Expand description
Connectionist Temporal Classification (CTC).
CTC (Graves et al. 2006) enables training and decoding of sequence models whose output length differs from, and is not pre-aligned with, the target label sequence. This module provides:
ctc_loss— the forward-backward CTC negative-log-likelihood loss over a blank-augmented label lattice, computed in log-space.ctc_log_occupancy— per-symbol posterior responsibilities (the basis of the CTC gradient).ctc_greedy_decode— best-path (arg-max) decoding with CTC collapse.ctc_prefix_beam_search— prefix-beam-search decoding that sums alignment multiplicities (Hannun et al. 2014).
Re-exports§
pub use ctc_decode::CtcHypothesis;pub use ctc_decode::ctc_greedy_decode;pub use ctc_decode::ctc_prefix_beam_search;pub use ctc_loss::ctc_log_occupancy;pub use ctc_loss::ctc_loss;pub use ctc_loss::ctc_loss_batch;
Modules§
- ctc_
decode - CTC decoding: best-path (greedy) and prefix-beam search (Graves 2006; Hannun 2014).
- ctc_
loss - Connectionist Temporal Classification (CTC) loss (Graves et al. 2006).