def setup_training(args):
"""Parse args, load dataset, and load model trainer."""
if not torch.cuda.is_available():
raise NotImplementedError("Training on CPU is not supported")
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Setup task and load dataset
task = tasks.setup_task(args)
task.load_dataset(
args.train_subset,
args.train_source_binary_path,
args.train_target_binary_path,
weights_file=getattr(args, "train_weights_path", None),
)
task.load_dataset(args.valid_subset, args.eval_source_binary_path,
args.eval_target_binary_path)
# Build model and criterion
model = task.build_model(args)
print("| building criterion")
criterion = task.build_criterion(args)
print(f"| model {args.arch}, criterion {criterion.__class__.__name__}")
print(f"| num. model params: \
{sum(p.numel() for p in model.parameters())}")
# Build trainer
if args.fp16:
trainer = FP16Trainer(args, task, model, criterion)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
"| NOTICE: your device may support faster training with --fp16"
)
trainer = Trainer(args, task, model, criterion)
print(f"| training on {args.distributed_world_size} GPUs")
print(
f"| max tokens per GPU = {args.max_tokens} and \
max sentences per GPU = {args.max_sentences}",
flush=True,
)
os.makedirs(args.save_dir, exist_ok=True)
# If --restore-file is already present under --save-dir, use that one
# instead of --pretrained-checkpoint-file. The idea is that
# --pretrained-checkpoint-file allows the user to specify restoring from a
# different run's checkpoint (possibly with different training params),
# while not polluting the previous run's checkpoint directory
# with new checkpoints. However, if training gets interrupted
# and the user restarts training, we want to resume from
# the checkpoints under --save-dir, instead of
# restarting again from the old run's checkpoint at
# --pretrained-checkpoint-file.
#
# Note that if args.restore_file is an absolute path, os.path.join() will
# ignore previous directory args and just use the absolute path as is.
checkpoint_path = os.path.join(args.save_dir, args.restore_file)
restore_state = True
if os.path.exists(checkpoint_path):
print(
f"| Using --save-dir={args.save_dir}, --restore-file={args.restore_file}."
)
elif args.pretrained_checkpoint_file and os.path.exists(
args.pretrained_checkpoint_file):
checkpoint_path = args.pretrained_checkpoint_file
restore_state = args.load_pretrained_checkpoint_state
print(
f"| Using --pretrained-checkpoint-file={args.pretrained_checkpoint_file}, "
f"--load-pretrained-checkpoint-state={args.load_pretrained_checkpoint_state}."
)
extra_state = default_extra_state(args)
if not os.path.isfile(checkpoint_path) and args.multi_model_restore_files:
print(
f"| Restoring individual models from {args.multi_model_restore_files}"
)
multi_model.import_individual_models(args.multi_model_restore_files,
trainer)
else:
loaded, loaded_extra_state = load_existing_checkpoint(
checkpoint_path=checkpoint_path,
trainer=trainer,
restore_state=restore_state,
)
if loaded_extra_state:
extra_state.update(loaded_extra_state)
if loaded:
args.path = [checkpoint_path]
calculate_bleu_on_subset(
args=args,
task=task,
epoch_str="initial loaded checkpoint",
offset=None,
dataset_split=args.valid_subset,
)
print(f"| extra_state: {extra_state}")
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences,
max_positions=trainer.get_model().max_positions(),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
)
epoch = extra_state["epoch"]
if extra_state["batch_offset"] == 0:
epoch -= 1 # this will be incremented when we call epoch_itr.next_epoch_itr()
epoch_itr.load_state_dict({
"epoch":
epoch,
"iterations_in_epoch":
extra_state["batch_offset"]
})
