def test_specaugment_2d_input_raises_error():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
feats = torch.from_numpy(cuts[0].load_features())
tfnm = SpecAugment(p=1.0, time_warp_factor=10)
with pytest.raises(AssertionError):
augmented = tfnm(feats)
assert (feats != augmented).any()
def libri_cut_set():
cuts = CutSet.from_json("test/fixtures/libri/cuts.json")
return CutSet.from_cuts([
cuts[0],
cuts[0].with_id("copy-1"),
cuts[0].with_id("copy-2"),
cuts[0].append(cuts[0]),
])
def libri_cut_set():
cs = CutSet.from_json('test/fixtures/libri/cuts.json')
return CutSet.from_cuts([
cs[0],
cs[0].with_id('copy-1'),
cs[0].with_id('copy-2'),
cs[0].append(cs[0])
])
def test_collate_audio_padding():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
assert len(set(cut.num_samples for cut in cuts)) > 1
correct_pad = max(cut.num_samples for cut in cuts)
audio, audio_lens = collate_audio(cuts)
assert audio.shape[-1] == correct_pad
assert max(audio_lens).item() == correct_pad
def test_collate_feature_padding():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
assert len(set(cut.num_frames for cut in cuts)) > 1
correct_pad = max(cut.num_frames for cut in cuts)
features, features_lens = collate_features(cuts)
assert features.shape[1] == correct_pad
assert max(features_lens).item() == correct_pad
def test_collate_audio_padding_fault_tolerant_return_vals():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
assert len(set(cut.num_samples for cut in cuts)) > 1
correct_pad = max(cut.num_samples for cut in cuts)
audio, audio_lens, cuts_ok = collate_audio(cuts, fault_tolerant=True)
assert len(cuts) == len(cuts_ok)
assert audio.shape[-1] == correct_pad
assert max(audio_lens).item() == correct_pad
def test_specaugment_batch(num_feature_masks, num_frame_masks):
cuts = CutSet.from_json('test/fixtures/ljspeech/cuts.json')
feats, feat_lens = collate_features(cuts)
tfnm = SpecAugment(p=1.0,
time_warp_factor=10,
features_mask_size=5,
frames_mask_size=20,
num_feature_masks=num_feature_masks,
num_frame_masks=num_frame_masks)
augmented = tfnm(feats)
assert (feats != augmented).any()
def test_cut_set_serialization(cut_set, format, compressed):
with NamedTemporaryFile(suffix=".gz" if compressed else "") as f:
if format == "yaml":
cut_set.to_yaml(f.name)
restored = CutSet.from_yaml(f.name)
if format == "json":
cut_set.to_json(f.name)
restored = CutSet.from_json(f.name)
if format == "jsonl":
cut_set.to_jsonl(f.name)
restored = CutSet.from_jsonl(f.name)
assert cut_set == restored
def test_cut_set_serialization(cut_set, format, compressed):
with NamedTemporaryFile(suffix='.gz' if compressed else '') as f:
if format == 'yaml':
cut_set.to_yaml(f.name)
restored = CutSet.from_yaml(f.name)
if format == 'json':
cut_set.to_json(f.name)
restored = CutSet.from_json(f.name)
if format == 'jsonl':
cut_set.to_jsonl(f.name)
restored = CutSet.from_jsonl(f.name)
assert cut_set == restored
def test_collate_custom_numbers():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
expected_snrs = []
for cut in cuts:
expected_snrs.append(random.random() * 20)
cut.snr = expected_snrs[-1]
snrs = collate_custom_field(cuts, "snr")
assert isinstance(snrs, torch.Tensor)
assert snrs.dtype == torch.float32
assert snrs.shape == (len(cuts), )
for idx, snr in enumerate(expected_snrs):
assert isclose(snrs[idx], snr, abs_tol=1e-5)
def test_collate_custom_temporal_array_ints(pad_value):
CODEBOOK_SIZE = 512
FRAME_SHIFT = 0.04
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
max_num_frames = max(
