def train(args):
"""Train with the given args.
Args:
args (namespace): The program arguments.
"""
set_deterministic_pytorch(args)
# check cuda availability
if not torch.cuda.is_available():
logging.warning('cuda is not available')
# get input and output dimension info
with open(args.valid_json, 'rb') as f:
valid_json = json.load(f)['utts']
utts = list(valid_json.keys())
idim = int(valid_json[utts[0]]['input'][0]['shape'][-1])
odim = int(valid_json[utts[0]]['output'][0]['shape'][-1])
logging.info('#input dims : ' + str(idim))
logging.info('#output dims: ' + str(odim))
# Initialize with pre-trained ASR encoder and MT decoder
if args.enc_init is not None or args.dec_init is not None:
model = load_trained_modules(idim, odim, args, interface=STInterface)
else:
model_class = dynamic_import(args.model_module)
model = model_class(idim, odim, args)
assert isinstance(model, STInterface)
subsampling_factor = model.subsample[0]
if args.rnnlm is not None:
rnnlm_args = get_model_conf(args.rnnlm, args.rnnlm_conf)
rnnlm = lm_pytorch.ClassifierWithState(
lm_pytorch.RNNLM(
len(args.char_list), rnnlm_args.layer, rnnlm_args.unit,
getattr(rnnlm_args, "embed_unit", None), # for backward compatibility
)
)
torch_load(args.rnnlm, rnnlm)
model.rnnlm = rnnlm
# write model config
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
model_conf = args.outdir + '/model.json'
with open(model_conf, 'wb') as f:
logging.info('writing a model config file to ' + model_conf)
f.write(json.dumps((idim, odim, vars(args)),
indent=4, ensure_ascii=False, sort_keys=True).encode('utf_8'))
for key in sorted(vars(args).keys()):
logging.info('ARGS: ' + key + ': ' + str(vars(args)[key]))
reporter = model.reporter
# check the use of multi-gpu
if args.ngpu > 1:
if args.batch_size != 0:
logging.warning('batch size is automatically increased (%d -> %d)' % (
args.batch_size, args.batch_size * args.ngpu))
args.batch_size *= args.ngpu
# set torch device
device = torch.device("cuda" if args.ngpu > 0 else "cpu")
if args.train_dtype in ("float16", "float32", "float64"):
dtype = getattr(torch, args.train_dtype)
else:
dtype = torch.float32
model = model.to(device=device, dtype=dtype)
# Setup an optimizer
if args.opt == 'adadelta':
optimizer = torch.optim.Adadelta(
model.parameters(), rho=0.95, eps=args.eps,
weight_decay=args.weight_decay)
elif args.opt == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.opt == 'noam':
from espnet.nets.pytorch_backend.transformer.optimizer import get_std_opt
optimizer = get_std_opt(model, args.adim, args.transformer_warmup_steps, args.transformer_lr)
else:
raise NotImplementedError("unknown optimizer: " + args.opt)
# setup apex.amp
if args.train_dtype in ("O0", "O1", "O2", "O3"):
try:
from apex import amp
except ImportError as e:
logging.error(f"You need to install apex for --train-dtype {args.train_dtype}. "
"See https://github.com/NVIDIA/apex#linux")
raise e
if args.opt == 'noam':
model, optimizer.optimizer = amp.initialize(model, optimizer.optimizer, opt_level=args.train_dtype)
else:
model, optimizer = amp.initialize(model, optimizer, opt_level=args.train_dtype)
use_apex = True
else:
use_apex = False
# FIXME: TOO DIRTY HACK
setattr(optimizer, "target", reporter)
setattr(optimizer, "serialize", lambda s: reporter.serialize(s))
# Setup a converter
converter = MultimodalConverter(subsampling_factor=subsampling_factor, dtype=dtype)
# read json data
with open(args.train_json, 'rb') as f:
train_json = json.load(f)['utts']
with open(args.valid_json, 'rb') as f:
valid_json = json.load(f)['utts']
use_sortagrad = args.sortagrad == -1 or args.sortagrad > 0
# make minibatch list (variable length)
train = make_batchset(train_json, args.batch_size,
args.maxlen_in, args.maxlen_out, args.minibatches,
min_batch_size=args.ngpu if args.ngpu > 1 else 1,
shortest_first=use_sortagrad,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout)
valid = make_batchset(valid_json, args.batch_size,
args.maxlen_in, args.maxlen_out, args.minibatches,
min_batch_size=args.ngpu if args.ngpu > 1 else 1,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout)
load_tr = LoadInputsAndTargets(
mode='asr', load_output=True, preprocess_conf=args.preprocess_conf,
preprocess_args={'train': True} # Switch the mode of preprocessing
)
load_cv = LoadInputsAndTargets(
mode='asr', load_output=True, preprocess_conf=args.preprocess_conf,
preprocess_args={'train': False} # Switch the mode of preprocessing
)
# hack to make batchsize argument as 1
# actual bathsize is included in a list
# default collate function converts numpy array to pytorch tensor
# we used an empty collate function instead which returns list
train_iter = {'main': ChainerDataLoader(
dataset=TransformDataset(train, lambda data: converter([load_tr(data)])),
batch_size=1, num_workers=args.n_iter_processes,
shuffle=not use_sortagrad, collate_fn=lambda x: x[0])}
valid_iter = {'main': ChainerDataLoader(
dataset=TransformDataset(valid, lambda data: converter([load_cv(data)])),
batch_size=1, shuffle=False, collate_fn=lambda x: x[0],
num_workers=args.n_iter_processes)}
# Set up a trainer
updater = CustomUpdater(
model, args.grad_clip, train_iter, optimizer,
device, args.ngpu, args.grad_noise, args.accum_grad, use_apex=use_apex)
trainer = training.Trainer(
updater, (args.epochs, 'epoch'), out=args.outdir)
if use_sortagrad:
trainer.extend(ShufflingEnabler([train_iter]),
trigger=(args.sortagrad if args.sortagrad != -1 else args.epochs, 'epoch'))
# Resume from a snapshot
if args.resume:
