def test_empirical_baseline_construction(self):
"""Tests that EmpiricalBaseline construction is correct."""
domain = domains.FixedLengthDiscreteDomain(vocab=domains.Vocabulary(
tokens=range(3), include_bos=True),
length=2)
train_data = np.array([[0, 1], [1, 0]])
train_ds = tf.data.Dataset.from_tensor_slices((train_data, ))
eb = evaluation.EmpiricalBaseline(domain, train_ds, alpha=0)
self.assertAllEqual(eb._empirical_dist, [0.5, 0.5, 0])
def test_empirical_baseline_evaluation(self):
"""Tests that EmpiricalBaseline evaluation is correct."""
domain = domains.FixedLengthDiscreteDomain(vocab=domains.Vocabulary(
tokens=range(2), include_bos=True),
length=1)
train_data = np.array([[0, 1], [1, 0]])
train_ds = tf.data.Dataset.from_tensor_slices((train_data, ))
eval_data = np.array([[0, 1], [1, 0]])
eval_ds = tf.data.Dataset.from_tensor_slices((eval_data, ))
eb = evaluation.EmpiricalBaseline(domain, train_ds)
metrics = evaluation.evaluate(eb, eval_ds)
self.assertAllEqual(np.asarray(metrics['accuracy']), 0.5)
self.assertAllClose(np.asarray(metrics['perplexity']), 2)
self.assertAllClose(np.asarray(metrics['loss']), 0.69, atol=0.1)