def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.benchmark = args.cudnn_benchmark
print("CUDNN BENCHMARK ", args.cudnn_benchmark)
if not args.cpu_run:
assert(torch.cuda.is_available())
if args.local_rank is not None:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
multi_gpu = args.local_rank is not None
if multi_gpu:
print("DISTRIBUTED with ", torch.distributed.get_world_size())
if args.fp16:
optim_level = 3
else:
optim_level = 0
jasper_model_definition = toml.load(args.model_toml)
dataset_vocab = jasper_model_definition['labels']['labels']
ctc_vocab = add_ctc_labels(dataset_vocab)
val_manifest = args.val_manifest
featurizer_config = jasper_model_definition['input_eval']
featurizer_config["optimization_level"] = optim_level
featurizer_config["fp16"] = args.fp16
args.use_conv_mask = jasper_model_definition['encoder'].get('convmask', True)
if args.masked_fill is not None:
print("{} masked_fill".format("Enabling" if args.masked_fill else "Disabling"))
jasper_model_definition["encoder"]["conv_mask"] = args.masked_fill
if args.max_duration is not None:
featurizer_config['max_duration'] = args.max_duration
if args.pad_to is not None:
featurizer_config['pad_to'] = args.pad_to
if featurizer_config['pad_to'] == "max":
featurizer_config['pad_to'] = -1
print('=== model_config ===')
print_dict(jasper_model_definition)
print()
print('=== feature_config ===')
print_dict(featurizer_config)
print()
data_layer = None
if args.wav is None:
data_layer = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=args.batch_size,
pad_to_max=featurizer_config['pad_to'] == -1,
shuffle=False,
multi_gpu=multi_gpu)
audio_preprocessor = AudioPreprocessing(**featurizer_config)
encoderdecoder = JasperEncoderDecoder(jasper_model_definition=jasper_model_definition, feat_in=1024, num_classes=len(ctc_vocab))
if args.ckpt is not None:
print("loading model from ", args.ckpt)
if os.path.isdir(args.ckpt):
exit(0)
else:
checkpoint = torch.load(args.ckpt, map_location="cpu")
for k in audio_preprocessor.state_dict().keys():
checkpoint['state_dict'][k] = checkpoint['state_dict'].pop("audio_preprocessor." + k)
audio_preprocessor.load_state_dict(checkpoint['state_dict'], strict=False)
encoderdecoder.load_state_dict(checkpoint['state_dict'], strict=False)
greedy_decoder = GreedyCTCDecoder()
# print("Number of parameters in encoder: {0}".format(model.jasper_encoder.num_weights()))
if args.wav is None:
N = len(data_layer)
step_per_epoch = math.ceil(N / (args.batch_size * (1 if not torch.distributed.is_initialized() else torch.distributed.get_world_size())))
if args.steps is not None:
print('-----------------')
print('Have {0} examples to eval on.'.format(args.steps * args.batch_size * (1 if not torch.distributed.is_initialized() else torch.distributed.get_world_size())))
print('Have {0} steps / (gpu * epoch).'.format(args.steps))
print('-----------------')
else:
print('-----------------')
print('Have {0} examples to eval on.'.format(N))
print('Have {0} steps / (gpu * epoch).'.format(step_per_epoch))
print('-----------------')
print ("audio_preprocessor.normalize: ", audio_preprocessor.featurizer.normalize)
if not args.cpu_run:
audio_preprocessor.cuda()
encoderdecoder.cuda()
if args.fp16:
encoderdecoder = amp.initialize( models=encoderdecoder,
opt_level=AmpOptimizations[optim_level])
encoderdecoder = model_multi_gpu(encoderdecoder, multi_gpu)
audio_preprocessor.eval()
encoderdecoder.eval()
greedy_decoder.eval()
eval(
data_layer=data_layer,
audio_processor=audio_preprocessor,
encoderdecoder=encoderdecoder,
greedy_decoder=greedy_decoder,
labels=ctc_vocab,
args=args,
multi_gpu=multi_gpu)
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
assert(torch.cuda.is_available())
torch.backends.cudnn.benchmark = args.cudnn
