def _main(cfg: DictConfig, output_file):
logging.basicConfig(
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
level=os.environ.get("LOGLEVEL", "INFO").upper(),
stream=output_file,
)
logger = logging.getLogger('fairseq_cli.predict')
utils.import_user_module(cfg.common)
if cfg.dataset.max_tokens is None and cfg.dataset.batch_size is None:
cfg.dataset.max_tokens = 12000
logger.info(cfg)
# Fix seed for stochastic decoding
if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
np.random.seed(cfg.common.seed)
utils.set_torch_seed(cfg.common.seed)
use_cuda = torch.cuda.is_available() and not cfg.common.cpu
# Load dataset splits
task = tasks.setup_task(cfg.task)
task.load_dataset(cfg.dataset.gen_subset)
# Set dictionaries
src_dict = getattr(task, 'source_dictionary', None)
tag_dict = task.tag_dictionary
overrides = ast.literal_eval(cfg.common_eval.model_overrides)
# Load ensemble
logger.info("loading model(s) from {}".format(cfg.common_eval.path))
models, saved_cfg = checkpoint_utils.load_model_ensemble(
utils.split_paths(cfg.common_eval.path),
arg_overrides=overrides,
task=task,
suffix=cfg.checkpoint.checkpoint_suffix,
strict=(cfg.checkpoint.checkpoint_shard_count == 1),
num_shards=cfg.checkpoint.checkpoint_shard_count,
)
# Optimize ensemble for generation
for model in models:
model.prepare_for_inference_(cfg)
if cfg.common.fp16:
model.half()
if use_cuda:
model.cuda()
# Load dataset (possibly sharded)
itr = task.get_batch_iterator(
dataset=task.dataset(cfg.dataset.gen_subset),
max_tokens=cfg.dataset.max_tokens,
max_sentences=cfg.dataset.batch_size,
max_positions=utils.resolve_max_positions(
task.max_positions(),
*[model.max_positions() for model in models]
),
ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=cfg.dataset.required_batch_size_multiple,
seed=cfg.common.seed,
num_shards=cfg.distributed_training.distributed_world_size,
shard_id=cfg.distributed_training.distributed_rank,
num_workers=cfg.dataset.num_workers,
data_buffer_size=cfg.dataset.data_buffer_size,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=cfg.common.log_format,
log_interval=cfg.common.log_interval,
default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"),
)
# Initialize generator
gen_timer = StopwatchMeter()
# Handle tokenization and BPE
tokenizer = encoders.build_tokenizer(cfg.tokenizer)
bpe = encoders.build_bpe(cfg.bpe)
def decode_fn(x): # decode tag
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
num_sentences = 0
wps_meter = TimeMeter()
for sample in progress:
sample = utils.move_to_cuda(sample) if use_cuda else sample
if 'net_input' not in sample:
continue
gen_timer.start()
hypos = task.inference_step(models, sample, cfg.task.tagging_head_name)
num_generated_tokens = sample['ntokens']
gen_timer.stop(num_generated_tokens)
for i, sample_id in enumerate(sample['id'].tolist()):
hypo = hypos[i]
hypo_tokens = np.array(
hypo) + tag_dict.nspecial # can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first
hypo_str = tag_dict.string(hypo_tokens)
has_target = sample['target'] is not None
# Remove padding
if 'src_tokens' in sample['net_input']:
src_tokens = utils.strip_pad(sample['net_input']['src_tokens'][i, :], src_dict.pad())
src_str = src_dict.string(src_tokens, cfg.common_eval.post_process)
assert len(hypo) == src_tokens.numel()
if has_target:
tag_offset = 1
tag_tokens = utils.strip_pad(sample['target'][i, :],
src_dict.pad()).int().cpu() - tag_offset + tag_dict.nspecial
tag_str = tag_dict.string(tag_tokens)
src_str = decode_fn(src_str)
tag_str = decode_fn(tag_str)
if not cfg.common_eval.quiet:
if src_dict is not None:
print('S-{}\t{}'.format(sample_id, src_str), file=output_file)
if has_target:
print('T-{}\t{}'.format(sample_id, tag_str), file=output_file)
print('H-{}\t{}'.format(sample_id, hypo_str), file=output_file)
wps_meter.update(num_generated_tokens)
progress.log({'wps': round(wps_meter.avg)})
num_sentences += sample["nsentences"] if "nsentences" in sample else sample['id'].numel()
logger.info('NOTE: hypothesis and token scores are output in base 2')
logger.info('Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)'.format(
num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg))
def validate(args, trainer, task, epoch_itr, subsets, experiment=None):
"""Evaluate the model on the validation set(s) and return the losses."""
