decode.py

import tensorflow.keras.backend as K

def _decode(y_pred, input_length, greedy=True, beam_width=100, top_paths=1):

    decoded = K.ctc_decode(y_pred=y_pred, input_length=input_length,
                           greedy=greedy, beam_width=beam_width, top_paths=top_paths)

    paths = [path.numpy() for path in decoded[0]]
    logprobs  = decoded[1].numpy()

    return (paths, logprobs)

def decode(y_pred, input_length, greedy=True, beam_width=100, top_paths=1, **kwargs):
    language_model = kwargs.get('language_model', None)

    paths, logprobs = _decode(y_pred=y_pred, input_length=input_length,
                              greedy=greedy, beam_width=beam_width, top_paths=top_paths)
    if language_model is not None:
        # TODO: compute using language model
        raise NotImplementedError("Language model search is not implemented yet")
    else:
        # simply output highest probability sequence
        # paths has been sorted from the start
        result = paths[0]
    return result

class Decoder(object):
    def __init__(self, greedy=True, beam_width=100, top_paths=1, **kwargs):
        self.greedy         = greedy
        self.beam_width     = beam_width
        self.top_paths      = top_paths
        self.language_model = kwargs.get('language_model', None)
        self.postprocessors = kwargs.get('postprocessors', [])

    def decode(self, y_pred, input_length):
        decoded = decode(y_pred, input_length, greedy=self.greedy, beam_width=self.beam_width,
                         top_paths=self.top_paths, language_model=self.language_model)
        preprocessed = []
        for output in decoded:
            out = output
            for postprocessor in self.postprocessors:
                out = postprocessor(out)
            preprocessed.append(out)

        return preprocessed