def __init__(
self,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.,
unk_penalty=0.,
retain_dropout=False,
temperature=1.,
match_source_len=False,
no_repeat_ngram_size=0,
search_strategy=None,
):
"""Generates translations of a given source sentence.
Args:
tgt_dict (~fairseq.data.Dictionary): target dictionary
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
assert temperature > 0, '--temperature must be greater than 0'
self.search = (search.BeamSearch(tgt_dict)
if search_strategy is None else search_strategy)
def __init__(self, task, sentence_avg, label_smoothing, lambda_ratio, rl_k,
max_decoding_len, rl_len_penalty, gpt2_encoder_json,
gpt2_vocab_bpe, mtdnn_max_len, mtdnn_path):
super().__init__(task)
assert hasattr(task, 'target_dictionary'
), "only support task with attr target_dictionary"
self.tgt_dict = task.target_dictionary
self.src_dict = task.source_dictionary
self.sentence_avg = sentence_avg
self.eps = label_smoothing
self.rl_k = rl_k
self.max_decoding_len = max_decoding_len
self.lambda_ratio = lambda_ratio
self.len_penalty = rl_len_penalty
self.eos = self.tgt_dict.eos()
self.pad = self.tgt_dict.pad()
self.unk = self.tgt_dict.unk()
self.vocab_size = len(self.tgt_dict)
kwargs = {
"gpt2_encoder_json": gpt2_encoder_json,
"gpt2_vocab_bpe": gpt2_vocab_bpe
}
self.bpe = GPT2BPE(kwargs)
print("bpe:", self.bpe)
self.greedy_searcher = search.BeamSearch(self.tgt_dict)
self.sampling_searcher = search.Sampling(self.tgt_dict)
self.evaluator_tokenizer = EvaluatorTokenizer(mtdnn_max_len)
self.evaluator_model = EvaluatorModel(mtdnn_path)
def _generate_hypotheses(self, model, sample):
# initialize generator
if self.args.seq_sampling:
search_strategy = search.Sampling(self.target_dictionary)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
if self._generator is None:
self._generator = SequenceGenerator(
tgt_dict=self.target_dictionary,
models=[model],
beam_size=self.args.seq_beam,
max_len_a=self.args.seq_max_len_a,
max_len_b=self.args.seq_max_len_b,
unk_penalty=self.args.seq_unkpen,
search_strategy=search_strategy)
# generate hypotheses
sample['hypos'] = self._generator.generate(
[model],
sample,
)
# add reference to the set of hypotheses
if self.args.seq_keep_reference:
self.add_reference_to_hypotheses(sample)
def build_generator(self, args):
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, 'sampling', False)
sampling_topk = getattr(args, 'sampling_topk', -1)
sampling_topp = getattr(args, 'sampling_topp', -1.0)
diverse_beam_groups = getattr(args, 'diverse_beam_groups', -1)
diverse_beam_strength = getattr(args, 'diverse_beam_strength', 0.5),
match_source_len = getattr(args, 'match_source_len', False)
diversity_rate = getattr(args, 'diversity_rate', -1)
if (
sum(
int(cond)
for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]
)
> 1
):
raise ValueError('Provided Search parameters are mutually exclusive.')
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
assert sampling_topp < 0 or sampling, '--sampling-topp requires --sampling'
if sampling:
search_strategy = search.Sampling(self.target_dictionary, sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(
self.target_dictionary, diverse_beam_groups, diverse_beam_strength)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary, min_len_a=1, min_len_b=0, max_len_a=1, max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(self.target_dictionary, diversity_rate)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
if args.context_type == 'src':
seq_cls = AudioContextAwareSequenceGenerator
else:
seq_cls = TargetContextAwareSequenceGenerator
return seq_cls(
self.target_dictionary,
beam_size=getattr(args, 'beam', 5),
max_len_a=getattr(args, 'max_len_a', 0),
max_len_b=getattr(args, 'max_len_b', 200),
min_len=getattr(args, 'min_len', 1),
normalize_scores=(not getattr(args, 'unnormalized', False)),
len_penalty=getattr(args, 'lenpen', 1),
unk_penalty=getattr(args, 'unkpen', 0),
temperature=getattr(args, 'temperature', 1.),
match_source_len=getattr(args, 'match_source_len', False),
no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0),
search_strategy=search_strategy,
)
def __init__(
self,
models,
tgt_dict,
beam_size=1,
minlen=1,
maxlen=None,
stop_early=True,
normalize_scores=True,
len_penalty=1,
unk_penalty=0,
retain_dropout=False,
sampling=False,
sampling_topk=-1,
sampling_temperature=1,
diverse_beam_groups=-1,
diverse_beam_strength=0.5,
desired_length=-1,
):
"""Generates translations of a given source sentence.
Args:
min/maxlen: The length of the generated output will be bounded by
minlen and maxlen (not including the end-of-sentence marker).
stop_early: Stop generation immediately after we finalize beam_size
hypotheses, even though longer hypotheses might have better
normalized scores.
normalize_scores: Normalize scores by the length of the output.