if args.fixed_validation_seed is not None:
# set fixed seed for every validation
utils.set_torch_seed(args.fixed_validation_seed)
valid_losses = []
for subset in subsets:
# Initialize data iterator
itr = task.get_batch_iterator(
dataset=task.dataset(subset),
max_tokens=args.max_tokens_valid,
max_sentences=args.max_sentences_valid,
max_positions=utils.resolve_max_positions(
task.max_positions(),
trainer.get_model().max_positions(),
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
epoch=epoch_itr.epoch,
prefix=f"valid on '{subset}' subset",
tensorboard_logdir=(
args.tensorboard_logdir if distributed_utils.is_master(args) else None
),
default_log_format=('tqdm' if not args.no_progress_bar else 'simple'),
)
# create a new root metrics aggregator so validation metrics
# don't pollute other aggregators (e.g., train meters)
all_preds, all_labels = [], []
with metrics.aggregate(new_root=True) as agg:
for sample in progress:
logging_outputs, preds, labels = trainer.valid_step(sample)
if preds is not None:
all_preds.extend(preds)
all_labels.extend(labels)
else:
all_preds, all_labels = None, None
if all_preds is not None:
all_preds = torch.cat(all_preds).cpu().numpy()
all_labels = torch.cat(all_labels).cpu().numpy()
# log validation stats
stats = get_valid_stats(args, trainer, agg.get_smoothed_values(), all_preds, all_labels)
progress.print(stats, tag=subset, step=trainer.get_num_updates())
if experiment:
# metrics = {'val_{}'.format(k): v for (k, v) in stats.items()}
# experiment.log_metrics(metrics, step=trainer.get_num_updates())
for key, val in stats.items():
if isinstance(val, meters.AverageMeter):
val = val.avg
if val is not None:
experiment.log_metric("val_{}".format(key), val, step=trainer.get_num_updates())
valid_losses.append(stats[args.best_checkpoint_metric])
return valid_losses
def main(args, init_distributed=False):
utils.import_user_module(args)
assert args.max_tokens is not None or args.max_sentences is not None, \
'Must specify batch size either with --max-tokens or --max-sentences'
metrics.reset()
# Initialize CUDA and distributed training
if torch.cuda.is_available() and not args.cpu and not getattr(args, 'tpu', False):
torch.cuda.set_device(args.device_id)
np.random.seed(args.seed)
utils.set_torch_seed(args.seed)
if init_distributed:
args.distributed_rank = distributed_utils.distributed_init(args)
if distributed_utils.is_master(args):
checkpoint_utils.verify_checkpoint_directory(args.save_dir)
# Print args
logger.info(args)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(args)
# Load valid dataset (we load training data below, based on the latest checkpoint)
for valid_sub_split in args.valid_subset.split(','):
task.load_dataset(valid_sub_split, combine=False, epoch=1)
# Build model and criterion
model = task.build_model(args)
criterion = task.build_criterion(args)
logger.info(model)
logger.info('model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))
logger.info('num. model params: {} (num. trained: {})'.format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
))
# (optionally) Configure quantization
if args.quantization_config_path is not None:
quantizer = quantization_utils.Quantizer(
config_path=args.quantization_config_path,
max_epoch=args.max_epoch,
max_update=args.max_update,
)
else:
quantizer = None
# Build trainer
if args.model_parallel_size == 1:
trainer = Trainer(args, task, model, criterion, quantizer)
else:
trainer = MegatronTrainer(args, task, model, criterion)
logger.info('training on {} devices (GPUs/TPUs)'.format(args.distributed_world_size))
logger.info('max tokens per GPU = {} and max sentences per GPU = {}'.format(
args.max_tokens,
args.max_sentences,
))
# Load the latest checkpoint if one is available and restore the
# corresponding train iterator
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(args, trainer)
if args.tpu:
import torch_xla.core.xla_model as xm
xm.rendezvous('load_checkpoint') # wait for all workers
xm.mark_step()
# 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 = meters.StopwatchMeter()
train_meter.start()
while (
lr > args.min_lr
and epoch_itr.next_epoch_idx <= max_epoch
):
# train for one epoch
valid_losses = train(args, trainer, task, epoch_itr, max_update)
if should_stop_early(args, valid_losses[0]) or trainer.get_num_updates() >= max_update:
break
# only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
epoch_itr = trainer.get_train_iterator(
epoch_itr.next_epoch_idx,
# sharded data: get train iterator for next epoch
load_dataset=(os.pathsep in getattr(args, 'data', '')),
)
train_meter.stop()
logger.info('done training in {:.1f} seconds'.format(train_meter.sum))
def main(cfg: DictConfig) -> None:
if isinstance(cfg, argparse.Namespace):
cfg = convert_namespace_to_omegaconf(cfg)
utils.import_user_module(cfg.common)
assert (
cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None
), "Must specify batch size either with --max-tokens or --batch-size"
metrics.reset()
np.random.seed(cfg.common.seed)
utils.set_torch_seed(cfg.common.seed)
if distributed_utils.is_master(cfg.distributed_training):
checkpoint_utils.verify_checkpoint_directory(cfg.checkpoint.save_dir)
# Print args
logger.info(cfg)
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(cfg.task)
# Load valid dataset (we load training data below, based on the latest checkpoint)
for valid_sub_split in cfg.dataset.valid_subset.split(","):
