def testWordErrorRateMetric(self):
ref = np.asarray([
# a b c
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
])
hyp = np.asarray([
[97, 34, 98, 34, 99], # a b c
[97, 34, 98, 0, 0], # a b
[97, 34, 98, 34, 100], # a b d
[0, 0, 0, 0, 0] # empty
])
labels = np.reshape(ref, ref.shape + (1, 1))
predictions = np.zeros((len(ref), np.max([len(s) for s in hyp]), 1, 1, 256))
for i, sample in enumerate(hyp):
for j, idx in enumerate(sample):
predictions[i, j, 0, 0, idx] = 1
with self.test_session() as session:
actual_wer, unused_actual_ref_len = session.run(
metrics.word_error_rate(predictions, labels))
expected_wer = 0.417
places = 3
self.assertAlmostEqual(round(actual_wer, places), expected_wer, places)
def testWordErrorRateMetric(self):
# Ensure availability of the WER metric function in the dictionary.
assert metrics.Metrics.WORD_ERROR_RATE in metrics.METRICS_FNS
# Test if WER is computed correctly.
ref = np.asarray([
# a b c
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
[97, 34, 98, 34, 99],
])
hyp = np.asarray([
[97, 34, 98, 34, 99], # a b c
[97, 34, 98, 0, 0], # a b
[97, 34, 98, 34, 100], # a b d
[0, 0, 0, 0, 0] # empty
])
labels = np.reshape(ref, ref.shape + (1, 1))
predictions = np.zeros(
(len(ref), np.max([len(s) for s in hyp]), 1, 1, 256))
for i, sample in enumerate(hyp):
for j, idx in enumerate(sample):
predictions[i, j, 0, 0, idx] = 1
with self.test_session() as session:
actual_wer, unused_actual_ref_len = session.run(
metrics.word_error_rate(predictions, labels))
expected_wer = 0.417
places = 3
self.assertAlmostEqual(round(actual_wer, places), expected_wer, places)