return extra_state, trainer, task, epoch_itr
def main(args):
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load dataset splits
load_dataset_splits(args, task, ['train', 'valid'])
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
trainer = FP16Trainer(args, task, model, criterion)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args, task, model, criterion)
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Initialize dataloader
max_positions = trainer.get_model().max_positions()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
)
# Load the latest checkpoint if one is available
load_checkpoint(args, trainer, epoch_itr)
# Send a dummy batch to warm the caching allocator
dummy_batch = task.dataset('train').get_dummy_batch(
args.max_tokens, max_positions)
trainer.dummy_train_step(dummy_batch)
# Train until the learning rate gets too small
max_epoch = args.max_epoch or math.inf
max_update = args.max_update or math.inf
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_losses = [None]
valid_subsets = args.valid_subset.split(',')
while lr > args.min_lr and epoch_itr.epoch <= max_epoch and trainer.get_num_updates(
) < max_update:
# train for one epoch
train(args, trainer, task, epoch_itr)
if epoch_itr.epoch % args.validate_interval == 0:
valid_losses = validate(args, trainer, task, epoch_itr,
valid_subsets)
# only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# save checkpoint
if epoch_itr.epoch % args.save_interval == 0:
save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
mlperf_compliance.mlperf_log.LOGGER.propagate = False
# framework = f'Pytorch NGC {os.environ["NVIDIA_PYTORCH_VERSION"]}'
# mlperf_submission_log(
# benchmark=mlperf_compliance.constants.TRANSFORMER,
# framework=framework)
mlperf_compliance.mlperf_log.setdefault(
root_dir=os.path.dirname(os.path.abspath(__file__)),
benchmark=mlperf_compliance.constants.TRANSFORMER,
stack_offset=1,
extra_print=False)
mlperf_print(key=mlperf_compliance.constants.INIT_START,
log_all_ranks=True)
torch.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
random.seed(args.seed)
np.random.seed(args.seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
# preinit and warmup streams/groups for allreduce communicators
allreduce_communicators = None
if args.distributed_world_size > 1 and args.enable_parallel_backward_allred_opt:
allreduce_groups = [
torch.distributed.new_group()
for _ in range(args.parallel_backward_allred_cuda_nstreams)
]
allreduce_streams = [
torch.cuda.Stream()
for _ in range(args.parallel_backward_allred_cuda_nstreams)
]
for group, stream in zip(allreduce_groups, allreduce_streams):
with torch.cuda.stream(stream):
torch.distributed.all_reduce(torch.cuda.FloatTensor(1),
group=group)
allreduce_communicators = (allreduce_groups, allreduce_streams)
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
mlperf_print(key=mlperf_compliance.constants.GLOBAL_BATCH_SIZE,
value=args.max_tokens * args.distributed_world_size)
mlperf_print(key=mlperf_compliance.constants.OPT_NAME,
value=args.optimizer)
assert (len(args.lr) == 1)
mlperf_print(key=mlperf_compliance.constants.OPT_BASE_LR,
value=args.lr[0] if len(args.lr) == 1 else args.lr)
mlperf_print(key=mlperf_compliance.constants.OPT_LR_WARMUP_STEPS,
value=args.warmup_updates)
assert (args.max_source_positions == args.max_target_positions)
mlperf_print(key=mlperf_compliance.constants.MAX_SEQUENCE_LENGTH,
value=args.max_target_positions)
mlperf_print(key=mlperf_compliance.constants.OPT_ADAM_BETA_1,
value=eval(args.adam_betas)[0])
mlperf_print(key=mlperf_compliance.constants.OPT_ADAM_BETA_2,
value=eval(args.adam_betas)[1])
mlperf_print(key=mlperf_compliance.constants.OPT_ADAM_EPSILON,
value=args.adam_eps)
