def get_dataloader(model_args_list):
''' return dataloader for inference '''
from inference import get_parser
from common.helpers import add_ctc_blank
from jasper import config
from common.dataset import (AudioDataset, FilelistDataset, get_data_loader,
SingleAudioDataset)
from common.features import FilterbankFeatures
parser = get_parser()
parser.add_argument('--component',
type=str,
default="model",
choices=["feature-extractor", "model", "decoder"],
help='Component to convert')
args = parser.parse_args(model_args_list)
if args.component == "decoder":
return None
cfg = config.load(args.model_config)
if args.max_duration is not None:
cfg['input_val']['audio_dataset']['max_duration'] = args.max_duration
cfg['input_val']['filterbank_features'][
'max_duration'] = args.max_duration
if args.pad_to_max_duration:
assert cfg['input_train']['audio_dataset']['max_duration'] > 0
cfg['input_train']['audio_dataset']['pad_to_max_duration'] = True
symbols = add_ctc_blank(cfg['labels'])
dataset_kw, features_kw = config.input(cfg, 'val')
dataset = AudioDataset(args.dataset_dir, args.val_manifests, symbols,
**dataset_kw)
data_loader = get_data_loader(dataset,
args.batch_size,
multi_gpu=False,
shuffle=False,
num_workers=4,
drop_last=False)
feature_proc = None
if args.component == "model":
feature_proc = FilterbankFeatures(**features_kw)
data_loader.collate_fn = FeatureCollate(feature_proc)
return data_loader
def main():
parser = get_parser()
args = parser.parse_args()
log_fpath = args.log_file or str(Path(args.output_dir, 'nvlog_infer.json'))
log_fpath = unique_log_fpath(log_fpath)
dllogger.init(backends=[JSONStreamBackend(Verbosity.DEFAULT, log_fpath),
StdOutBackend(Verbosity.VERBOSE,
metric_format=stdout_metric_format)])
[dllogger.log("PARAMETER", {k: v}) for k, v in vars(args).items()]
for step in ['DNN', 'data+DNN', 'data']:
for c in [0.99, 0.95, 0.9, 0.5]:
cs = 'avg' if c == 0.5 else f'{int(100*c)}%'
dllogger.metadata(f'{step.lower()}_latency_{c}',
{'name': f'{step} latency {cs}',
'format': ':>7.2f', 'unit': 'ms'})
dllogger.metadata(
'eval_wer', {'name': 'WER', 'format': ':>3.2f', 'unit': '%'})
if args.cpu:
device = torch.device('cpu')
else:
assert torch.cuda.is_available()
device = torch.device('cuda')
torch.backends.cudnn.benchmark = args.cudnn_benchmark
if args.seed is not None:
torch.manual_seed(args.seed + args.local_rank)
np.random.seed(args.seed + args.local_rank)
random.seed(args.seed + args.local_rank)
# set up distributed training
multi_gpu = not args.cpu and int(os.environ.get('WORLD_SIZE', 1)) > 1
if multi_gpu:
torch.cuda.set_device(args.local_rank)
distrib.init_process_group(backend='nccl', init_method='env://')
print_once(f'Inference with {distrib.get_world_size()} GPUs')
cfg = config.load(args.model_config)
config.apply_config_overrides(cfg, args)
symbols = helpers.add_ctc_blank(cfg['labels'])
use_dali = args.dali_device in ('cpu', 'gpu')
dataset_kw, features_kw = config.input(cfg, 'val')
measure_perf = args.steps > 0
# dataset
if args.transcribe_wav or args.transcribe_filelist:
if use_dali:
print("DALI supported only with input .json files; disabling")
use_dali = False
assert not args.pad_to_max_duration
assert not (args.transcribe_wav and args.transcribe_filelist)
if args.transcribe_wav:
dataset = SingleAudioDataset(args.transcribe_wav)
else:
dataset = FilelistDataset(args.transcribe_filelist)
data_loader = get_data_loader(dataset,
batch_size=1,
multi_gpu=multi_gpu,
shuffle=False,
num_workers=0,
drop_last=(True if measure_perf else False))
_, features_kw = config.input(cfg, 'val')
feat_proc = FilterbankFeatures(**features_kw)
elif use_dali:
# pad_to_max_duration is not supported by DALI - have simple padders
if features_kw['pad_to_max_duration']:
feat_proc = BaseFeatures(
pad_align=features_kw['pad_align'],
pad_to_max_duration=True,
max_duration=features_kw['max_duration'],
