def main():
"""
Launches data-parallel multi-gpu training.
"""
mlperf_log.ROOT_DIR_GNMT = os.path.dirname(os.path.abspath(__file__))
mlperf_log.LOGGER.propagate = False
args = parse_args()
device = utils.set_device(args.cuda, args.local_rank)
distributed = utils.init_distributed(args.cuda)
gnmt_print(key=mlperf_log.RUN_START, sync=True)
args.rank = utils.get_rank()
if not args.cudnn:
torch.backends.cudnn.enabled = False
# create directory for results
save_path = os.path.join(args.results_dir, args.save)
args.save_path = save_path
os.makedirs(save_path, exist_ok=True)
# setup logging
log_filename = f'log_rank_{utils.get_rank()}.log'
utils.setup_logging(os.path.join(save_path, log_filename))
if args.env:
utils.log_env_info()
logging.info(f'Saving results to: {save_path}')
logging.info(f'Run arguments: {args}')
# automatically set train_iter_size based on train_global_batch_size,
# world_size and per-worker train_batch_size
if args.train_global_batch_size is not None:
global_bs = args.train_global_batch_size
bs = args.train_batch_size
world_size = utils.get_world_size()
assert global_bs % (bs * world_size) == 0
args.train_iter_size = global_bs // (bs * world_size)
logging.info(f'Global batch size was set in the config, '
f'Setting train_iter_size to {args.train_iter_size}')
worker_seeds, shuffling_seeds = utils.setup_seeds(args.seed, args.epochs,
device)
worker_seed = worker_seeds[args.rank]
logging.info(f'Worker {args.rank} is using worker seed: {worker_seed}')
torch.manual_seed(worker_seed)
# build tokenizer
pad_vocab = utils.pad_vocabulary(args.math)
tokenizer = Tokenizer(os.path.join(args.dataset_dir, config.VOCAB_FNAME),
pad_vocab)
# build datasets
gnmt_print(key=mlperf_log.PREPROC_TOKENIZE_TRAINING, sync=False)
gnmt_print(key=mlperf_log.TRAIN_HP_MAX_SEQ_LEN,
value=args.max_length_train,
sync=False)
train_data = LazyParallelDataset(
src_fname=os.path.join(args.dataset_dir, config.SRC_TRAIN_FNAME),
tgt_fname=os.path.join(args.dataset_dir, config.TGT_TRAIN_FNAME),
tokenizer=tokenizer,
min_len=args.min_length_train,
max_len=args.max_length_train,
sort=False,
max_size=args.max_size)
gnmt_print(key=mlperf_log.PREPROC_NUM_TRAIN_EXAMPLES,
value=len(train_data),
sync=False)
val_data = ParallelDataset(src_fname=os.path.join(args.dataset_dir,
config.SRC_VAL_FNAME),
tgt_fname=os.path.join(args.dataset_dir,
config.TGT_VAL_FNAME),
tokenizer=tokenizer,
min_len=args.min_length_val,
max_len=args.max_length_val,
sort=True)
gnmt_print(key=mlperf_log.PREPROC_TOKENIZE_EVAL, sync=False)
test_data = TextDataset(src_fname=os.path.join(args.dataset_dir,
config.SRC_TEST_FNAME),
tokenizer=tokenizer,
min_len=args.min_length_test,
max_len=args.max_length_test,
sort=True)
gnmt_print(key=mlperf_log.PREPROC_NUM_EVAL_EXAMPLES,
value=len(test_data),
sync=False)
vocab_size = tokenizer.vocab_size
gnmt_print(key=mlperf_log.PREPROC_VOCAB_SIZE, value=vocab_size, sync=False)
# build GNMT model
model_config = {
'hidden_size': args.hidden_size,
'num_layers': args.num_layers,
'dropout': args.dropout,
'batch_first': False,
'share_embedding': args.share_embedding
}
model = GNMT(vocab_size=vocab_size, **model_config)
logging.info(model)
batch_first = model.batch_first
# define loss function (criterion) and optimizer
criterion = build_criterion(vocab_size, config.PAD, args.smoothing)
opt_config = {'optimizer': args.optimizer, 'lr': args.lr}