task.load_dataset(valid_sub_split, combine=False, epoch=1)
assert cfg.criterion, "Please specify criterion to train a model"
# Build model and criterion
model = task.build_model(cfg.model)
criterion = task.build_criterion(cfg.criterion)
logger.info(model)
logger.info("task: {}".format(task.__class__.__name__))
logger.info("model: {}".format(model.__class__.__name__))
logger.info("criterion: {})".format(criterion.__class__.__name__))
logger.info(
"num. model params: {} (num. trained: {})".format(
sum(p.numel() for p in model.parameters()),
sum(p.numel() for p in model.parameters() if p.requires_grad),
)
)
# (optionally) Configure quantization
if cfg.common.quantization_config_path is not None:
quantizer = quantization_utils.Quantizer(
config_path=cfg.common.quantization_config_path,
max_epoch=cfg.optimization.max_epoch,
max_update=cfg.optimization.max_update,
)
else:
quantizer = None
# Build trainer
if cfg.common.model_parallel_size == 1:
trainer = Trainer(cfg, task, model, criterion, quantizer)
else:
trainer = MegatronTrainer(cfg, task, model, criterion)
logger.info(
"training on {} devices (GPUs/TPUs)".format(
cfg.distributed_training.distributed_world_size
)
)
logger.info(
"max tokens per GPU = {} and batch size per GPU = {}".format(
cfg.dataset.max_tokens,
cfg.dataset.batch_size,
)
)
# Load the latest checkpoint if one is available and restore the
# corresponding train iterator
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(
cfg.checkpoint,
trainer,
# don't cache epoch iterators for sharded datasets
disable_iterator_cache=task.has_sharded_data("train"),
)
max_epoch = cfg.optimization.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = meters.StopwatchMeter()
train_meter.start()
while lr > cfg.optimization.min_lr and epoch_itr.next_epoch_idx <= max_epoch:
# train for one epoch
valid_losses, should_stop = train(cfg, trainer, task, epoch_itr)
if should_stop:
break
# only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
epoch_itr = trainer.get_train_iterator(
epoch_itr.next_epoch_idx,
# sharded data: get train iterator for next epoch
load_dataset=task.has_sharded_data("train"),
# don't cache epoch iterators for sharded datasets
disable_iterator_cache=task.has_sharded_data("train"),
)
train_meter.stop()
logger.info("done training in {:.1f} seconds".format(train_meter.sum))
def _main(args, output_file):
logging.basicConfig(
format='%(asctime)s | %(levelname)s | %(name)s | %(message)s',
datefmt='%Y-%m-%d %H:%M:%S',
level=logging.INFO,
stream=output_file,
)
logger = logging.getLogger('fairseq_cli.generate')
utils.import_user_module(args)
if args.max_tokens is None and args.max_sentences is None:
args.max_tokens = 12000
logger.info(args)
# Fix seed for stochastic decoding
if args.seed is not None and not args.no_seed_provided:
np.random.seed(args.seed)
utils.set_torch_seed(args.seed)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load dataset splits
task = tasks.setup_task(args)
task.load_dataset(args.gen_subset)
# Set dictionaries
try:
src_dict = getattr(task, 'source_dictionary', None)
except NotImplementedError:
src_dict = None
tgt_dict = task.target_dictionary
# Load ensemble
logger.info('loading model(s) from {}'.format(args.path))
models, _model_args = checkpoint_utils.load_model_ensemble(
utils.split_paths(args.path),
arg_overrides=eval(args.model_overrides),
task=task,
suffix=getattr(args, "checkpoint_suffix", ""),
)
# Optimize ensemble for generation
for model in models:
model.prepare_for_inference_(args)
if args.fp16:
model.half()
if use_cuda:
model.cuda()
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
# Load dataset (possibly sharded)
itr = task.get_batch_iterator(
dataset=task.dataset(args.gen_subset),
max_tokens=args.max_tokens,
max_sentences=args.max_sentences,
max_positions=utils.resolve_max_positions(
task.max_positions(),
*[model.max_positions() for model in models]
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
num_shards=args.num_shards,
shard_id=args.shard_id,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
default_log_format=('tqdm' if not args.no_progress_bar else 'none'),
)
# Initialize generator
gen_timer = StopwatchMeter()
generator = task.build_generator(models, args)
# Handle tokenization and BPE
tokenizer = encoders.build_tokenizer(args)
bpe = encoders.build_bpe(args)
def decode_fn(x):
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
scorer = scoring.build_scorer(args, tgt_dict)
num_sentences = 0
has_target = True
wps_meter = TimeMeter()
for sample in progress:
sample = utils.move_to_cuda(sample) if use_cuda else sample
if 'net_input' not in sample:
continue
prefix_tokens = None
if args.prefix_size > 0:
prefix_tokens = sample['target'][:, :args.prefix_size]
gen_timer.start()
hypos = task.inference_step(generator, models, sample, prefix_tokens)
num_generated_tokens = sum(len(h[0]['tokens']) for h in hypos)
gen_timer.stop(num_generated_tokens)
for i, sample_id in enumerate(sample['id'].tolist()):
has_target = sample['target'] is not None
# Remove padding
if 'src_tokens' in sample['net_input']:
src_tokens = utils.strip_pad(sample['net_input']['src_tokens'][i, :], tgt_dict.pad())
else:
src_tokens = None
target_tokens = None
if has_target:
target_tokens = utils.strip_pad(sample['target'][i, :], tgt_dict.pad()).int().cpu()
