guided_alignment_cost
- opennmt.utils.guided_alignment_cost(attention_probs, gold_alignment, sequence_length=None, cost_type='ce', weight=1)[source]
Computes the guided alignment cost.
- Parameters
attention_probs – The attention probabilities, a float
tf.Tensor
of shape \([B, T_t, T_s]\).gold_alignment – The true alignment matrix, a float
tf.Tensor
of shape \([B, T_t, T_s]\).sequence_length – The length of each sequence.
cost_type – The type of the cost function to compute (can be: ce, mse).
weight – The weight applied to the cost.
- Returns
The guided alignment cost.
- Raises
ValueError – if
cost_type
is invalid.