def main():
if not opt.multi_dataset:
if opt.data_format in ['bin', 'raw']:
start = time.time()
if opt.data.endswith(".train.pt"):
print("Loading data from '%s'" % opt.data)
dataset = torch.load(opt.data)
else:
print("Loading data from %s" % opt.data + ".train.pt")
dataset = torch.load(opt.data + ".train.pt")
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
dicts = dataset['dicts']
# For backward compatibility
train_dict = defaultdict(lambda: None, dataset['train'])
valid_dict = defaultdict(lambda: None, dataset['valid'])
if train_dict['src_lang'] is not None:
assert 'langs' in dicts
train_src_langs = train_dict['src_lang']
train_tgt_langs = train_dict['tgt_lang']
else:
# allocate new languages
dicts['langs'] = {'src': 0, 'tgt': 1}
train_src_langs = list()
train_tgt_langs = list()
# Allocation one for the bilingual case
train_src_langs.append(torch.Tensor([dicts['langs']['src']]))
train_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if not opt.streaming:
train_data = onmt.Dataset(
numpy_to_torch(train_dict['src']),
numpy_to_torch(train_dict['tgt']),
train_dict['src_sizes'],
train_dict['tgt_sizes'],
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
augment=opt.augment_speech,
upsampling=opt.upsampling,
num_split=len(opt.gpus))
else:
train_data = onmt.StreamDataset(
train_dict['src'],
train_dict['tgt'],
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
augment=opt.augment_speech,
upsampling=opt.upsampling)
if valid_dict['src_lang'] is not None:
assert 'langs' in dicts
valid_src_langs = valid_dict['src_lang']
valid_tgt_langs = valid_dict['tgt_lang']
else:
# allocate new languages
valid_src_langs = list()
valid_tgt_langs = list()
# Allocation one for the bilingual case
valid_src_langs.append(torch.Tensor([dicts['langs']['src']]))
valid_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if not opt.streaming:
valid_data = onmt.Dataset(
numpy_to_torch(valid_dict['src']),
numpy_to_torch(valid_dict['tgt']),
valid_dict['src_sizes'],
valid_dict['tgt_sizes'],
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
upsampling=opt.upsampling,
num_split=len(opt.gpus))
else:
valid_data = onmt.StreamDataset(
numpy_to_torch(valid_dict['src']),
numpy_to_torch(valid_dict['tgt']),
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
upsampling=opt.upsampling)
print(' * number of training sentences. %d' %
len(dataset['train']['src']))
print(' * maximum batch size (words per batch). %d' %
opt.batch_size_words)
elif opt.data_format in ['scp', 'scpmem', 'mmem']:
print("Loading memory mapped data files ....")
start = time.time()
from onmt.data.mmap_indexed_dataset import MMapIndexedDataset
from onmt.data.scp_dataset import SCPIndexDataset
dicts = torch.load(opt.data + ".dict.pt")
if opt.data_format in ['scp', 'scpmem']:
audio_data = torch.load(opt.data + ".scp_path.pt")
# allocate languages if not
if 'langs' not in dicts:
dicts['langs'] = {'src': 0, 'tgt': 1}
else:
print(dicts['langs'])
train_path = opt.data + '.train'
if opt.data_format in ['scp', 'scpmem']:
train_src = SCPIndexDataset(audio_data['train'],
concat=opt.concat)
else:
train_src = MMapIndexedDataset(train_path + '.src')
train_tgt = MMapIndexedDataset(train_path + '.tgt')
# check the lang files if they exist (in the case of multi-lingual models)
if os.path.exists(train_path + '.src_lang.bin'):
assert 'langs' in dicts
train_src_langs = MMapIndexedDataset(train_path + '.src_lang')
train_tgt_langs = MMapIndexedDataset(train_path + '.tgt_lang')
else:
train_src_langs = list()
train_tgt_langs = list()
# Allocate a Tensor(1) for the bilingual case
train_src_langs.append(torch.Tensor([dicts['langs']['src']]))
train_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
# check the length files if they exist
if os.path.exists(train_path + '.src_sizes.npy'):
train_src_sizes = np.load(train_path + '.src_sizes.npy')
train_tgt_sizes = np.load(train_path + '.tgt_sizes.npy')
else:
train_src_sizes, train_tgt_sizes = None, None
if opt.encoder_type == 'audio':
data_type = 'audio'
else:
data_type = 'text'
if not opt.streaming:
train_data = onmt.Dataset(
train_src,
train_tgt,
train_src_sizes,
train_tgt_sizes,
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