# Either retrieve the original sentences or regenerate them from tokens.
if align_dict is not None:
src_str = task.dataset(args.gen_subset).src.get_original_text(sample_id)
target_str = task.dataset(args.gen_subset).tgt.get_original_text(sample_id)
else:
if src_dict is not None:
src_str = src_dict.string(src_tokens, args.remove_bpe)
else:
src_str = ""
if has_target:
target_str = tgt_dict.string(
target_tokens,
args.remove_bpe,
escape_unk=True,
extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator),
)
src_str = decode_fn(src_str)
if has_target:
target_str = decode_fn(target_str)
if not args.quiet:
if src_dict is not None:
print('S-{}\t{}'.format(sample_id, src_str), file=output_file)
if has_target:
print('T-{}\t{}'.format(sample_id, target_str), file=output_file)
# Process top predictions
for j, hypo in enumerate(hypos[i][:args.nbest]):
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo['tokens'].int().cpu(),
src_str=src_str,
alignment=hypo['alignment'],
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator),
)
detok_hypo_str = decode_fn(hypo_str)
if not args.quiet:
score = hypo['score'] / math.log(2) # convert to base 2
# original hypothesis (after tokenization and BPE)
print('H-{}\t{}\t{}'.format(sample_id, score, hypo_str), file=output_file)
# detokenized hypothesis
print('D-{}\t{}\t{}'.format(sample_id, score, detok_hypo_str), file=output_file)
print('P-{}\t{}'.format(
sample_id,
' '.join(map(
lambda x: '{:.4f}'.format(x),
# convert from base e to base 2
hypo['positional_scores'].div_(math.log(2)).tolist(),
))
), file=output_file)
if args.print_alignment:
print('A-{}\t{}'.format(
sample_id,
' '.join(['{}-{}'.format(src_idx, tgt_idx) for src_idx, tgt_idx in alignment])
), file=output_file)
if args.print_step:
print('I-{}\t{}'.format(sample_id, hypo['steps']), file=output_file)
if getattr(args, 'retain_iter_history', False):
for step, h in enumerate(hypo['history']):
_, h_str, _ = utils.post_process_prediction(
hypo_tokens=h['tokens'].int().cpu(),
src_str=src_str,
alignment=None,
align_dict=None,
tgt_dict=tgt_dict,
remove_bpe=None,
)
print('E-{}_{}\t{}'.format(sample_id, step, h_str), file=output_file)
# Score only the top hypothesis
if has_target and j == 0:
if align_dict is not None or args.remove_bpe is not None:
# Convert back to tokens for evaluation with unk replacement and/or without BPE
target_tokens = tgt_dict.encode_line(target_str, add_if_not_exist=True)
hypo_tokens = tgt_dict.encode_line(detok_hypo_str, add_if_not_exist=True)
if hasattr(scorer, 'add_string'):
scorer.add_string(target_str, detok_hypo_str)
else:
scorer.add(target_tokens, hypo_tokens)
wps_meter.update(num_generated_tokens)
progress.log({'wps': round(wps_meter.avg)})
num_sentences += sample["nsentences"] if "nsentences" in sample else sample['id'].numel()
logger.info('NOTE: hypothesis and token scores are output in base 2')
logger.info('Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)'.format(
num_sentences, gen_timer.n, gen_timer.sum, num_sentences / gen_timer.sum, 1. / gen_timer.avg))
if has_target:
if args.bpe and not args.sacrebleu:
if args.remove_bpe:
logger.warning("BLEU score is being computed by splitting detokenized string on spaces, this is probably not what you want. Use --sacrebleu for standard 13a BLEU tokenization")
else:
logger.warning("If you are using BPE on the target side, the BLEU score is computed on BPE tokens, not on proper words. Use --sacrebleu for standard 13a BLEU tokenization")
# use print to be consistent with other main outputs: S-, H-, T-, D- and so on
print(
'Generate {} with beam={}: {}'.format(args.gen_subset, args.beam, scorer.result_string()),
file=output_file)
return scorer
def main(cfg: FairseqConfig) -> None:
if isinstance(cfg, argparse.Namespace):
cfg = convert_namespace_to_omegaconf(cfg)
utils.import_user_module(cfg.common)
if distributed_utils.is_master(cfg.distributed_training) and "job_logging_cfg" in cfg:
# make hydra logging work with ddp (see # see https://github.com/facebookresearch/hydra/issues/1126)
logging.config.dictConfig(OmegaConf.to_container(cfg.job_logging_cfg))
assert (
cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None
), "Must specify batch size either with --max-tokens or --batch-size"
metrics.reset()
np.random.seed(cfg.common.seed)
utils.set_torch_seed(cfg.common.seed)
if distributed_utils.is_master(cfg.distributed_training):
checkpoint_utils.verify_checkpoint_directory(cfg.checkpoint.save_dir)
# Print args
logger.info(cfg)
if cfg.checkpoint.write_checkpoints_asynchronously:
try:
import iopath # noqa: F401
except ImportError:
logging.exception(
"Asynchronous checkpoint writing is specified but iopath is "
"not installed: `pip install iopath`"
)