opt_config.update(literal_eval(args.optimizer_extra))
logging.info(f'Training optimizer config: {opt_config}')
scheduler_config = {
'warmup_steps': args.warmup_steps,
'remain_steps': args.remain_steps,
'decay_interval': args.decay_interval,
'decay_steps': args.decay_steps,
'decay_factor': args.decay_factor
}
logging.info(f'Training LR schedule config: {scheduler_config}')
num_parameters = sum([l.nelement() for l in model.parameters()])
logging.info(f'Number of parameters: {num_parameters}')
batching_opt = {
'shard_size': args.shard_size,
'num_buckets': args.num_buckets
}
# get data loaders
train_loader = train_data.get_loader(batch_size=args.train_batch_size,
seeds=shuffling_seeds,
batch_first=batch_first,
shuffle=True,
batching=args.batching,
batching_opt=batching_opt,
num_workers=args.train_loader_workers)
gnmt_print(key=mlperf_log.INPUT_BATCH_SIZE,
value=args.train_batch_size * utils.get_world_size(),
sync=False)
gnmt_print(key=mlperf_log.INPUT_SIZE,
value=train_loader.sampler.num_samples,
sync=False)
val_loader = val_data.get_loader(batch_size=args.val_batch_size,
batch_first=batch_first,
shuffle=False,
num_workers=args.val_loader_workers)
test_loader = test_data.get_loader(batch_size=args.test_batch_size,
batch_first=batch_first,
shuffle=False,
pad=True,
num_workers=args.test_loader_workers)
gnmt_print(key=mlperf_log.EVAL_SIZE,
value=len(test_loader.dataset),
sync=False)
translator = Translator(model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=args.beam_size,
max_seq_len=args.max_length_test,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
cuda=args.cuda,
print_freq=args.print_freq,
dataset_dir=args.dataset_dir,
target_bleu=args.target_bleu,
save_path=args.save_path)
# create trainer
total_train_iters = len(train_loader) // args.train_iter_size * args.epochs
save_info = {
'model_config': model_config,
'config': args,
'tokenizer': tokenizer.get_state()
}
trainer_options = dict(criterion=criterion,
grad_clip=args.grad_clip,
iter_size=args.train_iter_size,
save_path=save_path,
save_freq=args.save_freq,
save_info=save_info,
opt_config=opt_config,
scheduler_config=scheduler_config,
train_iterations=total_train_iters,
batch_first=batch_first,
keep_checkpoints=args.keep_checkpoints,
math=args.math,
print_freq=args.print_freq,
cuda=args.cuda,
distributed=distributed,
intra_epoch_eval=args.intra_epoch_eval,
translator=translator)
trainer_options['model'] = model
trainer = trainers.Seq2SeqTrainer(**trainer_options)
# optionally resume from a checkpoint
if args.resume:
checkpoint_file = args.resume
if os.path.isdir(checkpoint_file):
checkpoint_file = os.path.join(checkpoint_file, 'model_best.pth')
if os.path.isfile(checkpoint_file):
trainer.load(checkpoint_file)
else:
logging.error(f'No checkpoint found at {args.resume}')
# training loop
best_loss = float('inf')
break_training = False
test_bleu = None
gnmt_print(key=mlperf_log.TRAIN_LOOP, sync=True)
for epoch in range(args.start_epoch, args.epochs):
logging.info(f'Starting epoch {epoch}')
gnmt_print(key=mlperf_log.TRAIN_EPOCH, value=epoch, sync=True)
train_loader.sampler.set_epoch(epoch)
trainer.epoch = epoch
train_loss, train_perf = trainer.optimize(train_loader)
# evaluate on validation set
if args.eval:
logging.info(f'Running validation on dev set')
val_loss, val_perf = trainer.evaluate(val_loader)
# remember best prec@1 and save checkpoint
gnmt_print(key=mlperf_log.TRAIN_CHECKPOINT, sync=False)