logging.info('resumed from %s' % args.resume)
torch_resume(args.resume, trainer)
# Evaluate the model with the test dataset for each epoch
if args.save_interval_iters > 0:
trainer.extend(CustomEvaluator(model, valid_iter, reporter, device, args.ngpu),
trigger=(args.save_interval_iters, 'iteration'))
else:
trainer.extend(CustomEvaluator(model, valid_iter, reporter, device, args.ngpu))
# Save attention weight each epoch
if args.num_save_attention > 0:
data = sorted(list(valid_json.items())[:args.num_save_attention],
key=lambda x: int(x[1]['input'][0]['shape'][1]), reverse=True)
if hasattr(model, "module"):
att_vis_fn = model.module.calculate_all_attentions
plot_class = model.module.attention_plot_class
else:
att_vis_fn = model.calculate_all_attentions
plot_class = model.attention_plot_class
att_reporter = plot_class(
att_vis_fn, data, args.outdir + "/att_ws",
converter=converter, transform=load_cv, device=device)
trainer.extend(att_reporter, trigger=(1, 'epoch'))
else:
att_reporter = None
# Make a plot for training and validation values
trainer.extend(extensions.PlotReport(['main/loss', 'validation/main/loss',
'main/loss_asr', 'validation/main/loss_asr',
'main/loss_st', 'validation/main/loss_st'],
'epoch', file_name='loss.png'))
trainer.extend(extensions.PlotReport(['main/acc', 'validation/main/acc',
'main/acc_asr', 'validation/main/acc_asr'],
'epoch', file_name='acc.png'))
trainer.extend(extensions.PlotReport(['main/bleu', 'validation/main/bleu'],
'epoch', file_name='bleu.png'))
# Save best models
trainer.extend(snapshot_object(model, 'model.loss.best'),
trigger=training.triggers.MinValueTrigger('validation/main/loss'))
trainer.extend(snapshot_object(model, 'model.acc.best'),
trigger=training.triggers.MaxValueTrigger('validation/main/acc'))
# save snapshot which contains model and optimizer states
if args.save_interval_iters > 0:
trainer.extend(torch_snapshot(filename='snapshot.iter.{.updater.iteration}'),
trigger=(args.save_interval_iters, 'iteration'))
else:
trainer.extend(torch_snapshot(), trigger=(1, 'epoch'))
# epsilon decay in the optimizer
if args.opt == 'adadelta':
if args.criterion == 'acc':
trainer.extend(restore_snapshot(model, args.outdir + '/model.acc.best', load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/acc',
lambda best_value, current_value: best_value > current_value))
trainer.extend(adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
'validation/main/acc',
lambda best_value, current_value: best_value > current_value))
elif args.criterion == 'loss':
trainer.extend(restore_snapshot(model, args.outdir + '/model.loss.best', load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/loss',
lambda best_value, current_value: best_value < current_value))
trainer.extend(adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
'validation/main/loss',
lambda best_value, current_value: best_value < current_value))
elif args.opt == 'adam':
if args.criterion == 'acc':
trainer.extend(restore_snapshot(model, args.outdir + '/model.acc.best', load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/acc',
lambda best_value, current_value: best_value > current_value))
trainer.extend(adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
'validation/main/acc',
lambda best_value, current_value: best_value > current_value))
elif args.criterion == 'loss':
trainer.extend(restore_snapshot(model, args.outdir + '/model.loss.best', load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/loss',
lambda best_value, current_value: best_value < current_value))
trainer.extend(adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
'validation/main/loss',
lambda best_value, current_value: best_value < current_value))
# Write a log of evaluation statistics for each epoch
trainer.extend(extensions.LogReport(trigger=(args.report_interval_iters, 'iteration')))
report_keys = ['epoch', 'iteration', 'main/loss', 'main/loss_st', 'main/loss_asr',
'validation/main/loss', 'validation/main/loss_st', 'validation/main/loss_asr',
'main/acc', 'validation/main/acc']
if args.asr_weight > 0:
report_keys.append('main/acc_asr')
report_keys.append('validation/main/acc_asr')
report_keys += ['elapsed_time']
if args.opt == 'adadelta':
trainer.extend(extensions.observe_value(
'eps', lambda trainer: trainer.updater.get_optimizer('main').param_groups[0]["eps"]),
trigger=(args.report_interval_iters, 'iteration'))
report_keys.append('eps')
elif args.opt in ['adam', 'noam']:
trainer.extend(extensions.observe_value(
'lr', lambda trainer: trainer.updater.get_optimizer('main').param_groups[0]["lr"]),
trigger=(args.report_interval_iters, 'iteration'))
report_keys.append('lr')
if args.asr_weight > 0:
if args.mtlalpha > 0:
report_keys.append('main/cer_ctc')
report_keys.append('validation/main/cer_ctc')
if args.mtlalpha < 1:
if args.report_cer:
report_keys.append('validation/main/cer')
if args.report_wer:
report_keys.append('validation/main/wer')
if args.report_bleu:
report_keys.append('validation/main/bleu')
trainer.extend(extensions.PrintReport(
report_keys), trigger=(args.report_interval_iters, 'iteration'))
trainer.extend(extensions.ProgressBar(update_interval=args.report_interval_iters))
set_early_stop(trainer, args)
if args.tensorboard_dir is not None and args.tensorboard_dir != "":
trainer.extend(TensorboardLogger(SummaryWriter(args.tensorboard_dir), att_reporter),
trigger=(args.report_interval_iters, "iteration"))
# Run the training
trainer.run()
check_early_stop(trainer, args.epochs)
def train(args):
"""Train with the given args.
