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 test_stft_conv(self):
with open(
os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"../data/quartznet_speech_recognition.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=2,
)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
'stft_conv': True,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
jasper_encoder = nemo_asr.JasperEncoder(
**jasper_model_definition['JasperEncoder'], )
jasper_decoder = nemo_asr.JasperDecoderForClassification(
feat_in=jasper_model_definition['JasperEncoder']['jasper'][-1]
['filters'],
num_classes=len(self.labels))
ce_loss = nemo_asr.CrossEntropyLossNM()
# DAG
audio_signal, a_sig_length, targets, targets_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
logits = jasper_decoder(encoder_output=encoded)
loss = ce_loss(logits=logits, labels=targets)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss],
print_func=lambda x: logging.info(str(x[0].item())))
# Instantiate an optimizer to perform `train` action
optimizer = nemo.backends.pytorch.actions.PtActions()
optimizer.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"num_epochs": 10,
"lr": 0.0003
},
)
def test_jasper_eval(self):
with open(
os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"../data/quartznet_speech_recognition.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=2,
)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
jasper_encoder = nemo_asr.JasperEncoder(
**jasper_model_definition['JasperEncoder'], )
jasper_decoder = nemo_asr.JasperDecoderForClassification(
feat_in=jasper_model_definition['JasperEncoder']['jasper'][-1]
['filters'],
num_classes=len(self.labels))
ce_loss = nemo_asr.CrossEntropyLossNM()
# DAG
audio_signal, a_sig_length, targets, targets_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
logits = jasper_decoder(encoder_output=encoded)
loss = ce_loss(
logits=logits,
labels=targets,
)
from nemo.collections.asr.helpers import (
process_classification_evaluation_batch,
process_classification_evaluation_epoch,
)
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[loss, logits, targets],
user_iter_callback=lambda x, y:
process_classification_evaluation_batch(x, y, top_k=[1]),
user_epochs_done_callback=process_classification_evaluation_epoch,
)
# Instantiate an optimizer to perform `train` action
self.nf.eval(callbacks=[eval_callback])
def __init__(
self,
n_vocab: int,
d_char: int,
pad_id: int,
jasper_kwargs: Dict[str, Any],
d_out: int,
d_speaker_emb: Optional[int] = None,
d_speaker_x: Optional[int] = None,
d_speaker_o: Optional[int] = None,
pad16: bool = False,
poly_span: bool = False,
doubling: bool = False,
):
"""Creates TalkNet backbone instance.
Args:
n_vocab: Size of input vocabulary.
d_char: Dimension of char embedding.
pad_id: Id of padding symbol.
jasper_kwargs: Kwargs to instantiate QN encoder.
d_out: Dimension of output.
d_speaker_emb: Dimension of speaker embedding.
d_speaker_x: Dimension of pre speaker embedding.
d_speaker_o: Dimension of post speaker embedding.
pad16: True if pad tensors to 16.
poly_span: True if assign polynomial span embeddings for blanks.
doubling: True if using mel channels doubling trick.
"""
super().__init__()
# Embedding for input text
self.text_emb = nn.Embedding(n_vocab, d_char, padding_idx=pad_id).to(self._device)
self.text_emb.weight.data.uniform_(-1, 1)
# PolySpan
self.ps = PolySpanEmb(self.text_emb)
self._poly_span = poly_span
# Embedding for speaker
if d_speaker_emb is not None:
self.speaker_in = nn.Linear(d_speaker_emb, d_speaker_x).to(self._device)
self.speaker_out = nn.Linear(d_speaker_emb, d_speaker_o).to(self._device)
else:
self.speaker_in, self.speaker_out = None, None
jasper_params = jasper_kwargs['jasper']
d_enc_out = jasper_params[-1]["filters"]
d_x = d_char + (int(d_speaker_x or 0) if d_speaker_emb else 0)
self.jasper = nemo_asr.JasperEncoder(feat_in=d_x, **jasper_kwargs).to(self._device)
d_o = d_enc_out + (int(d_speaker_o or 0) if d_speaker_emb else 0)
self.out = nn.Conv1d(d_o, d_out * (1 + int(doubling)), kernel_size=1, bias=True).to(self._device)
self._pad16 = pad16
self._doubling = doubling
def test_freeze_unfreeze_TrainableNM(self):
path = os.path.abspath(os.path.join(os.path.dirname(__file__), "../data/jasper_smaller.yaml"))
with open(path) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
# featurizer_config=self.featurizer_config,
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
#'int_values': False,
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(**pre_process_params)
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(self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
jasper_encoder.freeze()
jasper_encoder.unfreeze(set(['encoder.4.mconv.0.conv.weight']))
frozen_weight = jasper_encoder.encoder[1].mconv[0].conv.weight.detach().cpu().numpy()
unfrozen_weight = jasper_encoder.encoder[4].mconv[0].conv.weight.detach().cpu().numpy()
# jasper_decoder.unfreeze()
# DAG
audio_signal, a_sig_length, transcript, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal, length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs, targets=transcript, input_length=encoded_len, target_length=transcript_len,