"""
self.models = models
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
self.desired_length = desired_length
self.minlen = minlen
max_decoder_len = min(m.max_decoder_positions() for m in self.models)
max_decoder_len -= 1 # we define maxlen not including the EOS marker
self.maxlen = max_decoder_len if maxlen is None else min(
maxlen, max_decoder_len)
self.stop_early = stop_early
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk,
sampling_temperature)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict,
diverse_beam_groups,
diverse_beam_strength)
else:
self.search = search.BeamSearch(tgt_dict)
def __init__(self,
tgt_dict,
beam_size=1,
tgt_len=200,
retain_dropout=False,
sampling=False,
sampling_topk=-1,
sampling_topp=-1.0,
temperature=1.,
diverse_beam_groups=-1,
diverse_beam_strength=0.5):
"""Generates translations of a given source sentence.
Args:
tgt_dict (~fairseq.data.Dictionary): target dictionary
beam_size (int, optional): beam width (default: 1)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
retain_dropout (bool, optional): use dropout when generating
(default: False)
sampling (bool, optional): sample outputs instead of beam search
(default: False)
sampling_topk (int, optional): only sample among the top-k choices
at each step (default: -1)
sampling_topp (float, optional): only sample among the smallest set
of words whose cumulative probability mass exceeds p
at each step (default: -1.0)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
diverse_beam_groups/strength (float, optional): parameters for
Diverse Beam Search sampling
"""
self.pad = tgt_dict.pad()
self.bos = tgt_dict.bos()
self.vocab_size = len(tgt_dict)
assert beam_size == 1, "Only support beam_size=1 currently"
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.tgt_len = tgt_len
self.retain_dropout = retain_dropout
self.temperature = temperature
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
assert sampling_topp < 0 or sampling, '--sampling-topp requires --sampling'
assert temperature > 0, '--temperature must be greater than 0'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk,
sampling_topp)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict,
diverse_beam_groups,
diverse_beam_strength)
else:
self.search = search.BeamSearch(tgt_dict)
def __init__(
self,
models,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.0,
unk_penalty=0.0,
retain_dropout=False,
temperature=1.0,
match_source_len=False,
no_repeat_ngram_size=0,
search_strategy=None,
eos=None,
waitk=1,
without_caching=True,
):
super().__init__()
if isinstance(models, EnsembleModel):
self.model = models
else:
self.model = EnsembleModel(models)
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos() if eos is None else eos
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
assert temperature > 0, "--temperature must be greater than 0"
self.search = (search.BeamSearch(tgt_dict)
if search_strategy is None else search_strategy)
if not self.retain_dropout:
self.model.eval()
self.waitk = waitk
self.without_caching = without_caching
def __init__(self, models, task, beam_size, max_lens):
self.args = task.args
self.models = models
self.tokenizer = task.tokenizer
self.eos = self.tokenizer.eos()
self.start_idx = None
self.minlen = 1
self.pad = self.tokenizer.pad()
self.vocab_size = len(self.tokenizer)
self.normalize_scores = True
self.len_penalty = 1
self.beam_size = beam_size
self.max_lens = max_lens
self.search = search.BeamSearch(self.tokenizer)
def test_prefix_beam_search(self):
search_strategy = search.BeamSearch(self.tgt_dict)
generator = SequenceGenerator(
[self.model],
self.tgt_dict,
beam_size=self.beam_size,
search_strategy=search_strategy,
)
sample = {
"net_input": {
"src_tokens": self.tokens,
"src_lengths": self.token_lengths,
}
}
# make sure test sample doesn't break any assertion
generator.forward(sample, prefix_tokens=self.tokens[:, :-1])
def _build_generator(self, src_lang, tgt_lang, args):
return SequenceGenerator(
ModuleList([
model.models[f"{src_lang}-{tgt_lang}"] for model in self.models
]),
self.dicts[tgt_lang],
beam_size=getattr(args, "beam", 5),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.0),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search.BeamSearch(self.dicts[tgt_lang]),
)
def __init__(
self, models, src_dict, tgt_dict, beam_size=1, minlen=1, maxlen=None, stop_early=True,
normalize_scores=True, len_penalty=1., unk_penalty=0., retain_dropout=False,
sampling=False, sampling_topk=-1, sampling_temperature=1.,
diverse_beam_groups=-1, diverse_beam_strength=0.5, args=None,
):
self.models = models
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
self.minlen = minlen
max_decoder_len = min(m.max_decoder_positions() for m in self.models)
max_decoder_len -= 1 # we define maxlen not including the EOS marker
self.maxlen = max_decoder_len if maxlen is None else min(maxlen, max_decoder_len)
self.stop_early = stop_early
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.tgt_dict = tgt_dict
self.args = args
punc = string.punctuation
self.tgt_punc = []
for w in range(len(tgt_dict)):
if tgt_dict[w] in punc:
self.tgt_punc += [w]
self.src_punc = [2,3,4] ## mask special symbols
for w in range(len(src_dict)):
if src_dict[w] in punc:
self.src_punc += [w]
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk, sampling_temperature)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict, diverse_beam_groups, diverse_beam_strength)
else:
self.search = search.BeamSearch(tgt_dict)
def build_generator(self, args):
search_strategy = search.BeamSearch(self.target_dictionary)
from ..constrained_decoding import ConstrainedSequenceGenerator
seq_gen_cls = ConstrainedSequenceGenerator
dicts = (self.source_dictionary, self.target_dictionary)
return seq_gen_cls(
*dicts,
order_constr=getattr(args, 'order_constr', False),