pValue = ctypes.cast((ctypes.c_int * 1)(), ctypes.POINTER(ctypes.c_int))
result = torch.cuda.cudart().cudaDeviceSetLimit(ctypes.c_int(0x05),
ctypes.c_int(128))
result = torch.cuda.cudart().cudaDeviceGetLimit(pValue, ctypes.c_int(0x05))
# torch.manual_seed(args.seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
trainer = FP16Trainer(args,
task,
model,
criterion,
allreduce_communicators=allreduce_communicators)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args,
task,
model,
criterion,
allreduce_communicators=None)
#if (args.online_eval or args.target_bleu) and not args.remove_bpe:
# args.remove_bpe='@@ '
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Initialize dataloader
max_positions = trainer.get_model().max_positions()
# Send a dummy batch to warm the caching allocator
dummy_batch = language_pair_dataset.get_dummy_batch_isolated(
args.max_tokens, max_positions, 8)
trainer.dummy_train_step(dummy_batch)
# Train until the learning rate gets too small or model reaches target score
max_epoch = args.max_epoch if args.max_epoch >= 0 else math.inf
max_update = args.max_update or math.inf
tgt_bleu = args.target_bleu or math.inf
current_bleu = 0.0
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_losses = [None]
valid_subsets = args.valid_subset.split(',')
# mlperf compliance synchronization
if args.distributed_world_size > 1:
assert (torch.distributed.is_initialized())
torch.distributed.all_reduce(torch.cuda.FloatTensor(1))
torch.cuda.synchronize()
mlperf_print(key=mlperf_compliance.constants.INIT_STOP, sync=True)
mlperf_print(key=mlperf_compliance.constants.RUN_START, sync=True)
# second sync after RUN_START tag is printed.
# this ensures no rank touches data until after RUN_START tag is printed.
barrier()
# Load dataset splits
load_dataset_splits(task, ['train', 'test'])
ctr = 0
class DummyEpochBatchIterator:
def __init__(self, epoch=0):
self.epoch = epoch
epoch_itr = DummyEpochBatchIterator(0)
# Main training loop
while lr >= args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates(
) < max_update and current_bleu < tgt_bleu:
first_epoch = epoch_itr.epoch + 1
mlperf_print(key=mlperf_compliance.constants.BLOCK_START,
metadata={
'first_epoch_num': first_epoch,
'epoch_count': 1
},
sync=True)
mlperf_print(key=mlperf_compliance.constants.EPOCH_START,
metadata={'epoch_num': first_epoch},
sync=True)
start = time.time()
gc.disable()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
dataloader_num_workers=args.dataloader_num_workers,
dataloader_pin_memory=args.enable_dataloader_pin_memory,
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
epoch=epoch_itr.epoch if ctr is not 0 else 0,
bucket_growth_factor=args.bucket_growth_factor,
seq_len_multiple=args.seq_len_multiple,
batching_scheme=args.batching_scheme,
batch_multiple_strategy=args.batch_multiple_strategy,
)
print("got epoch iterator", time.time() - start)
# Load the latest checkpoint if one is available
if ctr is 0:
load_checkpoint(args, trainer, epoch_itr)
# train for one epoch
start = time.time()
#exit(1)
train(args, trainer, task, epoch_itr)
print("epoch time ", time.time() - start)
start = time.time()
mlperf_print(key=mlperf_compliance.constants.EPOCH_STOP,
metadata={'epoch_num': first_epoch},
sync=True)
#if epoch_itr.epoch % args.validate_interval == 0:
# valid_losses = validate(args, trainer, task, epoch_itr, valid_subsets)
# Eval BLEU score
if args.online_eval or (not tgt_bleu is math.inf):
current_bleu = score(args, trainer, task, epoch_itr,
args.gen_subset)
mlperf_print(key=mlperf_compliance.tags.EVAL_ACCURACY,
value=str(current_bleu),
metadata={'epoch_num': first_epoch})
gc.enable()