return
# Setup task, e.g., translation, language modeling, etc.
task = tasks.setup_task(cfg.task)
assert cfg.criterion, "Please specify criterion to train a model"
# Build model and criterion
if cfg.distributed_training.ddp_backend == "fully_sharded":
with fsdp_enable_wrap(cfg.distributed_training):
model = fsdp_wrap(task.build_model(cfg.model))
else:
model = task.build_model(cfg.model)
criterion = task.build_criterion(cfg.criterion)
logger.info(model)
logger.info("task: {}".format(task.__class__.__name__))
logger.info("model: {}".format(model.__class__.__name__))
logger.info("criterion: {}".format(criterion.__class__.__name__))
logger.info(
"num. shared model params: {:,} (num. trained: {:,})".format(
sum(p.numel() for p in model.parameters() if not getattr(p, "expert", False)),
sum(p.numel() for p in model.parameters() if not getattr(p, "expert", False) and p.requires_grad)
)
)
logger.info(
"num. expert model params: {} (num. trained: {})".format(
sum(p.numel() for p in model.parameters() if getattr(p, "expert", False)),
sum(p.numel() for p in model.parameters() if getattr(p, "expert", False) and p.requires_grad),
)
)
# Load valid dataset (we load training data below, based on the latest checkpoint)
# We load the valid dataset AFTER building the model
for valid_sub_split in cfg.dataset.valid_subset.split(","):
task.load_dataset(valid_sub_split, combine=False, epoch=1)
# (optionally) Configure quantization
if cfg.common.quantization_config_path is not None:
quantizer = quantization_utils.Quantizer(
config_path=cfg.common.quantization_config_path,
max_epoch=cfg.optimization.max_epoch,
max_update=cfg.optimization.max_update,
)
else:
quantizer = None
# Build trainer
if cfg.common.model_parallel_size == 1:
trainer = Trainer(cfg, task, model, criterion, quantizer)
else:
trainer = MegatronTrainer(cfg, task, model, criterion)
logger.info(
"training on {} devices (GPUs/TPUs)".format(
cfg.distributed_training.distributed_world_size
)
)
logger.info(
"max tokens per device = {} and max sentences per device = {}".format(
cfg.dataset.max_tokens,
cfg.dataset.batch_size,
)
)
# Load the latest checkpoint if one is available and restore the
# corresponding train iterator
extra_state, epoch_itr = checkpoint_utils.load_checkpoint(
cfg.checkpoint,
trainer,
# don't cache epoch iterators for sharded datasets
disable_iterator_cache=task.has_sharded_data("train"),
)
if cfg.common.tpu:
import torch_xla.core.xla_model as xm
xm.rendezvous("load_checkpoint") # wait for all workers
max_epoch = cfg.optimization.max_epoch or math.inf
lr = trainer.get_lr()
train_meter = meters.StopwatchMeter()
train_meter.start()
while epoch_itr.next_epoch_idx <= max_epoch:
if lr <= cfg.optimization.stop_min_lr:
logger.info(
f"stopping training because current learning rate ({lr}) is smaller "
"than or equal to minimum learning rate "
f"(--stop-min-lr={cfg.optimization.stop_min_lr})"
)
break
# train for one epoch
valid_losses, should_stop = train(cfg, trainer, task, epoch_itr)
if should_stop:
break
# only use first validation loss to update the learning rate
lr = trainer.lr_step(epoch_itr.epoch, valid_losses[0])
epoch_itr = trainer.get_train_iterator(
epoch_itr.next_epoch_idx,
# sharded data: get train iterator for next epoch
load_dataset=task.has_sharded_data("train"),
# don't cache epoch iterators for sharded datasets
disable_iterator_cache=task.has_sharded_data("train"),
)
train_meter.stop()
logger.info("done training in {:.1f} seconds".format(train_meter.sum))
# ioPath implementation to wait for all asynchronous file writes to complete.
if cfg.checkpoint.write_checkpoints_asynchronously:
logger.info(
"ioPath PathManager waiting for all asynchronous checkpoint "
"writes to finish."
