cross_entropy_sequence_loss

opennmt.utils.cross_entropy_sequence_loss(logits, labels, sequence_length=None, label_smoothing=0.0, average_in_time=False, training=None, sequence_weight=None, mask_outliers=False)[source]

Computes the cross entropy loss of sequences.

Parameters

  • logits – The unscaled probabilities with shape \([B, T, V]\).

  • labels – The true labels with shape \([B, T]\).

  • sequence_length – The length of each sequence with shape \([B]\).

  • label_smoothing – The label smoothing value.

  • average_in_time – If True, also average the loss in the time dimension.

  • training – Compute training loss.

  • sequence_weight – The weight of each sequence with shape \([B]\).

  • mask_outliers – Mask large training loss values considered as outliers.

Returns

A tuple (cumulated loss, loss normalizer, token-level normalizer).