLayer(manifest_filepath=eval_datasets,
sample_rate=sample_rate,
labels=vocab,
batch_size=batch_size,
**eval_dl_params)
N = len(data_layer)
logger.info('Evaluating {0} examples'.format(N))
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate,
**jasper_params["AudioToMelSpectrogramPreprocessor"])
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))
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('================================')
audio_signal_e1, a_sig_length_e1, transcript_e1, transcript_len_e1 =\
data_layer()
processed_signal_e1, p_length_e1 = data_preprocessor(
input_signal=audio_signal_e1, length=a_sig_length_e1)
encoded_e1, encoded_len_e1 = jasper_encoder(
audio_signal=processed_signal_e1, length=p_length_e1)
log_probs_e1 = jasper_decoder(encoder_output=encoded_e1)
predictions_e1 = greedy_decoder(log_probs=log_probs_e1)
eval_tensors = [
log_probs_e1, predictions_e1, transcript_e1, transcript_len_e1,
encoded_len_e1
]
evaluated_tensors = neural_factory.infer(tensors=eval_tensors,
checkpoint_dir=load_dir,
cache=True)
greedy_hypotheses = post_process_predictions(evaluated_tensors[1], vocab)
references = post_process_transcripts(evaluated_tensors[2],
evaluated_tensors[3], vocab)
wer = word_error_rate(hypotheses=greedy_hypotheses, references=references)
logger.info("Greedy WER {:.2f}%".format(wer * 100))
if args.lm_path:
if args.alpha_max is None:
args.alpha_max = args.alpha
# include alpha_max in tuning range
args.alpha_max += args.alpha_step / 10.0
if args.beta_max is None:
args.beta_max = args.beta
# include beta_max in tuning range
args.beta_max += args.beta_step / 10.0
beam_wers = []
for alpha in np.arange(args.alpha, args.alpha_max, args.alpha_step):
for beta in np.arange(args.beta, args.beta_max, args.beta_step):
logger.info('================================')
logger.info(f'Infering with (alpha, beta): ({alpha}, {beta})')
beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM(
vocab=vocab,
beam_width=args.beam_width,
alpha=alpha,
beta=beta,
lm_path=args.lm_path,
num_cpus=max(os.cpu_count(), 1))
beam_predictions_e1 = beam_search_with_lm(
log_probs=log_probs_e1, log_probs_length=encoded_len_e1)
evaluated_tensors = neural_factory.infer(
tensors=[beam_predictions_e1],
use_cache=True,
verbose=False)
beam_hypotheses = []
# Over mini-batch
for i in evaluated_tensors[-1]:
# Over samples
for j in i:
beam_hypotheses.append(j[0][1])
wer = word_error_rate(hypotheses=beam_hypotheses,
references=references)
logger.info("Beam WER {:.2f}%".format(wer * 100))
beam_wers.append(((alpha, beta), wer * 100))
logger.info('Beam WER for (alpha, beta)')
logger.info('================================')
logger.info('\n' + '\n'.join([str(e) for e in beam_wers]))
logger.info('================================')
best_beam_wer = min(beam_wers, key=lambda x: x[1])
logger.info('Best (alpha, beta): '
f'{best_beam_wer[0]}, '
f'WER: {best_beam_wer[1]:.2f}%')
def create_dags(jasper_params, args, nf):
vocab = jasper_params['labels']
# build train and eval model
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"]
data_layer = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.train_dataset,
labels=vocab,
batch_size=args.batch_size,
**train_dl_params)
num_samples = len(data_layer)
steps_per_epoch = math.ceil(
num_samples / (args.batch_size * args.iter_per_step * nf.world_size))
total_steps = steps_per_epoch * args.num_epochs
print("Train samples=", num_samples, "num_steps=", total_steps)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
**jasper_params["AudioToMelSpectrogramPreprocessor"])
# data_augmentation = nemo_asr.SpectrogramAugmentation(
# **jasper_params['SpectrogramAugmentation']
# )
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_layer_eval = nemo_asr.AudioToTextDataLayer(
manifest_filepath=args.eval_datasets,
labels=vocab,
batch_size=args.eval_batch_size,
**eval_dl_params)
num_samples = len(data_layer_eval)
nemo.logging.info(f"Eval samples={num_samples}")
jasper_encoder = nemo_asr.JasperEncoder(**jasper_params["JasperEncoder"])
jasper_decoder = nemo_asr.JasperDecoderForCTC(
num_classes=len(vocab), **jasper_params["JasperDecoderForCTC"])
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# Training model
audio, audio_len, transcript, transcript_len = data_layer()
processed, processed_len = data_preprocessor(input_signal=audio,
length=audio_len)
encoded, encoded_len = jasper_encoder(audio_signal=processed,
length=processed_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)
# Evaluation model
audio_e, audio_len_e, transcript_e, transcript_len_e = data_layer_eval()
processed_e, processed_len_e = data_preprocessor(input_signal=audio_e,
length=audio_len_e)
encoded_e, encoded_len_e = jasper_encoder(audio_signal=processed_e,