seconds_to_frames(cut.duration, FRAME_SHIFT) for cut in cuts)
with NamedTemporaryFile(suffix=".h5") as f, NumpyHdf5Writer(
f.name) as writer:
expected_codebook_indices = []
for cut in cuts:
expected_codebook_indices.append(
np.random.randint(CODEBOOK_SIZE,
size=(seconds_to_frames(
cut.duration,
FRAME_SHIFT), )).astype(np.int16))
cut.codebook_indices = writer.store_array(
cut.id,
expected_codebook_indices[-1],
frame_shift=FRAME_SHIFT,
temporal_dim=0,
)
codebook_indices, codebook_indices_lens = collate_custom_field(
cuts, "codebook_indices", pad_value=pad_value)
assert isinstance(codebook_indices_lens, torch.Tensor)
assert codebook_indices_lens.dtype == torch.int32
assert codebook_indices_lens.shape == (len(cuts), )
assert codebook_indices_lens.tolist() == [
seconds_to_frames(c.duration, FRAME_SHIFT) for c in cuts
]
assert isinstance(codebook_indices, torch.Tensor)
assert codebook_indices.dtype == torch.int16
assert codebook_indices.shape == (len(cuts), max_num_frames)
for idx, cbidxs in enumerate(expected_codebook_indices):
exp_len = cbidxs.shape[0]
# PyTorch < 1.9.0 doesn't have an assert_equal function.
np.testing.assert_equal(codebook_indices[idx, :exp_len].numpy(),
cbidxs)
expected_pad_value = 0 if pad_value is None else pad_value
np.testing.assert_equal(codebook_indices[idx, exp_len:].numpy(),
expected_pad_value)
def test_collate_custom_array():
EMBEDDING_SIZE = 300
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
with NamedTemporaryFile(suffix=".h5") as f, NumpyHdf5Writer(
f.name) as writer:
expected_xvectors = []
for cut in cuts:
expected_xvectors.append(
np.random.randn(EMBEDDING_SIZE).astype(np.float32))
cut.xvector = writer.store_array(cut.id, expected_xvectors[-1])
xvectors = collate_custom_field(cuts, "xvector")
assert isinstance(xvectors, torch.Tensor)
assert xvectors.dtype == torch.float32
assert xvectors.shape == (len(cuts), EMBEDDING_SIZE)
for idx, xvec in enumerate(expected_xvectors):
torch.testing.assert_allclose(xvectors[idx], xvec)
def test_collate_custom_temporal_array_floats(pad_value):
VOCAB_SIZE = 500
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
max_num_frames = max(cut.num_frames for cut in cuts)
with NamedTemporaryFile(suffix=".h5") as f, NumpyHdf5Writer(
f.name) as writer:
expected_posteriors = []
for cut in cuts:
expected_posteriors.append(
np.random.randn(cut.num_frames, VOCAB_SIZE).astype(np.float32))
cut.posterior = writer.store_array(
cut.id,
expected_posteriors[-1],
frame_shift=cut.frame_shift,
temporal_dim=0,
)
posteriors, posterior_lens = collate_custom_field(cuts,
"posterior",
pad_value=pad_value)
assert isinstance(posterior_lens, torch.Tensor)
assert posterior_lens.dtype == torch.int32
assert posterior_lens.shape == (len(cuts), )
assert posterior_lens.tolist() == [c.num_frames for c in cuts]
assert isinstance(posteriors, torch.Tensor)
assert posteriors.dtype == torch.float32
assert posteriors.shape == (len(cuts), max_num_frames, VOCAB_SIZE)
for idx, post in enumerate(expected_posteriors):
exp_len = post.shape[0]
torch.testing.assert_allclose(posteriors[idx, :exp_len], post)
expected_pad_value = 0 if pad_value is None else pad_value
torch.testing.assert_allclose(
posteriors[idx, exp_len:],
expected_pad_value *
torch.ones_like(posteriors[idx, exp_len:]),
)
def main():
# load L, G, symbol_table
lang_dir = 'data/lang_nosp'
with open(lang_dir + '/L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
with open(lang_dir + '/G.fsa.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=True)