src_align_right=opt.src_align_right,
augment=opt.augment_speech,
upsampling=opt.upsampling,
cleaning=True,
verbose=True,
num_split=len(opt.gpus))
else:
train_data = onmt.StreamDataset(
train_src,
train_tgt,
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=False,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
upsampling=opt.upsampling)
valid_path = opt.data + '.valid'
if opt.data_format in ['scp', 'scpmem']:
valid_src = SCPIndexDataset(audio_data['valid'],
concat=opt.concat)
else:
valid_src = MMapIndexedDataset(valid_path + '.src')
valid_tgt = MMapIndexedDataset(valid_path + '.tgt')
if os.path.exists(valid_path + '.src_lang.bin'):
assert 'langs' in dicts
valid_src_langs = MMapIndexedDataset(valid_path + '.src_lang')
valid_tgt_langs = MMapIndexedDataset(valid_path + '.tgt_lang')
else:
valid_src_langs = list()
valid_tgt_langs = list()
# Allocation one for the bilingual case
valid_src_langs.append(torch.Tensor([dicts['langs']['src']]))
valid_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
# check the length files if they exist
if os.path.exists(valid_path + '.src_sizes.npy'):
valid_src_sizes = np.load(valid_path + '.src_sizes.npy')
valid_tgt_sizes = np.load(valid_path + '.tgt_sizes.npy')
else:
valid_src_sizes, valid_tgt_sizes = None, None
if not opt.streaming:
valid_data = onmt.Dataset(
valid_src,
valid_tgt,
valid_src_sizes,
valid_tgt_sizes,
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents,
src_align_right=opt.src_align_right,
cleaning=True,
verbose=True,
debug=True,
num_split=len(opt.gpus))
else:
# for validation data, we have to go through sentences (very slow but to ensure correctness)
valid_data = onmt.StreamDataset(
valid_src,
valid_tgt,
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents)
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
else:
raise NotImplementedError
print(' * number of sentences in training data: %d' %
train_data.size())
print(' * number of sentences in validation data: %d' %
valid_data.size())
else:
print("[INFO] Reading multiple dataset ...")
# raise NotImplementedError
dicts = torch.load(opt.data + ".dict.pt")
root_dir = os.path.dirname(opt.data)
print("Loading training data ...")
train_dirs, valid_dirs = dict(), dict()
# scan the data directory to find the training data
for dir_ in os.listdir(root_dir):
if os.path.isdir(os.path.join(root_dir, dir_)):
if str(dir_).startswith("train"):
idx = int(dir_.split(".")[1])
train_dirs[idx] = dir_
if dir_.startswith("valid"):
idx = int(dir_.split(".")[1])
valid_dirs[idx] = dir_
train_sets, valid_sets = list(), list()
for (idx_, dir_) in sorted(train_dirs.items()):
data_dir = os.path.join(root_dir, dir_)
print("[INFO] Loading training data %i from %s" % (idx_, dir_))
if opt.data_format in ['bin', 'raw']:
raise NotImplementedError
elif opt.data_format in ['scp', 'scpmem', 'mmem']:
from onmt.data.mmap_indexed_dataset import MMapIndexedDataset
from onmt.data.scp_dataset import SCPIndexDataset
if opt.data_format in ['scp', 'scpmem']:
audio_data = torch.load(
os.path.join(data_dir, "data.scp_path.pt"))
src_data = SCPIndexDataset(audio_data, concat=opt.concat)
else:
src_data = MMapIndexedDataset(
os.path.join(data_dir, "data.src"))
tgt_data = MMapIndexedDataset(
os.path.join(data_dir, "data.tgt"))
src_lang_data = MMapIndexedDataset(
os.path.join(data_dir, 'data.src_lang'))
tgt_lang_data = MMapIndexedDataset(
os.path.join(data_dir, 'data.tgt_lang'))
if os.path.exists(os.path.join(data_dir,
'data.src_sizes.npy')):
src_sizes = np.load(
os.path.join(data_dir, 'data.src_sizes.npy'))
tgt_sizes = np.load(
os.path.join(data_dir, 'data.tgt_sizes.npy'))
else:
src_sizes, sizes = None, None
if opt.encoder_type == 'audio':
data_type = 'audio'
else:
data_type = 'text'
if not opt.streaming:
train_data = onmt.Dataset(
src_data,
tgt_data,
src_sizes,
tgt_sizes,
src_lang_data,
tgt_lang_data,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
src_align_right=opt.src_align_right,
augment=opt.augment_speech,
upsampling=opt.upsampling,
cleaning=True,
verbose=True,
num_split=len(opt.gpus))
train_sets.append(train_data)
else:
print("Multi-dataset not implemented for Streaming tasks.")
