def create_trainer(
config_dict: Dict[str, Any],
output: Path,
):
# config
config = Config.from_dict(config_dict)
config.add_git_info()
assert_config(config)
output.mkdir(exist_ok=True, parents=True)
with (output / "config.yaml").open(mode="w") as f:
yaml.safe_dump(config.to_dict(), f)
# model
predictor = create_predictor(config.network)
model = Model(
loss_config=config.loss,
predictor=predictor,
local_padding_size=config.dataset.local_padding_size,
)
if config.train.weight_initializer is not None:
init_weights(model, name=config.train.weight_initializer)
device = torch.device("cuda")
model.to(device)
# dataset
_create_iterator = partial(
create_iterator,
batch_size=config.train.batchsize,
eval_batch_size=config.train.eval_batchsize,
num_processes=config.train.num_processes,
use_multithread=config.train.use_multithread,
)
datasets = create_dataset(config.dataset)
train_iter = _create_iterator(datasets["train"],
for_train=True,
for_eval=False)
test_iter = _create_iterator(datasets["test"],
for_train=False,
for_eval=False)
eval_iter = _create_iterator(datasets["eval"],
for_train=False,
for_eval=True)
valid_iter = None
if datasets["valid"] is not None:
valid_iter = _create_iterator(datasets["valid"],
for_train=False,
for_eval=True)
warnings.simplefilter("error", MultiprocessIterator.TimeoutWarning)
# optimizer
cp: Dict[str, Any] = copy(config.train.optimizer)
n = cp.pop("name").lower()
optimizer: Optimizer
if n == "adam":
optimizer = optim.Adam(model.parameters(), **cp)
elif n == "sgd":
optimizer = optim.SGD(model.parameters(), **cp)
else:
raise ValueError(n)
# updater
if not config.train.use_amp:
updater = StandardUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
else:
updater = AmpUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
# trainer
trigger_log = (config.train.log_iteration, "iteration")
trigger_eval = (config.train.eval_iteration, "iteration")
trigger_snapshot = (config.train.snapshot_iteration, "iteration")
trigger_stop = ((config.train.stop_iteration, "iteration")
if config.train.stop_iteration is not None else None)
trainer = Trainer(updater, stop_trigger=trigger_stop, out=output)
shift_ext = None
if config.train.linear_shift is not None:
shift_ext = extensions.LinearShift(**config.train.linear_shift)
if config.train.step_shift is not None:
shift_ext = extensions.StepShift(**config.train.step_shift)
if shift_ext is not None:
trainer.extend(shift_ext)
ext = extensions.Evaluator(test_iter, model, device=device)
trainer.extend(ext, name="test", trigger=trigger_log)
generator = Generator(
config=config,
predictor=predictor,
use_gpu=True,
max_batch_size=(config.train.eval_batchsize
if config.train.eval_batchsize is not None else
config.train.batchsize),
use_fast_inference=False,
)
generate_evaluator = GenerateEvaluator(
generator=generator,
time_length=config.dataset.time_length_evaluate,
local_padding_time_length=config.dataset.