with open(lang_dir + '/words.txt') as f:
symbol_table = k2.SymbolTable.from_str(f.read())
L = k2.arc_sort(L.invert_())
G = k2.arc_sort(G)
graph = k2.intersect(L, G)
graph = k2.arc_sort(graph)
# load dataset
feature_dir = 'exp/data1'
cuts_train = CutSet.from_json(feature_dir +
'/cuts_train-clean-100.json.gz')
cuts_dev = CutSet.from_json(feature_dir + '/cuts_dev-clean.json.gz')
train = K2SpeechRecognitionIterableDataset(cuts_train, shuffle=True)
validate = K2SpeechRecognitionIterableDataset(cuts_dev, shuffle=False)
train_dl = torch.utils.data.DataLoader(train,
batch_size=None,
num_workers=1)
valid_dl = torch.utils.data.DataLoader(validate,
batch_size=None,
num_workers=1)
dir = 'exp'
setup_logger('{}/log/log-train'.format(dir))
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
device_id = 0
device = torch.device('cuda', device_id)
model = Wav2Letter(num_classes=364, input_type='mfcc', num_features=40)
model.to(device)
learning_rate = 0.001
start_epoch = 0
num_epochs = 10
best_objf = 100000
best_epoch = start_epoch
best_model_path = os.path.join(dir, 'best_model.pt')
best_epoch_info_filename = os.path.join(dir, 'best-epoch-info')
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=5e-4)
# optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
for epoch in range(start_epoch, num_epochs):
curr_learning_rate = learning_rate * pow(0.4, epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = curr_learning_rate
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf = train_one_epoch(dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
device=device,
graph=graph,
symbols=symbol_table,
optimizer=optimizer,
current_epoch=epoch,
num_epochs=num_epochs)
if objf < best_objf:
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
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,
best_epoch=best_epoch)
# we always save the model for every epoch
model_path = os.path.join(dir, 'epoch-{}.pt'.format(epoch))
save_checkpoint(filename=model_path,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
epoch_info_filename = os.path.join(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,
best_epoch=best_epoch)
logging.warning('Done')
def main():
exp_dir = Path('exp-lstm-adam')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
# 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')
ctc_topo = build_ctc_topo(list(phone_symbol_table._id2sym.keys()))
ctc_topo = k2.arc_sort(ctc_topo)
if not os.path.exists(lang_dir / 'LG.pt'):
print("Loading L_disambig.fst.txt")
with open(lang_dir / 'L_disambig.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
print("Loading G.fsa.txt")
with open(lang_dir / 'G.fsa.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=True)
first_phone_disambig_id = find_first_disambig_symbol(
phone_symbol_table)
first_word_disambig_id = find_first_disambig_symbol(symbol_table)
LG = compile_LG(L=L,
G=G,
ctc_topo=ctc_topo,
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(LG.as_dict(), lang_dir / 'LG.pt')
else:
print("Loading pre-compiled LG")
d = torch.load(lang_dir / 'LG.pt')
LG = k2.Fsa.from_dict(d)
# load dataset
feature_dir = Path('exp/data')
print("About to get test cuts")
cuts_test = CutSet.from_json(feature_dir / 'cuts_test.json.gz')
print("About to create test dataset")
test = K2SpeechRecognitionIterableDataset(cuts_test,
max_frames=100000,
shuffle=False,
concat_cuts=False)
print("About to create test dataloader")
test_dl = torch.utils.data.DataLoader(test, batch_size=None, num_workers=1)
