Python源码示例：fairseq.sequence.SequenceGenerator()
示例1
def test_with_normalization(self):
generator = SequenceGenerator([self.model], self.tgt_dict)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
示例2
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator([self.model], self.tgt_dict, normalize_scores=False)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
示例3
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator([self.model], self.tgt_dict, len_penalty=lenpen)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
示例4
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator([self.model], self.tgt_dict, len_penalty=lenpen)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
示例5
def test_maxlen(self):
generator = SequenceGenerator([self.model], self.tgt_dict, maxlen=2)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.1, 0.6])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w2, w2, eos])
self.assertHypoScore(hypos[1][1], [0.3, 0.9, 0.01])
示例6
def test_with_normalization(self):
generator = SequenceGenerator([self.model], self.tgt_dict, beam_size=2)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
示例7
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, normalize_scores=False
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
示例8
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
示例9
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
示例10
def test_diverse_beam_search(self):
search_strategy = search.DiverseBeamSearch(self.tgt_dict, num_groups=2, diversity_strength=0.)
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, search_strategy=search_strategy,
)
sample = {'net_input': {'src_tokens': self.src_tokens, 'src_lengths': self.src_lengths}}
hypos = generator.forward(sample)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 0.6, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.9])
示例11
def build_generator(self, models, args):
if getattr(args, 'score_reference', False):
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(
self.target_dictionary,
eos=self.tgt_dict.index('[{}]'.format(self.args.target_lang))
)
else:
from fairseq.sequence_generator import SequenceGenerator
return SequenceGenerator(
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),
eos=self.tgt_dict.index('[{}]'.format(self.args.target_lang))
)
示例12
def test_with_normalization(self):
generator = SequenceGenerator([self.model], self.tgt_dict)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
示例13
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator([self.model], self.tgt_dict, normalize_scores=False)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
示例14
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator([self.model], self.tgt_dict, len_penalty=lenpen)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
示例15
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator([self.model], self.tgt_dict, len_penalty=lenpen)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
示例16
def test_maxlen(self):
generator = SequenceGenerator([self.model], self.tgt_dict, maxlen=2)
hypos = generator.generate(self.src_tokens, self.src_lengths, beam_size=2)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.1, 0.6])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w2, w2, eos])
self.assertHypoScore(hypos[1][1], [0.3, 0.9, 0.01])
示例17
def test_with_normalization(self):
generator = SequenceGenerator([self.model], self.tgt_dict, beam_size=2)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
示例18
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, normalize_scores=False
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
示例19
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
示例20
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, len_penalty=lenpen
)
hypos = generator.forward(self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
示例21
def test_diverse_beam_search(self):
search_strategy = search.DiverseBeamSearch(self.tgt_dict, num_groups=2, diversity_strength=0.)
generator = SequenceGenerator(
[self.model], self.tgt_dict, beam_size=2, search_strategy=search_strategy,
)
sample = {'net_input': {'src_tokens': self.src_tokens, 'src_lengths': self.src_lengths}}
hypos = generator.forward(sample)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 0.6, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.9])
示例22
def build_generator(self, models, args):
if getattr(args, 'score_reference', False):
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(
self.target_dictionary,
eos=self.tgt_dict.index('[{}]'.format(self.target_lang))
)
else:
from fairseq.sequence_generator import SequenceGenerator
return SequenceGenerator(
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),
eos=self.tgt_dict.index('[{}]'.format(self.args.target_lang))
)
示例23
def test_with_normalization(self):
generator = SequenceGenerator(self.tgt_dict, beam_size=2)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6])
示例24
def test_without_normalization(self):
# Sentence 1: unchanged from the normalized case
# Sentence 2: beams swap order
generator = SequenceGenerator(self.tgt_dict, beam_size=2, normalize_scores=False)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], normalized=False)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], normalized=False)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], normalized=False)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], normalized=False)
示例25
def test_with_lenpen_favoring_short_hypos(self):
lenpen = 0.6
generator = SequenceGenerator(self.tgt_dict, beam_size=2, len_penalty=lenpen)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.6], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
示例26
def test_with_lenpen_favoring_long_hypos(self):
lenpen = 5.0
generator = SequenceGenerator(self.tgt_dict, beam_size=2, len_penalty=lenpen)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][0], [0.1, 0.9, 0.9, 1.0], lenpen=lenpen)
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 1.0], lenpen=lenpen)
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.4, 1.0], lenpen=lenpen)
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.6], lenpen=lenpen)
示例27
def test_no_stop_early(self):
generator = SequenceGenerator(self.tgt_dict, stop_early=False, beam_size=2)
hypos = generator.generate([self.model], self.sample)
eos, w1, w2 = self.tgt_dict.eos(), self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w2, w1, w2, eos])
self.assertHypoScore(hypos[0][1], [0.1, 0.9, 0.9, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w2, w2, w2, w2, eos])
self.assertHypoScore(hypos[1][0], [0.3, 0.9, 0.99, 0.4, 1.0])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, w1, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.4, 1.0])
示例28
def test_diverse_beam_search(self):
generator = SequenceGenerator(
self.tgt_dict, beam_size=2, diverse_beam_groups=2, diverse_beam_strength=0.,
)
sample = {'net_input': {'src_tokens': self.src_tokens, 'src_lengths': self.src_lengths}}
hypos = generator.generate([self.model], sample)
eos, w1, w2 = self.eos, self.w1, self.w2
# sentence 1, beam 1
self.assertHypoTokens(hypos[0][0], [w1, w1, eos])
self.assertHypoScore(hypos[0][0], [0.9, 0.6, 1.0])
# sentence 1, beam 2
self.assertHypoTokens(hypos[0][1], [w1, w1, eos])
self.assertHypoScore(hypos[0][1], [0.9, 0.6, 1.0])
# sentence 2, beam 1
self.assertHypoTokens(hypos[1][0], [w1, w2, eos])
self.assertHypoScore(hypos[1][0], [0.7, 0.4, 0.9])
# sentence 2, beam 2
self.assertHypoTokens(hypos[1][1], [w1, w2, eos])
self.assertHypoScore(hypos[1][1], [0.7, 0.4, 0.9])
示例29
def build_generator(self, args):
if args.score_reference:
from fairseq.sequence_scorer import SequenceScorer
return SequenceScorer(self.target_dictionary)
else:
from fairseq.sequence_generator import SequenceGenerator
return SequenceGenerator(
self.target_dictionary,
beam_size=args.beam,
max_len_a=args.max_len_a,
max_len_b=args.max_len_b,
min_len=args.min_len,
stop_early=(not args.no_early_stop),
normalize_scores=(not args.unnormalized),
len_penalty=args.lenpen,
unk_penalty=args.unkpen,
sampling=args.sampling,
sampling_topk=args.sampling_topk,
sampling_temperature=args.sampling_temperature,
diverse_beam_groups=args.diverse_beam_groups,
diverse_beam_strength=args.diverse_beam_strength,
match_source_len=args.match_source_len,
no_repeat_ngram_size=args.no_repeat_ngram_size,
)
示例30
def test_ensemble_sequence_generator(self):
model = self.transformer_model
generator = SequenceGenerator(
[model], self.task.tgt_dict, beam_size=2, no_repeat_ngram_size=2
)
scripted_model = torch.jit.script(generator)
self._test_save_and_load(scripted_model)