local_padding_time_length_evaluate,
)
ext = extensions.Evaluator(eval_iter, generate_evaluator, device=device)
trainer.extend(ext, name="eval", trigger=trigger_eval)
if valid_iter is not None:
ext = extensions.Evaluator(valid_iter,
generate_evaluator,
device=device)
trainer.extend(ext, name="valid", trigger=trigger_eval)
if config.train.stop_iteration is not None:
saving_model_num = int(config.train.stop_iteration /
config.train.eval_iteration / 10)
else:
saving_model_num = 10
ext = extensions.snapshot_object(
predictor,
filename="predictor_{.updater.iteration}.pth",
n_retains=saving_model_num,
)
trainer.extend(
ext,
trigger=LowValueTrigger("eval/main/mcd", trigger=trigger_eval),
)
trainer.extend(extensions.FailOnNonNumber(), trigger=trigger_log)
trainer.extend(extensions.observe_lr(), trigger=trigger_log)
trainer.extend(extensions.LogReport(trigger=trigger_log))
trainer.extend(
extensions.PrintReport(["iteration", "main/loss", "test/main/loss"]),
trigger=trigger_log,
)
trainer.extend(TensorboardReport(writer=SummaryWriter(Path(output))),
trigger=trigger_log)
if config.project.category is not None:
ext = WandbReport(
config_dict=config.to_dict(),
project_category=config.project.category,
project_name=config.project.name,
output_dir=output.joinpath("wandb"),
)
trainer.extend(ext, trigger=trigger_log)
(output / "struct.txt").write_text(repr(model))
if trigger_stop is not None:
trainer.extend(extensions.ProgressBar(trigger_stop))
ext = extensions.snapshot_object(
trainer,
filename="trainer_{.updater.iteration}.pth",
n_retains=1,
autoload=True,
)
trainer.extend(ext, trigger=trigger_snapshot)
return trainer
def create_trainer(
config_dict: Dict[str, Any],
output: Path,
):
# config
config = Config.from_dict(config_dict)
config.add_git_info()
output.mkdir(exist_ok=True, parents=True)
with (output / "config.yaml").open(mode="w") as f:
yaml.safe_dump(config.to_dict(), f)
# model
networks = create_network(config.network)
model = Model(model_config=config.model, networks=networks)
if config.train.weight_initializer is not None:
init_weights(model, name=config.train.weight_initializer)
device = torch.device("cuda") if config.train.use_gpu else torch.device(
"cpu")
model.to(device)
# dataset
_create_iterator = partial(
create_iterator,
batch_size=config.train.batch_size,
eval_batch_size=config.train.eval_batch_size,
num_processes=config.train.num_processes,
use_multithread=config.train.use_multithread,
)
datasets = create_dataset(config.dataset)
train_iter = _create_iterator(datasets["train"], for_train=True)
test_iter = _create_iterator(datasets["test"], for_train=False)
eval_iter = _create_iterator(datasets["eval"], for_train=False)
warnings.simplefilter("error", MultiprocessIterator.TimeoutWarning)
# optimizer
optimizer = make_optimizer(config_dict=config.train.optimizer, model=model)
# updater
if not config.train.use_amp:
updater = StandardUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
else:
updater = AmpUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
# trainer
trigger_log = (config.train.log_iteration, "iteration")
trigger_eval = (config.train.eval_iteration, "iteration")
trigger_snapshot = (config.train.snapshot_iteration, "iteration")
trigger_stop = ((config.train.stop_iteration, "iteration")
if config.train.stop_iteration is not None else None)
trainer = Trainer(updater, stop_trigger=trigger_stop, out=output)
ext = extensions.Evaluator(test_iter, model, device=device)
trainer.extend(ext, name="test", trigger=trigger_log)
if config.train.stop_iteration is not None:
saving_model_num = int(config.train.stop_iteration /
config.train.eval_iteration / 10)
else:
saving_model_num = 10
ext = extensions.snapshot_object(
networks.predictor,
filename="predictor_{.updater.iteration}.pth",
n_retains=saving_model_num,
)
trainer.extend(
ext,
trigger=LowValueTrigger("test/main/loss", trigger=trigger_eval),