)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss], print_func=lambda x: logging.info(f'Train Loss: {str(x[0].item())}'),
)
optimizer = self.nf.get_trainer()
optimizer.train(
[loss], callbacks=[callback], optimizer="sgd", optimization_params={"max_steps": 5, "lr": 0.0003},
)
new_frozen_weight = jasper_encoder.encoder[1].mconv[0].conv.weight.data
new_unfrozen_weight = jasper_encoder.encoder[4].mconv[0].conv.weight.data
self.assertTrue(np.array_equal(frozen_weight, new_frozen_weight.detach().cpu().numpy()))
self.assertFalse(np.array_equal(unfrozen_weight, new_unfrozen_weight.detach().cpu().numpy()))
def test_asr_with_zero_ds(self):
logging.info("Testing ASR NMs with ZeroDS and without pre-processing")
path = os.path.abspath(os.path.join(os.path.dirname(__file__), "../data/jasper_smaller.yaml"))
with open(path) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo.backends.pytorch.common.ZerosDataLayer(
size=100,
dtype=torch.FloatTensor,
batch_size=4,
output_ports={
# "processed_signal": NeuralType(
# {
# 0: AxisType(BatchTag),
# 1: AxisType(SpectrogramSignalTag, dim=64),
# 2: AxisType(ProcessedTimeTag, dim=64),
# }
# ),
# "processed_length": NeuralType({0: AxisType(BatchTag)}),
# "transcript": NeuralType({0: AxisType(BatchTag), 1: AxisType(TimeTag, dim=64)}),
# "transcript_length": NeuralType({0: AxisType(BatchTag)}),
"processed_signal": NeuralType(
(AxisType(AxisKind.Batch), AxisType(AxisKind.Dimension, 64), AxisType(AxisKind.Time, 64)),
SpectrogramType(),
),
"processed_length": NeuralType(tuple('B'), LengthsType()),
"transcript": NeuralType((AxisType(AxisKind.Batch), AxisType(AxisKind.Time, 64)), LabelsType()),
"transcript_length": NeuralType(tuple('B'), LengthsType()),
},
)
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(self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
# DAG
processed_signal, p_length, transcript, transcript_len = dl()
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs, targets=transcript, input_length=encoded_len, target_length=transcript_len,
)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss], print_func=lambda x: logging.info(f'Train Loss: {str(x[0].item())}'),
)
# Instantiate an optimizer to perform `train` action
self.nf.train(
[loss], callbacks=[callback], optimization_params={"num_epochs": 2, "lr": 0.0003}, optimizer="sgd",
)
def test_quartznet_vad_training(self):
"""Integtaion test that instantiates a small QuartzNet model for vad and tests training with the
sample vad data.
Training is run for 3 forward and backward steps and asserts that loss after 3 steps is smaller than the loss
at the first step.
"""
with open(os.path.abspath(os.path.join(os.path.dirname(__file__), "../data/quartznet_vad.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToSpeechLabelDataLayer(
# featurizer_config=self.featurizer_config,
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=6,
)
pre_process_params = pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
}
# preprocessing = nemo_asr.AudioToMFCCPreprocessor(**pre_process_params)
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(**pre_process_params)
jasper_encoder = nemo_asr.JasperEncoder(**jasper_model_definition['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForClassification(
feat_in=jasper_model_definition['JasperEncoder']['jasper'][-1]['filters'], num_classes=len(self.labels)
)
ce_loss = nemo_asr.CrossEntropyLossNM()
# DAG
audio_signal, a_sig_length, targets, targets_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal, length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ce_loss(logits=log_probs, labels=targets)
loss_list = []
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss], print_func=partial(self.print_and_log_loss, loss_log_list=loss_list), step_freq=1
)
self.nf.train(
[loss], callbacks=[callback], optimizer="sgd", optimization_params={"max_steps": 3, "lr": 0.003},
)
self.nf.reset_trainer()
# Assert that training loss went down
assert loss_list[-1] < loss_list[0]
def test_freeze_unfreeze_TrainableNM(self):
with open("tests/data/jasper_smaller.yaml") as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
featurizer_config=self.featurizer_config,
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
'int_values': False,
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(**pre_process_params)
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(self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
jasper_encoder.freeze()
jasper_encoder.unfreeze(set(['encoder.4.conv.1.weight']))
jasper_decoder.unfreeze()
# DAG
audio_signal, a_sig_length, transcript, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal, length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal, length=p_length)
# print(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs, targets=transcript, input_length=encoded_len, target_length=transcript_len,
)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss], print_func=lambda x: print(f'Train Loss: {str(x[0].item())}'),
)
# Instantiate an optimizer to perform `train` action
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch, local_rank=None, create_tb_writer=False,
)
optimizer = neural_factory.get_trainer()
optimizer.train(