raise NotImplementedError
for (idx_, dir_) in sorted(valid_dirs.items()):
data_dir = os.path.join(root_dir, dir_)
print("[INFO] Loading validation data %i from %s" % (idx_, dir_))
if opt.data_format in ['bin', 'raw']:
raise NotImplementedError
elif opt.data_format in ['scp', 'scpmem', 'mmem']:
if opt.data_format in ['scp', 'scpmem']:
audio_data = torch.load(
os.path.join(data_dir, "data.scp_path.pt"))
src_data = SCPIndexDataset(audio_data, concat=opt.concat)
else:
src_data = MMapIndexedDataset(
os.path.join(data_dir, "data.src"))
tgt_data = MMapIndexedDataset(
os.path.join(data_dir, "data.tgt"))
src_lang_data = MMapIndexedDataset(
os.path.join(data_dir, 'data.src_lang'))
tgt_lang_data = MMapIndexedDataset(
os.path.join(data_dir, 'data.tgt_lang'))
if os.path.exists(os.path.join(data_dir,
'data.src_sizes.npy')):
src_sizes = np.load(
os.path.join(data_dir, 'data.src_sizes.npy'))
tgt_sizes = np.load(
os.path.join(data_dir, 'data.tgt_sizes.npy'))
else:
src_sizes, sizes = None, None
if opt.encoder_type == 'audio':
data_type = 'audio'
else:
data_type = 'text'
if not opt.streaming:
valid_data = onmt.Dataset(
src_data,
tgt_data,
src_sizes,
tgt_sizes,
src_lang_data,
tgt_lang_data,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents,
src_align_right=opt.src_align_right,
cleaning=True,
verbose=True,
debug=True,
num_split=len(opt.gpus))
valid_sets.append(valid_data)
else:
raise NotImplementedError
train_data = train_sets
valid_data = valid_sets
if opt.load_from:
checkpoint = torch.load(opt.load_from,
map_location=lambda storage, loc: storage)
print("* Loading dictionaries from the checkpoint")
dicts = checkpoint['dicts']
else:
dicts['tgt'].patch(opt.patch_vocab_multiplier)
checkpoint = None
# Put the vocab mask from dicts to the datasets
for data in [train_data, valid_data]:
if isinstance(data, list):
for i, data_ in enumerate(data):
data_.set_mask(dicts['tgt'].vocab_mask)
data[i] = data_
else:
data.set_mask(dicts['tgt'].vocab_mask)
if "src" in dicts:
print(' * vocabulary size. source = %d; target = %d' %
(dicts['src'].size(), dicts['tgt'].size()))
else:
print('[INFO] vocabulary size. target = %d' % (dicts['tgt'].size()))
print('* Building model...')
if not opt.fusion:
if opt.bayes_by_backprop:
model = build_bayesian_model(opt, dicts)
else:
model = build_model(opt, dicts)
""" Building the loss function """
# if opt.ctc_loss != 0:
# pass
# loss_function = NMTAndCTCLossFunc(dicts['tgt'].size(),
# label_smoothing=opt.label_smoothing,
# ctc_weight=opt.ctc_loss)
if opt.nce:
from onmt.modules.nce.nce_loss import NCELoss
loss_function = NCELoss(opt.model_size,
dicts['tgt'].size(),
noise_ratio=opt.nce_noise,
logz=9,
label_smoothing=opt.label_smoothing)
else:
loss_function = NMTLossFunc(opt.model_size,
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
mirror=opt.mirror_loss,
fast_xentropy=opt.fast_xentropy)
# This function replaces modules with the more optimized counterparts so that it can run faster
# Currently exp with LayerNorm
if not opt.memory_profiling:
optimize_model(model, fp16=opt.fp16)
else:
from onmt.model_factory import build_fusion
from onmt.modules.loss import FusionLoss
model = build_fusion(opt, dicts)
loss_function = FusionLoss(dicts['tgt'].size(),
label_smoothing=opt.label_smoothing)
n_params = sum([p.nelement() for p in model.parameters()])
print('* number of parameters: %d' % n_params)
if not opt.debugging and len(opt.gpus) == 1:
if opt.bayes_by_backprop:
from onmt.train_utils.bayes_by_backprop_trainer import BayesianTrainer
trainer = BayesianTrainer(model, loss_function, train_data,
valid_data, dicts, opt)
else:
trainer = XETrainer(model, loss_function, train_data, valid_data,
dicts, opt)
else:
from onmt.train_utils.new_trainer import Trainer
trainer = Trainer(model, loss_function, train_data, valid_data, dicts,
opt)
trainer.run(checkpoint=checkpoint)
def __init__(self, device, train_data, valid_data, dicts, opt, setup_optimizer=True):
"""
:param model:
:param device: int (GPU id)
:param loss_function:
:param train_data:
:param valid_data:
:param dicts:
:param opt:
"""
self.device = device
opt.node_rank = 0
opt.nodes = 1
self.world_size = len(opt.gpus)
# in the case of single node distributed, it should equal self.device
self.rank = self.device
# make a group to later use with self.all_reduce
self.group = dist.group.WORLD
self.print("[INFO] Training Options:", opt)
if self.world_size > 1:
dist.init_process_group(backend='nccl', init_method='env://', world_size=self.world_size, rank=self.rank)
self.model = None
if self.rank == 0:
self.train_data = train_data
self.valid_data = valid_data
else:
# Do we really need to deepcopy the data instances (which could cause memory leak easily)
self.train_data = copy.deepcopy(train_data)
self.valid_data = copy.deepcopy(valid_data)
self.dicts = dicts
self.opt = opt
self.cuda = (len(opt.gpus) >= 1 and opt.gpus[0] >= 0)
assert self.cuda, "[ERROR] Training is only available on GPUs."