# set up distributed training
if args.local_rank is not None:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
multi_gpu = torch.distributed.is_initialized()
if multi_gpu:
print_once("DISTRIBUTED TRAINING with {} gpus".format(torch.distributed.get_world_size()))
# define amp optimiation level
if args.fp16:
optim_level = Optimization.mxprO1
else:
optim_level = Optimization.mxprO0
jasper_model_definition = toml.load(args.model_toml)
dataset_vocab = jasper_model_definition['labels']['labels']
ctc_vocab = add_ctc_labels(dataset_vocab)
train_manifest = args.train_manifest
val_manifest = args.val_manifest
featurizer_config = jasper_model_definition['input']
featurizer_config_eval = jasper_model_definition['input_eval']
featurizer_config["optimization_level"] = optim_level
featurizer_config_eval["optimization_level"] = optim_level
sampler_type = featurizer_config.get("sampler", 'default')
perturb_config = jasper_model_definition.get('perturb', None)
if args.pad_to_max:
assert(args.max_duration > 0)
featurizer_config['max_duration'] = args.max_duration
featurizer_config_eval['max_duration'] = args.max_duration
featurizer_config['pad_to'] = "max"
featurizer_config_eval['pad_to'] = "max"
print_once('model_config')
print_dict(jasper_model_definition)
if args.gradient_accumulation_steps < 1:
raise ValueError('Invalid gradient accumulation steps parameter {}'.format(args.gradient_accumulation_steps))
if args.batch_size % args.gradient_accumulation_steps != 0:
raise ValueError('gradient accumulation step {} is not divisible by batch size {}'.format(args.gradient_accumulation_steps, args.batch_size))
data_layer = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
perturb_config=perturb_config,
manifest_filepath=train_manifest,
labels=dataset_vocab,
batch_size=args.batch_size // args.gradient_accumulation_steps,
multi_gpu=multi_gpu,
pad_to_max=args.pad_to_max,
sampler=sampler_type)
data_layer_eval = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config_eval,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=args.batch_size,
multi_gpu=multi_gpu,
pad_to_max=args.pad_to_max
)
model = Jasper(feature_config=featurizer_config, jasper_model_definition=jasper_model_definition, feat_in=1024, num_classes=len(ctc_vocab))
if args.ckpt is not None:
print_once("loading model from {}".format(args.ckpt))
checkpoint = torch.load(args.ckpt, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'], strict=True)
args.start_epoch = checkpoint['epoch']
else:
args.start_epoch = 0
ctc_loss = CTCLossNM( num_classes=len(ctc_vocab))
greedy_decoder = GreedyCTCDecoder()
print_once("Number of parameters in encoder: {0}".format(model.jasper_encoder.num_weights()))
print_once("Number of parameters in decode: {0}".format(model.jasper_decoder.num_weights()))
N = len(data_layer)
if sampler_type == 'default':
args.step_per_epoch = math.ceil(N / (args.batch_size * (1 if not torch.distributed.is_initialized() else torch.distributed.get_world_size())))
elif sampler_type == 'bucket':
args.step_per_epoch = int(len(data_layer.sampler) / args.batch_size )
print_once('-----------------')
print_once('Have {0} examples to train on.'.format(N))
print_once('Have {0} steps / (gpu * epoch).'.format(args.step_per_epoch))
print_once('-----------------')
fn_lr_policy = lambda s: lr_policy(args.lr, s, args.num_epochs * args.step_per_epoch)
model.cuda()
if args.optimizer_kind == "novograd":
optimizer = Novograd(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.optimizer_kind == "adam":
optimizer = AdamW(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
else:
raise ValueError("invalid optimizer choice: {}".format(args.optimizer_kind))
if optim_level in AmpOptimizations:
model, optimizer = amp.initialize(
#lnw block for error
#min_loss_scale=1.0,
models=model,
optimizers=optimizer,