[loss], callbacks=[callback], optimizer="sgd", optimization_params={"num_epochs": 2, "lr": 0.0003},
)
def test_quartz_encoder(self):
with open("tests/data/quartznet_test.yaml") as file:
yaml = YAML(typ="safe")
quartz_model_definition = yaml.load(file)
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=quartz_model_definition['AudioToMelSpectrogramPreprocessor']['features'],
**quartz_model_definition['JasperEncoder']
)
self.__test_export_route_all(
module=jasper_encoder,
out_name="quartz_encoder",
input_example=(torch.randn(16, 64, 256).cuda(), torch.randint(20, (16,)).cuda()),
)
def test_jasper_encoder(self):
with open("tests/data/jasper_smaller.yaml") as file:
yaml = YAML(typ="safe")
jasper_model_definition = yaml.load(file)
jasper_encoder = nemo_asr.JasperEncoder(
conv_mask=False,
feat_in=jasper_model_definition['AudioToMelSpectrogramPreprocessor']['features'],
**jasper_model_definition['JasperEncoder']
)
self.__test_export_route_all(
module=jasper_encoder,
out_name="jasper_encoder",
input_example=(torch.randn(16, 64, 256).cuda(), torch.randn(256).cuda()),
)
def test_quartz_encoder(self, df_type):
with open("tests/data/quartznet_test.yaml") as file:
yaml = YAML(typ="safe")
quartz_model_definition = yaml.load(file)
del quartz_model_definition['JasperEncoder']['conv_mask']
jasper_encoder = nemo_asr.JasperEncoder(
conv_mask=False,
feat_in=quartz_model_definition[
'AudioToMelSpectrogramPreprocessor']['features'],
**quartz_model_definition['JasperEncoder'],
)
self.__test_export_route(
module=jasper_encoder,
out_name="quartz_encoder",
mode=df_type,
input_example=torch.randn(16, 64, 256).cuda(),
)
def __init__(self,
model_yaml,
encoder_checkpoint,
decoder_checkpoint,
language_model=None):
super(JasperASR, self).__init__()
# Read model YAML
yaml = YAML(typ="safe")
with open(model_yaml) as f:
jasper_model_definition = yaml.load(f)
self.neural_factory = nemo.core.NeuralModuleFactory(
placement=nemo.core.DeviceType.GPU,
backend=nemo.core.Backend.PyTorch)
self.labels = jasper_model_definition["labels"]
self.data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor()
self.jasper_encoder = nemo_asr.JasperEncoder(
jasper=jasper_model_definition["JasperEncoder"]["jasper"],
activation=jasper_model_definition["JasperEncoder"]["activation"],
feat_in=jasper_model_definition[
"AudioToMelSpectrogramPreprocessor"]["features"],
)
self.jasper_encoder.restore_from(encoder_checkpoint, local_rank=0)
self.jasper_decoder = nemo_asr.JasperDecoderForCTC(feat_in=1024,
num_classes=len(
self.labels))
self.jasper_decoder.restore_from(decoder_checkpoint, local_rank=0)
self.greedy_decoder = nemo_asr.GreedyCTCDecoder()
self.beam_search_with_lm = None
if language_model:
self.beam_search_with_lm = nemo_asr.BeamSearchDecoderWithLM(
vocab=self.labels,
beam_width=64,
alpha=2.0,
beta=1.0,
lm_path=language_model,
num_cpus=max(os.cpu_count(), 1),
)
def main(
config_file,
nn_encoder,
nn_decoder,
nn_onnx_encoder,
nn_onnx_decoder,
pre_v09_model=False,
batch_size=1,
time_steps=256,
):
yaml = YAML(typ="safe")
logging.info("Loading config file...")
with open(config_file) as f:
jasper_model_definition = yaml.load(f)
logging.info("Determining model shape...")
if 'AudioPreprocessing' in jasper_model_definition:
num_encoder_input_features = jasper_model_definition['AudioPreprocessing']['features']
elif 'AudioToMelSpectrogramPreprocessor' in jasper_model_definition:
num_encoder_input_features = jasper_model_definition['AudioToMelSpectrogramPreprocessor']['features']
else:
num_encoder_input_features = 64
num_decoder_input_features = jasper_model_definition['JasperEncoder']['jasper'][-1]['filters']
logging.info(" Num encoder input features: {}".format(num_encoder_input_features))
logging.info(" Num decoder input features: {}".format(num_decoder_input_features))
nf = nemo.core.NeuralModuleFactory(create_tb_writer=False)
logging.info("Initializing models...")
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=num_encoder_input_features, **jasper_model_definition['JasperEncoder']
)
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=num_decoder_input_features, num_classes=len(jasper_model_definition['labels']),
)
# This is necessary if you are using checkpoints trained with NeMo
# version before 0.9
logging.info("Loading checkpoints...")
if pre_v09_model:
logging.info(" Converting pre v0.9 checkpoint...")
ckpt = torch.load(nn_encoder)
new_ckpt = {}
for k, v in ckpt.items():
new_k = k.replace('.conv.', '.mconv.')
if len(v.shape) == 3:
new_k = new_k.replace('.weight', '.conv.weight')
new_ckpt[new_k] = v
jasper_encoder.load_state_dict(new_ckpt)
else:
jasper_encoder.restore_from(nn_encoder)
jasper_decoder.restore_from(nn_decoder)
logging.info("Exporting encoder...")
nf.deployment_export(
jasper_encoder,
nn_onnx_encoder,
nemo.core.neural_factory.DeploymentFormat.ONNX,
torch.zeros(batch_size, num_encoder_input_features, time_steps, dtype=torch.float, device="cuda:0",),
)
logging.info("Exporting decoder...")
nf.deployment_export(
jasper_decoder,
nn_onnx_decoder,
nemo.core.neural_factory.DeploymentFormat.ONNX,
(torch.zeros(batch_size, num_decoder_input_features, time_steps // 2, dtype=torch.float, device="cuda:0",)),
)
logging.info("Export completed successfully.")
