def test_best_span(self):
bound = 5
start_pl = tf.placeholder(tf.float32, (None, None))
end_pl = tf.placeholder(tf.float32, (None, None))
best_span, best_val = best_span_from_bounds(start_pl, end_pl, bound)
sess = self.sess
for i in range(0, 20):
rng = np.random.RandomState(i)
l = rng.randint(50, 200)
batch = rng.randint(1, 60)
start = rng.uniform(size=(batch, l))
end = rng.uniform(size=(batch, l))
# exp since the tf version uses logits and the py version use probabilities
expected_span, expected_score = zip(*[
get_best_span_bounded(np.exp(start[i]), np.exp(end[i]), bound)
for i in range(batch)
])
actual_span, actuals_score = sess.run([best_span, best_val], {
start_pl: start,
end_pl: end
})
self.assertTrue(np.all(np.array(expected_span) == actual_span))
self.assertTrue(np.allclose(expected_score, np.exp(actuals_score)))
def get_best_span(self, bound: int):
if bound in self._bound_predictions:
return self._bound_predictions[bound]
else:
pred = best_span_from_bounds(self.start_logits, self.end_logits,
bound)
self._bound_predictions[bound] = pred
return pred
def get_best_span(self, bound: int):
return best_span_from_bounds(self.start_logits, self.end_logits, bound)