ctc_loss_op

→TensorImplPtr

int ctc_loss_op(const int & log_probs, const int & targets, const int & input_lengths, const int & target_lengths, int blank, bool zero_infinity)

source edit

Python wrappers

pylucid.nn.functional.ctc_lossfunction

Connectionist Temporal Classification (CTC) loss.

Computes the negative log-probability of all valid alignments between an input log-probability sequence and a flat target label sequence using the forward algorithm (Graves et al., 2006). The blank symbol allows the model to emit "no character" at any time step, and the loss is summed over all valid label expansions:

\mathcal{L} = -\log \sum_{\pi \in \mathcal{B}^{-1}(y)} \prod_{t=1}^{T} p_{\pi_t}^{(t)}

where $\mathcal{B}$ is the standard CTC collapse operation. No reduction is applied here — the caller is expected to apply mean / sum in Python.

References

Graves et al., "Connectionist Temporal Classification: Labelling Unsegmented Sequence Data with Recurrent Neural Networks" (ICML 2006).

Parameters

log_probsTensorImplPtr

Log-probability tensor of shape (T, N, C).

targetsTensorImplPtr

Flat int32 label tensor of shape (N * S,) where

S

is the maximum target length.

input_lengthsTensorImplPtr

int32 vector of shape (N,) giving the valid time-step count per sample.

target_lengthsTensorImplPtr

int32 vector of shape (N,) giving the true label length per sample.

blankint

Index of the CTC blank class.

zero_infinitybool

When true, infinite loss values (alignments with zero probability) are clamped to zero — both forward and backward.

Returns

TensorImplPtr

Per-sample loss tensor of shape (N,).