def create_all_dags(args, neural_factory):
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']
# 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"]
train_dl_params["normalize_transcripts"] = False
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.num_gpus))
nemo.logging.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"])
eval_dl_params["normalize_transcripts"] = False
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:
nemo.logging.warning("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))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
nemo.logging.info('================================')
nemo.logging.info(
f"Number of parameters in encoder: {jasper_encoder.num_weights}")
nemo.logging.info(
f"Number of parameters in decoder: {jasper_decoder.num_weights}")
nemo.logging.info(
f"Total number of parameters in model: "
f"{jasper_decoder.num_weights + jasper_encoder.num_weights}")
nemo.logging.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,
eval_metric='CER'),
step_freq=args.train_eval_freq,
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,
eval_metric='CER',
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 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:
spkr_params = yaml.load(f)
sample_rate = spkr_params["sample_rate"]
time_length = spkr_params.get("time_length", 8)
logging.info("max time length considered is {} sec".format(time_length))
# Calculate num_workers for dataloader
total_cpus = os.cpu_count()
cpu_per_traindl = max(int(total_cpus / neural_factory.world_size), 1) // 2
# create data layer for training
train_dl_params = copy.deepcopy(spkr_params["AudioToSpeechLabelDataLayer"])
train_dl_params.update(spkr_params["AudioToSpeechLabelDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
audio_augmentor = spkr_params.get("AudioAugmentor", None)
# del train_dl_params["normalize_transcripts"]
data_layer_train = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=args.train_dataset,
labels=None,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
augmentor=audio_augmentor,
time_length=time_length,
**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))
logging.info("Number of steps per epoch {}".format(steps_per_epoch))
# 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(spkr_params["AudioToSpeechLabelDataLayer"])
eval_dl_params.update(spkr_params["AudioToSpeechLabelDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
data_layers_test = []
for test_set in args.eval_datasets:
data_layer_test = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=test_set,
labels=data_layer_train.labels,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
time_length=time_length,
**eval_dl_params,
# normalize_transcripts=False
)
data_layers_test.append(data_layer_test)
# create shared modules
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate, **spkr_params["AudioToMelSpectrogramPreprocessor"],
)
spectr_augment_config = spkr_params.get("SpectrogramAugmentation", None)
if spectr_augment_config:
data_spectr_augmentation = nemo_asr.SpectrogramAugmentation(**spectr_augment_config)
# (QuartzNet uses the Jasper baseline encoder and decoder)
encoder = nemo_asr.JasperEncoder(**spkr_params["JasperEncoder"],)
decoder = nemo_asr.JasperDecoderForSpkrClass(
feat_in=spkr_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=data_layer_train.num_classes,
pool_mode=spkr_params["JasperDecoderForSpkrClass"]['pool_mode'],
emb_sizes=spkr_params["JasperDecoderForSpkrClass"]["emb_sizes"].split(","),
)
if os.path.exists(args.checkpoint_dir + "/JasperEncoder-STEP-100.pt"):
encoder.restore_from(args.checkpoint_dir + "/JasperEncoder-STEP-100.pt")
logging.info("Pretrained Encoder loaded")
weight = None
xent_loss = nemo_asr.CrossEntropyLossNM(weight=weight)
# assemble train DAG
audio_signal, audio_signal_len, label, label_len = data_layer_train()
processed_signal, processed_signal_len = data_preprocessor(input_signal=audio_signal, length=audio_signal_len)
if spectr_augment_config:
processed_signal = data_spectr_augmentation(input_spec=processed_signal)
encoded, encoded_len = encoder(audio_signal=processed_signal, length=processed_signal_len)
logits, _ = decoder(encoder_output=encoded)
loss = xent_loss(logits=logits, labels=label)
# create train callbacks
train_callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss, logits, label],
print_func=partial(monitor_classification_training_progress, eval_metric=[1]),
step_freq=args.print_freq,
get_tb_values=lambda x: [("train_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.checkpoint_dir, # load dir
step_freq=args.checkpoint_save_freq,
checkpoints_to_keep=125,
)
callbacks.append(chpt_callback)
# --- Assemble Validation DAG --- #
for i, eval_layer in enumerate(data_layers_test):
audio_signal_test, audio_len_test, label_test, _ = eval_layer()
processed_signal_test, processed_len_test = data_preprocessor(
input_signal=audio_signal_test, length=audio_len_test
)
encoded_test, encoded_len_test = encoder(audio_signal=processed_signal_test, length=processed_len_test)
logits_test, _ = decoder(encoder_output=encoded_test)
loss_test = xent_loss(logits=logits_test, labels=label_test)
tagname = os.path.dirname(args.eval_datasets[i]).split("/")[-1] + "_" + str(i)
print(tagname)
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[loss_test, logits_test, label_test],
user_iter_callback=partial(process_classification_evaluation_batch, top_k=1),
user_epochs_done_callback=partial(process_classification_evaluation_epoch, tag=tagname),
eval_step=args.eval_freq, # How often we evaluate the model on the test set
tb_writer=neural_factory.tb_writer,
)
callbacks.append(eval_callback)
return loss, callbacks, steps_per_epoch, loss_test, logits_test, label_test
def test_jasper_eval(self):
with open(
os.path.abspath(
os.path.join(os.path.dirname(__file__),
"../data/jasper_smaller.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
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(
self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# DAG
audio_signal, a_sig_length, transcript, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs,
targets=transcript,
input_length=encoded_len,
target_length=transcript_len,
)
predictions = greedy_decoder(log_probs=log_probs)
from nemo.collections.asr.helpers import (
process_evaluation_batch,
process_evaluation_epoch,
)
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[loss, predictions, transcript, transcript_len],
user_iter_callback=lambda x, y: process_evaluation_batch(
x, y, labels=self.labels),
user_epochs_done_callback=process_evaluation_epoch,
)
# Instantiate an optimizer to perform `train` action
self.nf.eval(callbacks=[eval_callback])
def test_clas(self):
with open('examples/asr/experimental/configs/garnet_an4.yaml') as file:
cfg = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
featurizer_config=self.featurizer_config,
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
'int_values': False,
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
'stft_conv': True,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
encoder = nemo_asr.JasperEncoder(
jasper=cfg['encoder']['jasper'],
activation=cfg['encoder']['activation'],
feat_in=cfg['input']['train']['features'],
)
connector = nemo_asr.JasperRNNConnector(
in_channels=cfg['encoder']['jasper'][-1]['filters'],
out_channels=cfg['decoder']['hidden_size'],
)
decoder = nemo.backends.pytorch.common.DecoderRNN(
voc_size=len(self.labels),
bos_id=0,
**cfg['decoder'] # fictive
)
loss = nemo.backends.pytorch.common.SequenceLoss()
# DAG
audio_signal, a_sig_length, transcripts, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = encoder(audio_signal=processed_signal,
length=p_length)
encoded = connector(tensor=encoded)
log_probs, _ = decoder(targets=transcripts, encoder_outputs=encoded)
loss = loss(log_probs=log_probs, targets=transcripts)
# Train
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss], print_func=lambda x: print(str(x[0].item())))
# Instantiate an optimizer to perform `train` action
neural_factory = nemo.core.NeuralModuleFactory(
backend=nemo.core.Backend.PyTorch,
local_rank=None,
create_tb_writer=False,
)
optimizer = neural_factory.get_trainer()
optimizer.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"num_epochs": 10,
"lr": 0.0003
},
)
def test_stft_conv_training(self):
"""Integtaion test that instantiates a small Jasper model and tests training with the sample asr data.