sample_rate=features_kw['sample_rate'],
window_size=features_kw['window_size'],
window_stride=features_kw['window_stride'])
features_kw['pad_to_max_duration'] = False
else:
feat_proc = None
data_loader = DaliDataLoader(
gpu_id=args.local_rank or 0,
dataset_path=args.dataset_dir,
config_data=dataset_kw,
config_features=features_kw,
json_names=args.val_manifests,
batch_size=args.batch_size,
pipeline_type=("train" if measure_perf else "val"), # no drop_last
device_type=args.dali_device,
symbols=symbols)
else:
dataset = AudioDataset(args.dataset_dir,
args.val_manifests,
symbols,
**dataset_kw)
data_loader = get_data_loader(dataset,
args.batch_size,
multi_gpu=multi_gpu,
shuffle=False,
num_workers=4,
drop_last=False)
feat_proc = FilterbankFeatures(**features_kw)
model = QuartzNet(encoder_kw=config.encoder(cfg),
decoder_kw=config.decoder(cfg, n_classes=len(symbols)))
if args.ckpt is not None:
print(f'Loading the model from {args.ckpt} ...')
checkpoint = torch.load(args.ckpt, map_location="cpu")
key = 'ema_state_dict' if args.ema else 'state_dict'
state_dict = checkpoint[key]
model.load_state_dict(state_dict, strict=True)
model.to(device)
model.eval()
if feat_proc is not None:
feat_proc.to(device)
feat_proc.eval()
if args.amp:
model = model.half()
if args.torchscript:
greedy_decoder = GreedyCTCDecoder()
feat_proc, model, greedy_decoder = torchscript_export(
data_loader, feat_proc, model, greedy_decoder, args.output_dir,
use_amp=args.amp, use_conv_masks=True, model_toml=args.model_toml,
device=device, save=args.torchscript_export)
if multi_gpu:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
agg = {'txts': [], 'preds': [], 'logits': []}
dur = {'data': [], 'dnn': [], 'data+dnn': []}
looped_loader = chain.from_iterable(repeat(data_loader))
greedy_decoder = GreedyCTCDecoder()
sync = lambda: torch.cuda.synchronize() if device.type == 'cuda' else None
steps = args.steps + args.warmup_steps or len(data_loader)
with torch.no_grad():
for it, batch in enumerate(tqdm(looped_loader, initial=1, total=steps)):
if use_dali:
feats, feat_lens, txt, txt_lens = batch
if feat_proc is not None:
feats, feat_lens = feat_proc(feats, feat_lens)
else:
batch = [t.to(device, non_blocking=True) for t in batch]
audio, audio_lens, txt, txt_lens = batch
feats, feat_lens = feat_proc(audio, audio_lens)
sync()
t1 = time.perf_counter()
if args.amp:
feats = feats.half()
if model.encoder.use_conv_masks:
log_probs, log_prob_lens = model(feats, feat_lens)
else:
log_probs = model(feats, feat_lens)
preds = greedy_decoder(log_probs)
sync()
t2 = time.perf_counter()
# burn-in period; wait for a new loader due to num_workers
if it >= 1 and (args.steps == 0 or it >= args.warmup_steps):
dur['data'].append(t1 - t0)
dur['dnn'].append(t2 - t1)
dur['data+dnn'].append(t2 - t0)
if txt is not None:
agg['txts'] += helpers.gather_transcripts([txt], [txt_lens],
symbols)
agg['preds'] += helpers.gather_predictions([preds], symbols)
agg['logits'].append(log_probs)
if it + 1 == steps:
break
sync()
t0 = time.perf_counter()
# communicate the results
if args.transcribe_wav:
for idx, p in enumerate(agg['preds']):
print_once(f'Prediction {idx+1: >3}: {p}')
elif args.transcribe_filelist:
pass
elif not multi_gpu or distrib.get_rank() == 0:
wer, _ = process_evaluation_epoch(agg)
dllogger.log(step=(), data={'eval_wer': 100 * wer})
if args.save_predictions:
with open(args.save_predictions, 'w') as f:
f.write('\n'.join(agg['preds']))
if args.save_logits:
logits = torch.cat(agg['logits'], dim=0).cpu()
torch.save(logits, args.save_logits)
# report timings
if len(dur['data']) >= 20:
ratios = [0.9, 0.95, 0.99]
for stage in dur:
lat = durs_to_percentiles(dur[stage], ratios)
for k in [0.99, 0.95, 0.9, 0.5]:
kk = str(k).replace('.', '_')
dllogger.log(step=(), data={f'{stage.lower()}_latency_{kk}': lat[k]})
else:
print_once('Not enough samples to measure latencies.')