Args:
args (namespace): The program arguments.
"""
set_deterministic_pytorch(args)
# check cuda availability
if not torch.cuda.is_available():
logging.warning("cuda is not available")
# get input and output dimension info
with open(args.valid_json, "rb") as f:
valid_json = json.load(f)["utts"]
utts = list(valid_json.keys())
idim = int(valid_json[utts[0]]["output"][1]["shape"][1])
odim = int(valid_json[utts[0]]["output"][0]["shape"][1])
logging.info("#input dims : " + str(idim))
logging.info("#output dims: " + str(odim))
# specify model architecture
model_class = dynamic_import(args.model_module)
model = model_class(idim, odim, args)
assert isinstance(model, MTInterface)
# write model config
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
model_conf = args.outdir + "/model.json"
with open(model_conf, "wb") as f:
logging.info("writing a model config file to " + model_conf)
f.write(
json.dumps((idim, odim, vars(args)),
indent=4,
ensure_ascii=False,
sort_keys=True).encode("utf_8"))
for key in sorted(vars(args).keys()):
logging.info("ARGS: " + key + ": " + str(vars(args)[key]))
reporter = model.reporter
# check the use of multi-gpu
if args.ngpu > 1:
if args.batch_size != 0:
logging.warning(
"batch size is automatically increased (%d -> %d)" %
(args.batch_size, args.batch_size * args.ngpu))
args.batch_size *= args.ngpu
# set torch device
device = torch.device("cuda" if args.ngpu > 0 else "cpu")
if args.train_dtype in ("float16", "float32", "float64"):
dtype = getattr(torch, args.train_dtype)
else:
dtype = torch.float32
model = model.to(device=device, dtype=dtype)
logging.warning(
"num. model params: {:,} (num. trained: {:,} ({:.1f}%))".format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
sum(p.numel() for p in model.parameters() if p.requires_grad) *
100.0 / sum(p.numel() for p in model.parameters()),
))
# Setup an optimizer
if args.opt == "adadelta":
optimizer = torch.optim.Adadelta(model.parameters(),
rho=0.95,
eps=args.eps,
weight_decay=args.weight_decay)
elif args.opt == "adam":
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.opt == "noam":
from espnet.nets.pytorch_backend.transformer.optimizer import get_std_opt
optimizer = get_std_opt(
model.parameters(),
args.adim,
args.transformer_warmup_steps,
args.transformer_lr,
)
else:
raise NotImplementedError("unknown optimizer: " + args.opt)
# setup apex.amp
if args.train_dtype in ("O0", "O1", "O2", "O3"):
try:
from apex import amp
except ImportError as e:
logging.error(
f"You need to install apex for --train-dtype {args.train_dtype}. "
"See https://github.com/NVIDIA/apex#linux")
raise e
if args.opt == "noam":
model, optimizer.optimizer = amp.initialize(
model, optimizer.optimizer, opt_level=args.train_dtype)
else:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.train_dtype)
use_apex = True
else:
use_apex = False
# FIXME: TOO DIRTY HACK
setattr(optimizer, "target", reporter)
setattr(optimizer, "serialize", lambda s: reporter.serialize(s))
# Setup a converter
converter = CustomConverter()
# read json data
with open(args.train_json, "rb") as f:
train_json = json.load(f)["utts"]
with open(args.valid_json, "rb") as f:
valid_json = json.load(f)["utts"]
use_sortagrad = args.sortagrad == -1 or args.sortagrad > 0
# make minibatch list (variable length)
train = make_batchset(
train_json,
args.batch_size,
args.maxlen_in,
args.maxlen_out,
args.minibatches,
min_batch_size=args.ngpu if args.ngpu > 1 else 1,
shortest_first=use_sortagrad,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout,
mt=True,
iaxis=1,
oaxis=0,
)
valid = make_batchset(
valid_json,
args.batch_size,
args.maxlen_in,
args.maxlen_out,
args.minibatches,
min_batch_size=args.ngpu if args.ngpu > 1 else 1,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout,
mt=True,
iaxis=1,
oaxis=0,
)
load_tr = LoadInputsAndTargets(mode="mt", load_output=True)
load_cv = LoadInputsAndTargets(mode="mt", load_output=True)
# hack to make batchsize argument as 1
# actual bathsize is included in a list
# default collate function converts numpy array to pytorch tensor
# we used an empty collate function instead which returns list
train_iter = ChainerDataLoader(
dataset=TransformDataset(train,
lambda data: converter([load_tr(data)])),
batch_size=1,
num_workers=args.n_iter_processes,
shuffle=not use_sortagrad,
collate_fn=lambda x: x[0],
)
valid_iter = ChainerDataLoader(
dataset=TransformDataset(valid,
lambda data: converter([load_cv(data)])),
batch_size=1,
shuffle=False,
collate_fn=lambda x: x[0],
num_workers=args.n_iter_processes,
)
# Set up a trainer
updater = CustomUpdater(
model,
args.grad_clip,
{"main": train_iter},