self.start_time = 0
# setting up models and others
if opt.lfv_multilingual:
from onmt.models.speech_recognizer.lid_loss import CrossEntropyLIDLoss
lid_loss = CrossEntropyLIDLoss(opt.n_languages, opt.label_smoothing, opt.fast_xentropy)
self.loss_function.add_loss_function(lid_loss, 'lid_loss')
torch.manual_seed(self.opt.seed)
# note: we must start creating models after ccreating the processes
# for some reason passing a pre-created model to a process creates a "pickle" error
if not opt.fusion:
if self.is_main():
print("[INFO] Building models .... ", flush=True)
model = build_model(opt, dicts)
""" Building the loss function """
if opt.ctc_loss > 0.0:
from onmt.speech.ctc_loss import CTC
self.ctc_loss_function = CTC(0.0, reduce=True)
if opt.nce:
from onmt.modules.nce.nce_loss import NCELoss
loss_function = NCELoss(opt.model_size, dicts['tgt'].size(), noise_ratio=opt.nce_noise,
logz=9, label_smoothing=opt.label_smoothing)
else:
loss_function = NMTLossFunc(opt.model_size, dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
mirror=opt.mirror_loss,
fast_xentropy=opt.fast_xentropy)
# This function replaces modules with the more optimized counterparts so that it can run faster
# Currently exp with LayerNorm
if not opt.memory_profiling:
# distributed is required to convert BatchNorm to SyncBatchNorm for DDP
optimize_model(model, distributed=(self.world_size > 1))
init_model_parameters(model, opt)
self.model = model
self.loss_function = loss_function
self.grad_scaler = torch.cuda.amp.GradScaler()
if opt.load_from:
checkpoint = torch.load(opt.load_from, map_location=lambda storage, loc: storage)
self.model.load_state_dict(checkpoint['model'])
if 'scaler' in checkpoint and checkpoint['scaler'] is not None:
self.grad_scaler.load_state_dict(checkpoint['scaler'])
if self.cuda:
torch.cuda.set_device(self.device)
self.loss_function = self.loss_function.cuda(device=self.device)
self.model = self.model.cuda(device=self.device)
if opt.ctc_loss > 0.0:
self.ctc_loss_function = self.ctc_loss_function.cuda(device=self.device)
# Ensure that the distributed copies have the same initial parameters
# Manual seed may not work the same for different GPU models.
# if self.world_size > 1:
# params = [p for p in self.model.parameters()]
#
# with torch.no_grad():
# if not self.is_main():
# # zero everything except for the main model
# for p in params:
# p.zero_()
# else:
# for p in params:
# p.add_(0)
#
# # run all_reduce to ensure that all models have exactly the same parameters
# if self.world_size > 1:
# params = [p for p in self.model.parameters()]
# all_reduce_and_rescale_tensors(params, 1)
if setup_optimizer:
self.optim = onmt.Optim(opt)
self.optim.set_parameters(self.model.parameters())
if self.is_main():
print("[INFO] Optimizer: ", self.optim.optimizer)
if opt.load_from:
if 'optim' in checkpoint and checkpoint['optim'] is not None and not opt.reset_optim:
self.optim.load_state_dict(checkpoint['optim'])
if self.world_size > 1:
# find_unused_parameters may be required for dropped layer (parameters that are not connected to
# any particular graph)
find_unused_parameters = True
self.model = torch.nn.parallel.DistributedDataParallel(self.model, device_ids=[self.rank],
output_device=self.rank,
find_unused_parameters=find_unused_parameters)
print("[INFO] Process %d ready." % self.rank, flush=True)
def __init__(self,
device,
train_data,
valid_data,
dicts,
opt,
setup_optimizer=True):
"""
:param model:
:param device: int (GPU id)
:param loss_function:
:param train_data:
:param valid_data:
:param dicts:
:param opt:
"""
# self.model = model
# self.model = model
# self.loss_function = loss_function
self.device = device
opt.node_rank = 0
opt.nodes = 1
self.world_size = len(opt.gpus)
# in the case of single node distributed, it should equal self.device
self.rank = self.device
# make a group to later use with dist.all_reduce
self.group = dist.group.WORLD
self.print("[INFO] Training Options:", opt)
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=self.world_size,
rank=self.rank)
self.model = None
if self.rank == 0:
self.train_data = train_data
self.valid_data = valid_data
else:
self.train_data = copy.deepcopy(train_data)
self.valid_data = copy.deepcopy(valid_data)
self.dicts = dicts
self.opt = opt
self.cuda = (len(opt.gpus) >= 1 and opt.gpus[0] >= 0)
assert self.cuda, "[ERROR] Training is only available on GPUs."