def main():
args = parse_args()
assert (torch.cuda.is_available())
assert args.prediction_frequency % args.log_frequency == 0
torch.backends.cudnn.benchmark = args.cudnn_benchmark
# set up distributed training
multi_gpu = int(os.environ.get('WORLD_SIZE', 1)) > 1
if multi_gpu:
torch.cuda.set_device(args.local_rank)
dist.init_process_group(backend='nccl', init_method='env://')
world_size = dist.get_world_size()
print_once(f'Distributed training with {world_size} GPUs\n')
else:
world_size = 1
torch.manual_seed(args.seed + args.local_rank)
np.random.seed(args.seed + args.local_rank)
random.seed(args.seed + args.local_rank)
init_log(args)
cfg = config.load(args.model_config)
config.apply_config_overrides(cfg, args)
symbols = helpers.add_ctc_blank(cfg['labels'])
assert args.grad_accumulation_steps >= 1
assert args.batch_size % args.grad_accumulation_steps == 0
batch_size = args.batch_size // args.grad_accumulation_steps
print_once('Setting up datasets...')
train_dataset_kw, train_features_kw = config.input(cfg, 'train')
val_dataset_kw, val_features_kw = config.input(cfg, 'val')
use_dali = args.dali_device in ('cpu', 'gpu')
if use_dali:
assert train_dataset_kw['ignore_offline_speed_perturbation'], \
"DALI doesn't support offline speed perturbation"
# pad_to_max_duration is not supported by DALI - have simple padders
if train_features_kw['pad_to_max_duration']:
train_feat_proc = BaseFeatures(
pad_align=train_features_kw['pad_align'],
pad_to_max_duration=True,
max_duration=train_features_kw['max_duration'],
sample_rate=train_features_kw['sample_rate'],
window_size=train_features_kw['window_size'],
window_stride=train_features_kw['window_stride'])
train_features_kw['pad_to_max_duration'] = False
else:
train_feat_proc = None
if val_features_kw['pad_to_max_duration']:
val_feat_proc = BaseFeatures(
pad_align=val_features_kw['pad_align'],
pad_to_max_duration=True,
max_duration=val_features_kw['max_duration'],
sample_rate=val_features_kw['sample_rate'],
window_size=val_features_kw['window_size'],
window_stride=val_features_kw['window_stride'])
val_features_kw['pad_to_max_duration'] = False
else:
val_feat_proc = None
train_loader = DaliDataLoader(
gpu_id=args.local_rank,
dataset_path=args.dataset_dir,
config_data=train_dataset_kw,
config_features=train_features_kw,
json_names=args.train_manifests,
batch_size=batch_size,
grad_accumulation_steps=args.grad_accumulation_steps,
pipeline_type="train",
device_type=args.dali_device,
symbols=symbols)
val_loader = DaliDataLoader(gpu_id=args.local_rank,
dataset_path=args.dataset_dir,
config_data=val_dataset_kw,
config_features=val_features_kw,
json_names=args.val_manifests,
batch_size=batch_size,
pipeline_type="val",
device_type=args.dali_device,
symbols=symbols)
else:
train_dataset_kw, train_features_kw = config.input(cfg, 'train')
train_dataset = AudioDataset(args.dataset_dir, args.train_manifests,