beam_size=getattr(args, 'beam', 5),
max_len_a=getattr(args, 'max_len_a', 0),
max_len_b=getattr(args, 'max_len_b', 200),
min_len=getattr(args, 'min_len', 1),
normalize_scores=(not getattr(args, 'unnormalized', False)),
len_penalty=getattr(args, 'lenpen', 1),
unk_penalty=getattr(args, 'unkpen', 0),
temperature=getattr(args, 'temperature', 1.),
match_source_len=getattr(args, 'match_source_len', False),
no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0),
search_strategy=search_strategy,
)
def __init__(
self,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
stop_early=True,
normalize_scores=True,
len_penalty=1.,
unk_penalty=0.,
retain_dropout=False,
sampling=False,
sampling_topk=-1,
sampling_temperature=1.,
diverse_beam_groups=-1,
diverse_beam_strength=0.5,
match_source_len=False,
no_repeat_ngram_size=0,
):
"""Generates translations of a given source sentence.
Args:
tgt_dict (~fairseq.data.Dictionary): target dictionary
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
stop_early (bool, optional): stop generation immediately after we
finalize beam_size hypotheses, even though longer hypotheses
might have better normalized scores (default: True)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
sampling (bool, optional): sample outputs instead of beam search
(default: False)
sampling_topk (int, optional): only sample among the top-k choices
at each step (default: -1)
sampling_temperature (float, optional): temperature for sampling,
where values >1.0 produces more uniform sampling and values
<1.0 produces sharper sampling (default: 1.0)
diverse_beam_groups/strength (float, optional): parameters for
Diverse Beam Search sampling
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.stop_early = stop_early
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk,
sampling_temperature)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
self.search = search.LengthConstrainedBeamSearch(
tgt_dict,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
else:
self.search = search.BeamSearch(tgt_dict)
def build_generator(self, models, args):
from examples.waitk.generators.waitk_sequence_generator import WaitkSequenceGenerator
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, "sampling", False)
sampling_topk = getattr(args, "sampling_topk", -1)
sampling_topp = getattr(args, "sampling_topp", -1.0)
diverse_beam_groups = getattr(args, "diverse_beam_groups", -1)
diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5)
match_source_len = getattr(args, "match_source_len", False)
diversity_rate = getattr(args, "diversity_rate", -1)
if (sum(
int(cond) for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]) > 1):
raise ValueError(
"Provided Search parameters are mutually exclusive.")
assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling"
assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling"
if sampling:
search_strategy = search.Sampling(self.target_dictionary,
sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(self.target_dictionary,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
return WaitkSequenceGenerator(
models,
self.target_dictionary,
beam_size=getattr(args, "beam", 5),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.0),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search_strategy,
waitk=args.eval_waitk)
def build_generator(
self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None
):
if getattr(args, "score_reference", False):
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(
self.target_dictionary,
compute_alignment=getattr(args, "print_alignment", False),
)
from fairseq.sequence_generator import (
SequenceGenerator,
SequenceGeneratorWithAlignment,
)
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, "sampling", False)
sampling_topk = getattr(args, "sampling_topk", -1)
sampling_topp = getattr(args, "sampling_topp", -1.0)
diverse_beam_groups = getattr(args, "diverse_beam_groups", -1)
diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5)
match_source_len = getattr(args, "match_source_len", False)
diversity_rate = getattr(args, "diversity_rate", -1)
constrained = getattr(args, "constraints", False)
prefix_allowed_tokens_fn = getattr(args, "prefix_allowed_tokens_fn", None)
if (
sum(
int(cond)
for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]
)
> 1
):
raise ValueError("Provided Search parameters are mutually exclusive.")
assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling"
assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling"
if sampling:
search_strategy = search.Sampling(
self.target_dictionary, sampling_topk, sampling_topp
)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(
self.target_dictionary, diverse_beam_groups, diverse_beam_strength
)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate
)
elif constrained:
search_strategy = search.LexicallyConstrainedBeamSearch(
self.target_dictionary, args.constraints
)
elif prefix_allowed_tokens_fn:
search_strategy = search.PrefixConstrainedBeamSearch(
self.target_dictionary, prefix_allowed_tokens_fn
)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
extra_gen_cls_kwargs = extra_gen_cls_kwargs or {}
if seq_gen_cls is None:
if getattr(args, "print_alignment", False):
seq_gen_cls = SequenceGeneratorWithAlignment
extra_gen_cls_kwargs['print_alignment'] = args.print_alignment
else:
seq_gen_cls = SequenceGenerator
return seq_gen_cls(
models,
self.target_dictionary,
beam_size=getattr(args, "beam", 5),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.0),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search_strategy,
**extra_gen_cls_kwargs,
)
def build_generator(self, args):
if args.score_reference:
args.score_reference = False
logger.warning(
'--score-reference is not applicable to speech recognition, ignoring it.'