test_stft_conv_training tests the torch_stft path while test_jasper_training tests the torch.stft path inside
of AudioToMelSpectrogramPreprocessor.
Training is run for 3 forward and backward steps and asserts that loss after 3 steps is smaller than the loss
at the first step.
Note: Training is done with batch gradient descent as opposed to stochastic gradient descent due to CTC loss
"""
with open(
os.path.abspath(
os.path.join(os.path.dirname(__file__),
"../data/jasper_smaller.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=30)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
'stft_conv': True,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
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(
self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
# DAG
audio_signal, a_sig_length, transcript, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs,
targets=transcript,
input_length=encoded_len,
target_length=transcript_len,
)
loss_list = []
callback = SimpleLossLoggerCallback(tensors=[loss],
print_func=partial(
self.print_and_log_loss,
loss_log_list=loss_list),
step_freq=1)
self.nf.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"max_steps": 3,
"lr": 0.001
},
)
self.nf.reset_trainer()
# Assert that training loss went down
assert loss_list[-1] < loss_list[0]
def test_clas(self):
with open('examples/asr/experimental/configs/garnet_an4.yaml') as file:
cfg = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
'stft_conv': True,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
encoder = nemo_asr.JasperEncoder(
jasper=cfg['encoder']['jasper'],
activation=cfg['encoder']['activation'],
feat_in=cfg['input']['train']['features'],
)
connector = nemo_asr.JasperRNNConnector(
in_channels=cfg['encoder']['jasper'][-1]['filters'],
out_channels=cfg['decoder']['hidden_size'],
)
decoder = nemo.backends.pytorch.common.DecoderRNN(
voc_size=len(self.labels),
bos_id=0,
hidden_size=cfg['decoder']['hidden_size'],
attention_method=cfg['decoder']['attention_method'],
attention_type=cfg['decoder']['attention_type'],
in_dropout=cfg['decoder']['in_dropout'],
gru_dropout=cfg['decoder']['gru_dropout'],
attn_dropout=cfg['decoder']['attn_dropout'],
teacher_forcing=cfg['decoder']['teacher_forcing'],
curriculum_learning=cfg['decoder']['curriculum_learning'],
rnn_type=cfg['decoder']['rnn_type'],
n_layers=cfg['decoder']['n_layers'],
tie_emb_out_weights=cfg['decoder']['tie_emb_out_weights'],
)
loss = nemo.backends.pytorch.common.SequenceLoss()
# DAG
audio_signal, a_sig_length, transcripts, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = encoder(audio_signal=processed_signal,
length=p_length)
encoded = connector(tensor=encoded)
log_probs, _ = decoder(targets=transcripts, encoder_outputs=encoded)
loss = loss(log_probs=log_probs, targets=transcripts)
# Train
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss],
print_func=lambda x: logging.info(str(x[0].item())))
# Instantiate an optimizer to perform `train` action
optimizer = self.nf.get_trainer()
optimizer.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"num_epochs": 10,
"lr": 0.0003
},
)
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
logging.info("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)
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,
)
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,
)
sample_rate=sample_rate,
**jasper_params["AudioToMelSpectrogramPreprocessor"],
)
if args.model == 'Res8':
encoder = Res8(args.hidden_size).to('cuda')
encoder.restore_from('./runs/{0}/checkpoints/Res8-STEP-{1}.pt'.format(
args.name, str(args.enc_step)))
encoder.freeze()
elif args.model == 'Res15':
encoder = Res15(args.hidden_size).to('cuda')
encoder.restore_from('./runs/{0}/checkpoints/Res15-STEP-{1}.pt'.format(
args.name, str(args.enc_step)))
encoder.freeze()
elif args.model == 'Quartz':
encoder = nemo_asr.JasperEncoder(**jasper_params["JasperEncoder"])
fc = LinearLayer(64 * 256) # TODO find shape from jasper_params
fc.restore_from('./runs/{0}/checkpoints/LinearLayer-STEP-{1}.pt'.format(
args.name, str(args.enc_step)))
encoder.restore_from(
'./runs/{0}/checkpoints/JasperEncoder-STEP-{1}.pt'.format(
args.name, str(args.enc_step)))
encoder.freeze()
l2_regularizer = L2Regularizer()
N = len(train_data_layer)
steps_per_epoch = math.ceil(N / float(batch_size) + 1)
"""BUILDING TRAIN GRAPH"""
audio_signal, audio_signal_len, commands, command_len = train_data_layer()
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:
spkr_params = yaml.load(f)