self.start_time = 0
# setting up models and others
if opt.lfv_multilingual:
from onmt.models.speech_recognizer.lid_loss import CrossEntropyLIDLoss
lid_loss = CrossEntropyLIDLoss(opt.n_languages,
opt.label_smoothing,
opt.fast_xentropy)
self.loss_function.add_loss_function(lid_loss, 'lid_loss')
torch.manual_seed(self.opt.seed)
# note: we must start creating models after ccreating the processes
# for some reason passing a pre-created model to a process creates a "pickle" error
if not opt.fusion:
if self.is_main():
print("BUILDING MODEL .... ", flush=True)
model = build_model(opt, dicts)
""" Building the loss function """
if opt.ctc_loss != 0:
loss_function = NMTAndCTCLossFunc(
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
ctc_weight=opt.ctc_loss)
elif opt.nce:
from onmt.modules.nce.nce_loss import NCELoss
loss_function = NCELoss(opt.model_size,
dicts['tgt'].size(),
noise_ratio=opt.nce_noise,
logz=9,
label_smoothing=opt.label_smoothing)
else:
loss_function = NMTLossFunc(
opt.model_size,
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
mirror=opt.mirror_loss,
fast_xentropy=opt.fast_xentropy)
# This function replaces modules with the more optimized counterparts so that it can run faster
# Currently exp with LayerNorm
if not opt.memory_profiling:
# distributed is required to convert BatchNorm to SyncBatchNorm for DDP
optimize_model(model, distributed=(self.world_size > 1))
# optimize_model(model)
init_model_parameters(model, opt)
self.model = model
self.loss_function = loss_function
if self.cuda:
torch.cuda.set_device(self.device)
self.loss_function = self.loss_function.cuda(device=self.device)
self.model = self.model.cuda(device=self.device)
# Ensure that the distributed copies have the same initial parameters
# Manual seed may not work the same for different GPU models.
if self.world_size > 1:
params = [p for p in self.model.parameters()]
with torch.no_grad():
if not self.is_main():
for p in params:
p.zero_()
else:
for p in params:
p.add_(0)
if self.world_size > 1:
params = [p for p in self.model.parameters()]
all_reduce_and_rescale_tensors(params, 1)
if setup_optimizer:
self.optim = onmt.Optim(opt)
self.optim.set_parameters(self.model.parameters())
if self.is_main():
print("[INFO] Optimizer: ", self.optim.optimizer)
if not self.opt.fp16:
opt_level = "O0"
keep_batchnorm_fp32 = False
elif self.opt.fp16_mixed:
opt_level = "O1"
keep_batchnorm_fp32 = None
else:
opt_level = "O2"
keep_batchnorm_fp32 = False
if self.cuda:
self.model, self.optim.optimizer = amp.initialize(
self.model,
self.optim.optimizer,
opt_level=opt_level,
keep_batchnorm_fp32=keep_batchnorm_fp32,
loss_scale="dynamic",
verbosity=1 if self.opt.verbose else 0)
# wrap the model into DDP after initializing by amp
if self.world_size > 1:
"""
delay_allreduce is required to avoid allreduce error during backward pass
"""
self.model = DDP(self.model,
delay_allreduce=True,
gradient_average=False)
# torch DDP is more likely to work with the official amp autocast
# self.model = torch.nn.parallel.DistributedDataParallel(self.model, device_ids=[self.rank],
# output_device=self.rank,
# find_unused_parameters=True)
print("[INFO] Process %d ready." % self.rank, flush=True)
def __init__(self,
device,
train_data,
valid_data,
dicts,
opt,
setup_optimizer=True):
"""
:param model:
:param device: int (GPU id)
:param loss_function:
:param train_data:
:param valid_data:
:param dicts:
:param opt:
"""
self.device = device
opt.node_rank = 0
opt.nodes = 1
self.world_size = len(opt.gpus)
# in the case of single node distributed, it should equal self.device
self.rank = self.device
# make a group to later use with self.all_reduce
self.group = dist.group.WORLD
self.print("[INFO] Training Options:", opt)
if self.world_size > 1:
dist.init_process_group(backend='nccl',
init_method='env://',
world_size=self.world_size,
rank=self.rank)
self.model = None
if self.rank == 0:
self.train_data = train_data
self.valid_data = valid_data
else:
# Do we really need to deepcopy the data instances (which could cause memory leak easily)
self.train_data = copy.deepcopy(train_data)
self.valid_data = copy.deepcopy(valid_data)
self.dicts = dicts
self.opt = opt
self.cuda = (len(opt.gpus) >= 1 and opt.gpus[0] >= 0)
assert self.cuda, "[ERROR] Training is only available on GPUs."