)
trainer.extend(extensions.FailOnNonNumber(), trigger=trigger_log)
trainer.extend(extensions.observe_lr(), trigger=trigger_log)
trainer.extend(extensions.LogReport(trigger=trigger_log))
trainer.extend(
extensions.PrintReport(["iteration", "main/loss", "test/main/loss"]),
trigger=trigger_log,
)
ext = TensorboardReport(writer=SummaryWriter(Path(output)))
trainer.extend(ext, trigger=trigger_log)
if config.project.category is not None:
ext = WandbReport(
config_dict=config.to_dict(),
project_category=config.project.category,
project_name=config.project.name,
output_dir=output.joinpath("wandb"),
)
trainer.extend(ext, trigger=trigger_log)
(output / "struct.txt").write_text(repr(model))
if trigger_stop is not None:
trainer.extend(extensions.ProgressBar(trigger_stop))
ext = extensions.snapshot_object(
trainer,
filename="trainer_{.updater.iteration}.pth",
n_retains=1,
autoload=True,
)
trainer.extend(ext, trigger=trigger_snapshot)
return trainer
def create_trainer(
config_dict: Dict[str, Any],
output: Path,
):
# config
config = Config.from_dict(config_dict)
config.add_git_info()
output.mkdir(exist_ok=True, parents=True)
with (output / "config.yaml").open(mode="w") as f:
yaml.safe_dump(config.to_dict(), f)
# model
device = torch.device("cuda")
predictor = create_predictor(config.network)
model = Model(
model_config=config.model,
predictor=predictor,
local_padding_length=config.dataset.local_padding_length,
)
init_weights(model, "orthogonal")
model.to(device)
# dataset
_create_iterator = partial(
create_iterator,
batch_size=config.train.batchsize,
eval_batch_size=config.train.eval_batchsize,
num_processes=config.train.num_processes,
use_multithread=config.train.use_multithread,
)
datasets = create_dataset(config.dataset)
train_iter = _create_iterator(datasets["train"],
for_train=True,
for_eval=False)
test_iter = _create_iterator(datasets["test"],
for_train=False,
for_eval=False)
eval_iter = _create_iterator(datasets["eval"],
for_train=False,
for_eval=True)
warnings.simplefilter("error", MultiprocessIterator.TimeoutWarning)
# optimizer
cp: Dict[str, Any] = copy(config.train.optimizer)
n = cp.pop("name").lower()
optimizer: Optimizer
if n == "adam":
optimizer = optim.Adam(model.parameters(), **cp)
elif n == "sgd":
optimizer = optim.SGD(model.parameters(), **cp)
else:
raise ValueError(n)
# updater
use_amp = config.train.use_amp if config.train.use_amp is not None else amp_exist
if use_amp:
updater = AmpUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
else:
updater = StandardUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
device=device,
)
# trainer
trigger_log = (config.train.log_iteration, "iteration")
trigger_eval = (config.train.eval_iteration, "iteration")
trigger_stop = ((config.train.stop_iteration, "iteration")
if config.train.stop_iteration is not None else None)
trainer = Trainer(updater, stop_trigger=trigger_stop, out=output)
writer = SummaryWriter(Path(output))
# # error at randint
# sample_data = datasets["train"][0]
# writer.add_graph(
# model,
# input_to_model=(
# sample_data["wave"].unsqueeze(0).to(device),
# sample_data["local"].unsqueeze(0).to(device),
# sample_data["speaker_id"].unsqueeze(0).to(device)
# if predictor.with_speaker
# else None,
# ),
# )
if config.train.multistep_shift is not None:
trainer.extend(
extensions.MultistepShift(**config.train.multistep_shift))
if config.train.step_shift is not None:
trainer.extend(extensions.StepShift(**config.train.step_shift))
ext = extensions.Evaluator(test_iter, model, device=device)
trainer.extend(ext, name="test", trigger=trigger_log)
generator = Generator(
config=config,
noise_schedule_config=NoiseScheduleModelConfig(start=1e-4,
stop=0.05,
num=50),
predictor=predictor,
sampling_rate=config.dataset.sampling_rate,
use_gpu=True,
)
generate_evaluator = GenerateEvaluator(
generator=generator,
local_padding_time_second=config.dataset.
