def __call__(self, words, weights, vocabulary_max):
if len(words) < vocabulary_max * self.trigger_ratio:
return words, weights
if not isinstance(words, numpy.ndarray):
words = numpy.array(words)
# Tail optimization does not help with very large vocabularies
if len(words) > vocabulary_max * 2:
indices = numpy.argpartition(weights, len(weights) - vocabulary_max)
indices = indices[-vocabulary_max:]
words = words[indices]
weights = weights[indices]
return words, weights
# Vocabulary typically consists of these three parts:
# 1) the core - we found it's border - `core_end` - 15%
# 2) the body - 70%
# 3) the minor tail - 15%
# (1) and (3) are roughly the same size
# (3) can be safely discarded, (2) can be discarded with care,
# (1) shall never be discarded.
sorter = numpy.argsort(weights)[::-1]
weights = weights[sorter]
trend_start = int(len(weights) * 0.2)
trend_finish = int(len(weights) * 0.8)
z = numpy.polyfit(numpy.arange(trend_start, trend_finish),
numpy.log(weights[trend_start:trend_finish]),
1)
exp_z = numpy.exp(z[1] + z[0] * numpy.arange(len(weights)))
avg_error = numpy.abs(weights[trend_start:trend_finish] -
exp_z[trend_start:trend_finish]).mean()
tail_size = numpy.argmax((numpy.abs(weights - exp_z) < avg_error)[::-1])
weights = weights[:-tail_size][:vocabulary_max]
words = words[sorter[:-tail_size]][:vocabulary_max]
return words, weights
评论列表
文章目录