sample_rate = spkr_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 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(spkr_params["AudioToSpeechLabelDataLayer"])
eval_dl_params.update(spkr_params["AudioToSpeechLabelDataLayer"]["eval"])
del eval_dl_params["train"]
del eval_dl_params["eval"]
eval_dl_params[
'shuffle'] = False # To grab the file names without changing data_layer
data_layer_test = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=args.eval_datasets[0],
labels=None,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
**eval_dl_params,
# normalize_transcripts=False
)
# create shared modules
data_preprocessor = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate,
**spkr_params["AudioToMelSpectrogramPreprocessor"],
)
# (QuartzNet uses the Jasper baseline encoder and decoder)
encoder = nemo_asr.JasperEncoder(**spkr_params["JasperEncoder"], )
decoder = nemo_asr.JasperDecoderForSpkrClass(
feat_in=spkr_params['JasperEncoder']['jasper'][-1]['filters'],
num_classes=254,
emb_sizes=spkr_params['JasperDecoderForSpkrClass']['emb_sizes'].split(
','),
pool_mode=spkr_params["JasperDecoderForSpkrClass"]['pool_mode'],
)
# --- Assemble Validation DAG --- #
audio_signal_test, audio_len_test, label_test, _ = data_layer_test()
processed_signal_test, processed_len_test = data_preprocessor(
input_signal=audio_signal_test, length=audio_len_test)
encoded_test, _ = encoder(audio_signal=processed_signal_test,
length=processed_len_test)
_, embeddings = decoder(encoder_output=encoded_test)
return embeddings, label_test
def create_all_dags(args, neural_factory):
yaml = YAML(typ="safe")
with open(args.model_config) as f:
jasper_params = yaml.load(f)
labels = jasper_params['labels'] # Vocab of tokens
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["AudioToSpeechLabelDataLayer"])
train_dl_params.update(
jasper_params["AudioToSpeechLabelDataLayer"]["train"])
del train_dl_params["train"]
del train_dl_params["eval"]
# del train_dl_params["normalize_transcripts"]
# Look for augmentations
audio_augmentor = jasper_params.get('AudioAugmentor', None)
data_layer = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=args.train_dataset,
labels=labels,
sample_rate=sample_rate,
batch_size=args.batch_size,
num_workers=cpu_per_traindl,
augmentor=audio_augmentor,
**train_dl_params,
)
crop_pad_augmentation = nemo_asr.CropOrPadSpectrogramAugmentation(
audio_length=128)
N = len(data_layer)
steps_per_epoch = math.ceil(
N / (args.batch_size * args.iter_per_step * args.num_gpus))
logging.info('Steps per epoch : {0}'.format(steps_per_epoch))
logging.info('Have {0} examples to train on.'.format(N))
data_preprocessor = nemo_asr.AudioToMFCCPreprocessor(
sample_rate=sample_rate,
**jasper_params["AudioToMFCCPreprocessor"],
)
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["AudioToSpeechLabelDataLayer"])
eval_dl_params.update(jasper_params["AudioToSpeechLabelDataLayer"]["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.AudioToSpeechLabelDataLayer(
manifest_filepath=eval_datasets,
sample_rate=sample_rate,
labels=labels,
batch_size=args.eval_batch_size,
num_workers=cpu_per_traindl,
**eval_dl_params,
)
data_layers_eval.append(data_layer_eval)
else:
logging.warning("There were no val datasets passed")
jasper_encoder = nemo_asr.JasperEncoder(**jasper_params["JasperEncoder"], )
jasper_decoder = nemo_asr.JasperDecoderForClassification(
feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(labels),
**jasper_params['JasperDecoderForClassification'],
)
ce_loss = nemo_asr.CrossEntropyLossNM()
logging.info('================================')
logging.info(
f"Number of parameters in encoder: {jasper_encoder.num_weights}")
logging.info(
f"Number of parameters in decoder: {jasper_decoder.num_weights}")
logging.info(f"Total number of parameters in model: "
f"{jasper_decoder.num_weights + jasper_encoder.num_weights}")
logging.info('================================')
# Train DAG
# --- Assemble Training DAG --- #
audio_signal, audio_signal_len, commands, command_len = data_layer()
processed_signal, processed_signal_len = data_preprocessor(
input_signal=audio_signal, length=audio_signal_len)
processed_signal, processed_signal_len = crop_pad_augmentation(
input_signal=processed_signal, length=audio_signal_len)
if spectr_augment_config:
processed_signal = data_spectr_augmentation(
input_spec=processed_signal)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=processed_signal_len)
decoded = jasper_decoder(encoder_output=encoded)
loss = ce_loss(logits=decoded, labels=commands)