evaluate_local_padding_time_second,
)
ext = extensions.Evaluator(eval_iter, generate_evaluator, device=device)
trainer.extend(ext, name="eval", trigger=trigger_eval)
if config.train.stop_iteration is not None:
saving_model_num = int(config.train.stop_iteration /
config.train.eval_iteration / 10)
else:
saving_model_num = 10
ext = extensions.snapshot_object(
predictor,
filename="predictor_{.updater.iteration}.pth",
n_retains=saving_model_num,
)
trainer.extend(
ext,
trigger=LowValueTrigger("eval/main/mcd", trigger=trigger_eval),
)
trainer.extend(extensions.FailOnNonNumber(), trigger=trigger_log)
trainer.extend(extensions.observe_lr(), trigger=trigger_log)
trainer.extend(extensions.LogReport(trigger=trigger_log))
trainer.extend(
extensions.PrintReport(["iteration", "main/loss", "test/main/loss"]),
trigger=trigger_log,
)
trainer.extend(ext, trigger=TensorboardReport(writer=writer))
if config.project.category is not None:
ext = WandbReport(
config_dict=config.to_dict(),
project_category=config.project.category,
project_name=config.project.name,
output_dir=output.joinpath("wandb"),
)
trainer.extend(ext, trigger=trigger_log)
(output / "struct.txt").write_text(repr(model))
if trigger_stop is not None:
trainer.extend(extensions.ProgressBar(trigger_stop))
ext = extensions.snapshot_object(
trainer,
filename="trainer_{.updater.iteration}.pth",
n_retains=1,
autoload=True,
)
trainer.extend(ext, trigger=trigger_eval)
return trainer
def create_trainer(
config_dict: Dict[str, Any],
output: Path,
):
# config
config = Config.from_dict(config_dict)
config.add_git_info()
output.mkdir(exist_ok=True, parents=True)
with (output / "config.yaml").open(mode="w") as f:
yaml.safe_dump(config.to_dict(), f)
# model
predictor = create_predictor(config.network)
model = Model(model_config=config.model, predictor=predictor)
if config.train.weight_initializer is not None:
init_weights(model, name=config.train.weight_initializer)
device = torch.device("cuda")
model.to(device)
# dataset
_create_iterator = partial(
create_iterator,
batch_size=config.train.batch_size,
num_processes=config.train.num_processes,
use_multithread=config.train.use_multithread,
)
datasets = create_dataset(config.dataset)
train_iter = _create_iterator(datasets["train"], for_train=True)
test_iter = _create_iterator(datasets["test"], for_train=False)
warnings.simplefilter("error", MultiprocessIterator.TimeoutWarning)
# optimizer
cp: Dict[str, Any] = copy(config.train.optimizer)
n = cp.pop("name").lower()
optimizer: Optimizer
if n == "adam":
optimizer = optim.Adam(model.parameters(), **cp)
elif n == "sgd":
optimizer = optim.SGD(model.parameters(), **cp)
else:
raise ValueError(n)
# updater
updater = StandardUpdater(
iterator=train_iter,
optimizer=optimizer,
model=model,
converter=list_concat,
device=device,
)
# trainer
trigger_log = (config.train.log_iteration, "iteration")
trigger_eval = (config.train.snapshot_iteration, "iteration")
trigger_stop = ((config.train.stop_iteration, "iteration")
if config.train.stop_iteration is not None else None)
trainer = Trainer(updater, stop_trigger=trigger_stop, out=output)
writer = SummaryWriter(Path(output))
sample_data = datasets["train"][0]
writer.add_graph(
model,
input_to_model=(
[sample_data["f0"].to(device)],
[sample_data["phoneme"].to(device)],
[sample_data["phoneme_list"].to(device)],
([sample_data["speaker_id"].to(device)]
if predictor.with_speaker else None),
),
)
ext = extensions.Evaluator(test_iter,
model,
converter=list_concat,
device=device)
trainer.extend(ext, name="test", trigger=trigger_log)
if config.train.stop_iteration is not None:
saving_model_num = int(config.train.stop_iteration /
config.train.snapshot_iteration / 10)
else:
saving_model_num = 10
ext = extensions.snapshot_object(
predictor,
filename="predictor_{.updater.iteration}.pth",
n_retains=saving_model_num,
)
trainer.extend(
ext,
trigger=LowValueTrigger("test/main/loss", trigger=trigger_eval),
)
trainer.extend(extensions.FailOnNonNumber(), trigger=trigger_log)
trainer.extend(extensions.observe_lr(), trigger=trigger_log)
trainer.extend(extensions.LogReport(trigger=trigger_log))
trainer.extend(
extensions.PrintReport(["iteration", "main/loss", "test/main/loss"]),
trigger=trigger_log,
)
ext = TensorboardReport(writer=writer)
trainer.extend(ext, trigger=trigger_log)
if config.project.category is not None:
ext = WandbReport(
config_dict=config.to_dict(),
project_category=config.project.category,
project_name=config.project.name,
output_dir=output.joinpath("wandb"),
)
trainer.extend(ext, trigger=trigger_log)
(output / "struct.txt").write_text(repr(model))
if trigger_stop is not None:
trainer.extend(extensions.ProgressBar(trigger_stop))
ext = extensions.snapshot_object(
trainer,
filename="trainer_{.updater.iteration}.pth",
n_retains=1,
autoload=True,
)
trainer.extend(ext, trigger=trigger_eval)
return trainer