def score(self, lines):
batch = self.make_batches(lines)
sample_score_dict = {}
# with progress_bar.build_progress_bar(self.args, itr) as t:
for sample in batch:
sample_id_lst = sample['id']
sample = utils.move_to_cuda(sample) if self.use_cuda else sample
if 'net_input' not in sample:
continue
hypos = self.scorer.generate(self.models, sample)
# print(hypos)
for sample_id, hypos_i in zip(sample_id_lst, hypos):
hypo = hypos_i[0]
pos_scores = hypo['positional_scores']
inf_scores = pos_scores.eq(float('inf')) | pos_scores.eq(
float('-inf'))
if inf_scores.any():
print(
'| Skipping tokens with inf scores:',
self.task.target_dictionary.string(
hypo['tokens'][inf_scores.nonzero()]))
pos_scores = pos_scores[(~inf_scores).nonzero()]
sample_score = pos_scores.sum().cpu()
count = pos_scores.numel()
w_lst = []
word_prob = []
for i in range(len(hypo['tokens'])):
w_ind = hypo['tokens'][i].item()
w = self.task.dictionary[w_ind]
word_prob.append((w, pos_scores[i].item()))
w_lst.append(w)
sample_score = -sample_score / count
if not self.args.quiet:
if self.args.output_sent:
print('H-{}\t{}\t{}'.format(sample_id, sample_score,
' '.join(w_lst)))
else:
print('H-{}\t{}'.format(sample_id, sample_score))
sample_score_dict[sample_id.item()] = sample_score.item()
# print(sample_id, sample_score.item())
return sample_score_dict
def backtranslate_samples(samples, collate_fn, generate_fn, cuda=True):
"""Backtranslate a list of samples.
Given an input (*samples*) of the form:
[{'id': 1, 'source': 'hallo welt'}]
this will return:
[{'id': 1, 'source': 'hello world', 'target': 'hallo welt'}]
Args:
samples (List[dict]): samples to backtranslate. Individual samples are
expected to have a 'source' key, which will become the 'target'
after backtranslation.
collate_fn (callable): function to collate samples into a mini-batch
generate_fn (callable): function to generate backtranslations
cuda (bool): use GPU for generation (default: ``True``)
Returns:
List[dict]: an updated list of samples with a backtranslated source
"""
collated_samples = collate_fn(samples)
s = utils.move_to_cuda(collated_samples) if cuda else collated_samples
generated_sources = generate_fn(s['net_input'])
def update_sample(sample, generated_source):
sample['target'] = sample[
'source'] # the original source becomes the target
sample['source'] = generated_source
return sample
# Go through each tgt sentence in batch and its corresponding best
# generated hypothesis and create a backtranslation data pair
# {id: id, source: generated backtranslation, target: original tgt}
return [
update_sample(
sample=input_sample,
generated_source=hypos[0]
['tokens'].cpu(), # highest scoring hypo is first
) for input_sample, hypos in zip(samples, generated_sources)
]
def main(parsed_args):
assert parsed_args.path is not None, '--path required for evaluation!'
import_user_module(parsed_args)
print(parsed_args)
use_cuda = torch.cuda.is_available() and not parsed_args.cpu
task = tasks.setup_task(parsed_args)
# Load ensemble
print('| loading model(s) from {}'.format(parsed_args.path))
models, args = utils.load_ensemble_for_inference(
parsed_args.path.split(':'), task, model_arg_overrides=eval(parsed_args.model_overrides),
)
for arg in vars(parsed_args).keys():
if arg not in {'self_target', 'future_target', 'past_target', 'tokens_per_sample', 'output_size_dictionary'}:
setattr(args, arg, getattr(parsed_args, arg))
task = tasks.setup_task(args)
# Load dataset splits
task.load_dataset(args.gen_subset)
print('| {} {} {} examples'.format(args.data, args.gen_subset, len(task.dataset(args.gen_subset))))
# Optimize ensemble for generation and set the source and dest dicts on the model (required by scorer)
for model in models:
model.make_generation_fast_()
if args.fp16:
model.half()
if use_cuda:
model.cuda()
assert len(models) > 0
print('num. model params: {}'.format(sum(p.numel() for p in models[0].parameters())))
itr = task.get_batch_iterator(
dataset=task.dataset(args.gen_subset),
max_tokens=args.max_tokens or 36000,
max_sentences=args.max_sentences,
max_positions=utils.resolve_max_positions(*[
model.max_positions() for model in models
]),
ignore_invalid_inputs=True,
num_shards=args.num_shards,
shard_id=args.shard_id,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
gen_timer = StopwatchMeter()
scorer = SequenceScorer(task.target_dictionary)
score_sum = 0.