# Only use first validation loss to update the learning rate
#lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# Save checkpoint
#if epoch_itr.epoch % args.save_interval == 0:
# save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
ctr = ctr + 1
print("validation and scoring ", time.time() - start)
mlperf_print(key=mlperf_compliance.constants.BLOCK_STOP,
metadata={'first_epoch_num': first_epoch},
sync=True)
train_meter.stop()
status = 'success' if current_bleu >= tgt_bleu else 'aborted'
mlperf_print(key=mlperf_compliance.constants.RUN_STOP,
metadata={'status': status})
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
if args.distributed_world_size > 1:
assert (torch.distributed.is_initialized())
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
pValue = ctypes.cast((ctypes.c_int * 1)(), ctypes.POINTER(ctypes.c_int))
result = torch.cuda.cudart().cudaDeviceSetLimit(ctypes.c_int(0x05),
ctypes.c_int(128))
result = torch.cuda.cudart().cudaDeviceGetLimit(pValue, ctypes.c_int(0x05))
torch.manual_seed(args.seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load dataset splits
load_dataset_splits(task, ['train', 'valid'])
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
trainer = FP16Trainer(args, task, model, criterion)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args, task, model, criterion)
if (args.online_eval or args.target_bleu) and not args.remove_bpe:
args.remove_bpe = '@@ '
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
max_positions = trainer.get_model().max_positions()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
)
# Load the latest checkpoint if one is available
load_checkpoint(args, trainer, epoch_itr)
# Send a dummy batch to warm the caching allocator
dummy_batch = task.dataset('train').get_dummy_batch(
args.max_tokens, max_positions)
trainer.dummy_train_step(dummy_batch)
# Train until the learning rate gets too small or model reaches target score
max_epoch = args.max_epoch or math.inf
max_update = args.max_update or math.inf
tgt_bleu = args.target_bleu or math.inf
current_bleu = 0.0
best_bleu = 0.0
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_losses = [None]
valid_subsets = args.valid_subset.split(',')
while lr >= args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates(
) < max_update and current_bleu < tgt_bleu:
# train for one epoch
train(args, trainer, task, epoch_itr)
if epoch_itr.epoch % args.validate_interval == 0:
valid_losses = validate(args, trainer, task, epoch_itr,
valid_subsets)
# Eval BLEU score
if args.online_eval or (not tgt_bleu is math.inf):
current_bleu, current_sc_bleu = score(args, trainer, task,
epoch_itr, args.gen_subset)
if current_bleu > best_bleu:
best_bleu = current_bleu
save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
# Only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# Save checkpoint
if epoch_itr.epoch % args.save_interval == 0:
save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
train_meter.stop()
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
from mlperf_compliance.mlperf_log import transformer_print
transformer_print(
key=mlperf_log.RUN_CLEAR_CACHES
) #before this tag we should run clearing caches on the host
# mlperf compliance synchronization
if args.distributed_world_size > 1:
assert (torch.distributed.is_initialized())
torch.distributed.broadcast(torch.tensor([1], device="cuda"), 0)
torch.cuda.synchronize()
transformer_print(key=mlperf_log.RUN_START)
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
transformer_print(key=mlperf_log.OPT_NAME, value=args.optimizer)
transformer_print(key=mlperf_log.OPT_LR, value=args.lr)
transformer_print(key=mlperf_log.OPT_HP_ADAM_BETA1,
value=eval(args.adam_betas)[0])
transformer_print(key=mlperf_log.OPT_HP_ADAM_BETA2,
