def main():
parser = get_parser()
LibriSpeechAsrDataModule.add_arguments(parser)
args = parser.parse_args()
model_type = args.model_type
epoch = args.epoch
avg = args.avg
att_rate = args.att_rate
num_paths = args.num_paths
use_lm_rescoring = args.use_lm_rescoring
use_whole_lattice = False
if use_lm_rescoring and num_paths < 1:
# It doesn't make sense to use n-best list for rescoring
# when n is less than 1
use_whole_lattice = True
output_beam_size = args.output_beam_size
exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa-vgg')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
logging.info(f'output_beam_size: {output_beam_size}')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
phone_ids = get_phone_symbols(phone_symbol_table)
P = create_bigram_phone_lm(phone_ids)
phone_ids_with_blank = [0] + phone_ids
ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank))
logging.debug("About to load model")
# Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N
# device = torch.device('cuda', 1)
device = torch.device('cuda')
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,
vgg_frontend=True)
elif model_type == "conformer":
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,
vgg_frontend=True)
elif model_type == "contextnet":
model = ContextNet(num_features=80, num_classes=len(phone_ids) +
1) # +1 for the blank symbol
else:
raise NotImplementedError("Model of type " + str(model_type) +
" is not implemented")
model.P_scores = torch.nn.Parameter(P.scores.clone(), requires_grad=False)
if avg == 1:
checkpoint = os.path.join(exp_dir, 'epoch-' + str(epoch - 1) + '.pt')
load_checkpoint(checkpoint, model)
else:
checkpoints = [
os.path.join(exp_dir, 'epoch-' + str(avg_epoch) + '.pt')
for avg_epoch in range(epoch - avg, epoch)
]
average_checkpoint(checkpoints, model)
model.to(device)
model.eval()
assert P.requires_grad is False
P.scores = model.P_scores.cpu()
print_transition_probabilities(P,
phone_symbol_table,
phone_ids,
filename='model_P_scores.txt')
P.set_scores_stochastic_(model.P_scores)
print_transition_probabilities(P,
phone_symbol_table,
phone_ids,
filename='P_scores.txt')
if not os.path.exists(lang_dir / 'HLG.pt'):
logging.debug("Loading L_disambig.fst.txt")
with open(lang_dir / 'L_disambig.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
logging.debug("Loading G.fst.txt")
with open(lang_dir / 'G.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
HLG = compile_HLG(L=L,
G=G,
H=ctc_topo,
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(HLG.as_dict(), lang_dir / 'HLG.pt')
else:
logging.debug("Loading pre-compiled HLG")
d = torch.load(lang_dir / 'HLG.pt')
HLG = k2.Fsa.from_dict(d)
if use_lm_rescoring:
if use_whole_lattice:
logging.info('Rescoring with the whole lattice')
else:
logging.info(f'Rescoring with n-best list, n is {num_paths}')
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
if not os.path.exists(lang_dir / 'G_4_gram.pt'):
logging.debug('Loading G_4_gram.fst.txt')
with open(lang_dir / 'G_4_gram.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
# G.aux_labels is not needed in later computations, so
# remove it here.
del G.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.
