def test_score_batch(self):
# Network predicts <unk> probability.
scorer = TextScorer(self.dummy_network)
word_ids = numpy.arange(6).reshape((3, 2))
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:, 0].astype('float32') / 5))
assert_almost_equal(logprobs[1],
numpy.log(word_ids[1:, 1].astype('float32') / 5))
# <unk> is removed from the resulting logprobs.
scorer = TextScorer(self.dummy_network, ignore_unk=True)
word_ids = numpy.arange(6).reshape((3, 2))
word_ids[1, 1] = self.vocabulary.word_to_id['<unk>']
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:, 0].astype('float32') / 5))
assert_almost_equal(logprobs[1],
numpy.log(word_ids[2:, 1].astype('float32') / 5))
# <unk> is assigned a constant logprob.
scorer = TextScorer(self.dummy_network,
ignore_unk=False,
unk_penalty=-5)
word_ids = numpy.arange(6).reshape((3, 2))
word_ids[1, 1] = self.vocabulary.word_to_id['<unk>']
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:, 0].astype('float32') / 5))
assert_almost_equal(logprobs[1][0], -5)
assert_almost_equal(logprobs[1][1],
numpy.log(word_ids[2, 1].astype('float32') / 5))
def test_score_batch(self):
# Network predicts <unk> probability.
scorer = TextScorer(self.dummy_network)
word_ids = numpy.arange(6).reshape((3, 2))
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:,0].astype('float32') / 5))
assert_almost_equal(logprobs[1],
numpy.log(word_ids[1:,1].astype('float32') / 5))
# <unk> is removed from the resulting logprobs.
scorer = TextScorer(self.dummy_network, ignore_unk=True)
word_ids = numpy.arange(6).reshape((3, 2))
word_ids[1,1] = self.vocabulary.word_to_id['<unk>']
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:,0].astype('float32') / 5))
assert_almost_equal(logprobs[1],
numpy.log(word_ids[2:,1].astype('float32') / 5))
# <unk> is assigned a constant logprob.
scorer = TextScorer(self.dummy_network, ignore_unk=False, unk_penalty=-5)
word_ids = numpy.arange(6).reshape((3, 2))
word_ids[1,1] = self.vocabulary.word_to_id['<unk>']
class_ids = numpy.arange(6).reshape((3, 2))
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(logprobs[0],
numpy.log(word_ids[1:,0].astype('float32') / 5))
assert_almost_equal(logprobs[1][0], -5)
assert_almost_equal(logprobs[1][1],
numpy.log(word_ids[2,1].astype('float32') / 5))
def test_score_batch(self):
# Network predicts <unk> probability. Out-of-shortlist words are mapped
# to <unk> class by .
scorer = TextScorer(self.dummy_network, use_shortlist=False)
word_ids = numpy.arange(15).reshape((3, 5)).T
class_ids, _ = self.vocabulary.get_class_memberships(word_ids)
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(
logprobs[0], numpy.log(word_ids[1:, 0].astype('float32') / 100.0))
assert_almost_equal(
logprobs[1], numpy.log(word_ids[1:, 1].astype('float32') / 100.0))
self.assertAlmostEqual(logprobs[2][0],
numpy.log(11.0 / 100.0),
places=5) # </s>
self.assertAlmostEqual(logprobs[2][1],
numpy.log(12.0 / 100.0),
places=5) # <unk>
self.assertAlmostEqual(logprobs[2][2],
numpy.log(12.0 / 100.0),
places=5)
self.assertAlmostEqual(logprobs[2][3],
numpy.log(12.0 / 100.0),
places=5)
# Network predicts <unk> probability. This is distributed for
# out-of-shortlist words according to word frequency.
scorer = TextScorer(self.dummy_network, use_shortlist=True)
word_ids = numpy.arange(15).reshape((3, 5)).T
class_ids, _ = self.vocabulary.get_class_memberships(word_ids)
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(
logprobs[0], numpy.log(word_ids[1:, 0].astype('float32') / 100.0))
assert_almost_equal(
logprobs[1], numpy.log(word_ids[1:, 1].astype('float32') / 100.0))
self.assertAlmostEqual(logprobs[2][0],
numpy.log(11.0 / 100.0),
places=5) # </s>
self.assertIsNone(logprobs[2][1]) # <unk>
self.assertAlmostEqual(logprobs[2][2],
numpy.log(12.0 / 100.0 * 0.3),
places=5)
self.assertAlmostEqual(logprobs[2][3],
numpy.log(12.0 / 100.0 * 0.7),
places=5)
# OOV and OOS words are replaced with None.
scorer = TextScorer(self.dummy_network,
use_shortlist=False,
exclude_unk=True)
word_ids = numpy.arange(15).reshape((3, 5)).T
class_ids, _ = self.vocabulary.get_class_memberships(word_ids)
membership_probs = numpy.ones_like(word_ids).astype('float32')
mask = numpy.ones_like(word_ids)
logprobs = scorer.score_batch(word_ids, class_ids, membership_probs,
mask)
assert_almost_equal(
logprobs[0], numpy.log(word_ids[1:, 0].astype('float32') / 100.0))
assert_almost_equal(
logprobs[1], numpy.log(word_ids[1:, 1].astype('float32') / 100.0))
self.assertAlmostEqual(logprobs[2][0],
numpy.log(11.0 / 100.0),
places=5) # </s>
self.assertIsNone(logprobs[2][1]) # <unk>
self.assertIsNone(logprobs[2][2])
self.assertIsNone(logprobs[2][3])