optimizer,
device,
args.ngpu,
False,
args.accum_grad,
use_apex=use_apex,
)
trainer = training.Trainer(updater, (args.epochs, "epoch"),
out=args.outdir)
if use_sortagrad:
trainer.extend(
ShufflingEnabler([train_iter]),
trigger=(args.sortagrad if args.sortagrad != -1 else args.epochs,
"epoch"),
)
# Resume from a snapshot
if args.resume:
logging.info("resumed from %s" % args.resume)
torch_resume(args.resume, trainer)
# Evaluate the model with the test dataset for each epoch
if args.save_interval_iters > 0:
trainer.extend(
CustomEvaluator(model, {"main": valid_iter}, reporter, device,
args.ngpu),
trigger=(args.save_interval_iters, "iteration"),
)
else:
trainer.extend(
CustomEvaluator(model, {"main": valid_iter}, reporter, device,
args.ngpu))
# Save attention weight each epoch
if args.num_save_attention > 0:
# NOTE: sort it by output lengths
data = sorted(
list(valid_json.items())[:args.num_save_attention],
key=lambda x: int(x[1]["output"][0]["shape"][0]),
reverse=True,
)
if hasattr(model, "module"):
att_vis_fn = model.module.calculate_all_attentions
plot_class = model.module.attention_plot_class
else:
att_vis_fn = model.calculate_all_attentions
plot_class = model.attention_plot_class
att_reporter = plot_class(
att_vis_fn,
data,
args.outdir + "/att_ws",
converter=converter,
transform=load_cv,
device=device,
ikey="output",
iaxis=1,
)
trainer.extend(att_reporter, trigger=(1, "epoch"))
else:
att_reporter = None
# Make a plot for training and validation values
trainer.extend(
extensions.PlotReport(["main/loss", "validation/main/loss"],
"epoch",
file_name="loss.png"))
trainer.extend(
extensions.PlotReport(["main/acc", "validation/main/acc"],
"epoch",
file_name="acc.png"))
trainer.extend(
extensions.PlotReport(["main/ppl", "validation/main/ppl"],
"epoch",
file_name="ppl.png"))
trainer.extend(
extensions.PlotReport(["main/bleu", "validation/main/bleu"],
"epoch",
file_name="bleu.png"))
# Save best models
trainer.extend(
snapshot_object(model, "model.loss.best"),
trigger=training.triggers.MinValueTrigger("validation/main/loss"),
)
trainer.extend(
snapshot_object(model, "model.acc.best"),
trigger=training.triggers.MaxValueTrigger("validation/main/acc"),
)
# save snapshot which contains model and optimizer states
if args.save_interval_iters > 0:
trainer.extend(
torch_snapshot(filename="snapshot.iter.{.updater.iteration}"),
trigger=(args.save_interval_iters, "iteration"),
)
else:
trainer.extend(torch_snapshot(), trigger=(1, "epoch"))
# epsilon decay in the optimizer
if args.opt == "adadelta":
if args.criterion == "acc":
trainer.extend(
restore_snapshot(model,
args.outdir + "/model.acc.best",
load_fn=torch_load),
trigger=CompareValueTrigger(
"validation/main/acc",
lambda best_value, current_value: best_value >
current_value,
),
)
trainer.extend(
adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
"validation/main/acc",
lambda best_value, current_value: best_value >
current_value,
),
)
elif args.criterion == "loss":
trainer.extend(
restore_snapshot(model,
args.outdir + "/model.loss.best",
load_fn=torch_load),
trigger=CompareValueTrigger(
"validation/main/loss",
lambda best_value, current_value: best_value <
current_value,
),
)
trainer.extend(
adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
"validation/main/loss",
lambda best_value, current_value: best_value <
current_value,
),
)
elif args.opt == "adam":
if args.criterion == "acc":
trainer.extend(
restore_snapshot(model,
args.outdir + "/model.acc.best",
load_fn=torch_load),
trigger=CompareValueTrigger(
"validation/main/acc",
lambda best_value, current_value: best_value >
current_value,
),
)
trainer.extend(
adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
"validation/main/acc",
lambda best_value, current_value: best_value >
current_value,
),
)
elif args.criterion == "loss":
trainer.extend(
restore_snapshot(model,
args.outdir + "/model.loss.best",
load_fn=torch_load),
trigger=CompareValueTrigger(
"validation/main/loss",
lambda best_value, current_value: best_value <
current_value,
),
)
trainer.extend(
adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
"validation/main/loss",
lambda best_value, current_value: best_value <
current_value,
),
)
# Write a log of evaluation statistics for each epoch
trainer.extend(
extensions.LogReport(trigger=(args.report_interval_iters,
"iteration")))
report_keys = [
"epoch",
"iteration",
"main/loss",
"validation/main/loss",
"main/acc",
"validation/main/acc",
"main/ppl",
"validation/main/ppl",
"elapsed_time",
]
if args.opt == "adadelta":
trainer.extend(
extensions.observe_value(
"eps",
lambda trainer: trainer.updater.get_optimizer("main").