# if not torch.cuda.is_available():
# logging.error('No GPU detected!')
# sys.exit(-1)
print("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')
model = TdnnLstm1b(num_features=40, num_classes=len(phone_symbol_table))
checkpoint = os.path.join(exp_dir, 'epoch-9.pt')
load_checkpoint(checkpoint, model)
model.to(device)
model.eval()
print("convert LG to device")
LG = LG.to(device)
LG.aux_labels = k2.ragged.remove_values_eq(LG.aux_labels, 0)
LG.requires_grad_(False)
print("About to decode")
results = decode(dataloader=test_dl,
model=model,
device=device,
LG=LG,
symbols=symbol_table)
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'%WER {errors2["total"] / total_chars:.2%} '
f'[{errors2["total"]} / {total_chars}, {errors2["ins"]} ins, {errors2["del"]} del, {errors2["sub"]} sub ]'
)
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():
# load L, G, symbol_table
lang_dir = 'data/lang_nosp'
with open(lang_dir + '/words.txt') as f:
symbol_table = k2.SymbolTable.from_str(f.read())
## This commented code created LG. We don't need that there.
## There were problems with disambiguation symbols; the G has
## disambiguation symbols which L.fst doesn't support.
# if not os.path.exists(lang_dir + '/LG.pt'):
# print("Loading L.fst.txt")
# with open(lang_dir + '/L.fst.txt') as f:
# L = k2.Fsa.from_openfst(f.read(), acceptor=False)
# print("Loading G.fsa.txt")
# with open(lang_dir + '/G.fsa.txt') as f:
# G = k2.Fsa.from_openfst(f.read(), acceptor=True)
# print("Arc-sorting L...")
# L = k2.arc_sort(L.invert_())
# G = k2.arc_sort(G)
# print(k2.is_arc_sorted(k2.get_properties(L)))
# print(k2.is_arc_sorted(k2.get_properties(G)))
# print("Intersecting L and G")
# graph = k2.intersect(L, G)
# graph = k2.arc_sort(graph)
# print(k2.is_arc_sorted(k2.get_properties(graph)))
# torch.save(graph.as_dict(), lang_dir + '/LG.pt')
# else:
# d = torch.load(lang_dir + '/LG.pt')
# print("Loading pre-prepared LG")
# graph = k2.Fsa.from_dict(d)
print("Loading L.fst.txt")
with open(lang_dir + '/L.fst.txt') as f:
L = k2.Fsa.from_openfst(f.read(), acceptor=False)
L = k2.arc_sort(L.invert_())
# load dataset
feature_dir = 'exp/data1'
print("About to get train cuts")
#cuts_train = CutSet.from_json(feature_dir +
# '/cuts_train-clean-100.json.gz')
cuts_train = CutSet.from_json(feature_dir + '/cuts_dev-clean.json.gz')
print("About to get dev cuts")
cuts_dev = CutSet.from_json(feature_dir + '/cuts_dev-clean.json.gz')
print("About to create train dataset")
train = K2SpeechRecognitionIterableDataset(cuts_train,
max_frames=1000,
shuffle=True)
print("About to create dev dataset")
validate = K2SpeechRecognitionIterableDataset(cuts_dev,
max_frames=1000,
shuffle=False)
print("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
batch_size=None,
num_workers=1)
print("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
batch_size=None,
num_workers=1)
exp_dir = 'exp'
setup_logger('{}/log/log-train'.format(exp_dir))
if not torch.cuda.is_available():
logging.error('No GPU detected!')