G.labels[G.labels >= first_word_disambig_id] = 0
G = k2.create_fsa_vec([G]).to(device)
G = k2.arc_sort(G)
torch.save(G.as_dict(), lang_dir / 'G_4_gram.pt')
else:
logging.debug('Loading pre-compiled G_4_gram.pt')
d = torch.load(lang_dir / 'G_4_gram.pt')
G = k2.Fsa.from_dict(d).to(device)
if use_whole_lattice:
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
G = G.to(device)
else:
logging.debug('Decoding without LM rescoring')
G = None
logging.debug("convert HLG to device")
HLG = HLG.to(device)
HLG.aux_labels = k2.ragged.remove_values_eq(HLG.aux_labels, 0)
HLG.requires_grad_(False)
if not hasattr(HLG, 'lm_scores'):
HLG.lm_scores = HLG.scores.clone()
# load dataset
librispeech = LibriSpeechAsrDataModule(args)
test_sets = ['test-clean', 'test-other']
# test_sets = ['test-other']
for test_set, test_dl in zip(test_sets, librispeech.test_dataloaders()):
logging.info(f'* DECODING: {test_set}')
results = decode(dataloader=test_dl,
model=model,
device=device,
HLG=HLG,
symbols=symbol_table,
num_paths=num_paths,
G=G,
use_whole_lattice=use_whole_lattice,
output_beam_size=output_beam_size)
recog_path = exp_dir / f'recogs-{test_set}.txt'
store_transcripts(path=recog_path, texts=results)
logging.info(f'The transcripts are stored in {recog_path}')
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = exp_dir / f'errs-{test_set}.txt'
with open(errs_filename, 'w') as f:
write_error_stats(f, test_set, results)
logging.info('Wrote detailed error stats to {}'.format(errs_filename))
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
fix_random_seed(42)
setup_dist(rank, world_size, args.master_port)
exp_dir = Path('exp-' + model_type + '-noam-mmi-att-musan-sa-vgg')
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)
graph_compiler = MmiTrainingGraphCompiler(
lexicon=lexicon,
device=device,
)
phone_ids = lexicon.phone_symbols()
P = create_bigram_phone_lm(phone_ids)
P.scores = torch.zeros_like(P.scores)
P = P.to(device)
mls = MLSAsrDataModule(args)
train_dl = mls.train_dataloaders()
valid_dl = mls.valid_dataloaders()
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
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=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,
vgg_frontend=True)
elif model_type == "conformer":
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,
vgg_frontend=True)
elif model_type == "contextnet":
model = ContextNet(num_features=80, num_classes=len(phone_ids) +
1) # +1 for the blank symbol
else:
raise NotImplementedError("Model of type " + str(model_type) +
" is not implemented")
model.P_scores = nn.Parameter(P.scores.clone(), requires_grad=True)
model.to(device)
describe(model)
model = DDP(model, device_ids=[rank])
# Now for the aligment model, if any
if args.use_ali_model:
ali_model = TdnnLstm1b(
num_features=80,
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,
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,
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)
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)
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')
torch.distributed.barrier()
cleanup_dist()
def main():
parser = get_parser()
GigaSpeechAsrDataModule.add_arguments(parser)
args = parser.parse_args()
model_type = args.model_type
epoch = args.epoch
avg = args.avg
att_rate = args.att_rate
num_paths = args.num_paths
use_lm_rescoring = args.use_lm_rescoring
use_whole_lattice = False
if use_lm_rescoring and num_paths < 1:
# It doesn't make sense to use n-best list for rescoring
# when n is less than 1
use_whole_lattice = True
output_beam_size = args.output_beam_size
suffix = ''
if args.context_window is not None and args.context_window > 0:
suffix = f'ac{args.context_window}'
giga_subset = f'giga{args.subset}'
exp_dir = Path(
f'exp-{model_type}-mmi-att-sa-vgg-normlayer-{giga_subset}-{suffix}')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
logging.info(f'output_beam_size: {output_beam_size}')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
phone_ids = get_phone_symbols(phone_symbol_table)
phone_ids_with_blank = [0] + phone_ids
ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank))
logging.debug("About to load model")
# Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N
# device = torch.device('cuda', 1)
device = torch.device('cuda')
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,
vgg_frontend=args.vgg_fronted)
elif model_type == "conformer":
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,
vgg_frontend=args.vgg_frontend,
is_espnet_structure=args.is_espnet_structure)
elif model_type == "contextnet":
model = ContextNet(num_features=80, num_classes=len(phone_ids) +
1) # +1 for the blank symbol
else:
raise NotImplementedError("Model of type " + str(model_type) +
" is not implemented")
if avg == 1:
checkpoint = os.path.join(exp_dir, 'epoch-' + str(epoch - 1) + '.pt')
load_checkpoint(checkpoint, model)
else:
checkpoints = [
os.path.join(exp_dir, 'epoch-' + str(avg_epoch) + '.pt')
for avg_epoch in range(epoch - avg, epoch)
]
average_checkpoint(checkpoints, model)
if args.torchscript:
logging.info('Applying TorchScript to model...')