length=processed_len_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)
nemo.logging.info("Num of params in encoder: {0}".format(
jasper_encoder.num_weights))
# Callbacks to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss, predictions, transcript, transcript_len],
print_func=partial(monitor_asr_train_progress, labels=vocab),
get_tb_values=lambda x: [["loss", x[0]]],
tb_writer=nf.tb_writer,
)
checkpointer_callback = nemo.core.CheckpointCallback(
folder=nf.checkpoint_dir, step_freq=args.checkpoint_save_freq)
eval_tensors = [loss_e, predictions_e, transcript_e, transcript_len_e]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=eval_tensors,
user_iter_callback=partial(process_evaluation_batch, labels=vocab),
user_epochs_done_callback=process_evaluation_epoch,
eval_step=args.eval_freq,
tb_writer=nf.tb_writer)
callbacks = [train_callback, checkpointer_callback, eval_callback]
return (loss, eval_tensors, callbacks, total_steps, vocab, log_probs_e,
encoded_len_e)
neural_factory = nemo.core.NeuralModuleFactory(
placement=(DeviceType.GPU
if torch.cuda.is_available() else DeviceType.CPU),
backend=nemo.core.Backend.PyTorch)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
factory=neural_factory)
jasper_encoder = nemo_asr.JasperEncoder(
jasper=jasper_model_definition['JasperEncoder']['jasper'],
activation=jasper_model_definition['JasperEncoder']['activation'],
feat_in=jasper_model_definition['AudioToMelSpectrogramPreprocessor']
['features'])
jasper_encoder.restore_from(CHECKPOINT_ENCODER, local_rank=0)
jasper_decoder = nemo_asr.JasperDecoderForCTC(feat_in=1024,
num_classes=len(labels))
jasper_decoder.restore_from(CHECKPOINT_DECODER, local_rank=0)
greedy_decoder = nemo_asr.GreedyCTCDecoder()
if ENABLE_NGRAM and os.path.isfile(LM_PATH):
beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM(vocab=labels,
beam_width=64,
alpha=2.0,
beta=1.0,
lm_path=LM_PATH,
num_cpus=max(
os.cpu_count(),
1))
else:
print("Beam search is not enabled")
from app import routes # noqa
if __name__ == '__main__':
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 offline_inference(config, encoder, decoder, audio_file):
MODEL_YAML = config
CHECKPOINT_ENCODER = encoder
CHECKPOINT_DECODER = decoder
sample_rate, signal = wave.read(audio_file)
# get labels (vocab)
yaml = YAML(typ="safe")
with open(MODEL_YAML) as f:
jasper_model_definition = yaml.load(f)
labels = jasper_model_definition['labels']
# build neural factory and neural modules
neural_factory = nemo.core.NeuralModuleFactory(
placement=nemo.core.DeviceType.GPU,
backend=nemo.core.Backend.PyTorch)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
factory=neural_factory,
**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=jasper_model_definition["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(labels))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# load model
jasper_encoder.restore_from(CHECKPOINT_ENCODER)
jasper_decoder.restore_from(CHECKPOINT_DECODER)
# AudioDataLayer
class AudioDataLayer(DataLayerNM):
@staticmethod
def create_ports():
input_ports = {}
output_ports = {
"audio_signal": NeuralType({0: AxisType(BatchTag),
1: AxisType(TimeTag)}),
"a_sig_length": NeuralType({0: AxisType(BatchTag)}),
}
return input_ports, output_ports
def __init__(self, **kwargs):
DataLayerNM.__init__(self, **kwargs)
self.output_enable = False
def __iter__(self):
return self
def __next__(self):
if not self.output_enable:
raise StopIteration
self.output_enable = False
return torch.as_tensor(self.signal, dtype=torch.float32), \
torch.as_tensor(self.signal_shape, dtype=torch.int64)
def set_signal(self, signal):
self.signal = np.reshape(signal.astype(np.float32)/32768., [1, -1])
self.signal_shape = np.expand_dims(self.signal.size, 0).astype(np.int64)
self.output_enable = True
def __len__(self):
return 1
@property
def dataset(self):
return None
@property
def data_iterator(self):
return self
# Instantiate necessary neural modules
data_layer = AudioDataLayer()
# Define inference DAG
audio_signal, audio_signal_len = data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal,
length=audio_signal_len)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=processed_signal_len)
log_probs = jasper_decoder(encoder_output=encoded)
predictions = greedy_decoder(log_probs=log_probs)
# audio inference
data_layer.set_signal(signal)
tensors = neural_factory.infer([
audio_signal,
processed_signal,
encoded,
log_probs,
predictions], verbose=False)
# results
audio = tensors[0][0][0].cpu().numpy()
features = tensors[1][0][0].cpu().numpy()
encoded_features = tensors[2][0][0].cpu().numpy(),
probs = tensors[3][0][0].cpu().numpy()
preds = tensors[4][0]
transcript = post_process_predictions([preds], labels)
return transcript, audio, features, encoded_features, probs, preds
def convert(request):
"""
