def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
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"
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=False)
audio_preprocessor = AudioPreprocessing(**featurizer_config)
audio_preprocessor.eval()
eval_transforms = torchvision.transforms.Compose([
lambda xs: [*audio_preprocessor(xs[0:2]), *xs[2:]],
lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
])
eval(
data_layer=data_layer,
audio_processor=eval_transforms,
args=args)
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
#torch.set_default_dtype(torch.double)
torch.manual_seed(args.seed)
torch.backends.cudnn.benchmark = args.cudnn_benchmark
#print("CUDNN BENCHMARK ", args.cudnn_benchmark)
if args.cuda:
assert(torch.cuda.is_available())
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']
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
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,
sampler='bucket' #sort by duration
)
audio_preprocessor = AudioPreprocessing(**featurizer_config)
model = RNNT(
feature_config=featurizer_config,
rnnt=model_definition['rnnt'],
num_classes=len(ctc_vocab)
)
if args.ckpt is not None and args.mode in[3]:
#print("loading model from ", args.ckpt)
checkpoint = torch.load(args.ckpt, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'], strict=False)
audio_preprocessor.featurizer.normalize = "per_feature"
if args.cuda:
audio_preprocessor.cuda()
audio_preprocessor.eval()
eval_transforms = []
if args.cuda:
eval_transforms.append(lambda xs: [xs[0].cuda(),xs[1].cuda(), *xs[2:]])
eval_transforms.append(lambda xs: [*audio_preprocessor(xs[0:2]), *xs[2:]])
# These are just some very confusing transposes, that's all.
# BxFxT -> TxBxF
eval_transforms.append(lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]])
eval_transforms = torchvision.transforms.Compose(eval_transforms)
if args.cuda:
model.cuda()
# Ideally, I would jit this as well... But this is just the constructor...
greedy_decoder = RNNTGreedyDecoder(len(ctc_vocab) - 1, model)
eval(
data_layer=data_layer,
audio_processor=eval_transforms,
greedy_decoder=greedy_decoder,
labels=ctc_vocab,
args=args)
def get_pytorch_components_and_onnx(args):
'''Returns PyTorch components used for inference
'''
model_definition = toml.load(args.model_toml)
dataset_vocab = model_definition['labels']['labels']
# Set up global labels for future vocab calls
global _global_ctc_labels
_global_ctc_labels = add_ctc_labels(dataset_vocab)
featurizer_config = model_definition['input_eval']
optim_level = 3 if args.pyt_fp16 else 0
featurizer_config["optimization_level"] = optim_level
audio_preprocessor = None
onnx_path = None
data_layer = None
wav = None
seq_len = None
if args.max_duration is not None:
featurizer_config['max_duration'] = args.max_duration
if args.dataset_dir is not None:
data_layer = AudioToTextDataLayer(dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=args.val_manifest,
labels=dataset_vocab,
batch_size=args.batch_size,
shuffle=False)
if args.wav is not None:
args.batch_size = 1
wav, seq_len = audio_from_file(args.wav)
if args.seq_len is None or args.seq_len == 0:
args.seq_len = seq_len / (featurizer_config['sample_rate'] / 100)
if args.transpose:
featurizer_config["transpose_out"] = True
model_definition["transpose_in"] = True
model = JasperEncoderDecoder(jasper_model_definition=model_definition,
feat_in=1024,
num_classes=len(get_vocab()),
transpose_in=args.transpose)
model = model.cuda()
model.eval()
audio_preprocessor = AudioPreprocessing(**featurizer_config)
audio_preprocessor = audio_preprocessor.cuda()
audio_preprocessor.eval()
if args.ckpt_path is not None:
if os.path.isdir(args.ckpt_path):
d_checkpoint = torch.load(args.ckpt_path + "/decoder.pt",
map_location="cpu")
e_checkpoint = torch.load(args.ckpt_path + "/encoder.pt",
map_location="cpu")
model.jasper_encoder.load_state_dict(e_checkpoint, strict=False)
model.jasper_decoder.load_state_dict(d_checkpoint, strict=False)
else:
checkpoint = torch.load(args.ckpt_path, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'], strict=False)
# if we are to produce engine, not run/create ONNX, postpone AMP initialization
# (ONNX parser cannot handle mixed FP16 ONNX yet)
if args.pyt_fp16 and args.engine_path is None:
amp.initialize(models=model, opt_level=AmpOptimizations[optim_level])
if args.make_onnx:
if args.onnx_path is None or args.ckpt_path is None:
raise Exception(
"--ckpt_path, --onnx_path must be provided when using --make_onnx"
)
onnx_path = get_onnx(args.onnx_path, model, args)
if args.pyt_fp16 and args.engine_path is not None:
