def train_one_epoch(
cls,
model: torch.nn.Module,
iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
optimizers: Sequence[torch.optim.Optimizer],
schedulers: Sequence[Optional[AbsScheduler]],
scaler: Optional[GradScaler],
reporter: SubReporter,
summary_writer,
options: TrainerOptions,
distributed_option: DistributedOption,
) -> bool:
assert check_argument_types()
grad_noise = options.grad_noise
accum_grad = options.accum_grad
grad_clip = options.grad_clip
grad_clip_type = options.grad_clip_type
log_interval = options.log_interval
no_forward_run = options.no_forward_run
ngpu = options.ngpu
use_wandb = options.use_wandb
distributed = distributed_option.distributed
if log_interval is None:
try:
log_interval = max(len(iterator) // 20, 10)
except TypeError:
log_interval = 100
model.train()
all_steps_are_invalid = True
# [For distributed] Because iteration counts are not always equals between
# processes, send stop-flag to the other processes if iterator is finished
iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
start_time = time.perf_counter()
for iiter, (utt_id, batch) in enumerate(
reporter.measure_iter_time(iterator, "iter_time"), 1):
assert isinstance(batch, dict), type(batch)
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if iterator_stop > 0:
break
batch["utt_id"] = utt_id
batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
if no_forward_run:
all_steps_are_invalid = False
continue
with autocast(scaler is not None):
with reporter.measure_time("forward_time"):
retval = model(**batch)
# Note(kamo):
# Supporting two patterns for the returned value from the model
# a. dict type
if isinstance(retval, dict):
loss = retval["loss"]
stats = retval["stats"]
weight = retval["weight"]
optim_idx = retval.get("optim_idx")
if optim_idx is not None and not isinstance(
optim_idx, int):
if not isinstance(optim_idx, torch.Tensor):
raise RuntimeError(
"optim_idx must be int or 1dim torch.Tensor, "
f"but got {type(optim_idx)}")
if optim_idx.dim() >= 2:
raise RuntimeError(
"optim_idx must be int or 1dim torch.Tensor, "
f"but got {optim_idx.dim()}dim tensor")
if optim_idx.dim() == 1:
for v in optim_idx:
if v != optim_idx[0]:
raise RuntimeError(
"optim_idx must be 1dim tensor "
"having same values for all entries"
)
optim_idx = optim_idx[0].item()
else:
optim_idx = optim_idx.item()
# b. tuple or list type
else:
loss, stats, weight = retval
optim_idx = None
stats = {k: v for k, v in stats.items() if v is not None}
if ngpu > 1 or distributed:
# Apply weighted averaging for loss and stats
loss = (loss * weight.type(loss.dtype)).sum()
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight,
distributed)
# Now weight is summation over all workers
loss /= weight
if distributed:
# NOTE(kamo): Multiply world_size because DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= torch.distributed.get_world_size()
loss /= accum_grad
reporter.register(stats, weight)
with reporter.measure_time("backward_time"):
if scaler is not None:
# Scales loss. Calls backward() on scaled loss
# to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose
# for corresponding forward ops.
scaler.scale(loss).backward()
else:
loss.backward()
if iiter % accum_grad == 0:
if scaler is not None:
# Unscales the gradients of optimizer's assigned params in-place
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.unscale_(optimizer)
# gradient noise injection
if grad_noise:
add_gradient_noise(
model,
reporter.get_total_count(),
duration=100,
eta=1.0,
scale_factor=0.55,
)
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(),
max_norm=grad_clip,
norm_type=grad_clip_type,
)
# PyTorch<=1.4, clip_grad_norm_ returns float value
if not isinstance(grad_norm, torch.Tensor):
grad_norm = torch.tensor(grad_norm)
if not torch.isfinite(grad_norm):
logging.warning(
f"The grad norm is {grad_norm}. Skipping updating the model."
