def test_beam_decoders(self):
'''
Test on random data that custom decoder outputs the same transcript
as standard TF beam search decoder
'''
seq = np.random.uniform(size=self.seq.shape).astype(np.float32)
logits = tf.constant(seq)
seq_len = tf.constant([self.seq.shape[0]])
beam_search_decoded = tf.nn.ctc_beam_search_decoder(logits, seq_len,
beam_width=self.beam_width,
top_paths=1,
merge_repeated=False)
with tf.Session() as sess:
res_beam = sess.run(beam_search_decoded)
decoded_beam, prob_beam = res_beam
prob1 = prob_beam[0][0]
decoded_text1 = ''.join([self.vocab[c] for c in decoded_beam[0].values])
res = ctc_beam_search_decoder(softmax(seq.squeeze()), self.vocab[:-1],
beam_size=self.beam_width)
prob2, decoded_text2 = res[0]
if tf.__version__ >= '1.11':
# works for newer versions only (with CTC decoder fix)
self.assertTrue( abs(prob1 - prob2) < self.tol )
self.assertTrue( prob2 < 0 )
self.assertTrue( decoded_text1 == decoded_text2 )
def perform_beam_search(self, probs: np.ndarray, lm: bool = False):
decoded = ctc_beam_search_decoder(
probs_seq=probs,
vocabulary=self.text_featurizer.vocab_array,
beam_size=self.text_featurizer.decoder_config["beam_width"],
ext_scoring_func=self.text_featurizer.scorer if lm else None)
decoded = decoded[0][-1]
return tf.convert_to_tensor(decoded, dtype=tf.string)
def predict_lm(self, inputs, lm, beam_size: int = 100, **kwargs):
"""
Transcribe inputs using Beam Search + LM, will return list of strings.
This method will not able to utilise batch decoding, instead will do loop to decode for each elements.
Parameters
----------
input: List[np.array]
List[np.array] or List[malaya_speech.model.frame.Frame].
lm: ctc_decoders.Scorer
Returned from `malaya_speech.stt.language_model()`.
beam_size: int, optional (default=100)
beam size for beam decoder.
Returns
-------
result: List[str]
"""
try:
from ctc_decoders import ctc_beam_search_decoder
except:
raise ModuleNotFoundError(
'ctc_decoders not installed. Please install it by `pip install ctc-decoders` and try again.'
)
inputs = [
input.array if isinstance(input, Frame) else input
for input in inputs
]
padded, lens = sequence_1d(inputs, return_len=True)
logits, seq_lens = self._sess.run(
[self._softmax, self._seq_lens],
feed_dict={
self._X: padded,
self._X_len: lens
},
)
logits = np.transpose(logits, axes=(1, 0, 2))
results = []
for i in range(len(logits)):
d = ctc_beam_search_decoder(
logits[i][:seq_lens[i]],
self._vocab,
beam_size,
ext_scoring_func=lm,
**kwargs,
)
results.append(d[0][1])
return results
def _perform_beam_search(
self,
probs: np.ndarray,
lm: bool = False,
):
from ctc_decoders import ctc_beam_search_decoder
decoded = ctc_beam_search_decoder(
probs_seq=probs,
vocabulary=self.text_featurizer.non_blank_tokens,
beam_size=self.text_featurizer.decoder_config.beam_width,
ext_scoring_func=self.text_featurizer.scorer if lm else None,
)
decoded = decoded[0][-1]
return tf.convert_to_tensor(decoded, dtype=tf.string)
def test_decoders(self):
'''
Test all CTC decoders on a sample transcript ('ten seconds').
Standard TF decoders should output 'then seconds'.
Custom CTC decoder with LM rescoring should yield 'ten seconds'.
'''
logits = tf.constant(self.seq)
seq_len = tf.constant([self.seq.shape[0]])
greedy_decoded = tf.nn.ctc_greedy_decoder(logits, seq_len,
merge_repeated=True)
beam_search_decoded = tf.nn.ctc_beam_search_decoder(logits, seq_len,
beam_width=self.beam_width,
top_paths=1,
merge_repeated=False)
with tf.Session() as sess:
res_greedy, res_beam = sess.run([greedy_decoded,
beam_search_decoded])
decoded_greedy, prob_greedy = res_greedy
decoded_text = ''.join([self.vocab[c] for c in decoded_greedy[0].values])
self.assertTrue( abs(7079.117 + prob_greedy[0][0]) < self.tol )
self.assertTrue( decoded_text == 'then seconds' )
decoded_beam, prob_beam = res_beam
decoded_text = ''.join([self.vocab[c] for c in decoded_beam[0].values])
if tf.__version__ >= '1.11':
# works for newer versions only (with CTC decoder fix)
self.assertTrue( abs(1.1842 + prob_beam[0][0]) < self.tol )
self.assertTrue( decoded_text == 'then seconds' )
scorer = Scorer(alpha=2.0, beta=0.5,
model_path='ctc_decoder_with_lm/ctc-test-lm.binary',
vocabulary=self.vocab[:-1])
res = ctc_beam_search_decoder(softmax(self.seq.squeeze()), self.vocab[:-1],
beam_size=self.beam_width,
ext_scoring_func=scorer)
res_prob, decoded_text = res[0]
self.assertTrue( abs(4.0845 + res_prob) < self.tol )
self.assertTrue( decoded_text == self.label )
0.04139363,
],
[
0.15882358,
0.1235788,
0.23376776,
0.20510435,
0.00279306,
0.05294827,
0.22298418,
],
]
greedy_result = ["ac'bdc", "b'da"]
beam_search_result = ['acdc', "b'a"]
ctc_greedy_decoder(np.array(probs_seq1), vocab_list) == greedy_result[0]
ctc_greedy_decoder(np.array(probs_seq2), vocab_list) == greedy_result[1]
ctc_beam_search_decoder(
probs_seq=np.array(probs_seq1),
beam_size=beam_size,
vocabulary=vocab_list,
)
ctc_beam_search_decoder(
probs_seq=np.array(probs_seq2),
beam_size=beam_size,
vocabulary=vocab_list,
)