amp.initialize(models=model, opt_level=AmpOptimizations[optim_level])
return {
'data_layer': data_layer,
'audio_preprocessor': audio_preprocessor,
'acoustic_model': model,
'input_wav': (wav, seq_len)
}, onnx_path
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.benchmark = args.cudnn_benchmark
assert (args.steps is None or args.steps > 5)
print("CUDNN BENCHMARK ", args.cudnn_benchmark)
assert (torch.cuda.is_available())
if args.fp16:
optim_level = Optimization.mxprO3
else:
optim_level = Optimization.mxprO0
batch_size = args.batch_size
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
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(jasper_model_definition)
print('feature_config')
print_dict(featurizer_config)
data_layer = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=batch_size,
pad_to_max=featurizer_config['pad_to'] == "max",
shuffle=False,
multi_gpu=False)
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)
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()))
N = len(data_layer)
step_per_epoch = math.ceil(N / args.batch_size)
print('-----------------')
if args.steps is None:
print('Have {0} examples to eval on.'.format(N))
print('Have {0} steps / (gpu * epoch).'.format(step_per_epoch))
else:
print('Have {0} examples to eval on.'.format(args.steps *
args.batch_size))
print('Have {0} steps / (gpu * epoch).'.format(args.steps))
print('-----------------')
audio_preprocessor.cuda()
encoderdecoder.cuda()
if args.fp16:
encoderdecoder = amp.initialize(
models=encoderdecoder, opt_level=AmpOptimizations[optim_level])
eval(data_layer=data_layer,
audio_processor=audio_preprocessor,
encoderdecoder=encoderdecoder,
greedy_decoder=greedy_decoder,
labels=ctc_vocab,
args=args)
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
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,
# sampler='bucket',
sort_by_duration=args.sort_by_duration,
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)
if args.ipex:
import intel_extension_for_pytorch as ipex
from rnn import IPEXStackTime
model.joint_net.eval()
data_type = torch.bfloat16 if args.mix_precision else torch.float32
if model.encoder["stack_time"].factor == 2:
model.encoder["stack_time"] = IPEXStackTime(model.encoder["stack_time"].factor)
model.joint_net = ipex.optimize(model.joint_net, dtype=data_type, auto_kernel_selection=True)
model.prediction["embed"] = model.prediction["embed"].to(data_type)
if args.jit:
print("running jit path")
model.joint_net.eval()
if args.mix_precision:
with torch.cpu.amp.autocast(), torch.no_grad():
model.joint_net = torch.jit.trace(model.joint_net, torch.randn(args.batch_size, 1, 1, model_definition['rnnt']['encoder_n_hidden'] + model_definition['rnnt']['pred_n_hidden']), check_trace=False)
else:
with torch.no_grad():
model.joint_net = torch.jit.trace(model.joint_net, torch.randn(args.batch_size, 1, 1, model_definition['rnnt']['encoder_n_hidden'] + model_definition['rnnt']['pred_n_hidden']), check_trace=False)
model.joint_net = torch.jit.freeze(model.joint_net)
else:
model = model.to("cpu")
#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_available() else torch.distributed.get_world_size())))
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_available() else torch.distributed.get_world_size())))
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} warm up steps / (gpu * epoch).'.format(args.warm_up))
print('Have {0} measure steps / (gpu * epoch).'.format(args.steps))
print('-----------------')
else:
print('-----------------')
print('Have {0} examples to eval on.'.format(N))
print('Have {0} warm up steps / (gpu * epoch).'.format(args.warm_up))
print('Have {0} measure 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.eval()
# eval_transforms = torchvision.transforms.Compose([
# lambda xs: [x.to(ipex.DEVICE) if args.ipex else x.cpu() for x in xs],
# lambda xs: [*audio_preprocessor(xs[0:2]), *xs[2:]],
# lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
# ])
eval_transforms = torchvision.transforms.Compose([
lambda xs: [x.cpu() for x in xs],
lambda xs: [*audio_preprocessor(xs[0:2]), *xs[2:]],
lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
])
model.eval()
if args.ipex:
ipex.nn.utils._model_convert.replace_lstm_with_ipex_lstm(model)
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)
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 get_pytorch_components_and_onnx(args):
'''Returns PyTorch components used for inference
'''
model_definition = toml.load(args.model_toml)
dataset_vocab = model_definition['labels']['labels']