self.start_time = 0
# setting up models and others
if opt.lfv_multilingual:
from onmt.models.speech_recognizer.lid_loss import CrossEntropyLIDLoss
lid_loss = CrossEntropyLIDLoss(opt.n_languages,
opt.label_smoothing,
opt.fast_xentropy)
self.loss_function.add_loss_function(lid_loss, 'lid_loss')
torch.manual_seed(self.opt.seed)
# note: we must start creating models after ccreating the processes
# for some reason passing a pre-created model to a process creates a "pickle" error
if not opt.fusion:
if self.is_main():
print("[INFO] Building models .... ", flush=True)
model = build_model(opt, dicts)
""" Building the loss function """
if opt.ctc_loss > 0.0:
from onmt.speech.ctc_loss import CTC
self.ctc_loss_function = CTC(0.0, reduce=True)
if opt.nce:
from onmt.modules.nce.nce_loss import NCELoss
loss_function = NCELoss(opt.model_size,
dicts['tgt'].size(),
noise_ratio=opt.nce_noise,
logz=9,
label_smoothing=opt.label_smoothing)
else:
loss_function = NMTLossFunc(
opt.model_size,
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
mirror=opt.mirror_loss,
fast_xentropy=opt.fast_xentropy)
# This function replaces modules with the more optimized counterparts so that it can run faster
# Currently exp with LayerNorm
if not opt.memory_profiling:
# distributed is required to convert BatchNorm to SyncBatchNorm for DDP
optimize_model(model, distributed=(self.world_size > 1))
init_model_parameters(model, opt)
self.model = model
self.loss_function = loss_function
# self.grad_scaler = torch.cuda.amp.GradScaler()
if self.cuda:
torch.cuda.set_device(self.device)
self.loss_function = self.loss_function.cuda(device=self.device)
self.model = self.model.cuda(device=self.device)
if opt.ctc_loss > 0.0:
self.ctc_loss_function = self.ctc_loss_function.cuda(
device=self.device)
if opt.load_from:
checkpoint = torch.load(opt.load_from,
map_location=lambda storage, loc: storage)
if setup_optimizer:
self.optim = onmt.Optim(opt)
self.optim.set_parameters(self.model.parameters())
if self.is_main():
print("[INFO] Optimizer: ", self.optim.optimizer)
if opt.load_from:
if 'optim' in checkpoint and checkpoint[
'optim'] is not None and not opt.reset_optim:
self.optim.load_state_dict(checkpoint['optim'])
if not self.opt.fp16:
opt_level = "O0"
keep_batchnorm_fp32 = False
elif self.opt.fp16_mixed:
opt_level = "O1"
keep_batchnorm_fp32 = None
else:
opt_level = "O2"
keep_batchnorm_fp32 = False
self.opt_level = opt_level
if self.cuda:
self.model, self.optim.optimizer = amp.initialize(
self.model,
self.optim.optimizer,
opt_level=opt_level,
keep_batchnorm_fp32=keep_batchnorm_fp32,
loss_scale="dynamic",
verbosity=1 if self.opt.verbose else 1)
if opt.load_from:
self.model.load_state_dict(checkpoint['model'])
if prec_opt is not None and hasattr(prec_opt, "fp16_mixed"):
# Only load amp information if the mode is the same
# Maybe its better to change between optimization mode?
if opt.fp16_mixed == prec_opt.fp16_mixed and opt.fp16 == prec_opt.fp16:
if 'amp' in checkpoint:
try:
amp.load_state_dict(checkpoint['amp'])
except Exception:
# loading the amp state can fail
pass
if self.world_size > 1:
# find_unused_parameters may be required for dropped layer (parameters that are not connected to
# any particular graph)
# find_unused_parameters = True
self.model = DDP(self.model,
delay_allreduce=True,
gradient_average=False)
print("[INFO] Process %d ready." % self.rank, flush=True)
def main():
if opt.data_format in ['bin', 'raw']:
start = time.time()
if opt.data.endswith(".train.pt"):
print("Loading data from '%s'" % opt.data)
dataset = torch.load(opt.data)
else:
print("Loading data from %s" % opt.data + ".train.pt")
dataset = torch.load(opt.data + ".train.pt")
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
dicts = dataset['dicts']
# For backward compatibility
train_dict = defaultdict(lambda: None, dataset['train'])
valid_dict = defaultdict(lambda: None, dataset['valid'])
if train_dict['src_lang'] is not None:
assert 'langs' in dicts
train_src_langs = train_dict['src_lang']
train_tgt_langs = train_dict['tgt_lang']
else:
# allocate new languages
dicts['langs'] = {'src': 0, 'tgt': 1}
train_src_langs = list()
train_tgt_langs = list()
# Allocation one for the bilingual case
train_src_langs.append(torch.Tensor([dicts['langs']['src']]))
train_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if not opt.streaming:
train_data = onmt.Dataset(train_dict['src'],
train_dict['tgt'],
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
augment=opt.augment_speech,
upsampling=opt.upsampling)
else:
train_data = onmt.StreamDataset(
train_dict['src'],
train_dict['tgt'],
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
augment=opt.augment_speech,
upsampling=opt.upsampling)
if valid_dict['src_lang'] is not None:
assert 'langs' in dicts
valid_src_langs = valid_dict['src_lang']
valid_tgt_langs = valid_dict['tgt_lang']
else:
# allocate new languages
valid_src_langs = list()
valid_tgt_langs = list()
# Allocation one for the bilingual case
valid_src_langs.append(torch.Tensor([dicts['langs']['src']]))
valid_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if not opt.streaming:
valid_data = onmt.Dataset(valid_dict['src'],
valid_dict['tgt'],
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
upsampling=opt.upsampling)
else:
valid_data = onmt.StreamDataset(
valid_dict['src'],
valid_dict['tgt'],
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=dataset.get("type", "text"),
sorting=True,
batch_size_sents=opt.batch_size_sents,
upsampling=opt.upsampling)
print(' * number of training sentences. %d' %
len(dataset['train']['src']))
print(' * maximum batch size (words per batch). %d' %
opt.batch_size_words)
elif opt.data_format == 'mmem':
print("Loading memory mapped data files ....")