)
from fairseq.sequence_generator import SequenceGenerator
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, 'sampling', False)
sampling_topk = getattr(args, 'sampling_topk', -1)
sampling_topp = getattr(args, 'sampling_topp', -1.0)
diverse_beam_groups = getattr(args, 'diverse_beam_groups', -1)
diverse_beam_strength = getattr(args, 'diverse_beam_strength', 0.5),
match_source_len = getattr(args, 'match_source_len', False)
diversity_rate = getattr(args, 'diversity_rate', -1)
if (sum(
int(cond) for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]) > 1):
raise ValueError(
'Provided Search parameters are mutually exclusive.')
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
assert sampling_topp < 0 or sampling, '--sampling-topp requires --sampling'
if sampling:
search_strategy = search.Sampling(self.target_dictionary,
sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(self.target_dictionary,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
return SequenceGenerator(
self.target_dictionary,
beam_size=getattr(args, 'beam', 5),
max_len_a=getattr(args, 'max_len_a', 0),
max_len_b=getattr(args, 'max_len_b', 200),
min_len=getattr(args, 'min_len', 1),
normalize_scores=(not getattr(args, 'unnormalized', False)),
len_penalty=getattr(args, 'lenpen', 1),
unk_penalty=getattr(args, 'unkpen', 0),
temperature=getattr(args, 'temperature', 1.),
match_source_len=getattr(args, 'match_source_len', False),
no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0),
search_strategy=search_strategy,
eos_factor=getattr(args, 'eos_factor', None),
)
def build_generator(
self, models, args, seq_gen_cls=None, extra_gen_cls_kwargs=None, prefix_allowed_tokens_fn=None,
):
"""
Build a :class:`~fairseq.SequenceGenerator` instance for this
task.
Args:
models (List[~fairseq.models.FairseqModel]): ensemble of models
args (fairseq.dataclass.configs.GenerationConfig):
configuration object (dataclass) for generation
extra_gen_cls_kwargs (Dict[str, Any]): extra options to pass
through to SequenceGenerator
prefix_allowed_tokens_fn (Callable[[int, torch.Tensor], List[int]]):
If provided, this function constrains the beam search to
allowed tokens only at each step. The provided function
should take 2 arguments: the batch ID (`batch_id: int`)
and a unidimensional tensor of token ids (`inputs_ids:
torch.Tensor`). It has to return a `List[int]` with the
allowed tokens for the next generation step conditioned
on the previously generated tokens (`inputs_ids`) and
the batch ID (`batch_id`). This argument is useful for
constrained generation conditioned on the prefix, as
described in "Autoregressive Entity Retrieval"
(https://arxiv.org/abs/2010.00904) and
https://github.com/facebookresearch/GENRE.
"""
if getattr(args, "score_reference", False):
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(
self.target_dictionary,
compute_alignment=getattr(args, "print_alignment", False),
)
from fairseq.sequence_generator import (
SequenceGenerator,
SequenceGeneratorWithAlignment,
)
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, "sampling", False)
sampling_topk = getattr(args, "sampling_topk", -1)
sampling_topp = getattr(args, "sampling_topp", -1.0)
diverse_beam_groups = getattr(args, "diverse_beam_groups", -1)
diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5)
match_source_len = getattr(args, "match_source_len", False)
diversity_rate = getattr(args, "diversity_rate", -1)
constrained = getattr(args, "constraints", False)
if prefix_allowed_tokens_fn is None:
prefix_allowed_tokens_fn = getattr(args, "prefix_allowed_tokens_fn", None)
if (
sum(
int(cond)
for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]
)
> 1
):
raise ValueError("Provided Search parameters are mutually exclusive.")
assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling"
assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling"
if sampling:
search_strategy = search.Sampling(
self.target_dictionary, sampling_topk, sampling_topp
)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(
self.target_dictionary, diverse_beam_groups, diverse_beam_strength
)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate
)
elif constrained:
search_strategy = search.LexicallyConstrainedBeamSearch(
self.target_dictionary, args.constraints
)
elif prefix_allowed_tokens_fn:
search_strategy = search.PrefixConstrainedBeamSearch(
self.target_dictionary, prefix_allowed_tokens_fn
)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
extra_gen_cls_kwargs = extra_gen_cls_kwargs or {}
if seq_gen_cls is None:
if getattr(args, "print_alignment", False):
seq_gen_cls = SequenceGeneratorWithAlignment
extra_gen_cls_kwargs["print_alignment"] = args.print_alignment
else:
seq_gen_cls = SequenceGenerator
return seq_gen_cls(
models,
self.target_dictionary,
beam_size=getattr(args, "beam", 5),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.0),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search_strategy,
**extra_gen_cls_kwargs,
)
def __init__(
self,
models,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.0,
unk_penalty=0.0,
retain_dropout=False,
temperature=1.0,
match_source_len=False,
no_repeat_ngram_size=0,
search_strategy=None,
eos=None,
exclude_prefix_max_len=None,
):
"""Generates translations of a given source sentence.