)
PathManager.async_close()
logger.info("ioPath PathManager finished waiting.")
def _set_seed(self):
# Set seed based on args.seed and the update number so that we get
# reproducible results when resuming from checkpoints
seed = self.cfg.common.seed + self.get_num_updates()
utils.set_torch_seed(seed)
def _main(args, output_file):
logging.basicConfig(
format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
datefmt="%Y-%m-%d %H:%M:%S",
level=os.environ.get("LOGLEVEL", "INFO").upper(),
stream=output_file,
)
logger = logging.getLogger("fairseq_cli.generate")
utils.import_user_module(args)
if args.max_tokens is None and args.batch_size is None:
args.max_tokens = 12000
logger.info(args)
# Fix seed for stochastic decoding
if args.seed is not None and not args.no_seed_provided:
np.random.seed(args.seed)
utils.set_torch_seed(args.seed)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load dataset splits
task = tasks.setup_task(args)
task.load_dataset(args.gen_subset)
# Set dictionaries
try:
src_dict = getattr(task, "source_dictionary", None)
except NotImplementedError:
src_dict = None
tgt_dict = task.target_dictionary
overrides = ast.literal_eval(args.model_overrides)
# Load ensemble
logger.info("loading model(s) from {}".format(args.path))
models, _model_args = checkpoint_utils.load_model_ensemble(
utils.split_paths(args.path),
arg_overrides=overrides,
task=task,
suffix=getattr(args, "checkpoint_suffix", ""),
strict=(args.checkpoint_shard_count == 1),
num_shards=args.checkpoint_shard_count,
)
if args.lm_path is not None:
overrides["data"] = args.data
try:
lms, _ = checkpoint_utils.load_model_ensemble(
[args.lm_path],
arg_overrides=overrides,
task=None,
)
except:
logger.warning(
f"Failed to load language model! Please make sure that the language model dict is the same "
f"as target dict and is located in the data dir ({args.data})"
)
raise
assert len(lms) == 1
else:
lms = [None]
# Optimize ensemble for generation
for model in chain(models, lms):
if model is None:
continue
if args.fp16:
model.half()
if use_cuda and not args.pipeline_model_parallel:
model.cuda()
model.prepare_for_inference_(args)
# Load alignment dictionary for unknown word replacement
# (None if no unknown word replacement, empty if no path to align dictionary)
align_dict = utils.load_align_dict(args.replace_unk)
# Load dataset (possibly sharded)
itr = task.get_batch_iterator(
dataset=task.dataset(args.gen_subset),
max_tokens=args.max_tokens,
max_sentences=args.batch_size,
max_positions=utils.resolve_max_positions(
task.max_positions(), *[model.max_positions() for model in models]
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
num_shards=args.num_shards,
shard_id=args.shard_id,
num_workers=args.num_workers,
data_buffer_size=args.data_buffer_size,
).next_epoch_itr(shuffle=False)
progress = progress_bar.progress_bar(
itr,
log_format=args.log_format,
log_interval=args.log_interval,
default_log_format=("tqdm" if not args.no_progress_bar else "none"),
)
# Initialize generator
gen_timer = StopwatchMeter()
extra_gen_cls_kwargs = {"lm_model": lms[0], "lm_weight": args.lm_weight}
generator = task.build_generator(
models, args, extra_gen_cls_kwargs=extra_gen_cls_kwargs
)
# Handle tokenization and BPE
tokenizer = task.build_tokenizer(args)
bpe = task.build_bpe(args)
def decode_fn(x):
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
scorer = scoring.build_scorer(args, tgt_dict)
num_sentences = 0
has_target = True
wps_meter = TimeMeter()
for sample in progress:
sample = utils.move_to_cuda(sample) if use_cuda else sample
if "net_input" not in sample:
continue
prefix_tokens = None
if args.prefix_size > 0:
prefix_tokens = sample["target"][:, : args.prefix_size]
constraints = None
if "constraints" in sample:
constraints = sample["constraints"]
gen_timer.start()
hypos = task.inference_step(
generator,
models,
sample,
prefix_tokens=prefix_tokens,
constraints=constraints,
)
num_generated_tokens = sum(len(h[0]["tokens"]) for h in hypos)
gen_timer.stop(num_generated_tokens)
for i, sample_id in enumerate(sample["id"].tolist()):
has_target = sample["target"] is not None
# Remove padding
if "src_tokens" in sample["net_input"]:
src_tokens = utils.strip_pad(
sample["net_input"]["src_tokens"][i, :], tgt_dict.pad()
)
else:
src_tokens = None
target_tokens = None
if has_target:
target_tokens = (
utils.strip_pad(sample["target"][i, :], tgt_dict.pad()).int().cpu()
)
