lookup_embedder.yaml

# Stores explicitly an embedding for each object in a lookup table. See
# https://pytorch.org/docs/stable/nn.html#torch.nn.Embedding
lookup_embedder:
  class_name: LookupEmbedder

  # Dimensionality of the embedding
  dim: 100

  # vector space of embedding: complex or euclidean
  space: euclidean

  # The initializer used to initialzed the embeddings. See
  # https://pytorch.org/docs/stable/nn.init.html for details. Example values
  # include xavier_uniform_, xavier_normal_, uniform_, and normal_.
  initialize: normal_

  # Arguments to the initializer. If this field contains a subkey that matches
  # the value of option lookup_embedder.initialize, pass the options under this
  # subkey to the corresponding pytorch functions. Otherwise, pass all options
  # to the initializer.
  #
  # Examples for specifying args for specific initializers (those are the defaults):
  #   normal_:
  #     mean: 0.0
  #     std: 1.0
  #   uniform_:
  #     a: 0.0
  #     b: 1.0
  #    xavier_normal_:
  #      gain: 1.0
  #    xavier_uniform_:
  #      gain: 1.0
  initialize_args:
    +++: +++

  # Initialize the model with embeddings stored in a packaged model
  pretrain:
      # path to packaged model containing entity/relation embeddings
      # leave empty to not pretrain
      model_filename: "" 
      # ensure that all entities/relations  can be initialized  with embeddings of 
      #  the packaged model
      #  if false initialize other embeddings normally
      ensure_all: False
 
  # Dropout used for the embeddings.
  dropout: 0.

  # Whether embeddings should be normalized. Normalization takes place before
  # each batch is processed.
  normalize:
    # l_p norm to use. Negative numbers mean do not normalize.
    p: -1.                    # common choices: 1., 2.

  # Whether and how embeddings should be regularized. Possible values are '' (do
  # not regularize) or lp (defaults to p=2; else set regularize_args.p) or 
  # n3 (which is l3 applied to the element-wise absolute value of complex embeddings)
  regularize: 'lp'

  # Weight used for regularization. Interpreted w.r.t. to the empirical risk.
  regularize_weight: 0.0

  # Further arguments for regulariztaion
  regularize_args:
    # If true, penalty for the embedding of an object is weighted by the
    # object's *relative* frequency in the training data. Note that this
    # generally reduces the magnitude of the penalty (for random embeddings: by
    # roughly a factor corresponding to the number of objects in the lookup
    # table) compared to unweighted regularization.
    weighted: False
    
    # Which norm to use for lp regularization.
    p: 2
    
    # Other options for the regularizer.
    +++: +++

  # Use a sparse tensor for the gradient. See
  # https://pytorch.org/docs/stable/nn.html#torch.nn.Embedding.
  # Sparse tensors might require special settings, e.g. SparseAdam instead of Adam
  sparse: False

  # Will be deprecated.
  round_dim_to: []