sys.exit(-1)
print("About to create model")
device_id = 0
device = torch.device('cuda', device_id)
model = Wav2Letter(num_classes=364, input_type='mfcc', num_features=40)
model.to(device)
learning_rate = 0.001
start_epoch = 0
num_epochs = 10
best_objf = 100000
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')
optimizer = optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=5e-4)
# optimizer = optim.SGD(net.parameters(), lr=learning_rate, momentum=0.9)
for epoch in range(start_epoch, num_epochs):
curr_learning_rate = learning_rate * pow(0.4, epoch)
for param_group in optimizer.param_groups:
param_group['lr'] = curr_learning_rate
logging.info('epoch {}, learning rate {}'.format(
epoch, curr_learning_rate))
objf = train_one_epoch(dataloader=train_dl,
valid_dataloader=valid_dl,
model=model,
device=device,
L=L,
symbols=symbol_table,
optimizer=optimizer,
current_epoch=epoch,
num_epochs=num_epochs)
if objf < best_objf:
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
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,
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,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
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,
best_epoch=best_epoch)
logging.warning('Done')
def test_collate_custom_attribute_missing():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
with pytest.raises(AttributeError):
collate_custom_field(cuts, "nonexistent_attribute")
def test_collate_custom_temporal_array_ints(pad_direction):
CODEBOOK_SIZE = 512
FRAME_SHIFT = 0.04
EXPECTED_PAD_VALUE = 0
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
max_num_frames = max(
seconds_to_frames(cut.duration, FRAME_SHIFT) for cut in cuts)
with NamedTemporaryFile(suffix=".h5") as f, NumpyHdf5Writer(
f.name) as writer:
expected_codebook_indices = []
for cut in cuts:
expected_codebook_indices.append(
np.random.randint(CODEBOOK_SIZE,
size=(seconds_to_frames(
cut.duration,
FRAME_SHIFT), )).astype(np.int16))
cut.codebook_indices = writer.store_array(
cut.id,
expected_codebook_indices[-1],
frame_shift=FRAME_SHIFT,
temporal_dim=0,
)
codebook_indices, codebook_indices_lens = collate_custom_field(
cuts, "codebook_indices", pad_direction=pad_direction)
assert isinstance(codebook_indices_lens, torch.Tensor)
assert codebook_indices_lens.dtype == torch.int32
assert codebook_indices_lens.shape == (len(cuts), )
assert codebook_indices_lens.tolist() == [
seconds_to_frames(c.duration, FRAME_SHIFT) for c in cuts
]
assert isinstance(codebook_indices, torch.Tensor)
assert (codebook_indices.dtype == torch.int64
) # the dtype got promoted by default
assert codebook_indices.shape == (len(cuts), max_num_frames)
for idx, cbidxs in enumerate(expected_codebook_indices):
exp_len = cbidxs.shape[0]
# PyTorch < 1.9.0 doesn't have an assert_equal function.
if pad_direction == "right":
np.testing.assert_equal(
codebook_indices[idx, :exp_len].numpy(), cbidxs)
np.testing.assert_equal(
codebook_indices[idx, exp_len:].numpy(),
EXPECTED_PAD_VALUE)
if pad_direction == "left":
np.testing.assert_equal(
codebook_indices[idx, -exp_len:].numpy(), cbidxs)
np.testing.assert_equal(
codebook_indices[idx, :-exp_len].numpy(),
EXPECTED_PAD_VALUE)
if pad_direction == "both":
half = (max_num_frames - exp_len) // 2
np.testing.assert_equal(codebook_indices[idx, :half].numpy(),
EXPECTED_PAD_VALUE)
np.testing.assert_equal(
codebook_indices[idx, half:half + exp_len].numpy(), cbidxs)
if half > 0:
