def test_cutmix(preserve_id: bool):
speech_cuts = DummyManifest(CutSet, begin_id=0, end_id=10)
for c in speech_cuts:
c.duration = 10.0
noise_cuts = DummyManifest(CutSet, begin_id=100, end_id=102)
for c in noise_cuts:
c.duration = 1.5
tfnm = CutMix(noise_cuts, snr=None, prob=1.0, preserve_id=preserve_id)
tfnm_cuts = tfnm(speech_cuts)
for c in tfnm_cuts:
assert isinstance(c, MixedCut)
assert c.tracks[0].cut.duration == 10.0
assert sum(t.cut.duration for t in c.tracks[1:]) == 10.0
if preserve_id:
assert all(
cut.id == cut_noisy.id for cut, cut_noisy in zip(speech_cuts, tfnm_cuts)
)
else:
assert all(
cut.id != cut_noisy.id for cut, cut_noisy in zip(speech_cuts, tfnm_cuts)
)
def test_cutmix():
speech_cuts = DummyManifest(CutSet, begin_id=0, end_id=10)
for c in speech_cuts:
c.duration = 10.0
noise_cuts = DummyManifest(CutSet, begin_id=100, end_id=102)
for c in noise_cuts:
c.duration = 1.5
tfnm = CutMix(noise_cuts, snr=None, prob=1.0)
tfnm_cuts = tfnm(speech_cuts)
for c in tfnm_cuts:
assert isinstance(c, MixedCut)
assert c.tracks[0].cut.duration == 10.0
assert sum(t.cut.duration for t in c.tracks[1:]) == 10.0
def main():
fix_random_seed(42)
start_epoch = 0
num_epochs = 8
exp_dir = 'exp-lstm-adam-ctc-musan'
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
logging.info("Loading L.fst")
if (lang_dir / 'Linv.pt').exists():
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
with open(lang_dir / 'L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_inv = k2.arc_sort(L.invert_())
torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt')
graph_compiler = CtcTrainingGraphCompiler(
L_inv=L_inv,
phones=phone_symbol_table,
words=word_symbol_table
)
phone_ids = get_phone_symbols(phone_symbol_table)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir /
'cuts_train-clean-100.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
train = K2SpeechRecognitionDataset(
cuts_train,
cut_transforms=[
CutConcatenate(),
CutMix(
cuts=cuts_musan,
prob=0.5,
snr=(10, 20)
)
]
)
train_sampler = SingleCutSampler(
cuts_train,
max_frames=90000,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(
train,
sampler=train_sampler,
batch_size=None,
num_workers=4
)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(cuts_dev, max_frames=90000)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(
validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1
)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=3)
model.to(device)
describe(model)
learning_rate = 1e-3
optimizer = optim.AdamW(model.parameters(),
lr=learning_rate,
weight_decay=5e-4)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path, model=model, optimizer=optimizer)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}")
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
curr_learning_rate = 1e-3
# curr_learning_rate = learning_rate * pow(0.4, epoch)
# for param_group in optimizer.param_groups:
# param_group['lr'] = curr_learning_rate
tb_writer.add_scalar('learning_rate', curr_learning_rate, epoch)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train)
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
epoch=epoch,
optimizer=None,
scheduler=None,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=best_objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=None,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
def main():
args = get_parser().parse_args()
model_type = args.model_type
start_epoch = args.start_epoch
num_epochs = args.num_epochs
max_duration = args.max_duration
accum_grad = args.accum_grad
att_rate = args.att_rate
fix_random_seed(42)
exp_dir = Path('exp-' + model_type + '-noam-ctc-att-musan-sa')
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(
log_dir=f'{exp_dir}/tensorboard') if args.tensorboard else None
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
logging.info("Loading L.fst")
if (lang_dir / 'Linv.pt').exists():
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
with open(lang_dir / 'L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_inv = k2.arc_sort(L.invert_())
torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt')
graph_compiler = CtcTrainingGraphCompiler(L_inv=L_inv,
phones=phone_symbol_table,
words=word_symbol_table)
phone_ids = get_phone_symbols(phone_symbol_table)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = load_manifest(feature_dir / 'cuts_train-clean-100.json.gz')
if args.full_libri:
cuts_train = (
cuts_train +
load_manifest(feature_dir / 'cuts_train-clean-360.json.gz') +
load_manifest(feature_dir / 'cuts_train-other-500.json.gz'))
logging.info("About to get dev cuts")
cuts_dev = (load_manifest(feature_dir / 'cuts_dev-clean.json.gz') +
load_manifest(feature_dir / 'cuts_dev-other.json.gz'))
logging.info("About to get Musan cuts")
cuts_musan = load_manifest(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
if args.concatenate_cuts:
logging.info(
f'Using cut concatenation with duration factor {args.duration_factor} and gap {args.gap}.'