count = 0
if args.remove_bpe is not None:
if args.remove_bpe == 'sentencepiece':
raise NotImplementedError
else:
bpe_cont = args.remove_bpe.rstrip()
bpe_toks = set(i for i in range(len(task.dictionary)) if task.dictionary[i].endswith(bpe_cont))
bpe_len = len(bpe_cont)
else:
bpe_toks = None
bpe_len = 0
word_stats = dict()
with progress_bar.build_progress_bar(args, itr) as t:
wps_meter = TimeMeter()
for sample in t:
sample = utils.move_to_cuda(sample) if use_cuda else sample
if 'net_input' not in sample:
continue
gen_timer.start()
hypos = scorer.generate(models, sample)
gen_timer.stop(sample['ntokens'])
for hypos_i in hypos:
hypo = hypos_i[0]
pos_scores = hypo['positional_scores']
skipped_toks = 0
if bpe_toks is not None:
for i in range(len(hypo['tokens']) - 1):
if hypo['tokens'][i].item() in bpe_toks:
skipped_toks += 1
pos_scores[i + 1] += pos_scores[i]
pos_scores[i] = 0
inf_scores = pos_scores.eq(float('inf')) | pos_scores.eq(float('-inf'))
if inf_scores.any():
print('| Skipping tokens with inf scores:',
task.target_dictionary.string(hypo['tokens'][inf_scores.nonzero()]))
pos_scores = pos_scores[(~inf_scores).nonzero()]
score_sum += pos_scores.sum().cpu().item() # custom fix to work with fp16
count += pos_scores.numel() - skipped_toks
if args.output_word_probs or args.output_word_stats:
w = ''
word_prob = []
is_bpe = False
for i in range(len(hypo['tokens'])):
w_ind = hypo['tokens'][i].item()
w += task.dictionary[w_ind]
if bpe_toks is not None and w_ind in bpe_toks:
w = w[:-bpe_len]
is_bpe = True
else:
word_prob.append((w, pos_scores[i].item()))
next_prob = None
ind = i + 1
while ind < len(hypo['tokens']):
if pos_scores[ind].item() != 0:
next_prob = pos_scores[ind]
break
ind += 1
word_stats.setdefault(w, WordStat(w, is_bpe)).add(pos_scores[i].item(), next_prob)
is_bpe = False
w = ''
if args.output_word_probs:
print('\t'.join('{} [{:2f}]'.format(x[0], x[1]) for x in word_prob))
wps_meter.update(sample['ntokens'])
t.log({'wps': round(wps_meter.avg)})
avg_nll_loss = -score_sum / count
print('| Evaluated {} tokens in {:.1f}s ({:.2f} tokens/s)'.format(gen_timer.n, gen_timer.sum, 1. / gen_timer.avg))
print('| Loss: {:.4f}, Perplexity: {:.2f}'.format(avg_nll_loss, np.exp(avg_nll_loss)))
if args.output_word_stats:
for ws in sorted(word_stats.values(), key=lambda x: x.count, reverse=True):
print(ws)
def main(args):
assert args.path is not None, '--path required for generation!'
assert not args.sampling or args.nbest == args.beam, \
'--sampling requires --nbest to be equal to --beam'
assert args.replace_unk is None or args.raw_text, \
'--replace-unk requires a raw text dataset (--raw-text)'
import_user_module(args)
"""
MODIFIED: The GEC task uses token-labeled raw text datasets, which
require raw text to be used.