if args.rank == 0:
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
trainer.save(save_all=args.save_all, is_best=is_best)
if args.eval:
gnmt_print(key=mlperf_log.EVAL_START, value=epoch, sync=True)
test_bleu, break_training = translator.run(calc_bleu=True,
epoch=epoch)
gnmt_print(key=mlperf_log.EVAL_ACCURACY,
value={
"epoch": epoch,
"value": round(test_bleu, 2)
},
sync=False)
gnmt_print(key=mlperf_log.EVAL_TARGET,
value=args.target_bleu,
sync=False)
gnmt_print(key=mlperf_log.EVAL_STOP, sync=True)
acc_log = []
acc_log += [f'Summary: Epoch: {epoch}']
acc_log += [f'Training Loss: {train_loss:.4f}']
if args.eval:
acc_log += [f'Validation Loss: {val_loss:.4f}']
acc_log += [f'Test BLEU: {test_bleu:.2f}']
perf_log = []
perf_log += [f'Performance: Epoch: {epoch}']
perf_log += [f'Training: {train_perf:.0f} Tok/s']
if args.eval:
perf_log += [f'Validation: {val_perf:.0f} Tok/s']
if args.rank == 0:
logging.info('\t'.join(acc_log))
logging.info('\t'.join(perf_log))
logging.info(f'Finished epoch {epoch}')
if break_training:
break
gnmt_print(key=mlperf_log.RUN_STOP,
value={"success": bool(break_training)},
sync=True)
gnmt_print(key=mlperf_log.RUN_FINAL, sync=False)
def main():
"""
Launches data-parallel multi-gpu training.
"""
training_start = time.time()
args = parse_args()
device = utils.set_device(args.cuda, args.local_rank)
utils.init_distributed(args.cuda)
args.rank = utils.get_rank()
if not args.cudnn:
torch.backends.cudnn.enabled = False
# create directory for results
os.makedirs(args.save_dir, exist_ok=True)
# setup logging
log_filename = f'log_rank_{utils.get_rank()}.log'
utils.setup_logging(args.log_all_ranks,
os.path.join(args.save_dir, log_filename))
if args.env:
utils.log_env_info()
logging.info(f'Saving results to: {args.save_dir}')
logging.info(f'Run arguments: {args}')
args.train_iter_size = set_iter_size(args.train_iter_size,
args.train_global_batch_size,
args.train_batch_size)
worker_seeds, shuffling_seeds = utils.setup_seeds(args.seed, args.epochs,
device)
worker_seed = worker_seeds[args.rank]
logging.info(f'Worker {args.rank} is using worker seed: {worker_seed}')
torch.manual_seed(worker_seed)
# build tokenizer
pad_vocab = utils.pad_vocabulary(args.math)
tokenizer = Tokenizer(args.vocab, args.bpe_codes, args.lang, pad_vocab)
# build datasets
train_data = LazyParallelDataset(
src_fname=args.train_src,
tgt_fname=args.train_tgt,
tokenizer=tokenizer,
min_len=args.train_min_length,
max_len=args.train_max_length,
sort=False,
max_size=args.train_max_size,
)
val_data = ParallelDataset(
src_fname=args.val_src,
tgt_fname=args.val_tgt,
tokenizer=tokenizer,
min_len=args.val_min_length,
max_len=args.val_max_length,
sort=True,
)
test_data = TextDataset(
src_fname=args.test_src,
tokenizer=tokenizer,
min_len=args.test_min_length,
max_len=args.test_max_length,
sort=True,
)
vocab_size = tokenizer.vocab_size
# build GNMT model
model_config = {
'hidden_size': args.hidden_size,
'vocab_size': vocab_size,
'num_layers': args.num_layers,
'dropout': args.dropout,
'batch_first': False,
'share_embedding': args.share_embedding,
}
model = GNMT(**model_config).to(device)
logging.info(model)
batch_first = model.batch_first
# define loss function (criterion) and optimizer
criterion = build_criterion(vocab_size, config.PAD,
args.smoothing).to(device)
opt_config = {'optimizer': args.optimizer, 'lr': args.lr}