)
# Cut concatenation should be the first transform in the list,
# so that if we e.g. mix noise in, it will fill the gaps between different utterances.
transforms = [
CutConcatenate(duration_factor=args.duration_factor, gap=args.gap)
] + transforms
train = K2SpeechRecognitionDataset(cuts_train,
cut_transforms=transforms,
input_transforms=[
SpecAugment(num_frame_masks=2,
features_mask_size=27,
num_feature_masks=2,
frames_mask_size=100)
])
if args.on_the_fly_feats:
# NOTE: the PerturbSpeed transform should be added only if we remove it from data prep stage.
# # Add on-the-fly speed perturbation; since originally it would have increased epoch
# # size by 3, we will apply prob 2/3 and use 3x more epochs.
# # Speed perturbation probably should come first before concatenation,
# # but in principle the transforms order doesn't have to be strict (e.g. could be randomized)
# transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2 / 3)] + transforms
# Drop feats to be on the safe side.
cuts_train = cuts_train.drop_features()
from lhotse.features.fbank import FbankConfig
train = K2SpeechRecognitionDataset(
cuts=cuts_train,
cut_transforms=transforms,
input_strategy=OnTheFlyFeatures(Fbank(
FbankConfig(num_mel_bins=80))),
input_transforms=[
SpecAugment(num_frame_masks=2,
features_mask_size=27,
num_feature_masks=2,
frames_mask_size=100)
])
if args.bucketing_sampler:
logging.info('Using BucketingSampler.')
train_sampler = BucketingSampler(cuts_train,
max_duration=max_duration,
shuffle=True,
num_buckets=args.num_buckets)
else:
logging.info('Using SingleCutSampler.')
train_sampler = SingleCutSampler(
cuts_train,
max_duration=max_duration,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(
train,
sampler=train_sampler,
batch_size=None,
num_workers=4,
)
logging.info("About to create dev dataset")
if args.on_the_fly_feats:
cuts_dev = cuts_dev.drop_features()
validate = K2SpeechRecognitionDataset(
cuts_dev.drop_features(),
input_strategy=OnTheFlyFeatures(Fbank(
FbankConfig(num_mel_bins=80))))
else:
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(
cuts_dev,
max_duration=max_duration,
)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
if att_rate != 0.0:
num_decoder_layers = 6
else:
num_decoder_layers = 0
if model_type == "transformer":
model = Transformer(
num_features=80,
nhead=args.nhead,
d_model=args.attention_dim,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4,
num_decoder_layers=num_decoder_layers)
else:
model = Conformer(
num_features=80,
nhead=args.nhead,
d_model=args.attention_dim,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4,
num_decoder_layers=num_decoder_layers)
model.to(device)
describe(model)
optimizer = Noam(model.parameters(),
model_size=args.attention_dim,
factor=1.0,
warm_step=args.warm_step)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
curr_learning_rate = optimizer._rate
if tb_writer is not None:
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
accum_grad=accum_grad,
att_rate=att_rate,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
)
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
optimizer=None,
scheduler=None,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
optimizer=optimizer,
scheduler=None,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
def train_dataloaders(self) -> DataLoader:
logging.info("About to get train cuts")
cuts_train = self.train_cuts()
logging.info("About to get Musan cuts")
cuts_musan = load_manifest(self.args.feature_dir /
'cuts_musan.json.gz')
logging.info("About to create train dataset")
transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
if self.args.concatenate_cuts:
logging.info(
f'Using cut concatenation with duration factor '
f'{self.args.duration_factor} and gap {self.args.gap}.')