symbols, **train_dataset_kw)
train_loader = get_data_loader(train_dataset,
batch_size,
multi_gpu=multi_gpu,
shuffle=True,
num_workers=4)
train_feat_proc = FilterbankFeatures(**train_features_kw)
val_dataset_kw, val_features_kw = config.input(cfg, 'val')
val_dataset = AudioDataset(args.dataset_dir, args.val_manifests,
symbols, **val_dataset_kw)
val_loader = get_data_loader(val_dataset,
batch_size,
multi_gpu=multi_gpu,
shuffle=False,
num_workers=4,
drop_last=False)
val_feat_proc = FilterbankFeatures(**val_features_kw)
dur = train_dataset.duration / 3600
dur_f = train_dataset.duration_filtered / 3600
nsampl = len(train_dataset)
print_once(f'Training samples: {nsampl} ({dur:.1f}h, '
f'filtered {dur_f:.1f}h)')
if train_feat_proc is not None:
train_feat_proc.cuda()
if val_feat_proc is not None:
val_feat_proc.cuda()
steps_per_epoch = len(train_loader) // args.grad_accumulation_steps
# set up the model
model = Jasper(encoder_kw=config.encoder(cfg),
decoder_kw=config.decoder(cfg, n_classes=len(symbols)))
model.cuda()
ctc_loss = CTCLossNM(n_classes=len(symbols))
greedy_decoder = GreedyCTCDecoder()
print_once(f'Model size: {num_weights(model) / 10**6:.1f}M params\n')
# optimization
kw = {'lr': args.lr, 'weight_decay': args.weight_decay}
if args.optimizer == "novograd":
optimizer = Novograd(model.parameters(), **kw)
elif args.optimizer == "adamw":
optimizer = AdamW(model.parameters(), **kw)
else:
raise ValueError(f'Invalid optimizer "{args.optimizer}"')
scaler = torch.cuda.amp.GradScaler(enabled=args.amp)
adjust_lr = lambda step, epoch, optimizer: lr_policy(
step,
epoch,
args.lr,
optimizer,
steps_per_epoch=steps_per_epoch,
warmup_epochs=args.warmup_epochs,
hold_epochs=args.hold_epochs,
num_epochs=args.epochs,
policy=args.lr_policy,
min_lr=args.min_lr,
exp_gamma=args.lr_exp_gamma)
if args.ema > 0:
ema_model = copy.deepcopy(model)
else:
ema_model = None
if multi_gpu:
model = torch.nn.parallel.DistributedDataParallel(
model, device_ids=[args.local_rank], output_device=args.local_rank)
if args.pyprof:
pyprof.init(enable_function_stack=True)
# load checkpoint
meta = {'best_wer': 10**6, 'start_epoch': 0}
checkpointer = Checkpointer(args.output_dir, 'Jasper',
args.keep_milestones)
if args.resume:
args.ckpt = checkpointer.last_checkpoint() or args.ckpt
if args.ckpt is not None:
checkpointer.load(args.ckpt, model, ema_model, optimizer, scaler, meta)
start_epoch = meta['start_epoch']
best_wer = meta['best_wer']
epoch = 1
step = start_epoch * steps_per_epoch + 1
if args.pyprof:
torch.autograd.profiler.emit_nvtx().__enter__()
profiler.start()
# training loop
model.train()
# pre-allocate
if args.pre_allocate_range is not None:
n_feats = train_features_kw['n_filt']
pad_align = train_features_kw['pad_align']
a, b = args.pre_allocate_range
for n_frames in range(a, b + pad_align, pad_align):
print_once(
f'Pre-allocation ({batch_size}x{n_feats}x{n_frames})...')