value=eval(args.adam_betas)[1])
transformer_print(key=mlperf_log.OPT_HP_ADAM_EPSILON, value=args.adam_eps)
pValue = ctypes.cast((ctypes.c_int * 1)(), ctypes.POINTER(ctypes.c_int))
result = torch.cuda.cudart().cudaDeviceSetLimit(ctypes.c_int(0x05),
ctypes.c_int(128))
result = torch.cuda.cudart().cudaDeviceGetLimit(pValue, ctypes.c_int(0x05))
torch.manual_seed(args.seed)
transformer_print(key=mlperf_log.RUN_SET_RANDOM_SEED, value=args.seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
transformer_print(key=mlperf_log.MODEL_HP_SEQ_BEAM_SEARCH,
value={
'alpha': args.lenpen,
'beam_size': args.beam,
'extra_decode_length': args.max_len_b,
'vocab_size': task.target_dictionary.__len__()
})
# Load dataset splits
load_dataset_splits(task, ['train', 'valid'])
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
trainer = FP16Trainer(args, task, model, criterion)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args, task, model, criterion)
if (args.online_eval or args.target_bleu) and not args.remove_bpe:
args.remove_bpe = '@@ '
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
transformer_print(key=mlperf_log.INPUT_BATCH_SIZE, value=args.max_tokens)
transformer_print(key=mlperf_log.INPUT_ORDER)
# Initialize dataloader
max_positions = trainer.get_model().max_positions()
# Send a dummy batch to warm the caching allocator
dummy_batch = task.dataset('train').get_dummy_batch(
args.max_tokens, max_positions)
trainer.dummy_train_step(dummy_batch)
# Train until the learning rate gets too small or model reaches target score
max_epoch = args.max_epoch or math.inf
max_update = args.max_update or math.inf
tgt_bleu = args.target_bleu or math.inf
current_bleu = 0.0
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_losses = [None]
valid_subsets = args.valid_subset.split(',')
ctr = 0
class DummyEpochBatchIterator:
def __init__(self, epoch=0):
self.epoch = epoch
epoch_itr = DummyEpochBatchIterator(0)
transformer_print(key=mlperf_log.TRAIN_LOOP)
while lr >= args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates(
) < max_update and current_bleu < tgt_bleu:
transformer_print(key=mlperf_log.TRAIN_EPOCH, value=epoch_itr.epoch)
import time
start = time.time()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
epoch=epoch_itr.epoch if ctr is not 0 else 0)
print("got epoch iterator", time.time() - start)
# Load the latest checkpoint if one is available
if ctr is 0:
load_checkpoint(args, trainer, epoch_itr)
# train for one epoch
start = time.time()
train(args, trainer, task, epoch_itr)
print("epoch time ", time.time() - start)
start = time.time()
if epoch_itr.epoch % args.validate_interval == 0:
valid_losses = validate(args, trainer, task, epoch_itr,
valid_subsets)
# Eval BLEU score
transformer_print(key=mlperf_log.EVAL_START, value=epoch_itr.epoch)
if args.online_eval or (not tgt_bleu is math.inf):
current_bleu = score(args, trainer, task, epoch_itr,
args.gen_subset)
transformer_print(key=mlperf_log.EVAL_ACCURACY,
value={
'epoch': epoch_itr.epoch,
'value': current_bleu
})
transformer_print(key=mlperf_log.EVAL_TARGET, value=tgt_bleu)
transformer_print(key=mlperf_log.EVAL_STOP, value=epoch_itr.epoch)
# Only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# Save checkpoint
if epoch_itr.epoch % args.save_interval == 0:
save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
ctr = ctr + 1
print("validation and scoring ", time.time() - start)
train_meter.stop()
transformer_print(key=mlperf_log.RUN_STOP)
transformer_print(key=mlperf_log.RUN_FINAL)
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
mllog.config(filename=os.path.join(
os.path.dirname(os.path.abspath(__file__)), 'transformer.log'))
mllogger = mllog.get_mllogger()
