kensho-technologies / pyctcdecode

A fast and lightweight python-based CTC beam search decoder for speech recognition.
Apache License 2.0
422 stars 89 forks source link

increase test coverage #7

Closed poneill closed 3 years ago

poneill commented 3 years ago

Add some more tests, including fuzzing of the API.

Fuzzing revealed that _decode_beams_mp_safe was missing a required arg, so patch that through as well, adding DEFAULT_PRUNE_HISTORY to constants.

poneill commented 3 years ago

yeah for sure, doing that now.

On Wed, Jun 30, 2021 at 2:55 PM Georg Kucsko @.***> wrote:

@.**** commented on this pull request.

nice addition tyty, can you just make triple sure around test runtime and flakiness? very traumatized from using hopethesis

In pyctcdecode/decoder.py https://github.com/kensho-technologies/pyctcdecode/pull/7#discussion_r661732899 :

@@ -556,6 +557,7 @@ def decode_beams_batch( beam_width: int = DEFAULT_BEAM_WIDTH, beam_prune_logp: float = DEFAULT_PRUNE_LOGP, token_min_logp: float = DEFAULT_MIN_TOKEN_LOGP,

  • prune_history: bool = DEFAULT_PRUNE_BEAMS,

probably add the same constant in decode_beams?

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/kensho-technologies/pyctcdecode/pull/7#pullrequestreview-696513817, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAETLZZHYCWYDXXWAH4WG2DTVNSA3ANCNFSM47S5BIAA .