param_groups[0]["eps"],
),
trigger=(args.report_interval_iters, "iteration"),
)
report_keys.append("eps")
elif args.opt in ["adam", "noam"]:
trainer.extend(
extensions.observe_value(
"lr",
lambda trainer: trainer.updater.get_optimizer("main").
param_groups[0]["lr"],
),
trigger=(args.report_interval_iters, "iteration"),
)
report_keys.append("lr")
if args.report_bleu:
report_keys.append("main/bleu")
report_keys.append("validation/main/bleu")
trainer.extend(
extensions.PrintReport(report_keys),
trigger=(args.report_interval_iters, "iteration"),
)
trainer.extend(
extensions.ProgressBar(update_interval=args.report_interval_iters))
set_early_stop(trainer, args)
if args.tensorboard_dir is not None and args.tensorboard_dir != "":
from torch.utils.tensorboard import SummaryWriter
trainer.extend(
TensorboardLogger(SummaryWriter(args.tensorboard_dir),
att_reporter),
trigger=(args.report_interval_iters, "iteration"),
)
# Run the training
trainer.run()
check_early_stop(trainer, args.epochs)
def train(args):
"""Train with the given args.
Args:
args (namespace): The program arguments.
"""
set_deterministic_pytorch(args)
# check cuda availability
if not torch.cuda.is_available():
logging.warning('cuda is not available')
# get paths to data
lang_pairs = sorted(args.lang_pairs.split(','))
args.one_to_many = True if len(lang_pairs) > 1 else False
tgt_langs = sorted([p.split('-')[-1] for p in lang_pairs])
src_lang = lang_pairs[0].split('-')[0]
if args.one_to_many:
train_jpaths = [
os.path.join(args.train_json, fname)
for fname in sorted(os.listdir(args.train_json))
if fname.endswith('.json')
]
valid_jpaths = [
os.path.join(args.valid_json, fname)
for fname in sorted(os.listdir(args.valid_json))
if fname.endswith('.json')
]
all_langs = list(
sorted(set([l for p in lang_pairs for l in p.split('-')])))
args.langs_dict = {}
offset = 2 # for <blank> and <unk>
for i, lang in enumerate(all_langs):
args.langs_dict[f'<2{lang}>'] = offset + i
logging.info(f'| train_jpaths: {train_jpaths}')
logging.info(f'| valid_jpaths: {valid_jpaths}')
logging.info(f'| lang_pairs : {lang_pairs}')
logging.info(f'| langs_dict : {args.langs_dict}')
else:
train_jpaths = [args.train_json]
valid_jpaths = [args.valid_json]
args.langs_dict = None
# get input and output dimension info
idim = 0
odim = 0
for i, jpath in enumerate(valid_jpaths):
with open(jpath, 'rb') as f:
valid_json = json.load(f)['utts']
utts = list(valid_json.keys())
idim_tmp = int(valid_json[utts[0]]['input'][0]['shape'][-1])
odim_tmp = int(valid_json[utts[0]]['output'][0]['shape'][-1])
logging.info('| pair {}: idim={}, odim={}'.format(
lang_pairs[i], idim_tmp, odim_tmp))
if idim == 0:
idim = idim_tmp
else:
assert idim == idim_tmp
if odim < odim_tmp:
odim = odim_tmp
logging.info('#input dims : ' + str(idim))
logging.info('#output dims: ' + str(odim))
# Initialize with pre-trained ASR encoder and MT decoder
if args.enc_init is not None or args.dec_init is not None:
logging.info('Loading pretrained ASR encoder and/or MT decoder ...')