# Either retrieve the original sentences or regenerate them from tokens.
if align_dict is not None:
src_str = task.dataset(args.gen_subset).src.get_original_text(sample_id)
target_str = task.dataset(args.gen_subset).tgt.get_original_text(
sample_id
)
last_src_str = ""
else:
if src_dict is not None:
src_str = src_dict.string(src_tokens, args.remove_bpe)
# src_tokens may contain multiple sentence
# we extract the last sent to compute mutual information
last_src_str = src_dict.string(extract_last_sent(src_tokens, src_dict.eos_index),
args.remove_bpe)
else:
src_str = ""
last_src_str = ""
if has_target:
target_str = tgt_dict.string(
target_tokens,
args.remove_bpe,
escape_unk=True,
extra_symbols_to_ignore=get_symbols_to_strip_from_output(
generator
),
)
src_str = decode_fn(src_str)
last_src_str = decode_fn(last_src_str)
if has_target:
target_str = decode_fn(target_str)
if not args.quiet:
if src_dict is not None:
print("S-{}\t{}".format(sample_id, src_str), file=output_file)
if src_dict is not None:
print("L-{}\t{}".format(sample_id, last_src_str), file=output_file)
if has_target:
print("T-{}\t{}".format(sample_id, target_str), file=output_file)
# Process top predictions
for j, hypo in enumerate(hypos[i][: args.nbest]):
hypo_tokens, hypo_str, alignment = utils.post_process_prediction(
hypo_tokens=hypo["tokens"].int().cpu(),
src_str=src_str,
alignment=hypo["alignment"],
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
extra_symbols_to_ignore=get_symbols_to_strip_from_output(generator),
)
detok_hypo_str = decode_fn(hypo_str)
if not args.quiet:
score = hypo["score"] / math.log(2) # convert to base 2
# original hypothesis (after tokenization and BPE)
print(
"H-{}\t{}\t{}".format(sample_id, score, hypo_str),
file=output_file,
)
# detokenized hypothesis
print(
"D-{}\t{}\t{}".format(sample_id, score, detok_hypo_str),
file=output_file,
)
print(
"P-{}\t{}".format(
sample_id,
" ".join(
map(
lambda x: "{:.4f}".format(x),
# convert from base e to base 2
hypo["positional_scores"]
.div_(math.log(2))
.tolist(),
)
),
),
file=output_file,
)
if args.print_alignment:
print(
"A-{}\t{}".format(
sample_id,
" ".join(
[
"{}-{}".format(src_idx, tgt_idx)
for src_idx, tgt_idx in alignment
]
),
),
file=output_file,
)
if args.print_step:
print(
"I-{}\t{}".format(sample_id, hypo["steps"]),
file=output_file,
)
if getattr(args, "retain_iter_history", False):
for step, h in enumerate(hypo["history"]):
_, h_str, _ = utils.post_process_prediction(
hypo_tokens=h["tokens"].int().cpu(),
src_str=src_str,
alignment=None,
align_dict=None,
tgt_dict=tgt_dict,
remove_bpe=None,
)
print(
"E-{}_{}\t{}".format(sample_id, step, h_str),
file=output_file,
)
if args.print_attention:
attention = hypo["attention"]
attention_jsonl = json.dumps(attention.cpu().numpy().tolist())
print(
"Z-{}\t{}".format(sample_id, attention_jsonl)
)
# Score only the top hypothesis
if has_target and j == 0:
if align_dict is not None or args.remove_bpe is not None:
# Convert back to tokens for evaluation with unk replacement and/or without BPE
target_tokens = tgt_dict.encode_line(
target_str, add_if_not_exist=True
)
hypo_tokens = tgt_dict.encode_line(
detok_hypo_str, add_if_not_exist=True
)
if hasattr(scorer, "add_string"):
scorer.add_string(target_str, detok_hypo_str)
else:
scorer.add(target_tokens, hypo_tokens)
wps_meter.update(num_generated_tokens)
progress.log({"wps": round(wps_meter.avg)})
num_sentences += (
sample["nsentences"] if "nsentences" in sample else sample["id"].numel()
)
logger.info("NOTE: hypothesis and token scores are output in base 2")
logger.info(
"Translated {} sentences ({} tokens) in {:.1f}s ({:.2f} sentences/s, {:.2f} tokens/s)".format(
num_sentences,
gen_timer.n,
gen_timer.sum,
num_sentences / gen_timer.sum,
1.0 / gen_timer.avg,
)
)
if has_target:
if args.bpe and not args.sacrebleu:
if args.remove_bpe:
logger.warning(
"BLEU score is being computed by splitting detokenized string on spaces, this is probably not what you want. Use --sacrebleu for standard 13a BLEU tokenization"
)
else:
logger.warning(
"If you are using BPE on the target side, the BLEU score is computed on BPE tokens, not on proper words. Use --sacrebleu for standard 13a BLEU tokenization"
)
# use print to be consistent with other main outputs: S-, H-, T-, D- and so on
print(
"Generate {} with beam={}: {}".format(
args.gen_subset, args.beam, scorer.result_string()
),
file=output_file,
)
return scorer
def validate(args, trainer, task, epoch_itr, subsets):
"""Evaluate the model on the validation set(s) and return the losses."""