mllogger.logger.propagate = False
log_start(key=constants.INIT_START, log_all_ranks=True)
# preinit and warmup streams/groups for allreduce communicators
allreduce_communicators = None
if args.distributed_world_size > 1 and args.enable_parallel_backward_allred_opt:
allreduce_groups = [
torch.distributed.new_group()
for _ in range(args.parallel_backward_allred_cuda_nstreams)
]
allreduce_streams = [
torch.cuda.Stream()
for _ in range(args.parallel_backward_allred_cuda_nstreams)
]
for group, stream in zip(allreduce_groups, allreduce_streams):
with torch.cuda.stream(stream):
torch.distributed.all_reduce(torch.cuda.FloatTensor(1),
group=group)
allreduce_communicators = (allreduce_groups, allreduce_streams)
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
log_event(key=constants.GLOBAL_BATCH_SIZE,
value=args.max_tokens * args.distributed_world_size)
log_event(key=constants.OPT_NAME, value=args.optimizer)
assert (len(args.lr) == 1)
log_event(key=constants.OPT_BASE_LR,
value=args.lr[0] if len(args.lr) == 1 else args.lr)
log_event(key=constants.OPT_LR_WARMUP_STEPS, value=args.warmup_updates)
assert (args.max_source_positions == args.max_target_positions)
log_event(key=constants.MAX_SEQUENCE_LENGTH,
value=args.max_target_positions,
metadata={'method': 'discard'})
log_event(key=constants.OPT_ADAM_BETA_1, value=eval(args.adam_betas)[0])
log_event(key=constants.OPT_ADAM_BETA_2, value=eval(args.adam_betas)[1])
log_event(key=constants.OPT_ADAM_EPSILON, value=args.adam_eps)
log_event(key=constants.SEED, value=args.seed)
# L2 Sector Promotion
pValue = ctypes.cast((ctypes.c_int * 1)(), ctypes.POINTER(ctypes.c_int))
result = ctypes.CDLL('libcudart.so').cudaDeviceSetLimit(
ctypes.c_int(0x05), ctypes.c_int(128))
result = ctypes.CDLL('libcudart.so').cudaDeviceGetLimit(
pValue, ctypes.c_int(0x05))
worker_seeds, shuffling_seeds = setup_seeds(
args.seed,
args.max_epoch + 1,
torch.device('cuda'),
args.distributed_rank,
args.distributed_world_size,
)
worker_seed = worker_seeds[args.distributed_rank]
print(
f'Worker {args.distributed_rank} is using worker seed: {worker_seed}')
torch.manual_seed(worker_seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
if args.distributed_weight_update != 0:
from fairseq.fp16_trainer import DistributedFP16Trainer
trainer = DistributedFP16Trainer(
args,
task,
model,
criterion,
allreduce_communicators=allreduce_communicators)
else:
from fairseq.fp16_trainer import FP16Trainer
trainer = FP16Trainer(
args,
task,
model,
criterion,
allreduce_communicators=allreduce_communicators)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args,
task,
model,
criterion,
allreduce_communicators=None)
#if (args.online_eval or args.target_bleu) and not args.remove_bpe:
# args.remove_bpe='@@ '
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Initialize dataloader
max_positions = trainer.get_model().max_positions()
# Send a dummy batch to warm the caching allocator
dummy_batch = language_pair_dataset.get_dummy_batch_isolated(
args.max_tokens, max_positions, 8)
trainer.dummy_train_step(dummy_batch)
# Train until the learning rate gets too small or model reaches target score
max_epoch = args.max_epoch if args.max_epoch >= 0 else math.inf
max_update = args.max_update or math.inf
tgt_bleu = args.target_bleu or math.inf
current_bleu = 0.0
lr = trainer.get_lr()
train_meter = StopwatchMeter()
train_meter.start()
valid_losses = [None]
# mlperf compliance synchronization
if args.distributed_world_size > 1:
assert (torch.distributed.is_initialized())
torch.distributed.all_reduce(torch.cuda.FloatTensor(1))
torch.cuda.synchronize()
log_end(key=constants.INIT_STOP, sync=False)
log_start(key=constants.RUN_START, sync=True)