# Cut concatenation should be the first transform in the list,
# so that if we e.g. mix noise in, it will fill the gaps between different utterances.
transforms = [
CutConcatenate(duration_factor=self.args.duration_factor,
gap=self.args.gap)
] + transforms
input_transforms = [
SpecAugment(num_frame_masks=2,
features_mask_size=27,
num_feature_masks=2,
frames_mask_size=100)
]
train = K2SpeechRecognitionDataset(
cut_transforms=transforms,
input_transforms=input_transforms,
return_cuts=True,
)
if self.args.on_the_fly_feats:
# NOTE: the PerturbSpeed transform should be added only if we remove it from data prep stage.
# # Add on-the-fly speed perturbation; since originally it would have increased epoch
# # size by 3, we will apply prob 2/3 and use 3x more epochs.
# # Speed perturbation probably should come first before concatenation,
# # but in principle the transforms order doesn't have to be strict (e.g. could be randomized)
# transforms = [PerturbSpeed(factors=[0.9, 1.1], p=2 / 3)] + transforms
# Drop feats to be on the safe side.
cuts_train = cuts_train.drop_features()
train = K2SpeechRecognitionDataset(
cut_transforms=transforms,
input_strategy=OnTheFlyFeatures(
Fbank(FbankConfig(num_mel_bins=80))),
input_transforms=input_transforms,
return_cuts=True,
)
if self.args.bucketing_sampler:
logging.info('Using BucketingSampler.')
train_sampler = BucketingSampler(
cuts_train,
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
num_buckets=self.args.num_buckets)
else:
logging.info('Using SingleCutSampler.')
train_sampler = SingleCutSampler(
cuts_train,
max_duration=self.args.max_duration,
shuffle=self.args.shuffle,
)
logging.info("About to create train dataloader")
train_dl = DataLoader(
train,
sampler=train_sampler,
batch_size=None,
num_workers=4,
persistent_workers=True,
)
return train_dl
def main():
args = get_parser().parse_args()
print('World size:', args.world_size, 'Rank:', args.local_rank)
setup_dist(rank=args.local_rank, world_size=args.world_size)
fix_random_seed(42)
start_epoch = 0
num_epochs = 10
use_adam = True
exp_dir = f'exp-lstm-adam-mmi-bigram-musan-dist'
setup_logger('{}/log/log-train'.format(exp_dir),
use_console=args.local_rank == 0)
tb_writer = SummaryWriter(
log_dir=f'{exp_dir}/tensorboard') if args.local_rank == 0 else None
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
logging.info("Loading L.fst")
if (lang_dir / 'Linv.pt').exists():
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
with open(lang_dir / 'L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_inv = k2.arc_sort(L.invert_())
torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt')
graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv,
phones=phone_symbol_table,
words=word_symbol_table)
phone_ids = get_phone_symbols(phone_symbol_table)
P = create_bigram_phone_lm(phone_ids)
P.scores = torch.zeros_like(P.scores)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
if not args.bucketing_sampler:
# We don't mix concatenating the cuts and bucketing
# Here we insert concatenation before mixing so that the
# noises from Musan are mixed onto almost-zero-energy
# padding frames.
transforms = [CutConcatenate()] + transforms
train = K2SpeechRecognitionDataset(cuts_train, cut_transforms=transforms)
if args.bucketing_sampler:
logging.info('Using BucketingSampler.')
train_sampler = BucketingSampler(cuts_train,
max_frames=40000,
shuffle=True,
num_buckets=30)
else:
logging.info('Using regular sampler with cut concatenation.')