# Set up global labels for future vocab calls
global _global_ctc_labels
_global_ctc_labels = add_ctc_labels(dataset_vocab)
featurizer_config = model_definition['input_eval']
optim_level = Optimization.mxprO3 if args.pyt_fp16 else Optimization.mxprO0
featurizer_config["optimization_level"] = optim_level
acoustic_model = None
audio_preprocessor = None
onnx_path = None
data_layer = None
wav = None
seq_len = None
dtype = torch.float
if args.max_duration is not None:
featurizer_config['max_duration'] = args.max_duration
if args.dataset_dir is not None:
data_layer = AudioToTextDataLayer(dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=args.val_manifest,
labels=dataset_vocab,
batch_size=args.batch_size,
shuffle=False)
if args.wav is not None:
args.batch_size = 1
args.engine_batch_size = 1
wav, seq_len = audio_from_file(args.wav)
if args.seq_len is None or args.seq_len == 0:
args.seq_len = seq_len / (featurizer_config['sample_rate'] / 100)
model = Jasper(feature_config=featurizer_config,
jasper_model_definition=model_definition,
feat_in=1024,
num_classes=len(get_vocab()))
model.cuda()
model.eval()
acoustic_model = model.acoustic_model
audio_preprocessor = model.audio_preprocessor
if args.ckpt_path is not None:
checkpoint = torch.load(args.ckpt_path, map_location="cpu")
model.load_state_dict(checkpoint['state_dict'], strict=False)
if args.make_onnx:
if args.onnx_path is None or acoustic_model is None:
raise Exception(
"--ckpt_path, --onnx_path must be provided when using --make_onnx"
)
onnx_path = get_onnx(args.onnx_path,
acoustic_model,
signal_shape=(args.engine_batch_size, 64,
args.seq_len),
dtype=torch.float)
if args.pyt_fp16:
amp.initialize(models=acoustic_model,
opt_level=AmpOptimizations[optim_level])
return {
'data_layer': data_layer,
'audio_preprocessor': audio_preprocessor,
'acoustic_model': acoustic_model,
'input_wav': (wav, seq_len)
}, onnx_path
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)
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
args.local_rank = os.environ.get('LOCAL_RANK', args.local_rank)
# set up distributed training
cpu_distributed_training = False
if torch.distributed.is_available() and int(os.environ.get('PMI_SIZE', '0')) > 1:
print('Distributed training with DDP')
os.environ['RANK'] = os.environ.get('PMI_RANK', '0')
os.environ['WORLD_SIZE'] = os.environ.get('PMI_SIZE', '1')
if not 'MASTER_ADDR' in os.environ:
os.environ['MASTER_ADDR'] = args.master_addr
if not 'MASTER_PORT' in os.environ:
os.environ['MASTER_PORT'] = args.port
# Initialize the process group with ccl backend
if args.backend == 'ccl':
import torch_ccl
dist.init_process_group(
backend=args.backend
)
cpu_distributed_training = True
if torch.distributed.is_initialized():
print("Torch distributed is initialized.")
args.rank = torch.distributed.get_rank()
args.world_size = torch.distributed.get_world_size()
else:
print("Torch distributed is not initialized.")
args.rank = 0
args.world_size = 1
multi_gpu = False
if multi_gpu:
print_once("DISTRIBUTED TRAINING with {} gpus".format(torch.distributed.get_world_size()))
optim_level = Optimization.mxprO0
model_definition = toml.load(args.model_toml)
dataset_vocab = model_definition['labels']['labels']
ctc_vocab = add_blank_label(dataset_vocab)
train_manifest = args.train_manifest
val_manifest = args.val_manifest
tst_manifest = args.tst_manifest
featurizer_config = model_definition['input']
featurizer_config_eval = 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 = 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(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))
preprocessor = preprocessing.AudioPreprocessing(**featurizer_config)
if args.cuda:
preprocessor.cuda()
else:
preprocessor.cpu()
augmentations = preprocessing.SpectrogramAugmentation(**featurizer_config)
if args.cuda:
augmentations.cuda()
else:
augmentations.cpu()
train_transforms = torchvision.transforms.Compose([
lambda xs: [x.cpu() for x in xs],
lambda xs: [*preprocessor(xs[0:2]), *xs[2:]],
lambda xs: [augmentations(xs[0]), *xs[1:]],
lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
])
eval_transforms = torchvision.transforms.Compose([
lambda xs: [x.cpu() for x in xs],
lambda xs: [*preprocessor(xs[0:2]), *xs[2:]],
lambda xs: [xs[0].permute(2, 0, 1), *xs[1:]],
])
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,
cpu_distributed_training=cpu_distributed_training)
eval_datasets = [(
AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config_eval,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=args.eval_batch_size,
multi_gpu=multi_gpu,
pad_to_max=args.pad_to_max