model = load_trained_modules(idim, odim, args, interface=STInterface)
logging.info(f'*** Model *** \n {model}')
else:
model_class = dynamic_import(args.model_module)
model = model_class(idim, odim, args)
logging.info(f'*** Model *** \n {model}')
assert isinstance(model, STInterface)
logging.info(
f'| Number of model parameters: {sum(p.numel() for p in model.parameters() if p.requires_grad)}'
)
subsampling_factor = model.subsample[0]
logging.info(f'subsampling_factor={subsampling_factor}')
if args.rnnlm is not None:
rnnlm_args = get_model_conf(args.rnnlm, args.rnnlm_conf)
rnnlm = lm_pytorch.ClassifierWithState(
lm_pytorch.RNNLM(
len(args.char_list),
rnnlm_args.layer,
rnnlm_args.unit,
getattr(rnnlm_args, "embed_unit",
None), # for backward compatibility
))
torch_load(args.rnnlm, rnnlm)
model.rnnlm = rnnlm
# write model config
if not os.path.exists(args.outdir):
os.makedirs(args.outdir)
model_conf = args.outdir + '/model.json'
with open(model_conf, 'wb') as f:
logging.info('writing a model config file to ' + model_conf)
f.write(
json.dumps((idim, odim, vars(args)),
indent=4,
ensure_ascii=False,
sort_keys=True).encode('utf_8'))
for key in sorted(vars(args).keys()):
logging.info('ARGS: ' + key + ': ' + str(vars(args)[key]))
reporter = model.reporter
# check the use of multi-gpu
if args.ngpu > 1:
if args.batch_size != 0:
logging.warning(
'batch size is automatically increased (%d -> %d)' %
(args.batch_size, args.batch_size * args.ngpu))
args.batch_size *= args.ngpu
# set torch device
device = torch.device("cuda" if args.ngpu > 0 else "cpu")
if args.train_dtype in ("float16", "float32", "float64"):
dtype = getattr(torch, args.train_dtype)
else:
dtype = torch.float32
model = model.to(device=device, dtype=dtype)
# Setup an optimizer
if args.opt == 'adadelta':
optimizer = torch.optim.Adadelta(model.parameters(),
rho=0.95,
eps=args.eps,
weight_decay=args.weight_decay)
elif args.opt == 'adam':
optimizer = torch.optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
elif args.opt == 'noam':
from espnet.nets.pytorch_backend.transformer.optimizer import get_std_opt
optimizer = get_std_opt(model, args.adim,
args.transformer_warmup_steps,
args.transformer_lr)
else:
raise NotImplementedError("unknown optimizer: " + args.opt)
# setup apex.amp
if args.train_dtype in ("O0", "O1", "O2", "O3"):
try:
from apex import amp
except ImportError as e:
logging.error(
f"You need to install apex for --train-dtype {args.train_dtype}. "
"See https://github.com/NVIDIA/apex#linux")
raise e
if args.opt == 'noam':
model, optimizer.optimizer = amp.initialize(
model, optimizer.optimizer, opt_level=args.train_dtype)
else:
model, optimizer = amp.initialize(model,
optimizer,
opt_level=args.train_dtype)
use_apex = True
else:
use_apex = False
# FIXME: TOO DIRTY HACK
setattr(optimizer, "target", reporter)
setattr(optimizer, "serialize", lambda s: reporter.serialize(s))
use_sortagrad = args.sortagrad == -1 or args.sortagrad > 0
logging.info(f'use_sortagrad: {use_sortagrad}')
# read json data
num_langs = len(tgt_langs)
train_all_pairs = [None] * num_langs
valid_all_pairs = [None] * num_langs
# check_data = {}
batch_size = args.batch_size // num_langs if num_langs > 1 else args.batch_size
for i, jpath in enumerate(train_jpaths):
with open(jpath, 'rb') as f:
train_json = json.load(f)['utts']
train_all_pairs[i] = make_batchset(
train_json,
batch_size,
args.maxlen_in,
args.maxlen_out,
args.minibatches,
min_batch_size=1,
shortest_first=use_sortagrad,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout)
# check_data[lang_pairs[i]] = list(train_json.keys())
for i, jpath in enumerate(valid_jpaths):
with open(jpath, 'rb') as f:
valid_json = json.load(f)['utts']
valid_all_pairs[i] = make_batchset(
valid_json,
batch_size,
args.maxlen_in,
args.maxlen_out,
args.minibatches,
min_batch_size=1,
count=args.batch_count,
batch_bins=args.batch_bins,
batch_frames_in=args.batch_frames_in,
batch_frames_out=args.batch_frames_out,
batch_frames_inout=args.batch_frames_inout)
# check_data[lang_pairs[i]] = list(valid_json.keys())
# print(f'len(train_all_pairs) = {len(train_all_pairs)}')
# print(f'len(valid_all_pairs) = {len(valid_all_pairs)}')
# for i, batch_langs in enumerate(train_all_pairs):
# print(f'batch for lang {lang_pairs[i]}')
# for batch_lang in batch_langs:
# print(f'len(batch_lang) = {len(batch_lang)}')
# print('-'*5)
if num_langs > 1:
cycle_train = [cycle(x) for x in train_all_pairs]
cycle_valid = [cycle(x) for x in valid_all_pairs]
num_batches_train = max(len(i) for i in train_all_pairs)
num_batches_valid = max(len(i) for i in valid_all_pairs)
train = [None] * num_batches_train
valid = [None] * num_batches_valid
for i, s in enumerate(zip(*cycle_train)):
x = []
for y in s:
x.extend(y)
train[i] = x
if i >= num_batches_train - 1:
break
for i, s in enumerate(zip(*cycle_valid)):
x = []
for y in s:
x.extend(y)
valid[i] = x
if i >= num_batches_valid - 1:
break
else:
train = train_all_pairs[0]
valid = valid_all_pairs[0]
# print(f'num_batches_train = {num_batches_train}')
# print(f'num_batches_valid = {num_batches_valid}')
# print(f'len(train) = {len(train)}')
# print(f'len(valid) = {len(valid)}')
# print('*** Checking results of make_batchset() ***')
# for i, batch in enumerate(train):
# # if i == 0:
# # print(batch)
# ids = [sample[0] for sample in batch]
# langs = [sample[1]['lang'] for sample in batch]
# pairs = ['en-'+l for l in langs]
# for i in range(len(ids)):
# r = ids[i] in list(check_data[pairs[i]])
# print(f'ids[i]={ids[i]} in {check_data[pairs[i]]}: {r}')
# print('-')
# if r:
# check_data[pairs[i]].remove(ids[i])
# print(f'len(batch) = {len(batch)}')
# print(f'langs in batch: {langs}')
# print('-'*5)
# # if i > 5:
# # break
# print('*** Samples that are not used yet ***')
# for k, v in check_data.items():
# print(k, v)
# print('-'*5)
# print('-'*20)
load_tr = LoadInputsAndTargets(mode='asr',
load_output=True,
preprocess_conf=args.preprocess_conf,
preprocess_args={'train': True},
langs_dict=args.langs_dict,
src_lang=src_lang)
load_cv = LoadInputsAndTargets(mode='asr',
load_output=True,
preprocess_conf=args.preprocess_conf,
preprocess_args={'train': False},
langs_dict=args.langs_dict,
src_lang=src_lang)
# print('LoadInputsAndTargets()')
# features, targets = load_cv(train[0])
# print(f'*** features: {features} ***')
# for f in features:
# # print(f)
# print(f'len(f) = {len(f)}')
# print('---')
# print(f'*** targets : {targets} ***')
# y1, y2 = zip(*targets)
# # print(f'y1 = {y1}')
# # print(f'y2 = {y2}')
# for s in zip(y1, y2):
# print(len(s[0][1]), len(s[1][1]))
# print('-'*20)
# Setup a converter
converter = CustomConverter(subsampling_factor=subsampling_factor,
dtype=dtype,
asr_task=args.asr_weight > 0)
# hack to make batchsize argument as 1
# actual bathsize is included in a list
# default collate function converts numpy array to pytorch tensor
# we used an empty collate function instead which returns list
n_iter_processes = args.n_iter_processes
if n_iter_processes < 0:
n_iter_processes = multiprocessing.cpu_count()
elif n_iter_processes > 0:
n_iter_processes = min(n_iter_processes, multiprocessing.cpu_count())
print(f'n_iter_processes = {n_iter_processes}')
train_iter = {
'main':
ChainerDataLoader(dataset=TransformDataset(
train, lambda data: converter([load_tr(data)])),
batch_size=1,
num_workers=n_iter_processes,
shuffle=not use_sortagrad,
collate_fn=lambda x: x[0],
pin_memory=False)
}
valid_iter = {
'main':
ChainerDataLoader(dataset=TransformDataset(
valid, lambda data: converter([load_cv(data)])),
batch_size=1,
shuffle=False,
collate_fn=lambda x: x[0],
num_workers=n_iter_processes,
pin_memory=False)
}
# xs_pad, ilens, ys_pad, ys_pad_asr = converter([load_cv(valid[0])])
# print('*** xs_pad ***')
# # print(xs_pad)
# print(xs_pad.size())
# print('*** ilens ***')
# print(ilens)
# print('*** ys_pad ***')
# # print(ys_pad)
# print(ys_pad.size())
# print('*** ys_pad_asr ***')
# print(ys_pad_asr)
# print('-'*20)
# print(train_iter['main'])
# i=0
# for item in train_iter['main']:
# print(item)
# print('-'*5)
# if i > 8:
# break
# i += 1
# Set up a trainer
updater = CustomUpdater(model,
args.grad_clip,
train_iter,
optimizer,
device,
args.ngpu,
args.grad_noise,
args.accum_grad,
use_apex=use_apex)
# trainer = training.Trainer(
# updater, (args.epochs, 'epoch'), out=args.outdir)
time_limit_trigger = TimeLimitTrigger(args)
trainer = training.Trainer(updater, time_limit_trigger, out=args.outdir)
logging.info(f'updater: {updater}')
logging.info(f'trainer: {trainer}')
if use_sortagrad:
logging.info(f'use_sortagrad ...')