)
# Must invoke scaler.update() if unscale_() is used in the iteration
# to avoid the following error:
# RuntimeError: unscale_() has already been called
# on this optimizer since the last update().
# Note that if the gradient has inf/nan values,
# scaler.step skips optimizer.step().
if scaler is not None:
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.step(optimizer)
scaler.update()
else:
all_steps_are_invalid = False
with reporter.measure_time("optim_step_time"):
for iopt, (optimizer, scheduler) in enumerate(
zip(optimizers, schedulers)):
if optim_idx is not None and iopt != optim_idx:
continue
if scaler is not None:
# scaler.step() first unscales the gradients of
# the optimizer's assigned params.
scaler.step(optimizer)
# Updates the scale for next iteration.
scaler.update()
else:
optimizer.step()
if isinstance(scheduler, AbsBatchStepScheduler):
scheduler.step()
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
optimizer.zero_grad()
# Register lr and train/load time[sec/step],
# where step refers to accum_grad * mini-batch
reporter.register(
dict(
{
f"optim{i}_lr{j}": pg["lr"]
for i, optimizer in enumerate(optimizers)
for j, pg in enumerate(optimizer.param_groups)
if "lr" in pg
},
train_time=time.perf_counter() - start_time,
), )
start_time = time.perf_counter()
# NOTE(kamo): Call log_message() after next()
reporter.next()
if iiter % log_interval == 0:
logging.info(reporter.log_message(-log_interval))
if summary_writer is not None:
reporter.tensorboard_add_scalar(summary_writer,
-log_interval)
if use_wandb:
reporter.wandb_log()
else:
if distributed:
iterator_stop.fill_(1)
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
return all_steps_are_invalid
def train_one_epoch(
cls,
model: torch.nn.Module,
iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
optimizers: Sequence[torch.optim.Optimizer],
schedulers: Sequence[Optional[AbsScheduler]],
scaler: Optional[GradScaler],
reporter: SubReporter,
summary_writer: Optional[SummaryWriter],
options: TrainerOptions,
) -> bool:
assert check_argument_types()
# Note(kamo): assumes one optimizer
assert cls.num_optimizers == 1, cls.num_optimizers
assert len(optimizers) == 1, len(optimizers)
optimizer = optimizers[0]
scheduler = schedulers[0]
grad_noise = options.grad_noise
accum_grad = options.accum_grad
grad_clip = options.grad_clip
log_interval = options.log_interval
no_forward_run = options.no_forward_run
ngpu = options.ngpu
distributed = isinstance(model,
torch.nn.parallel.DistributedDataParallel)
if log_interval is None:
try:
log_interval = max(len(iterator) // 20, 10)
except TypeError:
log_interval = 100
model.train()
all_steps_are_invalid = True
# [For distributed] Because iteration counts are not always equals between
# processes, send stop-flag to the other processes if iterator is finished
iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
start_time = time.perf_counter()
for iiter, (_, batch) in enumerate(
reporter.measure_iter_time(iterator, "iter_time"), 1):
assert isinstance(batch, dict), type(batch)
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if iterator_stop > 0:
break
batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
if no_forward_run:
all_steps_are_invalid = False
continue
with autocast(scaler is not None):
with reporter.measure_time("forward_time"):
loss, stats, weight = model(**batch)
stats = {k: v for k, v in stats.items() if v is not None}
if ngpu > 1 or distributed:
# Apply weighted averaging for loss and stats
loss = (loss * weight.type(loss.dtype)).sum()
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight,
distributed)
# Now weight is summation over all workers
loss /= weight
if distributed:
# NOTE(kamo): Multiply world_size because DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= torch.distributed.get_world_size()
loss /= accum_grad
reporter.register(stats, weight)
with reporter.measure_time("backward_time"):
if scaler is not None:
# Scales loss. Calls backward() on scaled loss
# to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose
# for corresponding forward ops.
scaler.scale(loss).backward()
else:
loss.backward()
if iiter % accum_grad == 0:
if scaler is not None:
# Unscales the gradients of optimizer's assigned params in-place
scaler.unscale_(optimizer)
# gradient noise injection
if grad_noise:
add_gradient_noise(
model,
reporter.get_total_count(),
duration=100,
eta=1.0,
scale_factor=0.55,
)
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), grad_clip)
# PyTorch<=1.4, clip_grad_norm_ returns float value
if not isinstance(grad_norm, torch.Tensor):
grad_norm = torch.tensor(grad_norm)
if not torch.isfinite(grad_norm):
logging.warning(
f"The grad norm is {grad_norm}. Skipping updating the model."