Args:
models (List[~fairseq.models.FairseqModel]): ensemble of models,
currently support fairseq.models.TransformerModel for scripting
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
super().__init__()
if isinstance(models, EnsembleModel):
self.model = models
else:
self.model = EnsembleModel(models)
self.tgt_dict = tgt_dict
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos() if eos is None else eos
self.bos = tgt_dict.bos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.exclude_prefix_max_len = exclude_prefix_max_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
assert temperature > 0, "--temperature must be greater than 0"
self.search = (search.BeamSearch(tgt_dict)
if search_strategy is None else search_strategy)
if not self.retain_dropout:
self.model.eval()
def build_generator(self,
models,
args,
seq_gen_cls=None,
extra_gen_cls_kwargs=None):
if getattr(args, "score_reference", False):
args.score_reference = False
logger.warning(
"--score-reference is not applicable to speech recognition, ignoring it."
)
from fairseq.sequence_generator import SequenceGenerator
# Choose search strategy. Defaults to Beam Search.
sampling = getattr(args, "sampling", False)
sampling_topk = getattr(args, "sampling_topk", -1)
sampling_topp = getattr(args, "sampling_topp", -1.0)
diverse_beam_groups = getattr(args, "diverse_beam_groups", -1)
diverse_beam_strength = getattr(args, "diverse_beam_strength", 0.5)
match_source_len = getattr(args, "match_source_len", False)
diversity_rate = getattr(args, "diversity_rate", -1)
if (sum(
int(cond) for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]) > 1):
raise ValueError(
"Provided Search parameters are mutually exclusive.")
assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling"
assert sampling_topp < 0 or sampling, "--sampling-topp requires --sampling"
if sampling:
search_strategy = search.Sampling(self.target_dictionary,
sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(self.target_dictionary,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
if seq_gen_cls is None:
seq_gen_cls = SequenceGenerator
extra_gen_cls_kwargs = extra_gen_cls_kwargs or {}
extra_gen_cls_kwargs["lm_weight"] = getattr(args, "lm_weight", 0.0)
extra_gen_cls_kwargs["eos_factor"] = getattr(args, "eos_factor", None)
return seq_gen_cls(
models,
self.target_dictionary,
beam_size=getattr(args, "beam", 5),
max_len_a=getattr(args, "max_len_a", 0),
max_len_b=getattr(args, "max_len_b", 200),
min_len=getattr(args, "min_len", 1),
normalize_scores=(not getattr(args, "unnormalized", False)),
len_penalty=getattr(args, "lenpen", 1),
unk_penalty=getattr(args, "unkpen", 0),
temperature=getattr(args, "temperature", 1.),
match_source_len=getattr(args, "match_source_len", False),
no_repeat_ngram_size=getattr(args, "no_repeat_ngram_size", 0),
search_strategy=search_strategy,
**extra_gen_cls_kwargs,
)
def __init__(
self,
models,
src_dict,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.0,
unk_penalty=0.0,
retain_dropout=False,
temperature=1.0,
match_source_len=False,
no_repeat_ngram_size=0,
search_strategy=None,
eos=None,
):
"""Generates transcripts and translations of a given source audio.
Args:
models (List[~fairseq.models.FairseqModel]): ensemble of models,
currently support fairseq.models.TransformerModel for scripting
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
super().__init__(models,
tgt_dict,
beam_size=beam_size,
max_len_a=max_len_a,
max_len_b=max_len_b,
min_len=min_len,
normalize_scores=normalize_scores,
len_penalty=len_penalty,
unk_penalty=unk_penalty,
retain_dropout=retain_dropout,
temperature=temperature,
match_source_len=match_source_len,
no_repeat_ngram_size=no_repeat_ngram_size,
search_strategy=search_strategy,
eos=eos)
if isinstance(models, EnsembleTwoPhaseModel):
self.model = models
else:
self.model = EnsembleTwoPhaseModel(models)
self.src_pad = src_dict.pad()
self.src_unk = src_dict.unk()
self.src_eos = src_dict.eos() if eos is None else eos
self.src_vocab_size = len(src_dict)
self.src_search = search.BeamSearch(src_dict)
self.search = HierarchicalBeamSearch(tgt_dict)
def build_generator(self, args):
if getattr(args, 'score_reference', False):
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(
self.target_dictionary,
compute_alignment=getattr(args, 'print_alignment', False),
)
from fairseq.sequence_generator import SequenceGenerator, SequenceGeneratorWithAlignment