"""
assert args.raw_text, \
f"--raw-text option is required for copy-based generation."
if args.max_tokens is None and args.max_sentences is None:
args.max_tokens = 12000
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load dataset splits
task = tasks.setup_task(args)
task.load_dataset(args.gen_subset)
print('| {} {} {} examples'.format(args.data, args.gen_subset,
len(task.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
print('| loading model(s) from {}'.format(args.path))
models, _model_args = utils.load_ensemble_for_inference(
args.path.split(':'),
task,
model_arg_overrides=eval(args.model_overrides),
)
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
need_attn=args.print_alignment,
)
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=8,
num_shards=args.num_shards,
shard_id=args.shard_id,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
# Initialize generator
gen_timer = StopwatchMeter()
generator = task.build_generator(args)
# Generate and compute BLEU score
if args.sacrebleu:
scorer = bleu.SacrebleuScorer()
else:
scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk())
num_sentences = 0
has_target = True
has_copy_scores = True
with progress_bar.build_progress_bar(args, itr) as t:
wps_meter = TimeMeter()
for sample in t:
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()
"""
MODIFIED: Use copy scores to replace <unk>'s with raw source words.
use_copy_scores may be False with non-copy-based transformers that
only use edit labels (e.g., transformer_aux_el and transformer_el).
"""
hypos = task.inference_step(generator, models, sample,
prefix_tokens)
use_copy_scores = hypos[0][0].get('copy_scores', None) is not None
if has_copy_scores and not use_copy_scores:
print("| generate_or_copy.py | INFO | "
"Model does not include copy scores. "
"Generating hypotheses without replacing UNKs.")
has_copy_scores = False
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
src_tokens = utils.strip_pad(
sample['net_input']['src_tokens'][i, :], tgt_dict.pad())
target_tokens = None
if has_target:
target_tokens = utils.strip_pad(
sample['target'][i, :], tgt_dict.pad()).int().cpu()
"""
MODIFIED: Replace <unk>s with raw source tokens.
This is analogous to the case where align_dict is provided
in the original generate.py.
"""
rawtext_dataset = task.dataset(args.gen_subset)
src_str = rawtext_dataset.src.get_original_text(sample_id)
tokenized_src_str = rawtext_dataset.src_dict.string(
src_tokens, bpe_symbol=args.remove_bpe)
target_str = rawtext_dataset.tgt.get_original_text(sample_id)
if not args.quiet:
if src_dict is not None:
# Raw source text
print('S-{}\t{}'.format(sample_id, src_str))
# Tokenized source text
print('K-{}\t{}'.format(sample_id, tokenized_src_str))
if has_target:
print('T-{}\t{}'.format(sample_id, target_str))
# Process top predictions
for k, hypo in enumerate(
hypos[i][: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'].int().cpu()
if hypo['alignment'] is not None else None,
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
)
"""
MODIFIED: Replace predicted <unk>s with the source token
that received the highest score.
"""
raw_src_tokens = src_str.split()
final_hypo_tokens_str = []
for tgt_position, hypo_token in enumerate(hypo_tokens):
if use_copy_scores and hypo_token == tgt_dict.unk():
# See sequence_copygenerator.py#L292 for details.
copy_scores = hypo[
'copy_scores'][:, tgt_position].cpu()
assert len(copy_scores) - 1 == len(raw_src_tokens), \
f"length of copy scores do not match input source tokens " \
f"(copy_scores: {copy_scores}, raw_src_tokens: {raw_src_tokens})"
src_position = torch.argmax(copy_scores).item()
# Don't copy if attending to an EOS (not ideal).
if src_position == len(raw_src_tokens):
print("WARNING: copy score highest at EOS.")
else:
final_hypo_tokens_str.append(
raw_src_tokens[src_position])
print('U-{}\t{}\t{}'.format(
sample_id,
tgt_position,
' '.join(
map(
lambda x: '{:.4f}'.format(x),
copy_scores.tolist(),
)),
))
else:
final_hypo_tokens_str.append(tgt_dict[hypo_token])
