search

githubharald / CTCDecoder

Connectionist Temporal Classification (CTC) decoding algorithms: best path, beam search, lexicon search, prefix search, and token passing. Implemented in Python.
https://towardsdatascience.com/3797e43a86c
MIT License
817 stars 182 forks source link

question about prefix beam search #22

Closed HaoDreamlong closed 3 years ago

HaoDreamlong commented 3 years ago

In PrefixSearch.py , in the func ctcPrefixSearch ,what is prob_ext?

githubharald commented 3 years ago

Quite some time ago that I implemented this ... but I guess it is the "cumulative probability" p(y...|x), see PhD thesis from Graves, page 64.

HaoDreamlong commented 3 years ago

Yes I found it eventually. Your notes are very vague but still lead me to that Dissertation. But thank you anyway But I found that this method is difficult to get the answer for a random input.e.g. np.random.seed(11) input_data = np.random.random([20, 5])

I'm trying to find out why this problem happened.

HaoDreamlong commented 3 years ago

Of course Softmax before input

githubharald commented 3 years ago

But I found that this method is difficult to get the answer for a random input.e.g.

what do you mean with "difficult to get the answer"? Is it returning the wrong result?

HaoDreamlong commented 3 years ago

endless loop

githubharald commented 3 years ago

I can remeber that the decoder was very slow ... so is it really endless (in this case the code contains a bug), or is it just taking quite long?

githubharald commented 3 years ago

closing because of inactivity