model = torch.jit.script(model)
ts_path = exp_dir / f'model_ts_epoch{epoch}_avg{avg}.pt'
logging.info(f'Storing the TorchScripted model in {ts_path}')
model.save(ts_path)
model.to(device)
model.eval()
if not os.path.exists(lang_dir / 'HLG.pt'):
logging.debug("Loading L_disambig.fst.txt")
with open(lang_dir / 'L_disambig.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
logging.debug("Loading G.fst.txt")
with open(lang_dir / 'G.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
HLG = compile_HLG(L=L,
G=G,
H=ctc_topo,
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(HLG.as_dict(), lang_dir / 'HLG.pt')
else:
logging.debug("Loading pre-compiled HLG")
d = torch.load(lang_dir / 'HLG.pt')
HLG = k2.Fsa.from_dict(d)
if use_lm_rescoring:
if use_whole_lattice:
logging.info('Rescoring with the whole lattice')
else:
logging.info(f'Rescoring with n-best list, n is {num_paths}')
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
if not os.path.exists(lang_dir / 'G_4_gram.pt'):
logging.debug('Loading G_4_gram.fst.txt')
with open(lang_dir / 'G_4_gram.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
# G.aux_labels is not needed in later computations, so
# remove it here.
del G.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.
G.labels[G.labels >= first_word_disambig_id] = 0
G = k2.create_fsa_vec([G]).to(device)
G = k2.arc_sort(G)
torch.save(G.as_dict(), lang_dir / 'G_4_gram.pt')
else:
logging.debug('Loading pre-compiled G_4_gram.pt')
d = torch.load(lang_dir / 'G_4_gram.pt')
G = k2.Fsa.from_dict(d).to(device)
if use_whole_lattice:
# Add epsilon self-loops to G as we will compose
# it with the whole lattice later
G = k2.add_epsilon_self_loops(G)
G = k2.arc_sort(G)
G = G.to(device)
# G.lm_scores is used to replace HLG.lm_scores during
# LM rescoring.
G.lm_scores = G.scores.clone()
else:
logging.debug('Decoding without LM rescoring')
G = None
if num_paths > 1:
logging.debug(f'Use n-best list decoding, n is {num_paths}')
else:
logging.debug('Use 1-best decoding')
logging.debug("convert HLG to device")
HLG = HLG.to(device)
HLG.aux_labels = k2.ragged.remove_values_eq(HLG.aux_labels, 0)
HLG.requires_grad_(False)
if not hasattr(HLG, 'lm_scores'):
HLG.lm_scores = HLG.scores.clone()
# load dataset
gigaspeech = GigaSpeechAsrDataModule(args)
test_sets = ['DEV', 'TEST']
for test_set, test_dl in zip(
test_sets,
[gigaspeech.valid_dataloaders(),
gigaspeech.test_dataloaders()]):
logging.info(f'* DECODING: {test_set}')
test_set_wers = dict()
results_dict = decode(dataloader=test_dl,
model=model,
HLG=HLG,
symbols=symbol_table,
num_paths=num_paths,
G=G,
use_whole_lattice=use_whole_lattice,
output_beam_size=output_beam_size)
for key, results in results_dict.items():
recog_path = exp_dir / f'recogs-{test_set}-{key}.txt'
store_transcripts(path=recog_path, texts=results)
logging.info(f'The transcripts are stored in {recog_path}')
ref_path = exp_dir / f'ref-{test_set}.trn'
hyp_path = exp_dir / f'hyp-{test_set}.trn'
store_transcripts_for_sclite(ref_path=ref_path,
hyp_path=hyp_path,
texts=results)
logging.info(
f'The sclite-format transcripts are stored in {ref_path} and {hyp_path}'
)
cmd = f'python3 GigaSpeech/utils/gigaspeech_scoring.py {ref_path} {hyp_path} {exp_dir / "tmp_sclite"}'
logging.info(cmd)
try:
subprocess.run(cmd, check=True, shell=True)
except subprocess.CalledProcessError:
logging.error(
'Skipping sclite scoring as it failed to run: Is "sclite" registered in your $PATH?"'