train_sampler = SingleCutSampler(
cuts_train,
max_frames=30000,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
sampler=train_sampler,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
# Note: we explicitly set world_size to 1 to disable the auto-detection of
# distributed training inside the sampler. This way, every GPU will
# perform the computation on the full dev set. It is a bit wasteful,
# but unfortunately loss aggregation between multiple processes with
# torch.distributed.all_reduce() tends to hang indefinitely inside
# NCCL after ~3000 steps. With the current approach, we can still report
# the loss on the full validation set.
valid_sampler = SingleCutSampler(cuts_dev,
max_frames=90000,
world_size=1,
rank=0)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = args.local_rank
device = torch.device('cuda', device_id)
model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=3)
model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True)
model.to(device)
describe(model)
if use_adam:
learning_rate = 1e-3
weight_decay = 5e-4
optimizer = optim.AdamW(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
# Equivalent to the following in the epoch loop:
# if epoch > 6:
# curr_learning_rate *= 0.8
lr_scheduler = optim.lr_scheduler.LambdaLR(
optimizer, lambda ep: 1.0 if ep < 7 else 0.8**(ep - 6))
else:
learning_rate = 5e-5
weight_decay = 1e-5
momentum = 0.9
lr_schedule_gamma = 0.7
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=momentum,
weight_decay=weight_decay)
lr_scheduler = optim.lr_scheduler.ExponentialLR(
optimizer=optimizer, gamma=lr_schedule_gamma)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
if args.world_size > 1:
logging.info(
'Using DistributedDataParallel in training. '
'The reported loss, num_frames, etc. for training steps include '
'only the batches seen in the master process (the actual loss '
'includes batches from all GPUs, and the actual num_frames is '
f'approx. {args.world_size}x larger.')
# For now do not sync BatchNorm across GPUs due to NCCL hanging in all_gather...
# model = torch.nn.SyncBatchNorm.convert_sync_batchnorm(model)
model = DDP(model,
device_ids=[args.local_rank],
output_device=args.local_rank)
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
# LR scheduler can hold multiple learning rates for multiple parameter groups;
# For now we report just the first LR which we assume concerns most of the parameters.
curr_learning_rate = lr_scheduler.get_last_lr()[0]
if tb_writer is not None:
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
P=P,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
)
lr_scheduler.step()
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
optimizer=None,
scheduler=None,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
local_rank=args.local_rank,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
local_rank=args.local_rank,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
cleanup_dist()
def main():
fix_random_seed(42)
start_epoch = 0
num_epochs = 10
use_adam = True
exp_dir = f'exp-lstm-adam-mmi-bigram-musan'
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
logging.info("Loading L.fst")
if (lang_dir / 'Linv.pt').exists():
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
with open(lang_dir / 'L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_inv = k2.arc_sort(L.invert_())
torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt')
graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv,
phones=phone_symbol_table,
words=word_symbol_table)
phone_ids = get_phone_symbols(phone_symbol_table)
P = create_bigram_phone_lm(phone_ids)
P.scores = torch.zeros_like(P.scores)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir / 'cuts_train.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
train = K2SpeechRecognitionDataset(cuts_train,
cut_transforms=[
CutConcatenate(),
CutMix(cuts=cuts_musan,
prob=0.5,
snr=(10, 20))
])
train_sampler = SingleCutSampler(
cuts_train,
max_frames=12000,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
sampler=train_sampler,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(cuts_dev, max_frames=12000)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=3)
model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True)
model.to(device)
describe(model)
if use_adam:
learning_rate = 1e-3
weight_decay = 5e-4
optimizer = optim.AdamW(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
# Equivalent to the following in the epoch loop:
# if epoch > 6:
# curr_learning_rate *= 0.8
lr_scheduler = optim.lr_scheduler.LambdaLR(
optimizer, lambda ep: 1.0 if ep < 7 else 0.8**(ep - 6))
else:
learning_rate = 5e-5
weight_decay = 1e-5
momentum = 0.9
lr_schedule_gamma = 0.7
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=momentum,
weight_decay=weight_decay)
lr_scheduler = optim.lr_scheduler.ExponentialLR(
optimizer=optimizer, gamma=lr_schedule_gamma)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
# LR scheduler can hold multiple learning rates for multiple parameter groups;
# For now we report just the first LR which we assume concerns most of the parameters.
curr_learning_rate = lr_scheduler.get_last_lr()[0]
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
P=P,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
)
lr_scheduler.step()
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
optimizer=None,
scheduler=None,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
def main():
args = get_parser().parse_args()
print('World size:', args.world_size, 'Rank:', args.local_rank)
setup_dist(rank=args.local_rank,
world_size=args.world_size,
master_port=args.master_port)
fix_random_seed(42)
start_epoch = 0
num_epochs = 10
use_adam = True
exp_dir = f'exp-lstm-adam-mmi-bigram-musan'
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
lexicon = Lexicon(lang_dir)
device_id = args.local_rank
device = torch.device('cuda', device_id)
phone_ids = lexicon.phone_symbols()
if not Path(lang_dir / 'P.pt').is_file():
logging.debug(f'Loading P from {lang_dir}/P.fst.txt')
with open(lang_dir / 'P.fst.txt') as f:
