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)
valid_losses = []
for subset in subsets:
logger.info('begin validation on "{}" subset'.format(subset))
# Initialize data iterator
itr = trainer.get_valid_iterator(subset).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)
with metrics.aggregate(new_root=True) 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=subset, step=trainer.get_num_updates())
valid_losses.append(stats[args.best_checkpoint_metric])
return valid_losses
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('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]
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)
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 _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.args.seed + self.get_num_updates()
utils.set_torch_seed(seed)
def main(args):
start_time = time.time()
total_translate_time = 0
utils.import_user_module(args)
if args.buffer_size < 1:
args.buffer_size = 1
if args.max_tokens is None and args.max_sentences is None:
args.max_sentences = 1
assert not args.sampling or args.nbest == args.beam, \
'--sampling requires --nbest to be equal to --beam'
assert not args.max_sentences or args.max_sentences <= args.buffer_size, \
'--max-sentences/--batch-size cannot be larger than --buffer-size'
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
# Setup task, e.g., translation
task = tasks.setup_task(args)
# Load ensemble
logger.info('loading model(s) from {}'.format(args.path))
models, _model_args = checkpoint_utils.load_model_ensemble(
args.path.split(os.pathsep),
arg_overrides=eval(args.model_overrides),
task=task,
suffix=getattr(args, "checkpoint_suffix", ""),
)
# Set dictionaries
src_dict = task.source_dictionary
tgt_dict = task.target_dictionary
# Optimize ensemble for generation
for model in models:
model.prepare_for_inference_(args)
if args.fp16:
model.half()
if use_cuda:
model.cuda()
# Initialize generator
generator = task.build_generator(models, args)
# Handle tokenization and BPE
tokenizer = encoders.build_tokenizer(args)
bpe = encoders.build_bpe(args)
def encode_fn(x):
if tokenizer is not None:
x = tokenizer.encode(x)
if bpe is not None:
x = bpe.encode(x)
return x
def decode_fn(x):
if bpe is not None:
x = bpe.decode(x)
if tokenizer is not None:
x = tokenizer.decode(x)
return x
# 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)
max_positions = utils.resolve_max_positions(
task.max_positions(), *[model.max_positions() for model in models])
if args.constraints:
logger.warning(
"NOTE: Constrained decoding currently assumes a shared subword vocabulary."
)
if args.buffer_size > 1:
logger.info('Sentence buffer size: %s', args.buffer_size)
logger.info('NOTE: hypothesis and token scores are output in base 2')
logger.info('Type the input sentence and press return:')
start_id = 0
for inputs in buffered_read(args.input, args.buffer_size):
results = []
for batch in make_batches(inputs, args, task, max_positions,
encode_fn):
bsz = batch.src_tokens.size(0)
src_tokens = batch.src_tokens
src_lengths = batch.src_lengths
constraints = batch.constraints
if use_cuda:
src_tokens = src_tokens.cuda()
src_lengths = src_lengths.cuda()
if constraints is not None:
constraints = constraints.cuda()
sample = {
'net_input': {
'src_tokens': src_tokens,
'src_lengths': src_lengths,
},
}
translate_start_time = time.time()
translations = task.inference_step(generator,
models,
sample,
constraints=constraints)
translate_time = time.time() - translate_start_time
total_translate_time += translate_time
list_constraints = [[] for _ in range(bsz)]
if args.constraints:
list_constraints = [unpack_constraints(c) for c in constraints]
for i, (id,
hypos) in enumerate(zip(batch.ids.tolist(), translations)):
src_tokens_i = utils.strip_pad(src_tokens[i], tgt_dict.pad())
constraints = list_constraints[i]
results.append((start_id + id, src_tokens_i, hypos, {
"constraints": constraints,
"time": translate_time / len(translations)
}))
# sort output to match input order
for id_, src_tokens, hypos, info in sorted(results,
key=lambda x: x[0]):
if src_dict is not None:
src_str = src_dict.string(src_tokens, args.remove_bpe)
print('S-{}\t{}'.format(id_, src_str))
print("W-{}\t{:.3f}\tseconds".format(id_, info["time"]))
for constraint in info["constraints"]:
print("C-{}\t{}".format(
id_, tgt_dict.string(constraint, args.remove_bpe)))
# Process top predictions
for hypo in hypos[:min(len(hypos), 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)
score = hypo['score'] / math.log(2) # convert to base 2
# original hypothesis (after tokenization and BPE)
print('H-{}\t{}\t{}'.format(id_, score, hypo_str))
# detokenized hypothesis
print('D-{}\t{}\t{}'.format(id_, score, detok_hypo_str))
print('P-{}\t{}'.format(
id_,
' '.join(
map(
lambda x: '{:.4f}'.format(x),
# convert from base e to base 2
hypo['positional_scores'].div_(math.log(2)
).tolist(),
))))
if args.print_alignment:
alignment_str = " ".join(
["{}-{}".format(src, tgt) for src, tgt in alignment])
print('A-{}\t{}'.format(id_, alignment_str))
# update running id_ counter
start_id += len(inputs)
logger.info("Total time: {:.3f} seconds; translation time: {:.3f}".format(
time.time() - start_time, total_translate_time))
def main(args):
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()
np.random.seed(args.seed)
utils.set_torch_seed(args.seed)
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("task: {} ({})".format(args.task, task.__class__.__name__))
logger.info("model: {} ({})".format(args.arch, model.__class__.__name__))
logger.info(
"criterion: {} ({})".format(args.criterion, 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:
raise NotImplementedError('here')
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)
# Train until the learning rate gets too small
max_epoch = args.max_epoch 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, should_stop = train(args, 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"),
)
train_meter.stop()
logger.info("done training in {:.1f} seconds".format(train_meter.sum))