)
else:
all_steps_are_invalid = False
with reporter.measure_time("optim_step_time"):
if scaler is not None:
# scaler.step() first unscales the gradients of
# the optimizer's assigned params.
scaler.step(optimizer)
# Updates the scale for next iteration.
scaler.update()
else:
optimizer.step()
if isinstance(scheduler, AbsBatchStepScheduler):
scheduler.step()
optimizer.zero_grad()
# Register lr and train/load time[sec/step],
# where step refers to accum_grad * mini-batch
reporter.register(
dict(
{
f"lr_{i}": pg["lr"]
for i, pg in enumerate(optimizer.param_groups)
if "lr" in pg
},
train_time=time.perf_counter() - start_time,
), )
start_time = time.perf_counter()
# NOTE(kamo): Call log_message() after next()
reporter.next()
if iiter % log_interval == 0:
logging.info(reporter.log_message(-log_interval))
if summary_writer is not None:
reporter.tensorboard_add_scalar(summary_writer,
-log_interval)
else:
if distributed:
iterator_stop.fill_(1)
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
return all_steps_are_invalid
def train_one_epoch(
cls,
model: torch.nn.Module,
iterator: Iterable[Tuple[List[str], Dict[str, torch.Tensor]]],
optimizers: Sequence[torch.optim.Optimizer],
schedulers: Sequence[Optional[AbsScheduler]],
reporter: SubReporter,
options: TrainerOptions,
) -> bool:
assert check_argument_types()
# Note(kamo): assumes one optimizer
assert cls.num_optimizers == 1, cls.num_optimizers
assert len(optimizers) == 1, len(optimizers)
optimizer = optimizers[0]
scheduler = schedulers[0]
grad_noise = options.grad_noise
accum_grad = options.accum_grad
grad_clip = options.grad_clip
log_interval = options.log_interval
no_forward_run = options.no_forward_run
ngpu = options.ngpu
distributed = isinstance(model,
torch.nn.parallel.DistributedDataParallel)
use_apex = options.train_dtype in ("O0", "O1", "O2", "O3")
if log_interval is None:
try:
log_interval = max(len(iterator) // 20, 10)
except TypeError:
log_interval = 100
model.train()
all_steps_are_invalid = True
# [For distributed] Because iteration counts are not always equals between
# processes, send stop-flag to the other processes if iterator is finished
iterator_stop = torch.tensor(0).to("cuda" if ngpu > 0 else "cpu")
start_time = time.perf_counter()
for iiter, (_, batch) in enumerate(
reporter.measure_iter_time(iterator, "iter_time"), 1):
assert isinstance(batch, dict), type(batch)
if distributed:
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
if iterator_stop > 0:
break
batch = to_device(batch, "cuda" if ngpu > 0 else "cpu")
if no_forward_run:
all_steps_are_invalid = False
reporter.register({})
continue
with reporter.measure_time("forward_time"):
loss, stats, weight = model(**batch)
if ngpu > 1 or distributed:
# Apply weighted averaging for loss and stats
loss = (loss * weight.type(loss.dtype)).sum()
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight, distributed)
# Now weight is summation over all workers
loss /= weight
if distributed:
# NOTE(kamo): Multiply world_size because DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= torch.distributed.get_world_size()
reporter.register(stats, weight)
loss /= accum_grad
with reporter.measure_time("backward_time"):
if use_apex:
try:
from apex import amp
except ImportError:
logging.error(
"You need to install apex. "
"See https://github.com/NVIDIA/apex#linux")
with amp.scale_loss(loss, optimizers) as scaled_loss:
scaled_loss.backward()
else:
loss.backward()
if iiter % accum_grad == 0:
# gradient noise injection
if grad_noise:
add_gradient_noise(
model,
reporter.get_total_count(),
duration=100,
eta=1.0,
scale_factor=0.55,
)
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(), grad_clip)
if not np.isfinite(grad_norm):
logging.warning(
f"The grad norm is {grad_norm}. Skipping updating the model."