# P is not an acceptor because there is
# a back-off state, whose incoming arcs
# have label #0 and aux_label eps.
P = k2.Fsa.from_openfst(f.read(), acceptor=False)
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
# P.aux_labels is not needed in later computations, so
# remove it here.
del P.aux_labels
# CAUTION(fangjun): The following line is crucial.
# Arcs entering the back-off state have label equal to #0.
# We have to change it to 0 here.
P.labels[P.labels >= first_phone_disambig_id] = 0
P = k2.remove_epsilon(P)
P = k2.arc_sort(P)
torch.save(P.as_dict(), lang_dir / 'P.pt')
else:
logging.debug('Loading pre-compiled P')
d = torch.load(lang_dir / 'P.pt')
P = k2.Fsa.from_dict(d)
graph_compiler = MmiTrainingGraphCompiler(
lexicon=lexicon,
P=P,
device=device,
)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir / 'cuts_train.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
train = K2SpeechRecognitionDataset(cuts_train,
cut_transforms=[
CutConcatenate(),
CutMix(cuts=cuts_musan,
prob=0.5,
snr=(10, 20))
])
train_sampler = SingleCutSampler(
cuts_train,
max_frames=40000,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
sampler=train_sampler,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(cuts_dev, max_frames=12000)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=3)
model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True)
model.to(device)
describe(model)
if use_adam:
learning_rate = 1e-3
weight_decay = 5e-4
optimizer = optim.AdamW(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
# Equivalent to the following in the epoch loop:
# if epoch > 6:
# curr_learning_rate *= 0.8
lr_scheduler = optim.lr_scheduler.LambdaLR(
optimizer, lambda ep: 1.0 if ep < 7 else 0.8**(ep - 6))
else:
learning_rate = 5e-5
weight_decay = 1e-5
momentum = 0.9
lr_schedule_gamma = 0.7
optimizer = optim.SGD(model.parameters(),
lr=learning_rate,
momentum=momentum,
weight_decay=weight_decay)
lr_scheduler = optim.lr_scheduler.ExponentialLR(
optimizer=optimizer, gamma=lr_schedule_gamma)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
# LR scheduler can hold multiple learning rates for multiple parameter groups;
# For now we report just the first LR which we assume concerns most of the parameters.
curr_learning_rate = lr_scheduler.get_last_lr()[0]
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
)
lr_scheduler.step()
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
optimizer=None,
scheduler=None,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scheduler=lr_scheduler,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
def main():
args = get_parser().parse_args()
start_epoch = args.start_epoch
num_epochs = args.num_epochs
max_frames = args.max_frames
accum_grad = args.accum_grad
den_scale = args.den_scale
att_rate = args.att_rate
fix_random_seed(42)
exp_dir = Path('exp-transformer-noam-mmi-att-musan')
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
word_symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
logging.info("Loading L.fst")
if (lang_dir / 'Linv.pt').exists():
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
with open(lang_dir / 'L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_inv = k2.arc_sort(L.invert_())
torch.save(L_inv.as_dict(), lang_dir / 'Linv.pt')
graph_compiler = MmiTrainingGraphCompiler(L_inv=L_inv,
phones=phone_symbol_table,
words=word_symbol_table)
phone_ids = get_phone_symbols(phone_symbol_table)
P = create_bigram_phone_lm(phone_ids)
P.scores = torch.zeros_like(P.scores)
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir / 'cuts_train-clean-100.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev-clean.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
transforms = [CutMix(cuts=cuts_musan, prob=0.5, snr=(10, 20))]
if not args.bucketing_sampler:
# We don't mix concatenating the cuts and bucketing
# Here we insert concatenation before mixing so that the
# noises from Musan are mixed onto almost-zero-energy
# padding frames.
transforms = [CutConcatenate()] + transforms
train = K2SpeechRecognitionDataset(cuts_train, cut_transforms=transforms)
if args.bucketing_sampler:
logging.info('Using BucketingSampler.')
train_sampler = BucketingSampler(cuts_train,
max_frames=max_frames,
shuffle=True,
num_buckets=args.num_buckets)
else:
logging.info('Using regular sampler with cut concatenation.')