# Choose search strategy. Defaults to Beam Search.
#print("ARGS is",args)
sampling = getattr(args, 'sampling', False)
#print("sampling is",sampling)
sampling_topk = getattr(args, 'sampling_topk', -1)
sampling_topp = getattr(args, 'sampling_topp', -1.0)
diverse_beam_groups = getattr(args, 'diverse_beam_groups', -1)
diverse_beam_strength = getattr(args, 'diverse_beam_strength', 0.5),
match_source_len = getattr(args, 'match_source_len', False)
diversity_rate = getattr(args, 'diversity_rate', -1)
dedup = getattr(args, 'dedup', False)
verb_idxs = getattr(args, 'verb_idxs', [])
banned_toks = getattr(args, 'banned_toks', [])
coef_trainer = getattr(args, 'coef_trainer', None)
coefs = getattr(args, 'coefs', [])
learn = getattr(args, 'learn', False)
learn_every_token = getattr(args, 'learn_every_token', False)
if (sum(
int(cond) for cond in [
sampling,
diverse_beam_groups > 0,
match_source_len,
diversity_rate > 0,
]) > 1):
raise ValueError(
'Provided Search parameters are mutually exclusive.')
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
assert sampling_topp < 0 or sampling, '--sampling-topp requires --sampling'
if sampling:
search_strategy = search.Sampling(self.target_dictionary,
sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
search_strategy = search.DiverseBeamSearch(self.target_dictionary,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
# this is useful for tagging applications where the output
# length should match the input length, so we hardcode the
# length constraints for simplicity
search_strategy = search.LengthConstrainedBeamSearch(
self.target_dictionary,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
elif diversity_rate > -1:
search_strategy = search.DiverseSiblingsSearch(
self.target_dictionary, diversity_rate)
else:
search_strategy = search.BeamSearch(self.target_dictionary)
if getattr(args, 'print_alignment', False):
seq_gen_cls = SequenceGeneratorWithAlignment
else:
seq_gen_cls = SequenceGenerator
return seq_gen_cls(
self.target_dictionary,
beam_size=getattr(args, 'beam', 5),
max_len_a=getattr(args, 'max_len_a', 0),
max_len_b=getattr(args, 'max_len_b', 200),
min_len=getattr(args, 'min_len', 1),
normalize_scores=(not getattr(args, 'unnormalized', False)),
len_penalty=getattr(args, 'lenpen', 1),
unk_penalty=getattr(args, 'unkpen', 0),
temperature=getattr(args, 'temperature', 1.),
match_source_len=getattr(args, 'match_source_len', False),
no_repeat_ngram_size=getattr(args, 'no_repeat_ngram_size', 0),
search_strategy=search_strategy
) #dedup=dedup,verb=verb_idxs,banned_toks=banned_toks,coef_trainer=coef_trainer,coefs=coefs,learn=learn,learn_every_token=learn_every_token)
def __init__(
self,
models,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.0,
unk_penalty=0.0,
temperature=1.0,
match_source_len=False,
no_repeat_ngram_size=0,
search_strategy=None,
eos=None,
symbols_to_strip_from_output=None,
lm_model=None,
lm_weight=1.0,
**kwargs,
):
"""Generates translations of a given source sentence.
Args:
models (List[~fairseq.models.FairseqModel]): ensemble of models,
currently support fairseq.models.TransformerModel for scripting
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
super().__init__()
if isinstance(models, EnsembleModel):
self.model = models
else:
self.model = EnsembleModel(models)
self.tgt_dict = tgt_dict
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos() if eos is None else eos
self.symbols_to_strip_from_output = (
symbols_to_strip_from_output.union({self.eos})
if symbols_to_strip_from_output is not None else {self.eos})
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
self.repeat_ngram_blocker = NGramRepeatBlock(no_repeat_ngram_size)
self.eos_factor = kwargs.get("eos_factor", None)
assert temperature > 0, "--temperature must be greater than 0"
assert self.eos_factor is None or self.eos_factor >= 1.0, "--eos-factor must be >= 1.0 if set"
self.search = (search.BeamSearch(tgt_dict)
if search_strategy is None else search_strategy)
# We only need to set src_lengths in LengthConstrainedBeamSearch.
# As a module attribute, setting it would break in multithread
# settings when the model is shared.
self.should_set_src_lengths = (hasattr(self.search,
"needs_src_lengths")
and self.search.needs_src_lengths)
self.model.eval()
self.lm_model = lm_model
self.lm_weight = lm_weight
if self.lm_model is not None:
self.lm_model.eval()
def __init__(self,
models,
tgt_dict,
beam_size=1,
minlen=1,
maxlen=None,
stop_early=True,
normalize_scores=True,
len_penalty=1.,
unk_penalty=0.,
retain_dropout=False,
sampling=False,
sampling_topk=-1,
sampling_temperature=1.,
diverse_beam_groups=-1,
diverse_beam_strength=0.5,
match_source_len=False,
no_repeat_ngram_size=0):
"""Generates translations of a given source sentence.