opt_config.update(literal_eval(args.optimizer_extra))
logging.info(f'Training optimizer config: {opt_config}')
scheduler_config = {
'warmup_steps': args.warmup_steps,
'remain_steps': args.remain_steps,
'decay_interval': args.decay_interval,
'decay_steps': args.decay_steps,
'decay_factor': args.decay_factor
}
logging.info(f'Training LR schedule config: {scheduler_config}')
num_parameters = sum([l.nelement() for l in model.parameters()])
logging.info(f'Number of parameters: {num_parameters}')
batching_opt = {
'shard_size': args.shard_size,
'num_buckets': args.num_buckets
}
# get data loaders
train_loader = train_data.get_loader(batch_size=args.train_batch_size,
seeds=shuffling_seeds,
batch_first=batch_first,
shuffle=True,
batching=args.batching,
batching_opt=batching_opt,
num_workers=args.train_loader_workers)
val_loader = val_data.get_loader(batch_size=args.val_batch_size,
batch_first=batch_first,
shuffle=False,
num_workers=args.val_loader_workers)
test_loader = test_data.get_loader(batch_size=args.test_batch_size,
batch_first=batch_first,
shuffle=False,
pad=True,
num_workers=args.test_loader_workers)
translator = Translator(
model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=args.beam_size,
max_seq_len=args.test_max_length,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
print_freq=args.print_freq,
reference=args.test_tgt,
)
# create trainer
total_train_iters = len(train_loader) // args.train_iter_size * args.epochs
save_info = {
'model_config': model_config,
'config': args,
'tokenizer': tokenizer.get_state()
}
loss_scaling = {
'init_scale': args.init_scale,
'upscale_interval': args.upscale_interval
}
trainer_options = dict(
model=model,
criterion=criterion,
grad_clip=args.grad_clip,
iter_size=args.train_iter_size,
save_dir=args.save_dir,
save_freq=args.save_freq,
save_info=save_info,
opt_config=opt_config,
scheduler_config=scheduler_config,
train_iterations=total_train_iters,
keep_checkpoints=args.keep_checkpoints,
math=args.math,
loss_scaling=loss_scaling,
print_freq=args.print_freq,
intra_epoch_eval=args.intra_epoch_eval,
translator=translator,
prealloc_mode=args.prealloc_mode,
)
trainer = trainers.Seq2SeqTrainer(**trainer_options)
# optionally resume from a checkpoint
if args.resume:
checkpoint_file = args.resume
if os.path.isdir(checkpoint_file):
checkpoint_file = os.path.join(checkpoint_file, 'model_best.pth')
if os.path.isfile(checkpoint_file):
trainer.load(checkpoint_file)
else:
logging.error(f'No checkpoint found at {args.resume}')
# training loop
best_loss = float('inf')
training_perf = []
break_training = False
test_bleu = None
for epoch in range(args.start_epoch, args.epochs):
logging.info(f'Starting epoch {epoch}')
train_loader.sampler.set_epoch(epoch)
trainer.epoch = epoch
train_loss, train_perf = trainer.optimize(train_loader)
training_perf.append(train_perf)
# evaluate on validation set
if args.eval:
logging.info(f'Running validation on dev set')
val_loss, val_perf = trainer.evaluate(val_loader)
# remember best prec@1 and save checkpoint
if args.rank == 0:
is_best = val_loss < best_loss
best_loss = min(val_loss, best_loss)
trainer.save(save_all=args.save_all, is_best=is_best)
if args.eval:
utils.barrier()
eval_fname = f'eval_epoch_{epoch}'
eval_path = os.path.join(args.save_dir, eval_fname)
_, eval_stats = translator.run(
calc_bleu=True,
epoch=epoch,
