Can this work as a replacement for parlance/ctcdecode? about fast-ctc-decode HOT 3 CLOSED

Yes, if you are not using an external language model then fast-ctc-decode can be used in place of the parlance beam search implementation. Ease of installation is a priority for this project.

from fast-ctc-decode.

What do the various arguments map to?

from fast-ctc-decode.

The first argument to the beam search is the probability output from your network where the first dimension is the width and the second is the number of features/labels (WxC).

>>> from fast_ctc_decode import beam_search
>>> posteriors = np.random.rand(100, 5).astype(np.float32)
>>>
>>> seq, path = beam_search(posteriors, alphabet="NACGT", beam_size=5)

Full documentation can be read either here or by using help(beam_search) from the Python interpreter, also the benchmark script compares fast_ctc_decode against parlance. The key difference is parlance take a batch of network outputs (NxWxC).

from fast-ctc-decode.