train_sampler = SingleCutSampler(
cuts_train,
max_frames=max_frames,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
sampler=train_sampler,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(cuts_dev, max_frames=max_frames)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
logging.info("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
if att_rate != 0.0:
num_decoder_layers = 6
else:
num_decoder_layers = 0
model = Transformer(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4,
num_decoder_layers=num_decoder_layers)
model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True)
model.to(device)
describe(model)
optimizer = Noam(model.parameters(),
model_size=256,
factor=1.0,
warm_step=args.warm_step)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
for epoch in range(start_epoch, num_epochs):
train_sampler.set_epoch(epoch)
curr_learning_rate = optimizer._rate
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
P=P,
device=device,
graph_compiler=graph_compiler,
optimizer=optimizer,
accum_grad=accum_grad,
den_scale=den_scale,
att_rate=att_rate,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
)
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
optimizer=None,
scheduler=None,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
optimizer=optimizer,
scheduler=None,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch)
logging.warning('Done')
def run(rank, world_size, args):
'''
Args:
rank:
It is a value between 0 and `world_size-1`, which is
passed automatically by `mp.spawn()` in :func:`main`.
The node with rank 0 is responsible for saving checkpoint.
world_size:
Number of GPUs for DDP training.
args:
The return value of get_parser().parse_args()
'''
model_type = args.model_type
start_epoch = args.start_epoch
num_epochs = args.num_epochs
accum_grad = args.accum_grad
den_scale = args.den_scale
att_rate = args.att_rate
use_pruned_intersect = args.use_pruned_intersect
fix_random_seed(42)
if world_size > 1:
setup_dist(rank, world_size, args.master_port)
exp_dir = Path('exp-' + model_type + '-mmi-att-sa-vgg-normlayer')
setup_logger(f'{exp_dir}/log/log-train-{rank}')
if args.tensorboard and rank == 0:
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
else:
tb_writer = None
# tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard') if args.tensorboard and rank == 0 else None
logging.info("Loading lexicon and symbol tables")
lang_dir = Path('data/lang_nosp')
lexicon = Lexicon(lang_dir)
device_id = rank
device = torch.device('cuda', device_id)
if not Path(lang_dir / 'P.pt').is_file():
logging.debug(f'Loading P from {lang_dir}/P.fst.txt')
with open(lang_dir / 'P.fst.txt') as f:
# P is not an acceptor because there is
# a back-off state, whose incoming arcs
# have label #0 and aux_label eps.
P = k2.Fsa.from_openfst(f.read(), acceptor=False)
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
# P.aux_labels is not needed in later computations, so
# remove it here.
del P.aux_labels
# CAUTION(fangjun): The following line is crucial.
# Arcs entering the back-off state have label equal to #0.
# We have to change it to 0 here.
P.labels[P.labels >= first_phone_disambig_id] = 0
P = k2.remove_epsilon(P)
P = k2.arc_sort(P)
torch.save(P.as_dict(), lang_dir / 'P.pt')
else:
logging.debug('Loading pre-compiled P')
d = torch.load(lang_dir / 'P.pt')
P = k2.Fsa.from_dict(d)
graph_compiler = MmiTrainingGraphCompiler(
lexicon=lexicon,
P=P,
device=device,
)
phone_ids = lexicon.phone_symbols()
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get train cuts")
cuts_train = CutSet.from_json(feature_dir / 'cuts_train.json.gz')
logging.info("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir / 'cuts_dev.json.gz')
logging.info("About to get Musan cuts")
cuts_musan = CutSet.from_json(feature_dir / 'cuts_musan.json.gz')
logging.info("About to create train dataset")
train = K2SpeechRecognitionDataset(cuts_train,
cut_transforms=[
CutConcatenate(),
CutMix(cuts=cuts_musan,
prob=0.5,
snr=(10, 20))
])
train_sampler = SingleCutSampler(
cuts_train,
max_frames=90000,
shuffle=True,
)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
sampler=train_sampler,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionDataset(cuts_dev)
valid_sampler = SingleCutSampler(cuts_dev, max_frames=90000)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
sampler=valid_sampler,
batch_size=None,
num_workers=1)
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
if use_pruned_intersect:
logging.info('Use pruned intersect for den_lats')
else:
logging.info("Don't use pruned intersect for den_lats")
logging.info("About to create model")
if att_rate != 0.0:
num_decoder_layers = 6
else:
num_decoder_layers = 0
if model_type == "transformer":
model = Transformer(
num_features=40,
nhead=args.nhead,
d_model=args.attention_dim,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4,
num_decoder_layers=num_decoder_layers,
vgg_frontend=True)
elif model_type == "conformer":
model = Conformer(
num_features=40,
nhead=args.nhead,
d_model=args.attention_dim,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4,
num_decoder_layers=num_decoder_layers,
vgg_frontend=True,
is_espnet_structure=True)
elif model_type == "contextnet":
model = ContextNet(num_features=40, num_classes=len(phone_ids) +
1) # +1 for the blank symbol
else:
raise NotImplementedError("Model of type " + str(model_type) +
" is not implemented")
if args.torchscript:
logging.info('Applying TorchScript to model...')