)
# The following prints out WERs, per-word error statistics and aligned
# ref/hyp pairs.
errs_filename = exp_dir / f'errs-{test_set}-{key}.txt'
with open(errs_filename, 'w') as f:
wer = write_error_stats(f, f'{test_set}-{key}', results)
test_set_wers[key] = wer
logging.info(
'Wrote detailed error stats to {}'.format(errs_filename))
test_set_wers = sorted(test_set_wers.items(), key=lambda x: x[1])
errs_info = exp_dir / f'wer-summary-{test_set}.txt'
with open(errs_info, 'w') as f:
print('settings\tWER', file=f)
for key, val in test_set_wers:
print('{}\t{}'.format(key, val), file=f)
s = '\nFor {}, WER of different settings are:\n'.format(test_set)
note = '\tbest for {}'.format(test_set)
for key, val in test_set_wers:
s += '{}\t{}{}\n'.format(key, val, note)
note = ''
logging.info(s)
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)
suffix = ''
if args.context_window is not None and args.context_window > 0:
suffix = f'ac{args.context_window}'
giga_subset = f'giga{args.subset}'
exp_dir = Path(
f'exp-{model_type}-mmi-att-sa-vgg-normlayer-{giga_subset}-{suffix}')
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
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 / f'P_{args.subset}.pt').is_file():
logging.debug(f'Loading P from {lang_dir}/P_{args.subset}.fst.txt')
with open(lang_dir / f'P_{args.subset}.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 / f'P_{args.subset}.pt')
else:
logging.debug('Loading pre-compiled P')
d = torch.load(lang_dir / f'P_{args.subset}.pt')
P = k2.Fsa.from_dict(d)
graph_compiler = MmiTrainingGraphCompiler(
lexicon=lexicon,
P=P,
device=device,
)
phone_ids = lexicon.phone_symbols()
gigaspeech = GigaSpeechAsrDataModule(args)
train_dl = gigaspeech.train_dataloaders()
valid_dl = gigaspeech.valid_dataloaders()
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=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,
vgg_frontend=True)
elif model_type == "conformer":
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,
vgg_frontend=True,
is_espnet_structure=True)
elif model_type == "contextnet":
model = ContextNet(num_features=80, 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=80,
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()
def main():
parser = get_parser()
args = parser.parse_args()
model_type = args.model_type
epoch = args.epoch
avg = args.avg
att_rate = args.att_rate
num_paths = args.num_paths
use_lm_rescoring = args.use_lm_rescoring
use_whole_lattice = False
if use_lm_rescoring and num_paths < 1:
# It doesn't make sense to use n-best list for rescoring
# when n is less than 1
use_whole_lattice = True
output_beam_size = args.output_beam_size
exp_dir = Path('exp-' + model_type + '-mmi-att-sa-vgg-normlayer')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
logging.info(f'output_beam_size: {output_beam_size}')
# load L, G, symbol_table
lang_dir = Path('data/lang_nosp')
symbol_table = k2.SymbolTable.from_file(lang_dir / 'words.txt')
phone_symbol_table = k2.SymbolTable.from_file(lang_dir / 'phones.txt')
phone_ids = get_phone_symbols(phone_symbol_table)
phone_ids_with_blank = [0] + phone_ids
ctc_topo = k2.arc_sort(build_ctc_topo(phone_ids_with_blank))
logging.debug("About to load model")
