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

    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.


The following model types are configured to share embeddings by default:

  • TransformerBaseSharedEmbeddings

  • TransformerBigSharedEmbeddings

For custom models, see the values in EmbeddingsSharingLevel that can be passed to the SequenceToSequence constructor.