Raw logit shape

[JRMeyer]

>>> Abhinav
[April 18, 2019, 12:07pm]

Before the logits are sent to ctc_beam_search_decoder on
line
, I do print(np.shape(logits)). I get (129, 29) or (137, 29) etc. I
know 129 or 137 is the num_strides. I assume 29 should be for the 29
possible characters. But in my alphabet.txt there are 26 alphabets,
space and apostrophe a total of 28. So what does 29 mean?

[This is an archived TTS discussion thread from discourse.mozilla.org/t/raw-logit-shape]

[JRMeyer]

>>> Abhinav
[April 18, 2019, 12:26pm]

Answering myself, I observed the highest probability is never on the
29th index, so assumed its just a fake thingy. So now

logits = np.squeeze(logits)
#print(num_strides)
#print(np.shape(logits))
a = ['a','b','c','d','e','f','g','h','i','j','k','l','m','n','o','p','q','r','s','t','u','v','w','x','y','z',''',' ']
raw = ''
char_idxs = np.argmax(logits, axis=1)
print(char_idxs)
for char_id in char_idxs:
raw += a[char_id-1]
print(raw)

scorer = Scorer(FLAGS.lm_alpha, FLAGS.lm_beta,
FLAGS.lm_binary_path, FLAGS.lm_trie_path,
Config.alphabet)

prints:

why should o nee hh a l l t o onn the ww a y slash
and the decoded output is why should one hall to on the way

[Archived Post]

[JRMeyer]

>>> kdavis
[April 18, 2019, 12:48pm]

It's a 'blank', not a space, required for
CTC.

[Archived Post]

[JRMeyer]

>>> Abhinav
[April 18, 2019, 1:07pm]

Thanks, so the 29 items are: slash
a = [' ','a','b','c','d','e','f','g','h','i','j','k','l','m','n','o','p','q','r','s','t','u','v','w','x','y','z',''','ϵ']

where ϵ is the CTC 'blank'

[Archived Post]

[JRMeyer]

>>> reuben
[April 18, 2019, 1:37pm]

Here's a greedy best path algorithm that handles the blank index and
merges repeated characters (merge_repeated_ defaults to true):
https://github.com/tensorflow/tensorflow/blob/6612da89516247503f03ef76e974b51a434fb52e/tensorflow/core/util/ctc/ctc_decoder.h#L95-L107

[Archived Post]