ctc_grad_from_alpha_beta_default

burn::tensor::backend::ops::ctc

Function ctc_grad_from_alpha_beta_default 

pub fn ctc_grad_from_alpha_beta_default<B>(
    log_probs: <B as BackendTypes>::FloatTensorPrimitive,
    targets: <B as BackendTypes>::IntTensorPrimitive,
    input_lengths: <B as BackendTypes>::IntTensorPrimitive,
    grad_loss: <B as BackendTypes>::FloatTensorPrimitive,
    log_alpha_full: <B as BackendTypes>::FloatTensorPrimitive,
    log_beta_full: <B as BackendTypes>::FloatTensorPrimitive,
    nll: <B as BackendTypes>::FloatTensorPrimitive,
    blank: usize,
) -> <B as BackendTypes>::FloatTensorPrimitive
where
    B: Backend,
Expand description

Compose the CTC gradient w.r.t. log_probs from pre-computed alpha, beta, and nll.

The T-iteration alpha and beta recursions are the dominant cost of the backward pass. Backends that fuse those recursions into a single kernel launch can call this helper to reuse the gradient composition.

ยงArguments

  • log_probs - Log-probabilities [T, N, C]
  • targets - Target label indices [N, S]
  • input_lengths - Actual input sequence lengths per batch element [N]
  • grad_loss - Upstream gradient w.r.t. the per-sample loss [N]
  • log_alpha_full - Alpha recursion output [T, N, 2S+1]
  • log_beta_full - Beta recursion output [T, N, 2S+1]
  • nll - Per-sample negative log-likelihood (forward loss) [N]
  • blank - Index of the blank label