start = time.time()
from onmt.data.mmap_indexed_dataset import MMapIndexedDataset
dicts = torch.load(opt.data + ".dict.pt")
# allocate languages if not
if 'langs' not in dicts:
dicts['langs'] = {'src': 0, 'tgt': 1}
else:
print(dicts['langs'])
train_path = opt.data + '.train'
train_src = MMapIndexedDataset(train_path + '.src')
train_tgt = MMapIndexedDataset(train_path + '.tgt')
# check the lang files if they exist (in the case of multi-lingual models)
if os.path.exists(train_path + '.src_lang.bin'):
assert 'langs' in dicts
train_src_langs = MMapIndexedDataset(train_path + '.src_lang')
train_tgt_langs = MMapIndexedDataset(train_path + '.tgt_lang')
else:
train_src_langs = list()
train_tgt_langs = list()
# Allocate a Tensor(1) for the bilingual case
train_src_langs.append(torch.Tensor([dicts['langs']['src']]))
train_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if opt.encoder_type == 'audio':
data_type = 'audio'
else:
data_type = 'text'
if not opt.streaming:
train_data = onmt.Dataset(train_src,
train_tgt,
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=True,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
src_align_right=opt.src_align_right,
upsampling=opt.upsampling,
cleaning=True,
verbose=True)
else:
train_data = onmt.StreamDataset(
train_src,
train_tgt,
train_src_langs,
train_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type=data_type,
sorting=False,
batch_size_sents=opt.batch_size_sents,
multiplier=opt.batch_size_multiplier,
upsampling=opt.upsampling)
valid_path = opt.data + '.valid'
valid_src = MMapIndexedDataset(valid_path + '.src')
valid_tgt = MMapIndexedDataset(valid_path + '.tgt')
if os.path.exists(valid_path + '.src_lang.bin'):
assert 'langs' in dicts
valid_src_langs = MMapIndexedDataset(valid_path + '.src_lang')
valid_tgt_langs = MMapIndexedDataset(valid_path + '.tgt_lang')
else:
valid_src_langs = list()
valid_tgt_langs = list()
# Allocation one for the bilingual case
valid_src_langs.append(torch.Tensor([dicts['langs']['src']]))
valid_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
if not opt.streaming:
valid_data = onmt.Dataset(valid_src,
valid_tgt,
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type="text",
sorting=False,
batch_size_sents=opt.batch_size_sents,
src_align_right=opt.src_align_right,
cleaning=True,
verbose=True)
else:
# for validation data, we have to go through sentences (very slow but to ensure correctness)
valid_data = onmt.StreamDataset(
valid_src,
valid_tgt,
valid_src_langs,
valid_tgt_langs,
batch_size_words=opt.batch_size_words,
data_type="text",
sorting=True,
batch_size_sents=opt.batch_size_sents)
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
else:
raise NotImplementedError
# additional_data = []
# if opt.additional_data != "none":
# add_data = opt.additional_data.split(";")
# add_format = opt.additional_data_format.split(";")
# assert (len(add_data) == len(add_format))
# for i in range(len(add_data)):
# if add_format[i] == 'raw':
# if add_data[i].endswith(".train.pt"):
# print("Loading data from '%s'" % opt.data)
# add_dataset = torch.load(add_data[i])
# else:
# print("Loading data from %s" % opt.data + ".train.pt")
# add_dataset = torch.load(add_data[i] + ".train.pt")
#
# additional_data.append(onmt.Dataset(add_dataset['train']['src'],
# dataset['train']['tgt'], batch_size_words=opt.batch_size_words,
# data_type=dataset.get("type", "text"), sorting=True,
# batch_size_sents=opt.batch_size_sents,
# multiplier=opt.batch_size_multiplier,
# reshape_speech=opt.reshape_speech,
# augment=opt.augment_speech))
# elif add_format[i] == 'bin':
#
# from onmt.data.indexed_dataset import IndexedInMemoryDataset
#
# train_path = add_data[i] + '.train'
# train_src = IndexedInMemoryDataset(train_path + '.src')
# train_tgt = IndexedInMemoryDataset(train_path + '.tgt')
#
# additional_data.append(onmt.Dataset(train_src,
# train_tgt,
# batch_size_words=opt.batch_size_words,
# data_type=opt.encoder_type,
# batch_size_sents=opt.batch_size_sents,
# multiplier=opt.batch_size_multiplier))
if opt.load_from:
checkpoint = torch.load(opt.load_from,
map_location=lambda storage, loc: storage)
print("* Loading dictionaries from the checkpoint")
dicts = checkpoint['dicts']
else:
dicts['tgt'].patch(opt.patch_vocab_multiplier)
checkpoint = None
if "src" in dicts:
print(' * vocabulary size. source = %d; target = %d' %
(dicts['src'].size(), dicts['tgt'].size()))
else:
print(' * vocabulary size. target = %d' % (dicts['tgt'].size()))
print(' * number of sentences in training data: %d' % train_data.size())
print(' * number of sentences in validation data: %d' % valid_data.size())
print('* Building model...')