# indexing like [idx, -0:] would return the whole array rather
# than an empty slice.
np.testing.assert_equal(
codebook_indices[idx, -half:].numpy(),
EXPECTED_PAD_VALUE)
def cut_set():
return CutSet.from_json('test/fixtures/ljspeech/cuts.json')
def global_mvn():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
return GlobalMVN.from_cuts(cuts)
def main():
args = get_parser().parse_args()
epoch = args.epoch
max_frames = args.max_frames
avg = args.avg
att_rate = args.att_rate
# exp_dir = Path('/export/gpudisk2/data/hegc/audio_workspace/snowfall_aishell1/exp-transformer-noam-mmi-att-musan')
exp_dir = Path('exp-transformer-noam-mmi-att-musan')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
# 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
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 = 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 / 'LG.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)
LG = compile_LG(L=L,
G=G,
ctc_topo=ctc_topo,
labels_disambig_id_start=first_phone_disambig_id,
aux_labels_disambig_id_start=first_word_disambig_id)
torch.save(LG.as_dict(), lang_dir / 'LG.pt')
else:
logging.debug("Loading pre-compiled LG")
d = torch.load(lang_dir / 'LG.pt')
LG = k2.Fsa.from_dict(d)
# load dataset
# feature_dir = Path('/export/gpudisk2/data/hegc/audio_workspace/snowfall_aishell1/exp/data')
feature_dir = Path('exp/data')
logging.debug("About to get test cuts")
cuts_test = CutSet.from_json(feature_dir / 'cuts_test.json.gz')
logging.debug("About to create test dataset")
test = K2SpeechRecognitionDataset(cuts_test)
sampler = SingleCutSampler(cuts_test, max_frames=max_frames)
logging.debug("About to create test dataloader")
test_dl = torch.utils.data.DataLoader(test, batch_size=None, sampler=sampler, num_workers=1)
# if not torch.cuda.is_available():
# logging.error('No GPU detected!')
# sys.exit(-1)
logging.debug("convert LG to device")
LG = LG.to(device)
LG.aux_labels = k2.ragged.remove_values_eq(LG.aux_labels, 0)
LG.requires_grad_(False)
logging.debug("About to decode")
results = decode(dataloader=test_dl,
model=model,
device=device,
LG=LG,
symbols=symbol_table)
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'%WER {errors2["total"] / total_chars:.2%} '
f'[{errors2["total"]} / {total_chars}, {errors2["ins"]} ins, {errors2["del"]} del, {errors2["sub"]} sub ]'
)
def main():
exp_dir = Path('exp-lstm-adam-mmi-mbr-musan')
setup_logger('{}/log/log-decode'.format(exp_dir), log_level='debug')
# 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')
model = TdnnLstm1b(
num_features=40,
num_classes=len(phone_ids) + 1, # +1 for the blank symbol
subsampling_factor=3)
model.P_scores = torch.nn.Parameter(P.scores.clone(), requires_grad=False)
checkpoint = os.path.join(exp_dir, 'epoch-9.pt')
load_checkpoint(checkpoint, 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)
# load dataset
feature_dir = Path('exp/data')
logging.debug("About to get test cuts")
cuts_test = CutSet.from_json(feature_dir / 'cuts_test-clean.json.gz')
logging.info("About to create test dataset")
test = K2SpeechRecognitionDataset(cuts_test)
sampler = SingleCutSampler(cuts_test, max_frames=100000)
logging.info("About to create test dataloader")
test_dl = torch.utils.data.DataLoader(test,
batch_size=None,
sampler=sampler,
num_workers=1)
