axis-numeric / axis_numeric /ctc_frames.py
serety95's picture
Update numeric space
f1ba785 verified
Raw History Blame Contribute Delete
3.77 kB
"""CTC beam over preselected per-timestep candidates."""
from dataclasses import dataclass
import math
@dataclass(frozen=True)
class BeamDecodeResult:
text: str
probability: float
min_character_probability: float
beam_margin: float
character_confidences: tuple[float, ...]
character_alternatives: tuple[tuple[tuple[str, float], ...], ...]
@dataclass(frozen=True)
class CtcFrame:
candidates: tuple[tuple[int, float], ...]
alternatives: tuple[tuple[str, float], ...]
margin: float
def select_candidates(
ranked: list[tuple[int, float]],
count: int,
blank: int,
) -> list[tuple[int, float]]:
selected: list[tuple[int, float]] = []
for index, probability in ranked:
if index == blank and len(selected) >= count:
continue
selected.append((index, probability))
if len(selected) >= count + 1:
break
return selected
def search_ctc_frames(
frames: list[CtcFrame],
vocabulary: str,
beam_width: int = 4,
) -> list[BeamDecodeResult]:
blank = len(vocabulary)
beams: dict[tuple[str, ...], tuple[float, list[float], list]] = {(): (0.0, [], [])}
margin = 1.0
for frame in frames:
margin = frame.margin
next_beams: dict[tuple[str, ...], tuple[float, list[float], list]] = {}
for prefix, (log_prob, confidences, alt_history) in beams.items():
_extend(next_beams, prefix, log_prob, confidences, alt_history, frame, vocabulary, blank)
beams = dict(sorted(next_beams.items(), key=lambda item: item[1][0], reverse=True)[:beam_width])
return _results(beams, margin, beam_width)
def _extend(next_beams, prefix, log_prob, confidences, alt_history, frame: CtcFrame, vocabulary: str, blank: int) -> None:
step_alts = frame.alternatives
for index, probability in frame.candidates:
if probability <= 0:
continue
char = vocabulary[index] if index != blank else ""
score = log_prob + math.log(max(probability, 1e-9))
history = alt_history + [step_alts]
if index == blank:
_store(next_beams, prefix, score, confidences + [probability], history)
continue
if prefix and prefix[-1] == char:
_store(next_beams, prefix, score, confidences, history)
continue
_store(next_beams, prefix + (char,), score, confidences + [probability], history)
def _store(next_beams, key, score, confidences, alt_history) -> None:
stored = next_beams.get(key)
if stored is None or score > stored[0]:
next_beams[key] = (score, confidences, alt_history)
def _results(beams, margin: float, beam_width: int) -> list[BeamDecodeResult]:
results: list[BeamDecodeResult] = []
ranked = sorted(beams.items(), key=lambda item: item[1][0], reverse=True)
for prefix, (log_prob, confidences, alt_history) in ranked:
if not prefix:
continue
text = "".join(prefix)
selected = confidences[-len(text):] if len(confidences) >= len(text) else confidences
if not selected:
selected = [math.exp(log_prob / max(len(text), 1))]
aligned = list(alt_history[-len(text):]) if alt_history else []
while len(aligned) < len(text):
aligned.append(())
results.append(BeamDecodeResult(
text=text,
probability=sum(selected) / len(selected),
min_character_probability=min(selected),
beam_margin=margin,
character_confidences=tuple(selected[: len(text)]),
character_alternatives=tuple(aligned[: len(text)]),
))
return results[:beam_width]