def test_as_training_data_produces_correct_numpy_arrays(self):
instance = IndexedBackgroundInstance(self.base_instance,
self.background_instances)
instance.pad({'word_sequence_length': 3, 'background_sentences': 2})
(word_array, background_array), label = instance.as_training_data()
assert numpy.all(label == numpy.asarray([0, 1]))
assert numpy.all(word_array == numpy.asarray([0, 1, 2]))
assert numpy.all(
background_array == numpy.asarray([[2, 3, 4], [0, 4, 5]]))
instance.indexed_instance.label = False
_, label = instance.as_training_data()
assert numpy.all(label == numpy.asarray([1, 0]))
instance.label = True # we ignore this label, only using the indexed_instance label
_, label = instance.as_training_data()
assert numpy.all(label == numpy.asarray([1, 0]))
def test_as_training_data_produces_correct_numpy_arrays_with_complex_contained_instance(
self):
# We need the background array to always be _second_, not last.
instance = IndexedBackgroundInstance(self.qa_instance,
self.background_instances)
instance.pad({
'word_sequence_length': 2,
'answer_length': 2,
'num_options': 3,
'background_sentences': 2,
})
(question_array, background_array,
answer_array), label = instance.as_training_data()
assert numpy.all(label == numpy.asarray([0, 1, 0]))
assert numpy.all(question_array == numpy.asarray([2, 3]))
assert numpy.all(
answer_array == numpy.asarray([[2, 3], [0, 4], [5, 6]]))
assert numpy.all(background_array == numpy.asarray([[3, 4], [4, 5]]))