# if not torch.cuda.is_available():
# logging.error('No GPU detected!')
# sys.exit(-1)
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)
logging.debug("About to decode")
results = decode(dataloader=test_dl,
model=model,
device=device,
HLG=HLG,
symbols=symbol_table)
s = ''
for ref, hyp in results:
s += f'ref={ref}\n'
s += f'hyp={hyp}\n'
logging.info(s)
# compute WER
dists = [edit_distance(r, h) for r, h in results]
errors = {
key: sum(dist[key] for dist in dists)
for key in ['sub', 'ins', 'del', 'total']
}
total_words = sum(len(ref) for ref, _ in results)
# 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 ]'
)
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 test_global_mvn_from_cuts():
cuts = CutSet.from_json("test/fixtures/ljspeech/cuts.json")
stats1 = GlobalMVN.from_cuts(cuts)
stats2 = GlobalMVN.from_cuts(cuts, max_cuts=1)
assert isinstance(stats1, GlobalMVN)
assert isinstance(stats2, GlobalMVN)
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 ]'
)
def main():
fix_random_seed(42)
exp_dir = 'exp-lstm-adam'
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,
oov='<SPOKEN_NOISE>')
# 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 create train dataset")
train = K2SpeechRecognitionIterableDataset(cuts_train,
max_frames=90000,
shuffle=True)
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionIterableDataset(cuts_dev,
max_frames=90000,
shuffle=False,
concat_cuts=False)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
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_symbol_table),
subsampling_factor=3)
learning_rate = 0.00001
start_epoch = 0
num_epochs = 10
best_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
global_batch_idx_valid = 0 # for logging only
if start_epoch > 0:
model_path = os.path.join(exp_dir,
'epoch-{}.pt'.format(start_epoch - 1))
(epoch, learning_rate, objf) = load_checkpoint(filename=model_path,
model=model)
best_objf = objf
logging.info("epoch = {}, objf = {}".format(epoch, objf))
model.to(device)
describe(model)
# optimizer = optim.SGD(model.parameters(),
# lr=learning_rate,
# momentum=0.9,
# weight_decay=5e-4)
optimizer = optim.AdamW(
model.parameters(),
# lr=learning_rate,
weight_decay=5e-4)
for epoch in range(start_epoch, num_epochs):
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 = 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,
global_batch_idx_valid=global_batch_idx_valid)
# the lower, the better
if objf < best_objf:
best_objf = objf
best_epoch = epoch
save_checkpoint(filename=best_model_path,
model=model,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
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,
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,
epoch=epoch,
learning_rate=curr_learning_rate,
objf=objf)
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,
best_epoch=best_epoch)
logging.warning('Done')
def main():
if not torch.cuda.is_available():
logging.error("No GPU detected!")
sys.exit(-1)
device_id = 0
device = torch.device("cuda", device_id)
# Reserve the GPU with a dummy variable
reserve_variable = torch.ones(1).to(device)
exp_dir = Path("exp-tl1a-adam-xent")
setup_logger("{}/log/log-decode".format(exp_dir), log_level="debug")
if not os.path.exists(exp_dir / "HCLG.pt"):
logging.info("Preparing decoding graph")
# sym_str = """
# <eps> 0
# silence 1
# speech 2
# """
# symbol_table = k2.SymbolTable.from_str(sym_str)
HCLG = prepare_decoding_graph(
min_silence_duration=0.03,
min_speech_duration=0.3,
max_speech_duration=10.0,
)
# Arc sort the HCLG since it is needed for intersect
logging.info("Sorting decoding graph by outgoing arcs")
HCLG = k2.arc_sort(HCLG)
# HCLG.symbols = symbol_table
torch.save(HCLG.as_dict(), exp_dir / "HCLG.pt")
else:
logging.info("Loading pre-compiled decoding graph")
d = torch.load(exp_dir / "HCLG.pt")
HCLG = k2.Fsa.from_dict(d)
# 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 = K2VadDataset(cuts_test, return_cuts=True)
sampler = SingleCutSampler(cuts_test, max_frames=100000)
logging.info("About to create test dataloader")
test_dl = torch.utils.data.DataLoader(test,
batch_size=None,
sampler=sampler,
num_workers=1)
logging.info("About to load model")
model = TdnnLstm1a(
num_features=80,
num_classes=2, # speech/silence
subsampling_factor=1,
)
checkpoint = os.path.join(exp_dir, "best_model.pt")
load_checkpoint(checkpoint, model)
model.to(device)
model.eval()
logging.info("convert decoding graph to device")
HCLG = HCLG.to(device)
HCLG.requires_grad_(False)
logging.info("About to decode")
results = decode(dataloader=test_dl, model=model, device=device, HCLG=HCLG)
# Compute frame-level accuracy and precision-recall metrics
y_true = []
y_pred = []
for result in results:
cut, ref, hyp = result
y_true.append(ref)
y_pred.append(hyp)
y_true = torch.cat(y_true, dim=0).numpy()
y_pred = torch.cat(y_pred, dim=0).numpy()
logging.info("Results: \n{}".format(
classification_report(y_true,
y_pred,
target_names=["silence", "speech"])))
# Create output segments per recording
create_and_write_segments(
[result[0] for result in results], # cuts
[result[2] for result in results], # outputs
exp_dir / "segments", # segments file
)
def main():
fix_random_seed(42)
exp_dir = f'exp-lstm-adam-mmi-mbr-musan'
setup_logger('{}/log/log-train'.format(exp_dir))
tb_writer = SummaryWriter(log_dir=f'{exp_dir}/tensorboard')
if not torch.cuda.is_available():
logging.warn('No GPU detected!')