eval_path=eval_path,
)
test_bleu = eval_stats['bleu']
if args.target_bleu and test_bleu >= args.target_bleu:
logging.info(f'Target accuracy reached')
break_training = True
acc_log = []
acc_log += [f'Summary: Epoch: {epoch}']
acc_log += [f'Training Loss: {train_loss:.4f}']
if args.eval:
acc_log += [f'Validation Loss: {val_loss:.4f}']
acc_log += [f'Test BLEU: {test_bleu:.2f}']
perf_log = []
perf_log += [f'Performance: Epoch: {epoch}']
perf_log += [f'Training: {train_perf:.0f} Tok/s']
if args.eval:
perf_log += [f'Validation: {val_perf:.0f} Tok/s']
if args.rank == 0:
logging.info('\t'.join(acc_log))
logging.info('\t'.join(perf_log))
logging.info(f'Finished epoch {epoch}')
if break_training:
break
utils.barrier()
training_stop = time.time()
training_time = training_stop - training_start
logging.info(f'Total training time {training_time:.0f} s')
table = TrainingTable()
avg_training_perf = sum(training_perf) / len(training_perf)
table.add(utils.get_world_size(), args.train_batch_size, test_bleu,
avg_training_perf, training_time)
if utils.get_rank() == 0:
table.write('Training Summary', args.math)
passed = utils.benchmark(test_bleu, args.target_bleu, train_perf,
args.target_perf)
if not passed:
sys.exit(1)
def main():
"""
Launches data-parallel multi-gpu training.
"""
mlperf_compliance.mlperf_log.LOGGER.propagate = False
mlperf_compliance.mlperf_log.setdefault(
root_dir=os.path.dirname(os.path.abspath(__file__)),
benchmark=mlperf_compliance.constants.GNMT,
stack_offset=1,
extra_print=False
)
mlperf_print(key=mlperf_compliance.constants.INIT_START,
log_all_ranks=True)
args = parse_args()
device = utils.set_device(args.cuda, args.local_rank)
distributed = utils.init_distributed(args.cuda)
# preinit and warmup streams/ groups for apex DDP communicators
allreduce_communicators=None
if distributed and args.apex_num_allreduce_streams > 1:
bucket_pgs = [torch.distributed.new_group() for _ in range(args.apex_num_allreduce_streams)]
bucket_streams = [torch.cuda.Stream() for _ in range(args.apex_num_allreduce_streams)]
for pg, stream in zip(bucket_pgs,bucket_streams):
with torch.cuda.stream(stream):
torch.distributed.all_reduce(torch.cuda.FloatTensor(1), group=pg)
allreduce_communicators=(bucket_pgs,bucket_streams)
args.rank = utils.get_rank()
if not args.cudnn:
torch.backends.cudnn.enabled = False
# create directory for results
save_path = os.path.join(args.results_dir, args.save)
args.save_path = save_path
os.makedirs(save_path, exist_ok=True)
# setup logging
log_filename = f'log_rank_{utils.get_rank()}.log'
utils.setup_logging(args.log_all_ranks,
os.path.join(save_path, log_filename))
if args.env:
utils.log_env_info()
logging.info(f'Saving results to: {save_path}')
logging.info(f'Run arguments: {args}')
# automatically set train_iter_size based on train_global_batch_size,
# world_size and per-worker train_batch_size
if args.train_global_batch_size is not None:
global_bs = args.train_global_batch_size
bs = args.train_batch_size
world_size = utils.get_world_size()
assert global_bs % (bs * world_size) == 0
args.train_iter_size = global_bs // (bs * world_size)
logging.info(f'Global batch size was set in the config, '
f'Setting train_iter_size to {args.train_iter_size}')
# setup L2 promotion
if args.cuda:
utils.l2_promote()
worker_seeds, shuffling_seeds = utils.setup_seeds(args.seed, args.epochs,
device)