trainer.extend(
ShufflingEnabler([train_iter]),
trigger=(args.sortagrad if args.sortagrad != -1 else args.epochs,
'epoch'))
# Evaluate the model with the test dataset for each epoch
if args.save_interval_iters > 0:
trainer.extend(CustomEvaluator(model, valid_iter, reporter, device,
args.ngpu),
trigger=(args.save_interval_iters, 'iteration'))
else:
trainer.extend(
CustomEvaluator(model, valid_iter, reporter, device, args.ngpu))
# Save attention weight each epoch
if args.num_save_attention > 0:
data = sorted(list(valid_json.items())[:args.num_save_attention],
key=lambda x: int(x[1]['input'][0]['shape'][1]),
reverse=True)
if hasattr(model, "module"):
att_vis_fn = model.module.calculate_all_attentions
plot_class = model.module.attention_plot_class
else:
att_vis_fn = model.calculate_all_attentions
plot_class = model.attention_plot_class
att_reporter = plot_class(att_vis_fn,
data,
args.outdir + "/att_ws",
converter=converter,
transform=load_cv,
device=device)
trainer.extend(att_reporter, trigger=(1, 'epoch'))
else:
att_reporter = None
# Make a plot for training and validation values
trainer.extend(
extensions.PlotReport([
'main/loss', 'validation/main/loss', 'main/loss_asr',
'validation/main/loss_asr', 'main/loss_st',
'validation/main/loss_st'
],
'epoch',
file_name='loss.png'))
trainer.extend(
extensions.PlotReport([
'main/acc', 'validation/main/acc', 'main/acc_asr',
'validation/main/acc_asr'
],
'epoch',
file_name='acc.png'))
trainer.extend(
extensions.PlotReport(['main/bleu', 'validation/main/bleu'],
'epoch',
file_name='bleu.png'))
# Save best models
if args.report_interval_iters > 0:
trainer.extend(snapshot_object(model, 'model.loss.best'),
trigger=MinValueTrigger(
'validation/main/loss',
trigger=(args.report_interval_iters, 'iteration'),
best_value=None))
trainer.extend(snapshot_object(model, 'model.acc.best'),
trigger=MaxValueTrigger(
'validation/main/acc',
trigger=(args.report_interval_iters, 'iteration'),
best_value=None))
else:
trainer.extend(snapshot_object(model, 'model.loss.best'),
trigger=MinValueTrigger('validation/main/loss',
best_value=None))
trainer.extend(snapshot_object(model, 'model.acc.best'),
trigger=MaxValueTrigger('validation/main/acc',
best_value=None))
# save snapshot which contains model and optimizer states
if args.save_interval_iters > 0:
trainer.extend(
torch_snapshot(filename='snapshot.iter.{.updater.iteration}'),
trigger=(args.save_interval_iters, 'iteration'))
else:
trainer.extend(torch_snapshot(), trigger=(1, 'epoch'))
# epsilon decay in the optimizer
if args.opt == 'adadelta':
if args.criterion == 'acc':
trainer.extend(restore_snapshot(model,
args.outdir + '/model.acc.best',
load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/acc', lambda best_value,
current_value: best_value > current_value))
trainer.extend(adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
'validation/main/acc', lambda best_value,
current_value: best_value > current_value))
elif args.criterion == 'loss':
trainer.extend(restore_snapshot(model,
args.outdir + '/model.loss.best',
load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/loss', lambda best_value,
current_value: best_value < current_value))
trainer.extend(adadelta_eps_decay(args.eps_decay),
trigger=CompareValueTrigger(
'validation/main/loss', lambda best_value,
current_value: best_value < current_value))
elif args.opt == 'adam':
if args.criterion == 'acc':
trainer.extend(restore_snapshot(model,
args.outdir + '/model.acc.best',
load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/acc', lambda best_value,
current_value: best_value > current_value))
trainer.extend(adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
'validation/main/acc', lambda best_value,
current_value: best_value > current_value))
elif args.criterion == 'loss':
trainer.extend(restore_snapshot(model,
args.outdir + '/model.loss.best',
load_fn=torch_load),
trigger=CompareValueTrigger(
'validation/main/loss', lambda best_value,
current_value: best_value < current_value))
trainer.extend(adam_lr_decay(args.lr_decay),
trigger=CompareValueTrigger(
'validation/main/loss', lambda best_value,
current_value: best_value < current_value))
# Write a log of evaluation statistics for each epoch
trainer.extend(
extensions.LogReport(trigger=(args.report_interval_iters,
'iteration')))
report_keys = [
'epoch', 'iteration', 'main/loss', 'main/loss_st', 'main/loss_asr',
'validation/main/loss', 'validation/main/loss_st',
'validation/main/loss_asr', 'main/acc', 'validation/main/acc'
]
if args.asr_weight > 0:
report_keys.append('main/acc_asr')
report_keys.append('validation/main/acc_asr')
report_keys += ['elapsed_time']
if args.opt == 'adadelta':
trainer.extend(extensions.observe_value(
'eps', lambda trainer: trainer.updater.get_optimizer('main').
param_groups[0]["eps"]),
trigger=(args.report_interval_iters, 'iteration'))
report_keys.append('eps')
elif args.opt in ['adam', 'noam']:
trainer.extend(extensions.observe_value(
'lr', lambda trainer: trainer.updater.get_optimizer('main').
param_groups[0]["lr"]),
trigger=(args.report_interval_iters, 'iteration'))
report_keys.append('lr')
if args.asr_weight > 0:
if args.mtlalpha > 0:
report_keys.append('main/cer_ctc')
report_keys.append('validation/main/cer_ctc')
if args.mtlalpha < 1:
if args.report_cer:
report_keys.append('validation/main/cer')
if args.report_wer:
report_keys.append('validation/main/wer')
if args.report_bleu:
report_keys.append('validation/main/bleu')
trainer.extend(extensions.PrintReport(report_keys),
trigger=(args.report_interval_iters, 'iteration'))
trainer.extend(
extensions.ProgressBar(update_interval=args.report_interval_iters))
set_early_stop(trainer, args)
if args.tensorboard_dir is not None and args.tensorboard_dir != "":
trainer.extend(TensorboardLogger(SummaryWriter(args.tensorboard_dir),
att_reporter),
trigger=(args.report_interval_iters, "iteration"))
# Resume from a snapshot
if args.resume:
logging.info('resumed from %s' % args.resume)
torch_resume(args.resume, trainer)
# Run the training
trainer.run()
check_early_stop(trainer, args.epochs)