),
args.eval_frequency,
'Eval clean',
)]
if tst_manifest:
eval_datasets.append((
AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config_eval,
manifest_filepath=tst_manifest,
labels=dataset_vocab,
batch_size=args.eval_batch_size,
multi_gpu=multi_gpu,
pad_to_max=args.pad_to_max
),
args.test_frequency,
'Test other',
))
model = RNNT(
feature_config=featurizer_config,
rnnt=model_definition['rnnt'],
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
loss_fn = RNNTLoss(blank=len(ctc_vocab) - 1)
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('-----------------')
constant_lr_policy = lambda _: args.lr
fn_lr_policy = constant_lr_policy
if args.lr_decay:
pre_decay_policy = fn_lr_policy
fn_lr_policy = lambda s: lr_decay(args.num_epochs * args.step_per_epoch, s, pre_decay_policy(s))
if args.lr_warmup:
pre_warmup_policy = fn_lr_policy
fn_lr_policy = lambda s: lr_warmup(args.lr_warmup, s, pre_warmup_policy(s) )
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 args.cuda and optim_level in AmpOptimizations:
assert False, "not supported in ipex"
if args.ckpt is not None:
optimizer.load_state_dict(checkpoint['optimizer'])
if args.ipex:
if args.bf16:
model, optimizer = ipex.optimize(model, dtype=torch.bfloat16, optimizer=optimizer)
ipex.nn.utils._model_convert.replace_lstm_with_ipex_lstm(model)
else:
model, optimizer = ipex.optimize(model, dtype=torch.float32, optimizer=optimizer)
ipex.nn.utils._model_convert.replace_lstm_with_ipex_lstm(model)
if args.world_size > 1:
device_ids = None
model = torch.nn.parallel.DistributedDataParallel(model, device_ids=device_ids)
print_once(model)
print_once("# parameters: {}".format(sum(p.numel() for p in model.parameters())))
greedy_decoder = RNNTGreedyDecoder(len(ctc_vocab) - 1, model.module if multi_gpu else model)
if args.tb_path and args.local_rank == 0:
logger = TensorBoardLogger(args.tb_path, model.module if multi_gpu else model, args.histogram)
else:
logger = DummyLogger()
train(
data_layer=data_layer,
model=model,
loss_fn=loss_fn,
greedy_decoder=greedy_decoder,
optimizer=optimizer,
data_transforms=train_transforms,
labels=ctc_vocab,
optim_level=optim_level,
multi_gpu=multi_gpu,
fn_lr_policy=fn_lr_policy,
evalutaion=evaluator(model, eval_transforms, loss_fn, greedy_decoder, ctc_vocab, eval_datasets, logger),
logger=logger,
args=args)
def main(args):
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
assert (args.steps is None or args.steps > 5)
if args.cpu:
device = torch.device('cpu')
else:
assert (torch.cuda.is_available())
device = torch.device('cuda')
torch.backends.cudnn.benchmark = args.cudnn_benchmark
print("CUDNN BENCHMARK ", args.cudnn_benchmark)
optim_level = 3 if args.amp else 0
batch_size = args.batch_size
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
if args.max_duration is not None:
featurizer_config['max_duration'] = args.max_duration
# TORCHSCRIPT: Cant use mixed types. Using -1 for "max"
if args.pad_to is not None:
featurizer_config['pad_to'] = args.pad_to if args.pad_to >= 0 else -1
if featurizer_config['pad_to'] == "max":
featurizer_config['pad_to'] = -1
args.use_conv_mask = jasper_model_definition['encoder'].get(
'convmask', True)
if args.use_conv_mask and args.torch_script:
print(
'WARNING: Masked convs currently not supported for TorchScript. Disabling.'
)
jasper_model_definition['encoder']['convmask'] = False
print('model_config')
print_dict(jasper_model_definition)
print('feature_config')
print_dict(featurizer_config)
data_layer = AudioToTextDataLayer(
dataset_dir=args.dataset_dir,
featurizer_config=featurizer_config,
manifest_filepath=val_manifest,
labels=dataset_vocab,
batch_size=batch_size,
pad_to_max=featurizer_config['pad_to'] == -1,
shuffle=False,
multi_gpu=False)
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)
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()))
N = len(data_layer)
step_per_epoch = math.ceil(N / args.batch_size)
print('-----------------')
if args.steps is None:
print('Have {0} examples to eval on.'.format(N))
print('Have {0} steps / (epoch).'.format(step_per_epoch))
else:
print('Have {0} examples to eval on.'.format(args.steps *
args.batch_size))
print('Have {0} steps / (epoch).'.format(args.steps))
print('-----------------')
audio_preprocessor.to(device)
encoderdecoder.to(device)
if args.amp:
encoderdecoder = amp.initialize(models=encoderdecoder,
opt_level='O' + str(optim_level))
eval(data_layer=data_layer,
audio_processor=audio_preprocessor,
encoderdecoder=encoderdecoder,
greedy_decoder=greedy_decoder,
labels=ctc_vocab,
device=device,
args=args)