model = torch.jit.script(model)
model.to(device)
describe(model)
if world_size > 1:
model = DDP(model, device_ids=[rank])
# Now for the alignment model, if any
if args.use_ali_model:
ali_model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=4)
ali_model_fname = Path(
f'exp-lstm-adam-ctc-musan/epoch-{args.ali_model_epoch}.pt')
assert ali_model_fname.is_file(), \
f'ali model filename {ali_model_fname} does not exist!'
ali_model.load_state_dict(
torch.load(ali_model_fname, map_location='cpu')['state_dict'])
ali_model.to(device)
ali_model.eval()
ali_model.requires_grad_(False)
logging.info(f'Use ali_model: {ali_model_fname}')
else:
ali_model = None
logging.info('No ali_model')
optimizer = Noam(model.parameters(),
model_size=args.attention_dim,
factor=args.lr_factor,
warm_step=args.warm_step,
weight_decay=args.weight_decay)
scaler = GradScaler(enabled=args.amp)
best_objf = np.inf
best_valid_objf = np.inf
best_epoch = start_epoch
best_model_path = os.path.join(exp_dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(exp_dir, 'best-epoch-info')
global_batch_idx_train = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
ckpt = load_checkpoint(filename=model_path,
model=model,
optimizer=optimizer,
scaler=scaler)
best_objf = ckpt['objf']
best_valid_objf = ckpt['valid_objf']
global_batch_idx_train = ckpt['global_batch_idx_train']
logging.info(
f"epoch = {ckpt['epoch']}, objf = {best_objf}, valid_objf = {best_valid_objf}"
)
for epoch in range(start_epoch, num_epochs):
#train_dl.sampler.set_epoch(epoch)
curr_learning_rate = optimizer._rate
if tb_writer is not None:
tb_writer.add_scalar('train/learning_rate', curr_learning_rate,
global_batch_idx_train)
tb_writer.add_scalar('train/epoch', epoch, global_batch_idx_train)
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf, valid_objf, global_batch_idx_train = train_one_epoch(
dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
ali_model=ali_model,
device=device,
graph_compiler=graph_compiler,
use_pruned_intersect=use_pruned_intersect,
optimizer=optimizer,
accum_grad=accum_grad,
den_scale=den_scale,
att_rate=att_rate,
current_epoch=epoch,
tb_writer=tb_writer,
num_epochs=num_epochs,
global_batch_idx_train=global_batch_idx_train,
world_size=world_size,
scaler=scaler)
# the lower, the better
if valid_objf < best_valid_objf:
best_valid_objf = valid_objf
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
optimizer=None,
scheduler=None,
scaler=None,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train,
local_rank=rank,
torchscript=args.torchscript_epoch != -1
and epoch >= args.torchscript_epoch)
save_training_info(filename=best_epoch_info_filename,
model_path=best_model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch,
local_rank=rank)
# we always save the model for every epoch
model_path = os.path.join(exp_dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
optimizer=optimizer,
scheduler=None,
scaler=scaler,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
valid_objf=valid_objf,
global_batch_idx_train=global_batch_idx_train,
local_rank=rank,
torchscript=args.torchscript_epoch != -1
and epoch >= args.torchscript_epoch)
epoch_info_filename = os.path.join(exp_dir,
'epoch-{}-info'.format(epoch))
save_training_info(filename=epoch_info_filename,
model_path=model_path,
current_epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf,
best_objf=best_objf,
valid_objf=valid_objf,
best_valid_objf=best_valid_objf,
best_epoch=best_epoch,
local_rank=rank)
logging.warning('Done')
if world_size > 1:
torch.distributed.barrier()
cleanup_dist()