# Note: raw input tokens could be included here.
final_hypo_str = ' '.join([
token for token in final_hypo_tokens_str
if token != tgt_dict.eos_word
])
if not args.quiet:
print('H-{}\t{}\t{}'.format(sample_id, hypo['score'],
final_hypo_str))
print('P-{}\t{}'.format(
sample_id, ' '.join(
map(
lambda x: '{:.4f}'.format(x),
hypo['positional_scores'].tolist(),
))))
if args.print_alignment:
print('A-{}\t{}'.format(
sample_id, ' '.join(
map(lambda x: str(utils.item(x)),
alignment))))
# Score only the top hypothesis
if has_target and k == 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)
if hasattr(scorer, 'add_string'):
scorer.add_string(target_str, hypo_str)
else:
scorer.add(target_tokens, hypo_tokens)
wps_meter.update(num_generated_tokens)
t.log({'wps': round(wps_meter.avg)})
num_sentences += sample['nsentences']
print(
'| 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:
print('| Generate {} with beam={}: {}'.format(args.gen_subset,
args.beam,
scorer.result_string()))
return scorer
def _prepare_sample(self, sample):
if sample is None or len(sample) == 0:
return None
if self.cuda:
sample = utils.move_to_cuda(sample)
return sample
def main(args):
assert args.path is not None, '--path required for generation!'
assert not args.sampling or args.nbest == args.beam, \
'--sampling requires --nbest to be equal to --beam'
assert args.replace_unk is None or args.raw_text, \
'--replace-unk requires a raw text dataset (--raw-text)'
import_user_module(args)
if args.max_tokens is None and args.max_sentences is None:
args.max_tokens = 12000
print(args)
use_cuda = torch.cuda.is_available() and not args.cpu
# Load dataset splits
task = tasks.setup_task(args)
task.load_dataset(args.gen_subset)
print('| {} {} {} examples'.format(args.data, args.gen_subset,
len(task.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
print('| loading model(s) from {}'.format(args.path))
models, _model_args = utils.load_ensemble_for_inference(
args.path.split(':'),
task,
model_arg_overrides=eval(args.model_overrides),
)
# Optimize ensemble for generation
for model in models:
model.make_generation_fast_(
beamable_mm_beam_size=None if args.no_beamable_mm else args.beam,
need_attn=args.print_alignment,
)
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=8,
num_shards=args.num_shards,
shard_id=args.shard_id,
num_workers=args.num_workers,
).next_epoch_itr(shuffle=False)
# Initialize generator
gen_timer = StopwatchMeter()
generator = task.build_generator(args)
# Generate and compute BLEU score
if args.sacrebleu:
scorer = bleu.SacrebleuScorer()
else:
scorer = bleu.Scorer(tgt_dict.pad(), tgt_dict.eos(), tgt_dict.unk())
num_sentences = 0
has_target = True
with progress_bar.build_progress_bar(args, itr) as t:
wps_meter = TimeMeter()
for sample in t:
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
src_tokens = utils.strip_pad(
sample['net_input']['src_tokens'][i, :], tgt_dict.pad())
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)
if not args.quiet:
if src_dict is not None:
print('S-{}\t{}'.format(sample_id, src_str))
if has_target:
print('T-{}\t{}'.format(sample_id, target_str))
# Process top predictions
for i, hypo in enumerate(
hypos[i][: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'].int().cpu()
if hypo['alignment'] is not None else None,
align_dict=align_dict,
tgt_dict=tgt_dict,
remove_bpe=args.remove_bpe,
)
if not args.quiet:
print('H-{}\t{}\t{}'.format(sample_id, hypo['score'],
hypo_str))
print('P-{}\t{}'.format(
sample_id, ' '.join(
map(
lambda x: '{:.4f}'.format(x),
hypo['positional_scores'].tolist(),
))))
if args.print_alignment:
print('A-{}\t{}'.format(
sample_id, ' '.join(
map(lambda x: str(utils.item(x)),
alignment))))
# Score only the top hypothesis
if has_target and i == 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)
if hasattr(scorer, 'add_string'):
scorer.add_string(target_str, hypo_str)
else:
scorer.add(target_tokens, hypo_tokens)
wps_meter.update(num_generated_tokens)
t.log({'wps': round(wps_meter.avg)})
num_sentences += sample['nsentences']
print(
'| 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:
print('| Generate {} with beam={}: {}'.format(args.gen_subset,
args.beam,
scorer.result_string()))
return scorer