)
else:
all_steps_are_invalid = False
with reporter.measure_time("optim_step_time"):
optimizer.step()
if isinstance(scheduler, AbsBatchStepScheduler):
scheduler.step()
optimizer.zero_grad()
# Register lr and train/load time[sec/step],
# where step refers to accum_grad * mini-batch
reporter.register(
dict(
{
f"lr_{i}": pg["lr"]
for i, pg in enumerate(optimizer.param_groups)
if "lr" in pg
},
train_time=time.perf_counter() - start_time,
),
# Suppress to increment the internal counter.
not_increment_count=True,
)
start_time = time.perf_counter()
if iiter % log_interval == 0:
logging.info(reporter.log_message())
else:
if distributed:
iterator_stop.fill_(1)
torch.distributed.all_reduce(iterator_stop, ReduceOp.SUM)
return all_steps_are_invalid
def test_add_gradient_noise():
linear = torch.nn.Linear(1, 1)
linear(torch.rand(1, 1)).sum().backward()
add_gradient_noise(linear, 100)
def train_one_batch(cls,
batch,
model,
scaler,
ngpu,
distributed,
reporter,
iiter,
accum_grad,
grad_noise,
grad_clip,
grad_clip_type,
optimizers,
schedulers,
start_time
):
model.train()
with autocast(scaler is not None):
retval = model(**batch)
loss, stats, weight = retval
optim_idx = None
stats = {k: v for k, v in stats.items() if v is not None}
if ngpu > 1 or distributed:
# Apply weighted averaging for loss and stats
loss = (loss * weight.type(loss.dtype)).sum()
# if distributed, this method can also apply all_reduce()
stats, weight = recursive_average(stats, weight, distributed)
# Now weight is summation over all workers
loss /= weight
if distributed:
# NOTE(kamo): Multiply world_size because DistributedDataParallel
# automatically normalizes the gradient by world_size.
loss *= torch.distributed.get_world_size()
loss /= accum_grad
with reporter.measure_time("backward_time"):
if scaler is not None:
# Scales loss. Calls backward() on scaled loss
# to create scaled gradients.
# Backward passes under autocast are not recommended.
# Backward ops run in the same dtype autocast chose
# for corresponding forward ops.
scaler.scale(loss).backward()
else:
loss.backward()
loss.detach()
torch.cuda.empty_cache()
if iiter % accum_grad == 0:
if scaler is not None:
# Unscales the gradients of optimizer's assigned params in-place
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.unscale_(optimizer)
# gradient noise injection
if grad_noise:
add_gradient_noise(
model,
reporter.get_total_count(),
duration=100,
eta=1.0,
scale_factor=0.55,
)
# compute the gradient norm to check if it is normal or not
grad_norm = torch.nn.utils.clip_grad_norm_(
model.parameters(),
max_norm=grad_clip,
norm_type=grad_clip_type,
)
# PyTorch<=1.4, clip_grad_norm_ returns float value
if not isinstance(grad_norm, torch.Tensor):
grad_norm = torch.tensor(grad_norm)
if not torch.isfinite(grad_norm):
logging.warning(
f"The grad norm is {grad_norm}. Skipping updating the model."
)
# Must invoke scaler.update() if unscale_() is used in the iteration
# to avoid the following error:
# RuntimeError: unscale_() has already been called
# on this optimizer since the last update().
# Note that if the gradient has inf/nan values,
# scaler.step skips optimizer.step().
if scaler is not None:
for iopt, optimizer in enumerate(optimizers):
if optim_idx is not None and iopt != optim_idx:
continue
scaler.step(optimizer)
scaler.update()
else:
all_steps_are_invalid = False
with reporter.measure_time("optim_step_time"):
for iopt, (optimizer, scheduler) in enumerate(
zip(optimizers, schedulers)
):
if optim_idx is not None and iopt != optim_idx:
continue
if scaler is not None:
# scaler.step() first unscales the gradients of
# the optimizer's assigned params.
scaler.step(optimizer)
# Updates the scale for next iteration.
scaler.update()
else:
optimizer.step()
if isinstance(scheduler, AbsBatchStepScheduler):
scheduler.step()
optimizer.zero_grad()
reporter.register(stats, weight)
# Register lr and train/load time[sec/step],
# where step refers to accum_grad * mini-batch
reporter.register(
dict(
{
f"optim{i}_lr{j}": pg["lr"]
for i, optimizer in enumerate(optimizers)
for j, pg in enumerate(optimizer.param_groups)
if "lr" in pg
},
train_time=time.perf_counter() - start_time,
),
)
return all_steps_are_invalid