worker_seed = worker_seeds[args.rank]
logging.info(f'Worker {args.rank} is using worker seed: {worker_seed}')
torch.manual_seed(worker_seed)
# build tokenizer
# https://github.com/mlperf/policies/issues/201
pad_vocab = utils.pad_vocabulary(args.math)
tokenizer = Tokenizer(os.path.join(args.dataset_dir, config.VOCAB_FNAME),
pad_vocab)
vocab_size = tokenizer.vocab_size
# build GNMT model
model_config = {'hidden_size': args.hidden_size,
'num_layers': args.num_layers,
'dropout': args.dropout, 'batch_first': False,
'share_embedding': args.share_embedding,
'fusion': args.fused_attention}
model = GNMT(vocab_size=vocab_size, **model_config)
logging.info(model)
# define loss function (criterion) and optimizer
criterion = build_criterion(vocab_size, config.PAD, args.smoothing,
args.fused_xentropy)
opt_config = {'optimizer': args.optimizer, 'lr': args.lr}
opt_config.update(literal_eval(args.optimizer_extra))
logging.info(f'Training optimizer config: {opt_config}')
num_parameters = sum([l.nelement() for l in model.parameters()])
logging.info(f'Number of parameters: {num_parameters}')
# create trainer
save_info = {'model_config': model_config, 'config': args, 'tokenizer':
tokenizer.get_state()}
loss_scaling = {'init_scale': args.init_scale, 'upscale_interval':
args.upscale_interval}
trainer_options = dict(
criterion=criterion,
grad_clip=args.grad_clip,
iter_size=args.train_iter_size,
save_path=save_path,
save_freq=args.save_freq,
save_info=save_info,
opt_config=opt_config,
batch_first=model.batch_first,
keep_checkpoints=args.keep_checkpoints,
math=args.math,
loss_scaling=loss_scaling,
print_freq=args.print_freq,
cuda=args.cuda,
distributed=distributed,
distributed_overlap_allreduce=args.enable_apex_allreduce_overlap,
distributed_overlap_num_allreduce_streams=args.apex_num_allreduce_streams,
distributed_overlap_allreduce_messagesize=args.apex_message_size,
distributed_overlap_allreduce_communicators=allreduce_communicators,
intra_epoch_eval=args.intra_epoch_eval,
prealloc_mode=args.prealloc_mode)
trainer_options['model'] = model
trainer = trainers.Seq2SeqTrainer(**trainer_options)
trainer.preallocate(args.train_batch_size, args.max_length_train,
training=True)
mlperf_print(key=mlperf_compliance.constants.INIT_STOP,
sync=True)
mlperf_print(key=mlperf_compliance.constants.RUN_START,
sync=True)
utils.barrier()
mlperf_print(key=mlperf_compliance.constants.MAX_SEQUENCE_LENGTH,
value=args.max_length_train,
metadata={'method': 'discard'})
if args.use_preproc_data:
train_data = PreprocessedDataset(
min_len=args.min_length_train,
max_len=args.max_length_train,
vocab_size=tokenizer.vocab_size,
)
train_data.read_data(
os.path.join(args.preproc_data_dir, 'training.bin'),
tokenizer.vocab_size,
)
train_data.prepare()
else:
train_data = LazyParallelDataset(
src_fname=os.path.join(args.dataset_dir, config.SRC_TRAIN_FNAME),
tgt_fname=os.path.join(args.dataset_dir, config.TGT_TRAIN_FNAME),
tokenizer=tokenizer,
min_len=args.min_length_train,
max_len=args.max_length_train,
sort=False,
max_size=args.max_size,
)
test_data = TextDataset(
src_fname=os.path.join(args.dataset_dir, config.SRC_TEST_FNAME),
tokenizer=tokenizer,
min_len=args.min_length_test,
max_len=args.max_length_test,
sort=True)
batching_opt = {'shard_size': args.shard_size,
'num_buckets': args.num_buckets}
# get data loaders
train_loader = train_data.get_loader(batch_size=args.train_batch_size,