if args.fixed_validation_seed is not None:
# set fixed seed for every validation
utils.set_torch_seed(args.fixed_validation_seed)
# import pdb;pdb.set_trace()
valid_losses = []
for subset in subsets:
# Initialize data iterator
itr = task.get_batch_iterator(
dataset=task.dataset(subset),
max_tokens=args.max_tokens_valid,
max_sentences=args.max_sentences_valid,
max_positions=utils.resolve_max_positions(
task.max_positions(),
trainer.get_model().max_positions(),
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
# print(args.max_tokens_valid, args.max_sentences_valid, utils.resolve_max_positions(task.max_positions(), trainer.get_model().max_positions()))
# print(args.skip_invalid_size_inputs_valid_test,args.required_batch_size_multiple,)
progress = progress_bar.build_progress_bar(
args,
itr,
epoch_itr.epoch,
prefix='valid on \'{}\' subset'.format(subset),
no_progress_bar='simple')
# reset validation loss meters
for k in ['valid_loss', 'valid_nll_loss']:
meter = trainer.get_meter(k)
if meter is not None:
meter.reset()
extra_meters = collections.defaultdict(lambda: AverageMeter())
for sample in progress:
log_output = trainer.valid_step(sample)
for k, v in log_output.items():
if k in [
'loss', 'nll_loss', 'ntokens', 'nsentences',
'sample_size'
]:
continue
extra_meters[k].update(v)
# log validation stats
stats = get_valid_stats(trainer, args, extra_meters)
for k, meter in extra_meters.items():
stats[k] = meter.avg
progress.print(stats, tag=subset, step=trainer.get_num_updates())
valid_losses.append(stats[args.best_checkpoint_metric].avg if args.
best_checkpoint_metric ==
'loss' else stats[args.best_checkpoint_metric])
return valid_losses
def validate(args,
trainer,
task,
epoch_for_logging,
subsets,
valid_name=None,
num_updates=None,
global_epoch=None):
"""Evaluate the model on the validation set(s) and return the losses."""
task.split = 'valid'
valid_name_ = valid_name if valid_name is not None else 'valid'
num_updates = trainer.get_num_updates(
) if num_updates is None else num_updates
if args.fixed_validation_seed is not None:
# Set fixed seed for every validation
utils.set_torch_seed(args.fixed_validation_seed)
valid_losses = []
for subset in subsets:
# Initialize data iterator
itr = task.get_batch_iterator(
dataset=task.dataset(subset),
max_tokens=args.max_tokens_valid,
max_sentences=args.max_sentences_valid,
max_positions=utils.resolve_max_positions(
task.max_positions(),
trainer.get_model().max_positions(),
),
ignore_invalid_inputs=args.skip_invalid_size_inputs_valid_test,
required_batch_size_multiple=args.required_batch_size_multiple,
seed=args.seed,
num_shards=args.distributed_world_size,
shard_id=args.distributed_rank,
num_workers=args.num_workers,
epoch=1,
).next_epoch_itr(shuffle=False)
progress = progress_bar.build_progress_bar(
args,
itr,
epoch_for_logging,
prefix='valid on \'{}\' subset'.format(valid_name_),
no_progress_bar='simple')
progress = maybe_wrap_neptune_logging(progress, args)
# Add global epoch to beginning of progress bar description
if global_epoch is not None:
try:
progress.wrapped_bar.tqdm.set_description(
desc='epoch {:03d} | \'{}\' {}'.format(
global_epoch, valid_name_,
progress.wrapped_bar.prefix),
refresh=True)
except:
progress.tqdm.set_description(
desc='epoch {:03d} | \'{}\' {}'.format(
global_epoch, valid_name_, progress.tqdm.desc),
refresh=True)
# Reset validation meters
metrics.reset_meters(valid_name_)
with metrics.aggregate(valid_name) as agg:
for sample in progress:
trainer.valid_step(sample)
# Log validation stats
stats = get_valid_stats(args, trainer, agg.get_smoothed_values())
progress.print(stats, tag=valid_name_, step=num_updates)
valid_losses.append(stats[args.best_checkpoint_metric])
return valid_losses