# Callbacks needed to print info to console and Tensorboard
train_callback = nemo.core.SimpleLossLoggerCallback(
# Notice that we pass in loss, predictions, and the labels (commands).
# Of course we would like to see our training loss, but we need the
# other arguments to calculate the accuracy.
tensors=[loss, decoded, commands],
# The print_func defines what gets printed.
print_func=partial(monitor_classification_training_progress,
eval_metric=None),
get_tb_values=lambda x: [("loss", x[0])],
tb_writer=neural_factory.tb_writer,
)
chpt_callback = nemo.core.CheckpointCallback(
folder=neural_factory.checkpoint_dir,
load_from_folder=args.load_dir,
step_freq=args.checkpoint_save_freq,
)
callbacks = [train_callback, chpt_callback]
# assemble eval DAGs
for i, eval_dl in enumerate(data_layers_eval):
# --- Assemble Training DAG --- #
test_audio_signal, test_audio_signal_len, test_commands, test_command_len = eval_dl(
)
test_processed_signal, test_processed_signal_len = data_preprocessor(
input_signal=test_audio_signal, length=test_audio_signal_len)
test_processed_signal, test_processed_signal_len = crop_pad_augmentation(
input_signal=test_processed_signal,
length=test_processed_signal_len)
test_encoded, test_encoded_len = jasper_encoder(
audio_signal=test_processed_signal,
length=test_processed_signal_len)
test_decoded = jasper_decoder(encoder_output=test_encoded)
test_loss = ce_loss(logits=test_decoded, labels=test_commands)
# create corresponding eval callback
tagname = os.path.basename(args.eval_datasets[i]).split(".")[0]
eval_callback = nemo.core.EvaluatorCallback(
eval_tensors=[test_loss, test_decoded, test_commands],
user_iter_callback=partial(process_classification_evaluation_batch,
top_k=1),
user_epochs_done_callback=partial(
process_classification_evaluation_epoch,
eval_metric=1,
tag=tagname),
eval_step=args.
eval_freq, # How often we evaluate the model on the test set
tb_writer=neural_factory.tb_writer,
)
callbacks.append(eval_callback)
return loss, callbacks, steps_per_epoch
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)
try:
vocab = quartz_params['labels']
sample_rate = quartz_params['sample_rate']
except KeyError:
logging.error("Please make sure you are using older config format (the ones with -old suffix)")
exit(1)
# 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:
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 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,
)
if args.local_rank is not None:
logging.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)
logging.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()
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),
)
logging.info('================================')
logging.info(
f"Number of parameters in encoder: {jasper_encoder.num_weights}")
logging.info(
f"Number of parameters in decoder: {jasper_decoder.num_weights}")
logging.info(f"Total number of parameters in model: "
f"{jasper_decoder.num_weights + jasper_encoder.num_weights}")
logging.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)
logging.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)
logging.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 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],
)
def main(
config_file,
nn_encoder,
nn_decoder,
nn_onnx_encoder,
nn_onnx_decoder,
pre_v09_model=False,
batch_size=1,
time_steps=256,
decoder_type='ctc',
):
yaml = YAML(typ="safe")
logging.info("Loading config file...")
with open(config_file) as f:
jasper_model_definition = yaml.load(f)
logging.info("Determining model shape...")
num_encoder_input_features = 64
decoder_params = jasper_model_definition['init_params']['decoder_params']['init_params']
num_decoder_input_features = decoder_params['feat_in']
logging.info(" Num encoder input features: {}".format(num_encoder_input_features))
logging.info(" Num decoder input features: {}".format(num_decoder_input_features))
nf = nemo.core.NeuralModuleFactory(create_tb_writer=False)
logging.info("Initializing models...")
jasper_encoder = nemo_asr.JasperEncoder(**jasper_model_definition['init_params']['encoder_params']['init_params'])
if decoder_type == 'ctc':
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=num_decoder_input_features,
num_classes=decoder_params['num_classes'],
vocabulary=decoder_params['vocabulary'],
)
elif decoder_type == 'classification':
if 'labels' in jasper_model_definition:
num_classes = len(jasper_model_definition['labels'])
else:
raise ValueError("List of class labels must be defined in model config file with key 'labels'")
jasper_decoder = nemo_asr.JasperDecoderForClassification(
feat_in=num_decoder_input_features, num_classes=num_classes
)
else:
raise ValueError("`decoder_type` must be one of ['ctc', 'classification']")
# This is necessary if you are using checkpoints trained with NeMo
# version before 0.9
logging.info("Loading checkpoints...")
if pre_v09_model:
logging.info(" Converting pre v0.9 checkpoint...")
ckpt = torch.load(nn_encoder)
new_ckpt = {}
for k, v in ckpt.items():
new_k = k.replace('.conv.', '.mconv.')
if len(v.shape) == 3:
new_k = new_k.replace('.weight', '.conv.weight')
new_ckpt[new_k] = v
jasper_encoder.load_state_dict(new_ckpt)
else:
jasper_encoder.restore_from(nn_encoder)
jasper_decoder.restore_from(nn_decoder)
# Create export directories if they don't already exist
base_export_dir, export_fn = os.path.split(nn_onnx_encoder)
if base_export_dir and not os.path.exists(base_export_dir):
os.makedirs(base_export_dir)
base_export_dir, export_fn = os.path.split(nn_onnx_decoder)
if base_export_dir and not os.path.exists(base_export_dir):
os.makedirs(base_export_dir)
logging.info("Exporting encoder...")
nf.deployment_export(
jasper_encoder,
nn_onnx_encoder,
nemo.core.neural_factory.DeploymentFormat.ONNX,
torch.zeros(batch_size, num_encoder_input_features, time_steps, dtype=torch.float, device="cuda:0",),
)
del jasper_encoder
logging.info("Exporting decoder...")
nf.deployment_export(
jasper_decoder,
nn_onnx_decoder,
nemo.core.neural_factory.DeploymentFormat.ONNX,
(torch.zeros(batch_size, num_decoder_input_features, time_steps // 2, dtype=torch.float, device="cuda:0",)),
)
del jasper_decoder
logging.info("Export completed successfully.")