seeds=shuffling_seeds,
batch_first=model.batch_first,
shuffle=True,
batching=args.batching,
batching_opt=batching_opt,
num_workers=args.train_loader_workers)
mlperf_print(key=mlperf_compliance.constants.GLOBAL_BATCH_SIZE,
value=args.train_batch_size * utils.get_world_size(),
sync=False)
test_loader = test_data.get_loader(batch_size=args.test_batch_size,
batch_first=model.batch_first,
shuffle=False,
num_workers=args.test_loader_workers)
translator = Translator(model=model,
tokenizer=tokenizer,
loader=test_loader,
beam_size=args.beam_size,
max_seq_len=args.max_length_test,
len_norm_factor=args.len_norm_factor,
len_norm_const=args.len_norm_const,
cov_penalty_factor=args.cov_penalty_factor,
cuda=args.cuda,
print_freq=args.print_freq,
dataset_dir=args.dataset_dir,
target_bleu=args.target_bleu,
save_path=args.save_path)
total_train_iters = len(train_loader) // args.train_iter_size * args.epochs
scheduler_config = {'warmup_steps': args.warmup_steps,
'remain_steps': args.remain_steps,
'decay_interval': args.decay_interval,
'decay_steps': args.decay_steps,
'decay_factor': args.decay_factor}
logging.info(f'Training LR schedule config: {scheduler_config}')
scheduler = WarmupMultiStepLR(trainer.optimizer, total_train_iters,
**scheduler_config)
trainer.scheduler = scheduler
trainer.translator = translator
# optionally resume from a checkpoint
if args.resume:
checkpoint_file = args.resume
if os.path.isdir(checkpoint_file):
checkpoint_file = os.path.join(
checkpoint_file, 'model_best.pth')
if os.path.isfile(checkpoint_file):
trainer.load(checkpoint_file)
else:
logging.error(f'No checkpoint found at {args.resume}')
# training loop
break_training = False
test_bleu = None
for epoch in range(args.start_epoch, args.epochs):
mlperf_print(key=mlperf_compliance.constants.BLOCK_START,
metadata={'first_epoch_num': epoch + 1,
'epoch_count': 1},
sync=True)
mlperf_print(key=mlperf_compliance.constants.EPOCH_START,
metadata={'epoch_num': epoch + 1},
sync=True)
logging.info(f'Starting epoch {epoch}')
train_loader.sampler.set_epoch(epoch)
trainer.epoch = epoch
train_loss, train_perf = trainer.optimize(train_loader)
mlperf_print(key=mlperf_compliance.constants.EPOCH_STOP,
metadata={'epoch_num': epoch + 1},
sync=True)
if args.eval:
mlperf_print(key=mlperf_compliance.constants.EVAL_START,
metadata={'epoch_num': epoch + 1},
sync=True)
test_bleu, break_training = translator.run(calc_bleu=True,
epoch=epoch)
mlperf_print(key=mlperf_compliance.constants.EVAL_ACCURACY,
value=test_bleu,
metadata={'epoch_num': epoch + 1},
sync=False)
mlperf_print(key=mlperf_compliance.constants.EVAL_STOP,
metadata={'epoch_num': epoch + 1},
sync=True)
acc_log = []
acc_log += [f'Summary: Epoch: {epoch}']
acc_log += [f'Training Loss: {train_loss:.4f}']
if args.eval:
acc_log += [f'Test BLEU: {test_bleu:.2f}']
perf_log = []
perf_log += [f'Performance: Epoch: {epoch}']
perf_log += [f'Training: {train_perf:.0f} Tok/s']
if args.rank == 0:
logging.info('\t'.join(acc_log))
logging.info('\t'.join(perf_log))
logging.info(f'Finished epoch {epoch}')
mlperf_print(key=mlperf_compliance.constants.BLOCK_STOP,
metadata={'first_epoch_num': epoch + 1},
sync=True)
if break_training:
break
if args.use_preproc_data:
train_data.finalize()
status = 'success' if break_training else 'aborted'
mlperf_print(key=mlperf_compliance.constants.RUN_STOP,
metadata={'status': status},
sync=True)