opt_level=AmpOptimizations[optim_level])
if args.ckpt is not None:
optimizer.load_state_dict(checkpoint['optimizer'])
model = model_multi_gpu(model, multi_gpu)
train(
data_layer=data_layer,
data_layer_eval=data_layer_eval,
model=model,
ctc_loss=ctc_loss,
greedy_decoder=greedy_decoder,
optimizer=optimizer,
labels=ctc_vocab,
optim_level=optim_level,
multi_gpu=multi_gpu,
fn_lr_policy=fn_lr_policy if args.lr_decay else None,
args=args)
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.benchmark = args.cudnn_benchmark
print("CUDNN BENCHMARK ", args.cudnn_benchmark)
assert(torch.cuda.is_available())
if args.local_rank is not None:
torch.cuda.set_device(args.local_rank)
torch.distributed.init_process_group(backend='nccl', init_method='env://')
multi_gpu = args.local_rank is not None
if multi_gpu:
print("DISTRIBUTED with ", torch.distributed.get_world_size())
if args.fp16:
optim_level = Optimization.mxprO3
else:
optim_level = Optimization.mxprO0
model_definition = toml.load(args.model_toml)
dataset_vocab = model_definition['labels']['labels']
ctc_vocab = add_blank_label(dataset_vocab)
val_manifest = args.val_manifest
featurizer_config = model_definition['input_eval']
featurizer_config["optimization_level"] = optim_level
if args.max_duration is not None:
featurizer_config['max_duration'] = args.max_duration
if args.pad_to is not None:
featurizer_config['pad_to'] = args.pad_to if args.pad_to >= 0 else "max"
print('model_config')
print_dict(model_definition)
print('feature_config')
print_dict(featurizer_config)
data_layer = None
if args.wav is None:
data_layer = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=args.batch_size,
pad_to_max=featurizer_config['pad_to'] == "max",
shuffle=False,
multi_gpu=multi_gpu)
audio_preprocessor = AudioPreprocessing(**featurizer_config)
#encoderdecoder = JasperEncoderDecoder(jasper_model_definition=jasper_model_definition, feat_in=1024, num_classes=len(ctc_vocab))
model = RNNT(
feature_config=featurizer_config,
rnnt=model_definition['rnnt'],
num_classes=len(ctc_vocab)
)
if args.ckpt is not None:
print("loading model from ", args.ckpt)
checkpoint = torch.load(args.ckpt, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'], strict=False)
#greedy_decoder = GreedyCTCDecoder()
# print("Number of parameters in encoder: {0}".format(model.jasper_encoder.num_weights()))
if args.wav is None:
N = len(data_layer)
step_per_epoch = math.ceil(N / (args.batch_size * (1 if not torch.distributed.is_initialized() else torch.distributed.get_world_size())))
if args.steps is not None:
print('-----------------')
print('Have {0} examples to eval on.'.format(args.steps * args.batch_size * (1 if not torch.distributed.is_initialized() else torch.distributed.get_world_size())))
print('Have {0} steps / (gpu * epoch).'.format(args.steps))
print('-----------------')
else:
print('-----------------')
print('Have {0} examples to eval on.'.format(N))
print('Have {0} steps / (gpu * epoch).'.format(step_per_epoch))
print('-----------------')
else:
audio_preprocessor.featurizer.normalize = "per_feature"
print ("audio_preprocessor.normalize: ", audio_preprocessor.featurizer.normalize)
audio_preprocessor.cuda()
audio_preprocessor.eval()
eval_transforms = torchvision.transforms.Compose([
lambda xs: [x.cuda() for x in xs],
lambda xs: [*audio_preprocessor(xs[0:2]), *xs[2:]],
lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
])
model.cuda()
if args.fp16:
model = amp.initialize(
models=model,
opt_level=AmpOptimizations[optim_level])
model = model_multi_gpu(model, multi_gpu)
greedy_decoder = RNNTGreedyDecoder(len(ctc_vocab) - 1, model.module if multi_gpu else model)
eval(
data_layer=data_layer,
audio_processor=eval_transforms,
encoderdecoder=model,
greedy_decoder=greedy_decoder,
labels=ctc_vocab,
args=args,
multi_gpu=multi_gpu)