Args:
beam_size (int, optional): beam width (default: 1)
min/maxlen (int, optional): the length of the generated output will
be bounded by minlen and maxlen (not including end-of-sentence)
stop_early (bool, optional): stop generation immediately after we
finalize beam_size hypotheses, even though longer hypotheses
might have better normalized scores (default: True)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
sampling (bool, optional): sample outputs instead of beam search
(default: False)
sampling_topk (int, optional): only sample among the top-k choices
at each step (default: -1)
sampling_temperature (float, optional): temperature for sampling,
where values >1.0 produces more uniform sampling and values
<1.0 produces sharper sampling (default: 1.0)
diverse_beam_groups/strength (float, optional): parameters for
Diverse Beam Search sampling
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
self.models = models
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
self.minlen = minlen
max_decoder_len = min(m.max_decoder_positions() for m in self.models)
max_decoder_len -= 1 # we define maxlen not including the EOS marker
self.maxlen = max_decoder_len if maxlen is None else min(
maxlen, max_decoder_len)
self.stop_early = stop_early
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
self.js = {}
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk,
sampling_temperature)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict,
diverse_beam_groups,
diverse_beam_strength)
elif match_source_len:
self.search = search.LengthConstrainedBeamSearch(
tgt_dict,
min_len_a=1,
min_len_b=0,
max_len_a=1,
max_len_b=0,
)
else:
self.search = search.BeamSearch(tgt_dict)
def __init__(
self,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.,
unk_penalty=0.,
retain_dropout=False,
sampling=False,
sampling_topk=-1,
sampling_topp=-1.0,
temperature=1.,
diverse_beam_groups=-1,
diverse_beam_strength=0.5,
match_source_len=False,
no_repeat_ngram_size=0,
coverage_weight=0.0,
eos_factor=None,
):
"""Generates translations of a given source sentence.
Args:
tgt_dict (~fairseq.data.Dictionary): target dictionary
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
retain_dropout (bool, optional): use dropout when generating
(default: False)
sampling (bool, optional): sample outputs instead of beam search
(default: False)
sampling_topk (int, optional): only sample among the top-k choices
at each step (default: -1)
sampling_topp (float, optional): only sample among the smallest set
of words whose cumulative probability mass exceeds p
at each step (default: -1.0)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
diverse_beam_groups/strength (float, optional): parameters for
Diverse Beam Search sampling
match_source_len (bool, optional): outputs should match the source
length (default: False)
"""
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.retain_dropout = retain_dropout
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
self.coverage_weight = coverage_weight
self.eos_factor = eos_factor
assert sampling_topk < 0 or sampling, '--sampling-topk requires --sampling'
assert sampling_topp < 0 or sampling, '--sampling-topp requires --sampling'
assert temperature > 0, '--temperature must be greater than 0'
assert eos_factor is None or eos_factor >= 1.0, '--eos-factor must be >= 1.0 if set'
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk, sampling_topp)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict, diverse_beam_groups, diverse_beam_strength)
elif match_source_len:
self.search = search.LengthConstrainedBeamSearch(
tgt_dict, min_len_a=1, min_len_b=0, max_len_a=1, max_len_b=0,
)
else:
self.search = search.BeamSearch(tgt_dict)
def __init__(
self,
models,
tgt_dict,
beam_size=1,
minlen=1,
maxlen=None,
stop_early=True,
normalize_scores=True,
len_penalty=0,
unk_reward=0,
lexicon_reward=0,
retain_dropout=False,
word_reward=0,
model_weights=None,
use_char_source=False,
diverse_beam_groups=-1,
diverse_beam_strength=0.5,
diversity_sibling_gamma=0.0,
sampling=False,
sampling_topk=-1,
sampling_temperature=1,
):
"""Generates translations of a given source sentence.
Args:
models: List of FairseqModel objects. Each one must implement
reorder_encoder_output() method to replicate encoder outputs.
min/maxlen: The length of the generated output will be bounded by
minlen and maxlen (not including the end-of-sentence marker).
stop_early: Stop generation immediately after we finalize beam_size
hypotheses, even though longer hypotheses might have better
normalized scores.
normalize_scores: Normalize scores by the length of the output.
word_reward: add this value to score each token except EOS
(an alternative method to len_penalty for encouraging longer
output)
model_weights: None or list of Python floats of the same length as
`models` with ensemble interpolation weights.
use_char_source: if True, encoder inputs consist of (src_tokens,
src_lengths, char_inds, word_lengths)
diverse_beam_groups: number of groups for Diverse Beam Search
(-1 by default is vanilla beam search)
diverse_beam_strength: strength of diversity penalty for Diverse
Beam Search.