** Create new sound recognation object by convert audio to text .
** Use Case Exemple of Post:
{
"audio":"base64 format",
}
"""
import json
from ruamel.yaml import YAML
import nemo
import nemo_asr
import IPython.display as ipd
MODEL_YAML = "/home/docker/app/ai_models/quartznet15x5.yaml"
CHECKPOINT_ENCODER = "/home/docker/app/ai_models/JasperEncoder-STEP-243800.pt"
CHECKPOINT_DECODER = "/home/docker/app/ai_models/JasperDecoderForCTC-STEP-243800.pt"
ENABLE_NGRAM = False
yaml = YAML(typ="safe")
with open(MODEL_YAML) as f:
jasper_model_definition = yaml.load(f)
labels = jasper_model_definition['labels']
neural_factory = nemo.core.NeuralModuleFactory(
placement=nemo.core.DeviceType.CPU, backend=nemo.core.Backend.PyTorch)
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
factory=neural_factory)
jasper_encoder = nemo_asr.JasperEncoder(
jasper=jasper_model_definition['JasperEncoder']['jasper'],
activation=jasper_model_definition['JasperEncoder']['activation'],
feat_in=jasper_model_definition['AudioToMelSpectrogramPreprocessor']
['features'])
jasper_encoder.restore_from(CHECKPOINT_ENCODER, local_rank=0)
jasper_decoder = nemo_asr.JasperDecoderForCTC(feat_in=1024,
num_classes=len(labels))
jasper_decoder.restore_from(CHECKPOINT_DECODER, local_rank=0)
greedy_decoder = nemo_asr.GreedyCTCDecoder()
def wav_to_text(manifest, greedy=True):
from ruamel.yaml import YAML
yaml = YAML(typ="safe")
with open(MODEL_YAML) as f:
jasper_model_definition = yaml.load(f)
labels = jasper_model_definition['labels']
data_layer = nemo_asr.AudioToTextDataLayer(shuffle=False,
manifest_filepath=manifest,
labels=labels,
batch_size=1)
audio_signal, audio_signal_len, _, _ = data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=processed_signal_len)
log_probs = jasper_decoder(encoder_output=encoded)
predictions = greedy_decoder(log_probs=log_probs)
if ENABLE_NGRAM:
print('Running with beam search')
beam_predictions = beam_search_with_lm(
log_probs=log_probs, log_probs_length=encoded_len)
eval_tensors = [beam_predictions]
if greedy:
eval_tensors = [predictions]
tensors = neural_factory.infer(tensors=eval_tensors)
if greedy:
from nemo_asr.helpers import post_process_predictions
prediction = post_process_predictions(tensors[0], labels)
else:
prediction = tensors[0][0][0][0][1]
return prediction
def create_manifest(file_path):
# create manifest
manifest = dict()
manifest['audio_filepath'] = file_path
manifest['duration'] = 18000
manifest['text'] = 'todo'
with open(file_path + ".json", 'w') as fout:
fout.write(json.dumps(manifest))
return file_path + ".json"
data = request.FILES['audio']
path = "media/" + data.name
audio = Song.objects.create(audio_file=data)
transcription = wav_to_text(create_manifest(audio.audio_file.path))
return Response({'Output': transcription})