# Note: Use "export CUDA_VISIBLE_DEVICES=N" to setup device id to N
# device = torch.device('cuda', 1)
device = torch.device('cuda')
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=args.vgg_fronted)
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=args.vgg_frontend,
is_espnet_structure=args.is_espnet_structure)
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 avg == 1:
checkpoint = os.path.join(exp_dir, 'epoch-' + str(epoch - 1) + '.pt')
load_checkpoint(checkpoint, model)
else:
checkpoints = [
os.path.join(exp_dir, 'epoch-' + str(avg_epoch) + '.pt')
for avg_epoch in range(epoch - avg, epoch)
]
average_checkpoint(checkpoints, model)
model.to(device)
model.eval()
if not os.path.exists(lang_dir / 'HLG.pt'):
logging.debug("Loading L_disambig.fst.txt")
with open(lang_dir / 'L_disambig.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
logging.debug("Loading G.fst.txt")
with open(lang_dir / 'G.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
HLG = compile_HLG(L=L,
G=G,
H=ctc_topo,
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(HLG.as_dict(), lang_dir / 'HLG.pt')
else:
logging.debug("Loading pre-compiled HLG")
d = torch.load(lang_dir / 'HLG.pt')
HLG = k2.Fsa.from_dict(d)
logging.debug('Decoding without LM rescoring')
G = None
if num_paths > 1:
logging.debug(f'Use n-best list decoding, n is {num_paths}')
else:
logging.debug('Use 1-best decoding')
logging.debug("convert HLG to device")
HLG = HLG.to(device)
HLG.aux_labels = k2.ragged.remove_values_eq(HLG.aux_labels, 0)
HLG.requires_grad_(False)
if not hasattr(HLG, 'lm_scores'):
HLG.lm_scores = HLG.scores.clone()
# load dataset
feature_dir = Path('exp/data')
logging.info("About to get test cuts")
cuts_test = CutSet.from_json(feature_dir / 'cuts_test.json.gz')
logging.info("About to create test dataset")
test = K2SpeechRecognitionDataset(cuts_test)
test_sampler = SingleCutSampler(cuts_test, max_frames=12000)
logging.info("About to create test dataloader")
test_dl = torch.utils.data.DataLoader(test,
sampler=test_sampler,
batch_size=None,
num_workers=1)
logging.info("About to decode")
results = decode(dataloader=test_dl,
model=model,
HLG=HLG,
symbols=symbol_table,
num_paths=num_paths,
G=G,
use_whole_lattice=use_whole_lattice,
output_beam_size=output_beam_size)
s = ''
results2 = []
for ref, hyp in results:
s += f'ref={ref}\n'
s += f'hyp={hyp}\n'
results2.append((list(''.join(ref)), list(''.join(hyp))))
logging.info(s)
# compute WER
dists = [edit_distance(r, h) for r, h in results]
dists2 = [edit_distance(r, h) for r, h in results2]
errors = {
key: sum(dist[key] for dist in dists)
for key in ['sub', 'ins', 'del', 'total']
}
errors2 = {
key: sum(dist[key] for dist in dists2)
for key in ['sub', 'ins', 'del', 'total']
}
total_words = sum(len(ref) for ref, _ in results)
total_chars = sum(len(ref) for ref, _ in results2)
# Print Kaldi-like message:
# %WER 8.20 [ 4459 / 54402, 695 ins, 427 del, 3337 sub ]
logging.info(
f'%WER {errors["total"] / total_words:.2%} '
f'[{errors["total"]} / {total_words}, {errors["ins"]} ins, {errors["del"]} del, {errors["sub"]} sub ]'
)
logging.info(
f'%CER {errors2["total"] / total_chars:.2%} '
f'[{errors2["total"]} / {total_chars}, {errors2["ins"]} ins, {errors2["del"]} del, {errors2["sub"]} sub ]'
)