diversity_sibling_gamma: The diversity rate of sibling rank (-0.0 by default
to disable sibling rank penalty)
sampling (bool, optional): sample outputs instead of beam search
(default: False)
sampling_topk (int, optional): only sample among the top-k choices
at each step (default: -1)
sampling_temperature (float, optional): temperature for sampling,
where values >1.0 produces more uniform sampling and values
<1.0 produces sharper sampling (default: 1.0)
"""
self.models = models
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos()
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
self.minlen = minlen
max_decoder_len = min(m.max_decoder_positions() for m in self.models)
self.maxlen = (max_decoder_len if maxlen is None else min(
maxlen, max_decoder_len))
self.stop_early = stop_early
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_reward = unk_reward
self.lexicon_reward = lexicon_reward
self.lexicon_indices = tgt_dict.lexicon_indices_list()
self.retain_dropout = retain_dropout
self.word_reward = word_reward
if model_weights is not None:
assert len(models) == len(model_weights)
self.model_weights = model_weights
else:
self.model_weights = [1.0 / len(models)] * len(models)
self.use_char_source = use_char_source
assert sampling_topk < 0 or sampling, "--sampling-topk requires --sampling"
if sampling:
self.search = search.Sampling(tgt_dict, sampling_topk,
sampling_temperature)
elif diverse_beam_groups > 0:
self.search = search.DiverseBeamSearch(tgt_dict,
diverse_beam_groups,
diverse_beam_strength)
else:
self.search = search.BeamSearch(tgt_dict)
self.diversity_sibling_gamma = diversity_sibling_gamma
def __init__(
self,
models,
tgt_dict,
beam_size=1,
max_len_a=0,
max_len_b=200,
min_len=1,
normalize_scores=True,
len_penalty=1.0,
unk_penalty=0.0,
temperature=1.0,
match_source_len=False,
no_repeat_ngram_size=0,
repetition_penalty=1,
repetition_penalty_slope=0,
repetition_penalty_window_length=None,
ngram_repetition_penalty=None,
search_strategy=None,
eos=None,
symbols_to_strip_from_output=None,
):
"""Generates translations of a given source sentence.
Args:
models (List[~fairseq.models.FairseqModel]): ensemble of models,
currently support fairseq.models.TransformerModel for scripting
beam_size (int, optional): beam width (default: 1)
max_len_a/b (int, optional): generate sequences of maximum length
ax + b, where x is the source length
min_len (int, optional): the minimum length of the generated output
(not including end-of-sentence)
normalize_scores (bool, optional): normalize scores by the length
of the output (default: True)
len_penalty (float, optional): length penalty, where <1.0 favors
shorter, >1.0 favors longer sentences (default: 1.0)
unk_penalty (float, optional): unknown word penalty, where <0
produces more unks, >0 produces fewer (default: 0.0)
temperature (float, optional): temperature, where values
>1.0 produce more uniform samples and values <1.0 produce
sharper samples (default: 1.0)
match_source_len (bool, optional): outputs should match the source
length (default: False)
no_repeat_ngram_size (int, optional): forbid repetition of n-grams
this long (default: 0)
ngram_repetition_penalty (int, optional): penalty for repeating an
n-gram, combine with no_repeat_ngram_size
repetition_penalty (float, optional): CTRL-style repetition penalty
term (default: 1=no penalty)
repetition_penalty_slope (float, optional): slope for the power to
raise the repetition penalty to, (penalty)^(1+slope*(count-1))
repetition_penalty_window_length (int, optional): only look back
this long to calculate the repetition penalty
"""
super().__init__()
if isinstance(models, EnsembleModel):
self.model = models
else:
self.model = EnsembleModel(models)
self.tgt_dict = tgt_dict
self.pad = tgt_dict.pad()
self.unk = tgt_dict.unk()
self.eos = tgt_dict.eos() if eos is None else eos
self.symbols_to_strip_from_output = (
symbols_to_strip_from_output.union({self.eos})
if symbols_to_strip_from_output is not None else {self.eos})
self.vocab_size = len(tgt_dict)
self.beam_size = beam_size
# the max beam size is the dictionary size - 1, since we never select pad
self.beam_size = min(beam_size, self.vocab_size - 1)
self.max_len_a = max_len_a
self.max_len_b = max_len_b
self.min_len = min_len
self.normalize_scores = normalize_scores
self.len_penalty = len_penalty
self.unk_penalty = unk_penalty
self.temperature = temperature
self.match_source_len = match_source_len
self.no_repeat_ngram_size = no_repeat_ngram_size
self.repetition_penalty = repetition_penalty
self.repetition_penalty_slope = repetition_penalty_slope
self.repetition_penalty_window_length = repetition_penalty_window_length
self.ngram_repetition_penalty = ngram_repetition_penalty
assert temperature > 0, "--temperature must be greater than 0"
self.search = (
search.BeamSearch(tgt_dict) if search_strategy is None else search_strategy
)
# We only need to set src_lengths in LengthConstrainedBeamSearch.
# As a module attribute, setting it would break in multithread
# settings when the model is shared.
self.should_set_src_lengths = hasattr(self.search, 'needs_src_lengths') and self.search.needs_src_lengths
self.model.eval()