if not opt.fusion:
model = build_model(opt, dicts)
""" Building the loss function """
if opt.ctc_loss != 0:
loss_function = NMTAndCTCLossFunc(
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
ctc_weight=opt.ctc_loss)
else:
loss_function = NMTLossFunc(opt.model_size,
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing,
mirror=opt.mirror_loss)
# This function replaces modules with the more optimized counterparts so that it can run faster
# Currently exp with LayerNorm
optimize_model(model)
else:
from onmt.model_factory import build_fusion
from onmt.modules.loss import FusionLoss
model = build_fusion(opt, dicts)
loss_function = FusionLoss(dicts['tgt'].size(),
label_smoothing=opt.label_smoothing)
n_params = sum([p.nelement() for p in model.parameters()])
print('* number of parameters: %d' % n_params)
if len(opt.gpus) > 1 or opt.virtual_gpu > 1:
raise NotImplementedError(
"Multi-GPU training is not supported at the moment.")
else:
trainer = XETrainer(model, loss_function, train_data, valid_data,
dicts, opt)
trainer.run(checkpoint=checkpoint)
def main():
start = time.time()
print("Loading data from '%s'" % opt.data)
if opt.data_format == 'raw':
dataset = torch.load(opt.data)
elapse = str(datetime.timedelta(seconds=int(time.time() - start)))
print("Done after %s" % elapse)
dicts = dataset['dicts']
# For backward compatibility
train_dict = defaultdict(lambda: None, dataset['train'])
valid_dict = defaultdict(lambda: None, dataset['valid'])
if train_dict['src_lang'] is not None:
assert 'langs' in dicts
train_src_langs = train_dict['src_lang']
train_tgt_langs = train_dict['tgt_lang']
else:
# allocate new languages
dicts['langs'] = {'src': 0, 'tgt': 1}
train_src_langs = list()
train_tgt_langs = list()
# Allocation one for the bilingual case
train_src_langs.append(torch.Tensor([dicts['langs']['src']]))
train_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
train_data = LanguageModelDataset(
dataset['train']['tgt'],
train_tgt_langs,
batch_size_sents=opt.batch_size_sents,
seq_length=opt.lm_seq_length)
if valid_dict['src_lang'] is not None:
assert 'langs' in dicts
valid_src_langs = valid_dict['src_lang']
valid_tgt_langs = valid_dict['tgt_lang']
else:
# allocate new languages
valid_src_langs = list()
valid_tgt_langs = list()
# Allocation one for the bilingual case
valid_src_langs.append(torch.Tensor([dicts['langs']['src']]))
valid_tgt_langs.append(torch.Tensor([dicts['langs']['tgt']]))
valid_data = LanguageModelDataset(
dataset['valid']['tgt'],
valid_tgt_langs,
batch_size_sents=opt.batch_size_sents,
seq_length=opt.lm_seq_length)
if opt.load_from:
checkpoint = torch.load(opt.load_from,
map_location=lambda storage, loc: storage)
print("* Loading dictionaries from the checkpoint")
dicts = checkpoint['dicts']
else:
dicts['tgt'].patch(opt.patch_vocab_multiplier)
checkpoint = None
if "src" in dicts:
print(' * vocabulary size. source = %d; target = %d' %
(dicts['src'].size(), dicts['tgt'].size()))
else:
print(' * vocabulary size. target = %d' % (dicts['tgt'].size()))
print(' * number of training sentences. %d' % train_data.size())
print(' * maximum batch size (words per batch). %d' %
(opt.batch_size_sents * opt.lm_seq_length))
else:
raise NotImplementedError
print('Building model...')
model = build_language_model(opt, dicts)
optimize_model(model)
""" Building the loss function """
loss_function = NMTLossFunc(opt.model_size,
dicts['tgt'].size(),
label_smoothing=opt.label_smoothing)
n_params = sum([p.nelement() for p in model.parameters()])
print('* number of parameters: %d' % n_params)
if len(opt.gpus) > 1 or opt.virtual_gpu > 1:
raise NotImplementedError("Multi-GPU training is not supported ATM.")
else:
# if opt.fp16:
# trainer = FP16XETrainer(model, loss_function, train_data, valid_data, dicts, opt)
# else:
trainer = XETrainer(model, loss_function, train_data, valid_data,
dicts, opt)
trainer.run(checkpoint=checkpoint)