logging.warn('USE CPU (very slow)!')
device = torch.device('cpu')
else:
logging.info('Use GPU')
device_id = 0
device = torch.device('cuda', device_id)
# 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():
logging.info('Loading precompiled L')
L_inv = k2.Fsa.from_dict(torch.load(lang_dir / 'Linv.pt'))
else:
logging.info('Compiling L')
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')
logging.info("Loading L_disambig.fst")
if (lang_dir / 'L_disambig.pt').exists():
logging.info('Loading precompiled L_disambig')
L_disambig = k2.Fsa.from_dict(torch.load(lang_dir / 'L_disambig.pt'))
else:
logging.info('Compiling L_disambig')
with open(lang_dir / 'L_disambig.fst.txt') as f:
L_disambig = k2.Fsa.from_openfst(f.read(), acceptor=False)
L_disambig = k2.arc_sort(L_disambig)
torch.save(L_disambig.as_dict(), lang_dir / 'L_disambig.pt')
logging.info("Loading G.fst")
if (lang_dir / 'G_uni.pt').exists():
logging.info('Loading precompiled G')
G = k2.Fsa.from_dict(torch.load(lang_dir / 'G_uni.pt'))
else:
logging.info('Compiling G')
with open(lang_dir / 'G_uni.fst.txt') as f:
G = k2.Fsa.from_openfst(f.read(), acceptor=False)
G = k2.arc_sort(G)
torch.save(G.as_dict(), lang_dir / 'G_uni.pt')
graph_compiler = MmiMbrTrainingGraphCompiler(L_inv=L_inv,
L_disambig=L_disambig,
G=G,
device=device,
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")
train = K2SpeechRecognitionIterableDataset(cuts_train,
max_frames=30000,
shuffle=True,
aug_cuts=cuts_musan,
aug_prob=0.5,
aug_snr=(10, 20))
logging.info("About to create dev dataset")
validate = K2SpeechRecognitionIterableDataset(cuts_dev,
max_frames=60000,
shuffle=False,
concat_cuts=False)
logging.info("About to create train dataloader")
train_dl = torch.utils.data.DataLoader(train,
batch_size=None,
num_workers=4)
logging.info("About to create dev dataloader")
valid_dl = torch.utils.data.DataLoader(validate,
batch_size=None,
num_workers=1)
logging.info("About to create model")
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)
start_epoch = 0
num_epochs = 10
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
use_adam = True
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)
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}"
)
model.to(device)
describe(model)
P = P.to(device)
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,
last_epoch=start_epoch - 1)
for epoch in range(start_epoch, num_epochs):
# 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,
)
# 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,
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,
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)
lr_scheduler.step()
logging.warning('Done')
def test_specaugment_single():
cuts = CutSet.from_json('test/fixtures/ljspeech/cuts.json')
feats = torch.from_numpy(cuts[0].load_features())
tfnm = SpecAugment(p=1.0, time_warp_factor=10)
augmented = tfnm(feats)
assert (feats != augmented).any()
