Embeddings

Pretrained

Pretrained embeddings can be configured in the data section of the YAML configuration, e.g.:

data:
  source_embedding:
    path: data/glove/glove-100000.txt
    with_header: True
    case_insensitive: True
    trainable: False

  # target_embedding: ...

The format of the embedding file and the options are described in the opennmt.inputters.load_pretrained_embeddings function.

Pretrained embeddings are only loaded when initializing the model weights. If you continue the training with from a checkpoint, this configuration is ignored and the embeddings values are loaded from the checkpoint.

Sharing

SequenceToSequence models take a share_embeddings argument in the constructor that accepts a EmbeddingsSharingLevel value to enable different level of embeddings sharing.

See for example the custom model definition transformer_shared_embedding.py that shares all embeddings, including the output softmax weights, for a Transformer model.