# second sync after RUN_START tag is printed.
# this ensures no rank touches data until after RUN_START tag is printed.
barrier()
# Load dataset splits
load_dataset_splits(task, ['train', 'test'])
log_event(key=constants.TRAIN_SAMPLES,
value=len(task.dataset(args.train_subset)),
sync=False)
log_event(key=constants.EVAL_SAMPLES,
value=len(task.dataset(args.gen_subset)),
sync=False)
ctr = 0
start = time.time()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
dataloader_num_workers=args.dataloader_num_workers,
dataloader_pin_memory=args.enable_dataloader_pin_memory,
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seeds=shuffling_seeds,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
epoch=epoch_itr.epoch if ctr is not 0 else 0,
bucket_growth_factor=args.bucket_growth_factor,
seq_len_multiple=args.seq_len_multiple,
batching_scheme=args.batching_scheme,
batch_multiple_strategy=args.batch_multiple_strategy,
)
print("got epoch iterator", time.time() - start)
# Main training loop
while lr >= args.min_lr and epoch_itr.epoch < max_epoch and trainer.get_num_updates(
) < max_update and current_bleu < tgt_bleu:
first_epoch = epoch_itr.epoch + 1
log_start(key=constants.BLOCK_START,
metadata={
'first_epoch_num': first_epoch,
'epoch_count': 1
},
sync=False)
log_start(key=constants.EPOCH_START,
metadata={'epoch_num': first_epoch},
sync=False)
gc.disable()
# Load the latest checkpoint if one is available
if ctr is 0:
load_checkpoint(args, trainer, epoch_itr)
# train for one epoch
start = time.time()
#exit(1)
train(args, trainer, task, epoch_itr, shuffling_seeds)
print("epoch time ", time.time() - start)
start = time.time()
log_end(key=constants.EPOCH_STOP,
metadata={'epoch_num': first_epoch},
sync=False)
# Eval BLEU score
if args.online_eval or (not tgt_bleu is math.inf):
current_bleu = score(args, trainer, task, epoch_itr,
args.gen_subset)
log_event(key=constants.EVAL_ACCURACY,
value=float(current_bleu) / 100.0,
metadata={'epoch_num': first_epoch})
gc.enable()
# Only use first validation loss to update the learning rate
#lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
# Save checkpoint
#if epoch_itr.epoch % args.save_interval == 0:
# save_checkpoint(args, trainer, epoch_itr, valid_losses[0])
ctr = ctr + 1
print("validation and scoring ", time.time() - start)
log_end(key=constants.BLOCK_STOP,
metadata={'first_epoch_num': first_epoch},
sync=False)
train_meter.stop()
status = 'success' if current_bleu >= tgt_bleu else 'aborted'
log_end(key=constants.RUN_STOP, metadata={'status': status})
print('| done training in {:.1f} seconds'.format(train_meter.sum))
def main(args):
if args.max_tokens is None:
args.max_tokens = 6000
print(args)
if not torch.cuda.is_available():
raise NotImplementedError('Training on CPU is not supported')
torch.cuda.set_device(args.device_id)
torch.manual_seed(args.seed)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load dataset splits
load_dataset_splits(task, ['train', 'valid'])
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
print('| model {}, criterion {}'.format(args.arch,
criterion.__class__.__name__))
print('| num. model params: {}'.format(
sum(p.numel() for p in model.parameters())))
# Build trainer
if args.fp16:
trainer = FP16Trainer(args, task, model, criterion)
else:
if torch.cuda.get_device_capability(0)[0] >= 7:
print(
'| NOTICE: your device may support faster training with --fp16'
)
trainer = Trainer(args, task, model, criterion)
print('| training on {} GPUs'.format(args.distributed_world_size))
print('| max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Initialize dataloader
max_positions = trainer.get_model().max_positions()
epoch_itr = data.EpochBatchIterator(
dataset=task.dataset(args.train_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences_valid,
max_positions=max_positions,
ignore_invalid_inputs=True,
required_batch_size_multiple=8,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
)
# Load the latest checkpoint if one is available
load_checkpoint(args, trainer, epoch_itr)
embedding = model.decoder.embed_tokens.weight.data.cpu().numpy()
print(embedding.shape)
Ar, s = low_rank_approx(embedding, 2)
print(Ar.shape)
np.savetxt('svd', Ar, delimiter=' ')