def test_with_probs_1(self):
"""
This test is from https://github.com/PaddlePaddle/DeepSpeech/blob/develop/decoders/tests/test_decoders.py
This is not good test: with small beam should be "acdc", but with large "acb"
"""
labels = ["\'", ' ', 'a', 'b', 'c', 'd', '_']
decoder = CTCDecoder(beam_width=20,
after_logsoftmax=True,
blank_idx=6,
time_major=False,
labels=labels,
wip=0.0)
probs_seq = [[[
0.06390443, 0.21124858, 0.27323887, 0.06870235, 0.0361254,
0.18184413, 0.16493624
],
[
0.03309247, 0.22866108, 0.24390638, 0.09699597,
0.31895462, 0.0094893, 0.06890021
],
[
0.218104, 0.19992557, 0.18245131, 0.08503348,
0.14903535, 0.08424043, 0.08120984
],
[
0.12094152, 0.19162472, 0.01473646, 0.28045061,
0.24246305, 0.05206269, 0.09772094
],
[
0.1333387, 0.00550838, 0.00301669, 0.21745861,
0.20803985, 0.41317442, 0.01946335
],
[
0.16468227, 0.1980699, 0.1906545, 0.18963251,
0.19860937, 0.04377724, 0.01457421
]]]
log_probs = torch.log(torch.FloatTensor(probs_seq))
expected_greedy_result = ["ac'bdc"]
expected_beam_search_result = ["acdc"]
# print("-" * 50)
# best_computed, best_loss_computed = self._get_best_result_brute_force(log_probs, labels, blank_id=6)
# print(best_computed, best_loss_computed, np.exp(-best_loss_computed))
# expected_beam_search_result = [best_computed]
# print("-" * 50)
_, _, greedy_result = decoder.decode_greedy(log_probs)
self.assertListEqual(greedy_result, expected_greedy_result)
_, _, beam_search_result = decoder.decode(log_probs)
self.assertListEqual(beam_search_result, expected_beam_search_result)
def test_with_probs_2(self):
"""
This test is from https://github.com/PaddlePaddle/DeepSpeech/blob/develop/decoders/tests/test_decoders.py
"""
labels = ["\'", ' ', 'a', 'b', 'c', 'd', '_']
decoder = CTCDecoder(beam_width=20,
after_logsoftmax=True,
blank_idx=6,
time_major=False,
labels=labels,
wip=0.0)
probs_seq = [[[
0.08034842, 0.22671944, 0.05799633, 0.36814645, 0.11307441,
0.04468023, 0.10903471
],
[
0.09742457, 0.12959763, 0.09435383, 0.21889204,
0.15113123, 0.10219457, 0.20640612
],
[
0.45033529, 0.09091417, 0.15333208, 0.07939558,
0.08649316, 0.12298585, 0.01654384
],
[
0.02512238, 0.22079203, 0.19664364, 0.11906379,
0.07816055, 0.22538587, 0.13483174
],
[
0.17928453, 0.06065261, 0.41153005, 0.1172041,
0.11880313, 0.07113197, 0.04139363
],
[
0.15882358, 0.1235788, 0.23376776, 0.20510435,
0.00279306, 0.05294827, 0.22298418
]]]
log_probs = torch.log(torch.FloatTensor(probs_seq))
expected_greedy_result = ["b'da"]
expected_beam_search_result = ["b'a"]
# print("-" * 50)
# best_computed, best_loss_computed = self._get_best_result_brute_force(log_probs, labels, blank_id=6)
# print(best_computed, best_loss_computed, np.exp(-best_loss_computed))
# expected_beam_search_result = [best_computed]
# print("-" * 50)
_, _, greedy_result = decoder.decode_greedy(log_probs)
self.assertListEqual(greedy_result, expected_greedy_result)
_, _, beam_search_result = decoder.decode(log_probs)
self.assertListEqual(beam_search_result, expected_beam_search_result)
def test_random_with_lm(self):
blank_idx = 0
np.random.seed(534)
labels = ["_"] + list(string.ascii_lowercase) + [" "]
logits = torch.FloatTensor(np.random.rand(1, 50, len(labels)))
decoder = CTCDecoder(beam_width=100,
blank_idx=blank_idx,
labels=labels,
time_major=False,
lm_path=os.path.join(
os.path.dirname(__file__), "librispeech_data",
"librispeech_3-gram_pruned.3e-7.arpa.gz"),
alpha=1.0)
_, _, beam_search_result = decoder.decode(F.log_softmax(logits, -1))
print(beam_search_result) # random small words from lexicon
self.assertListEqual(beam_search_result, [
"m r i g a y u h h t m ",
]) # TODO: Test that in vocabulary
def test_with_probs_sm(self):
labels = ["_", "a"]
decoder = CTCDecoder(beam_width=20,
blank_idx=0,
time_major=False,
labels=labels)
probs_seq1 = [
[0.7, 0.3],
[0.7, 0.3],
]
log_probs = torch.log(torch.FloatTensor([
probs_seq1,
]))
expected_greedy_result = [""]
expected_beam_search_result = ["a"]
_, _, greedy_result = decoder.decode_greedy(log_probs)
self.assertListEqual(greedy_result, expected_greedy_result)
_, _, beam_search_result = decoder.decode(log_probs)
self.assertListEqual(beam_search_result, expected_beam_search_result)
def test_greedy_simple(self):
logits = torch.FloatTensor([[[1, 2, 4, 3, 10], [2, 1, 3, 8,
1]]]).contiguous()
logits_lengths = torch.LongTensor([2])
blank_idx = 0
labels = ["_", "a", "b", "c", "d"]
correct_targets = torch.LongTensor([[4, 3]])
correct_targets_lengths = torch.LongTensor([2])
correct_sentences = [
"dc",
]
decoder = CTCDecoder(beam_width=1,
blank_idx=blank_idx,
labels=labels,
time_major=False)
decoded_targets, decoded_targets_lengths, decoded_sentences = decoder.decode(
logits, logits_lengths=logits_lengths)
self.assertListEqual(decoded_targets_lengths.numpy().tolist(),
correct_targets_lengths.numpy().tolist())
self.assertListEqual(decoded_targets.numpy().tolist(),
correct_targets.numpy().tolist())
self.assertListEqual(decoded_sentences, correct_sentences)
def test_simple_lm(self):
blank_idx = 0
labels = ["_", "a", "b", "c", "d"]
decoder = CTCDecoder(beam_width=1,
blank_idx=blank_idx,
labels=labels,
time_major=False,
lm_path=os.path.join(
os.path.dirname(__file__), "librispeech_data",
"librispeech_3-gram_pruned.3e-7.arpa.gz"))
decoder._print_scores_for_sentence(
["mother", "washed", "the", "frame"])
print("=" * 50)
decoder._print_scores_for_sentence([
"mother".upper(), "washed".upper(), "the".upper(), "frame".upper()
])
print("=" * 50)
print("=" * 50)
decoder = CTCDecoder(beam_width=1,
blank_idx=blank_idx,
labels=labels,
time_major=False,
lm_path=os.path.join(
os.path.dirname(__file__), "librispeech_data",
"librispeech_3-gram_pruned.3e-7.arpa.gz"),
case_sensitive=True)
decoder._print_scores_for_sentence(
["mother", "washed", "the", "frame"])
print("=" * 50)
decoder._print_scores_for_sentence([
"mother".upper(), "washed".upper(), "the".upper(), "frame".upper()
])