# description: UMLS CUI embeddings

## Installer

#

cui2vec_rm_update:

  class_name: zensols.install.RemoveUpdate

  dry_run: false

  paths:

    - '{target}/../__MACOSX'

# The below URL might need to be updated from a new URL given by navigating to

# the embeddings link below

#

# paper: https://arxiv.org/pdf/1804.01486.pdf

# embeddings: https://figshare.com/s/00d69861786cd0156d81

cui2vec_resource:

  class_name: zensols.install.Resource

  url: 'https://figshare.com/ndownloader/files/10959626?private_link=00d69861786cd0156d81'

  name: 'cui2vec.zip'

  is_compressed: true

  check_path: 'cui2vec_pretrained.csv'

  updates: 'instance: list: cui2vec_rm_update'

  # if deleted, it will re-download on each access of the weights

  clean_up: false

cui2vec_installer:

  class_name: zensols.install.Installer

  package_resource: zensols.mednlp

  resources: 'instance: list: cui2vec_resource'

## Embedding

#

cui2vec_500_embedding:

  class_name: 'zensols.mednlp.cui2vec.Cui2VecEmbedModel'

  installer: 'instance: cui2vec_installer'

  resource: 'instance: cui2vec_resource'

  lowercase: false

# a vectorizer that turns tokens (TokensContainer) in to indexes given to the

# embedding layer

cui2vec_500_feature_vectorizer:

  class_name: zensols.deepnlp.vectorize.WordVectorEmbeddingFeatureVectorizer

  # the feature id is used to connect instance data with the vectorizer used to

  # generate the feature at run time

  feature_id: 'wvcui2vec500'

  # encode at the (feature) document level

  fold_method: 'concat_tokens'

  # the word embedding model

  embed_model: 'instance: cui2vec_500_embedding'

  # do not serialize (pickle) the decoded output to do the work up front

  encode_transformed: '${mednlp_default:cui2vec_encode_transformed}'

  # the FeatureToken attribute used to index the embedding vectors

  token_feature_id: 'cui_'

# a torch.nn.Module implementation that uses the an embedding model

cui2vec_500_embedding_layer:

  class_name: zensols.deepnlp.layer.WordVectorEmbeddingLayer

  embed_model: 'instance: cui2vec_500_embedding'

  feature_vectorizer_manager: 'instance: cui2vec_feature_vectorizer_manager'

  # freeze the embedding to train faster

  trainable: '${mednlp_default:cui2vec_trainable}'

## Vectorizer

#

cui2vec_feature_vectorizer_manager:

  class_name: zensols.deepnlp.vectorize.FeatureDocumentVectorizerManager

  torch_config: 'instance: torch_config'

  doc_parser: 'instance: ${mednlp_default:doc_parser}'

  # do not truncate tokens

  token_length: -1

  configured_vectorizers: 'list: cui2vec_500_feature_vectorizer'

# cui embeddings

cui2vec_feature_batch_mappings:

  manager_mappings:

    - vectorizer_manager_name: cui2vec_feature_vectorizer_manager

      fields:

        - attr: cui2vec_500_embedding

          feature_id: ${cui2vec_500_feature_vectorizer:feature_id}

          is_agg: true

          attr_access: doc