def test_natural_questions_open(self):
input_data = {
'question': ['What are the names of the Olsen Twins?'],
'answer': ['Mary-Kate', 'Ashley']
}
og_dataset = tf.data.Dataset.from_tensors(input_data)
dataset = preprocessors.natural_questions_open(og_dataset)
t5.data.assert_dataset(
dataset, {
'inputs':
'nq question: What are the names of the Olsen Twins?',
'targets': 'Mary-Kate',
'answers': ['Mary-Kate', 'Ashley'],
})
def test_natural_questions(self):
input_examples = [{
'question': {
'text': 'is the answer to this question no',
},
'annotations': {
'short_answers': {
'start_token': ([], [0, 0]),
'end_token': ([], [0, 0]),
'text': ([], [0, 0])
},
'yes_no_answer': [-1, -1]
}
}, {
'question': {
'text': 'is the answer to this question yes',
},
'annotations': {
'short_answers': {
'start_token': ([3, 3], [1, 0, 1, 0]),
'end_token': ([7, 5], [1, 0, 1, 0]),
'text': (['not sure sir', 'not sure'], [1, 0, 1, 0]),
},
'yes_no_answer': [-1, 0, -1, 1]
}
}, {
'question': {
'text': 'what are the names of the olsen twins',
},
'annotations': {
'short_answers': {
'start_token': ([0, 3], [2, 0]),
'end_token': ([3, 4], [2, 0]),
'text': (['Mary-Kate', 'Ashley'], [2, 0])
},
'yes_no_answer': [-1, -1]
}
}]
def _short_ans_to_ragged(ex):
for field in ['start_token', 'end_token', 'text']:
values, row_lengths = ex['annotations']['short_answers'][field]
ex['annotations']['short_answers'][field] = (
tf.RaggedTensor.from_row_lengths(values, row_lengths))
return ex
og_dataset = tf.data.Dataset.from_generator(
lambda: (x for x in input_examples),
output_types={
'question': {
'text': tf.string
},
'annotations': {
'short_answers': {
'start_token': (tf.int64, tf.int64),
'end_token': (tf.int64, tf.int64),
'text': (tf.string, tf.int64)
},
'yes_no_answer': tf.int64
}
},
output_shapes={
'question': {
'text': []
},
'annotations': {
'short_answers': {
'start_token': ([None], [None]),
'end_token': ([None], [None]),
'text': ([None], [None]),
},
'yes_no_answer': [None]
}
}).map(_short_ans_to_ragged)
dataset = preprocessors.natural_questions_nocontext(og_dataset)
test_utils.assert_dataset(
dataset,
[{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'answer: no answer: yes answer: not sure sir '
'answer: not sure',
'short_answers/values': ['not sure sir', 'not sure'],
'short_answers/row_starts': [0, 1, 1, 2],
'yes_no_answers': [-1, 0, -1, 1],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Mary-Kate answer: Ashley',
'short_answers/values': ['Mary-Kate', 'Ashley'],
'short_answers/row_starts': [0, 2],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_nocontext(og_dataset,
drop_yes_no=True)
test_utils.assert_dataset(
dataset, [{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'answer: not sure sir answer: not sure',
'short_answers/values': ['not sure sir', 'not sure'],
'short_answers/row_starts': [0, 1, 1, 2],
'yes_no_answers': [-1, -1, -1, -1],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Mary-Kate answer: Ashley',
'short_answers/values': ['Mary-Kate', 'Ashley'],
'short_answers/row_starts': [0, 2],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_nocontext(og_dataset,
max_tokens=2)
test_utils.assert_dataset(
dataset, [{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'answer: no answer: yes answer: not sure',
'short_answers/values': ['not sure'],
'short_answers/row_starts': [0, 0, 0, 1],
'yes_no_answers': [-1, 0, -1, 1],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Ashley',
'short_answers/values': ['Ashley'],
'short_answers/row_starts': [0, 1],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_nocontext(og_dataset,
max_answers=1)
test_utils.assert_dataset(
dataset, [{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'answer: no',
'short_answers/values': ['not sure sir', 'not sure'],
'short_answers/row_starts': [0, 1, 1, 2],
'yes_no_answers': [-1, 0, -1, 1],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Mary-Kate',
'short_answers/values': ['Mary-Kate', 'Ashley'],
'short_answers/row_starts': [0, 2],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_nocontext(og_dataset,
drop_yes_no=True,
max_tokens=2,
max_answers=1)
test_utils.assert_dataset(
dataset, [{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'answer: not sure',
'short_answers/values': ['not sure'],
'short_answers/row_starts': [0, 0, 0, 1],
'yes_no_answers': [-1, -1, -1, -1],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Ashley',
'short_answers/values': ['Ashley'],
'short_answers/row_starts': [0, 1],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_nocontext(og_dataset,
drop_yes_no=True,
max_tokens=1)
test_utils.assert_dataset(dataset, [{
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'answer: Ashley',
'short_answers/values': ['Ashley'],
'short_answers/row_starts': [0, 1],
'yes_no_answers': [-1, -1],
}])
dataset = preprocessors.natural_questions_open(og_dataset,
max_tokens=3)
test_utils.assert_dataset(dataset, [{
'inputs': 'nq question: is the answer to this question yes',
'targets': 'not sure',
'answers': ['not sure'],
}, {
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'Mary-Kate',
'answers': ['Mary-Kate', 'Ashley'],
}])
dataset = preprocessors.natural_questions_open(og_dataset,
max_tokens=1,
sample_answer=True)
test_utils.assert_dataset(dataset, [{
'inputs': 'nq question: what are the names of the olsen twins',
'targets': 'Ashley',
'answers': ['Ashley'],
}])