def test_collate_batch_idxs(self) -> None:
"""
Tests that (q|c)[len_sorted][orig_idxs] == q|c
"""
samples: List[EncodedSample] = [
self.make_sample("c1 c2 c3", [], "q0", "c1 c2"),
self.make_sample("c1", [], "q1", "c1 c2 c3"),
self.make_sample("c1 c2", [], "q2", "c1"),
]
batch: QABatch = collate_batch(samples)
self.assertTrue(np.allclose(batch.context_len_idxs, [0, 2, 1]))
self.assertTrue(np.allclose(batch.question_len_idxs, [1, 0, 2]))
self.assertTrue(
np.allclose(
batch.question_words,
(batch.question_words[batch.question_len_idxs]
)[batch.question_orig_idxs],
))
self.assertTrue(
np.allclose(
batch.context_words,
(batch.context_words[batch.context_len_idxs]
)[batch.context_orig_idxs],
))
def test_collate_batch_q_len_sorting(self) -> None:
"""
Tests that question lengths in batch are sorted, and
len_idxs indices map to correct indices
"""
samples: List[EncodedSample] = [
self.make_sample("c1", [], "q0", "c1"),
self.make_sample("c1", [], "q1", "c1 c2 c3"),
self.make_sample("c1", [], "q2", "c1 c2"),
]
batch: QABatch = collate_batch(samples)
self.assertTrue(
np.all(batch.question_words.numpy() == np.stack(
[[1, 0, 0], [1, 2, 3], [1, 2, 0]])),
"Batch questions: {0} Expected: {1}".format(
batch.question_words, [[1, 0, 0], [1, 2, 3], [1, 2, 0]]),
)
self.assertTrue(
np.allclose(batch.question_lens, [3, 2, 1]),
"Question lens: {0} expected: {1}".format(batch.question_lens,
[3, 2, 1]),
)
self.assertTrue(
np.allclose(batch.question_len_idxs, [1, 2, 0]),
"Question len idxs: {0} expected: {1}".format(
batch.question_len_idxs, [1, 2, 0]),
)
def test_collate_batch_different_context_word_lens(self) -> None:
"""
Tests that collate batch deals with questions with different
context word lengths
"""
samples: List[EncodedSample] = [
self.make_sample("c00", [], "q0", "c1"),
self.make_sample("c1", [], "q1", "c1"),
]
batch: QABatch = collate_batch(samples)
self.assertEqual(batch.context_chars.shape, t.Size([2, 1, 3]))
self.check_collated_chars(batch.context_chars, batch.context_words)
def test_collate_batch_context_chars(self) -> None:
"""
Tests that collate batch includes all context word characters that are parsed correctly
"""
samples: List[EncodedSample] = [
self.make_sample("c1", [], "q0", "c1"),
self.make_sample("c2", [], "q1", "c1"),
self.make_sample("c3", [], "q2", "c1"),
]
batch: QABatch = collate_batch(samples)
self.assertEqual(batch.question_chars.shape, t.Size([3, 1, 2]))
self.check_collated_chars(batch.context_chars, batch.context_words)
def test_collate_batch_different_question_word_numbers_and_lens(
self) -> None:
"""
Tests that collate batch deals with questions with different
question word lengths and numbers
"""
samples: List[EncodedSample] = [
self.make_sample("c1", [], "q0", "c1"),
self.make_sample("c1", [], "q1", "c00"),
self.make_sample("c1", [], "q2", "c1 c00"),
]
batch: QABatch = collate_batch(samples)
self.assertEqual(batch.question_chars.shape, t.Size([3, 2, 3]))
self.check_collated_chars(batch.question_chars, batch.question_words)
def test_collate_batch_ctx_len_sorting(self) -> None:
"""
Tests that question lengths in batch are sorted, and
len_idxs indices map to correct indices
"""
samples: List[EncodedSample] = [
self.make_sample("c1 c2 c3", [], "q0", "c1"),
self.make_sample("c1", [], "q1", "c1"),
self.make_sample("c1 c2", [], "q2", "c1"),
]
batch: QABatch = collate_batch(samples)
self.assertTrue(np.allclose(batch.context_lens, [3, 2, 1]))
self.assertTrue(np.allclose(batch.context_len_idxs, [0, 2, 1]))
def test_collate_batch_simple_words(self) -> None:
"""
Tests that collate batch includes all question ids in original order
"""
samples: List[EncodedSample] = [
self.make_sample("c1", [], "q0", "c1"),
self.make_sample("c1", [], "q1", "c2"),
self.make_sample("c1", [], "q2", "c3"),
]
batch: QABatch = collate_batch(samples)
self.assertEqual(
batch.question_ids,
[QuestionId("q0"),
QuestionId("q1"),
QuestionId("q2")])
self.assertTrue(
np.all(batch.question_words.numpy() == np.stack([[1], [2], [3]])),
"Batch questions: {0} Expected: {1}".format(
batch.question_words, [[1], [2], [3]]),
)