def test_stft_conv(self):
with open(
os.path.abspath(
os.path.join(os.path.dirname(__file__),
"../data/jasper_smaller.yaml"))) as file:
jasper_model_definition = self.yaml.load(file)
dl = nemo_asr.AudioToTextDataLayer(
manifest_filepath=self.manifest_filepath,
labels=self.labels,
batch_size=4,
)
pre_process_params = {
'frame_splicing': 1,
'features': 64,
'window_size': 0.02,
'n_fft': 512,
'dither': 1e-05,
'window': 'hann',
'sample_rate': 16000,
'normalize': 'per_feature',
'window_stride': 0.01,
'stft_conv': True,
}
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
**pre_process_params)
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(
self.labels))
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(self.labels))
# DAG
audio_signal, a_sig_length, transcript, transcript_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
log_probs = jasper_decoder(encoder_output=encoded)
loss = ctc_loss(
log_probs=log_probs,
targets=transcript,
input_length=encoded_len,
target_length=transcript_len,
)
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss],
print_func=lambda x: logging.info(str(x[0].item())))
# Instantiate an optimizer to perform `train` action
optimizer = self.nf.get_trainer()
optimizer.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"num_epochs": 10,
"lr": 0.0003
},
)
def create_all_dags(args, neural_factory):
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"]
if args.dataset:
d_path = Path(args.dataset)
if not args.train_dataset:
args.train_dataset = str(d_path / Path("train_manifest.json"))
if not args.eval_datasets:
args.eval_datasets = [str(d_path / Path("test_manifest.json"))]
data_loader_layer = nemo_asr.AudioToTextDataLayer
if args.remote_data:
train_dl_params["rpyc_host"] = args.remote_data
data_loader_layer = RpycAudioToTextDataLayer
# data_layer = data_loader_layer(
# 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 = math.ceil(
# N / (args.batch_size * args.iter_per_step * args.num_gpus)
# )
# logging.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"])
if args.remote_data:
eval_dl_params["rpyc_host"] = args.remote_data
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 = data_loader_layer(
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:
# logging.warning("There were no val datasets passed")
jasper_encoder = nemo_asr.JasperEncoder(
feat_in=jasper_params["AudioToMelSpectrogramPreprocessor"]["features"],
**jasper_params["JasperEncoder"],
)
jasper_encoder.restore_from(args.encoder_checkpoint, local_rank=0)
jasper_decoder = nemo_asr.JasperDecoderForCTC(
feat_in=jasper_params["JasperEncoder"]["jasper"][-1]["filters"],
num_classes=len(vocab),
)
jasper_decoder.restore_from(args.decoder_checkpoint, local_rank=0)
ctc_loss = nemo_asr.CTCLossNM(num_classes=len(vocab))
greedy_decoder = nemo_asr.GreedyCTCDecoder()
# logging.info("================================")
# logging.info(f"Number of parameters in encoder: {jasper_encoder.num_weights}")
# logging.info(f"Number of parameters in decoder: {jasper_decoder.num_weights}")
# logging.info(
# f"Total number of parameters in model: "
# f"{jasper_decoder.num_weights + jasper_encoder.num_weights}"
# )
# logging.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),
# get_tb_values=lambda x: [("loss", x[0])],
# tb_writer=neural_factory.tb_writer,
# )
#
# chpt_callback = nemo.core.CheckpointCallback(
# folder=neural_factory.checkpoint_dir,
# load_from_folder=args.load_dir,
# step_freq=args.checkpoint_save_freq,
# checkpoints_to_keep=30,
# )
#
# callbacks = [train_callback, chpt_callback]
callbacks = []
# 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),
eval_step=args.eval_freq,
tb_writer=neural_factory.tb_writer,
)
callbacks.append(eval_callback)
return callbacks
def test_quartznet_speaker_reco_training(self):
"""Integtaion test that instantiates a small QuartzNet model for speaker recognition and tests training with the
sample an4 data.
Training is run for 3 forward and backward steps and asserts that loss after 3 steps is smaller than the loss
at the first step.
"""
with open(
os.path.abspath(
os.path.join(os.path.dirname(__file__),
"../data/quartznet_spkr_test.yaml"))) as file:
spkr_params = self.yaml.load(file)
dl = nemo_asr.AudioToSpeechLabelDataLayer(
manifest_filepath=self.manifest_filepath,
labels=None,
batch_size=10,
)
sample_rate = 16000
preprocessing = nemo_asr.AudioToMelSpectrogramPreprocessor(
sample_rate=sample_rate,
**spkr_params["AudioToMelSpectrogramPreprocessor"],
)
jasper_encoder = nemo_asr.JasperEncoder(**spkr_params['JasperEncoder'])
jasper_decoder = nemo_asr.JasperDecoderForSpkrClass(
feat_in=spkr_params['JasperEncoder']['jasper'][-1]['filters'],
num_classes=dl.num_classes,
pool_mode=spkr_params['JasperDecoderForSpkrClass']['pool_mode'],
emb_sizes=spkr_params["JasperDecoderForSpkrClass"]
["emb_sizes"].split(","),
)
ce_loss = nemo_asr.CrossEntropyLossNM()
# DAG
audio_signal, a_sig_length, targets, targets_len = dl()
processed_signal, p_length = preprocessing(input_signal=audio_signal,
length=a_sig_length)
encoded, encoded_len = jasper_encoder(audio_signal=processed_signal,
length=p_length)
# logging.info(jasper_encoder)
log_probs, _ = jasper_decoder(encoder_output=encoded)
loss = ce_loss(logits=log_probs, labels=targets)
loss_list = []
callback = nemo.core.SimpleLossLoggerCallback(
tensors=[loss],
print_func=partial(self.print_and_log_loss,
loss_log_list=loss_list),
step_freq=1)
self.nf.random_seed = 42
self.nf.train(
[loss],
callbacks=[callback],
optimizer="sgd",
optimization_params={
"max_steps": 4,
"lr": 0.002
},
)
self.nf.reset_trainer()
# Assert that training loss went down
assert loss_list[-1] < loss_list[0]