feat = torch.randn(batch_size, n_feats, n_frames, device='cuda')
feat_lens = torch.ones(batch_size, device='cuda').fill_(n_frames)
txt = torch.randint(high=len(symbols) - 1,
size=(batch_size, 100),
device='cuda')
txt_lens = torch.ones(batch_size, device='cuda').fill_(100)
with torch.cuda.amp.autocast(enabled=args.amp):
log_probs, enc_lens = model(feat, feat_lens)
del feat
loss = ctc_loss(log_probs, txt, enc_lens, txt_lens)
loss.backward()
model.zero_grad()
torch.cuda.empty_cache()
bmark_stats = BenchmarkStats()
for epoch in range(start_epoch + 1, args.epochs + 1):
if multi_gpu and not use_dali:
train_loader.sampler.set_epoch(epoch)
epoch_utts = 0
epoch_loss = 0
accumulated_batches = 0
epoch_start_time = time.time()
epoch_eval_time = 0
for batch in train_loader:
if accumulated_batches == 0:
step_loss = 0
step_utts = 0
step_start_time = time.time()
if use_dali:
# with DALI, the data is already on GPU
feat, feat_lens, txt, txt_lens = batch
if train_feat_proc is not None:
feat, feat_lens = train_feat_proc(feat, feat_lens)
else:
batch = [t.cuda(non_blocking=True) for t in batch]
audio, audio_lens, txt, txt_lens = batch
feat, feat_lens = train_feat_proc(audio, audio_lens)
# Use context manager to prevent redundant accumulation of gradients
if (multi_gpu and
accumulated_batches + 1 < args.grad_accumulation_steps):
ctx = model.no_sync()
else:
ctx = empty_context()
with ctx:
with torch.cuda.amp.autocast(enabled=args.amp):
log_probs, enc_lens = model(feat, feat_lens)
loss = ctc_loss(log_probs, txt, enc_lens, txt_lens)
loss /= args.grad_accumulation_steps
if multi_gpu:
reduced_loss = reduce_tensor(loss.data, world_size)
else:
reduced_loss = loss
if torch.isnan(reduced_loss).any():
print_once(f'WARNING: loss is NaN; skipping update')
continue
else:
step_loss += reduced_loss.item()
step_utts += batch[0].size(0) * world_size
epoch_utts += batch[0].size(0) * world_size
accumulated_batches += 1
scaler.scale(loss).backward()
if accumulated_batches % args.grad_accumulation_steps == 0:
epoch_loss += step_loss
scaler.step(optimizer)
scaler.update()
adjust_lr(step, epoch, optimizer)
optimizer.zero_grad()
apply_ema(model, ema_model, args.ema)
if step % args.log_frequency == 0:
preds = greedy_decoder(log_probs)
wer, pred_utt, ref = greedy_wer(preds, txt, txt_lens,
symbols)
if step % args.prediction_frequency == 0:
print_once(f' Decoded: {pred_utt[:90]}')
print_once(f' Reference: {ref[:90]}')
step_time = time.time() - step_start_time
log(
(epoch, step % steps_per_epoch
or steps_per_epoch, steps_per_epoch), step, 'train', {
'loss': step_loss,
'wer': 100.0 * wer,
'throughput': step_utts / step_time,
'took': step_time,
'lrate': optimizer.param_groups[0]['lr']
})
step_start_time = time.time()
if step % args.eval_frequency == 0:
tik = time.time()
wer = evaluate(epoch, step, val_loader, val_feat_proc,
symbols, model, ema_model, ctc_loss,
greedy_decoder, args.amp, use_dali)
if wer < best_wer and epoch >= args.save_best_from:
checkpointer.save(model,
ema_model,
optimizer,
scaler,
epoch,
step,
best_wer,
is_best=True)
best_wer = wer
epoch_eval_time += time.time() - tik
step += 1
accumulated_batches = 0
# end of step
# DALI iterator need to be exhausted;
# if not using DALI, simulate drop_last=True with grad accumulation
if not use_dali and step > steps_per_epoch * epoch:
break
epoch_time = time.time() - epoch_start_time
epoch_loss /= steps_per_epoch
log(
(epoch, ), None, 'train_avg', {
'throughput': epoch_utts / epoch_time,
'took': epoch_time,
'loss': epoch_loss
})
bmark_stats.update(epoch_utts, epoch_time, epoch_loss)
if epoch % args.save_frequency == 0 or epoch in args.keep_milestones:
checkpointer.save(model, ema_model, optimizer, scaler, epoch, step,
best_wer)
if 0 < args.epochs_this_job <= epoch - start_epoch:
print_once(f'Finished after {args.epochs_this_job} epochs.')
break
# end of epoch
if args.pyprof:
profiler.stop()
torch.autograd.profiler.emit_nvtx().__exit__(None, None, None)
log((), None, 'train_avg', bmark_stats.get(args.benchmark_epochs_num))
if epoch == args.epochs:
evaluate(epoch, step, val_loader, val_feat_proc, symbols, model,
ema_model, ctc_loss, greedy_decoder, args.amp, use_dali)
checkpointer.save(model, ema_model, optimizer, scaler, epoch, step,
best_wer)
flush_log()