# 🤖 PIPELINE DEFINITION
nlp:
  "@core": pipeline
  lang: eds

  components:
    normalizer:
      '@factory': eds.normalizer

    sentencizer:
      '@factory': eds.sentences

    ner:
      '@factory': eds.ner_crf
      mode: "joint"
      target_span_getter: "gold_spans"
      # Set spans as both to ents and in separate `ent.label` groups
      span_setter: [ "ents", "*" ]
      infer_span_setter: true

      embedding:
        '@factory': eds.text_cnn
        kernel_sizes: [ 3 ]

        embedding:
          '@factory': eds.transformer
          model: hf-internal-testing/tiny-bert
          window: 128
          stride: 96
          new_tokens: [ [ "(?:\\n\\s*)*\\n", "⏎" ] ]

    qualifier:
      '@factory': eds.span_classifier
      attributes: { "_.negation": [ "sosy" ], "_.unit": [ "measure" ] }
      span_getter: ["ents", "gold_spans"]

      embedding:
        '@factory': eds.span_pooler

        embedding: # ${ nlp.components.ner.embedding }
          '@factory': eds.text_cnn
          kernel_sizes: [ 3 ]

          embedding: ${ nlp.components.ner.embedding.embedding }

# 📈 SCORERS
scorer:
  speed: true
  qual:
    '@metrics': eds.span_attributes
    span_getter: ${nlp.components.qualifier.span_getter}
    qualifiers: ${nlp.components.qualifier.attributes}
  ner:
    '@metrics': eds.ner_exact
    span_getter: ${nlp.components.ner.target_span_getter}

# 🎛️ OPTIMIZER
optimizer:
  "@core": optimizer
  optim: AdamW
  module: ${ nlp }
  groups:
    "^transformer": false
    ".*":
      lr: 1e-3

# 📚 DATA
train_data:
  - data:
      '@readers': standoff
      path: ./dataset/
      converter:
        - '@factory': eds.standoff_dict2doc
          span_setter : 'gold_spans'
          span_attributes : ['sosy', 'unit', 'negation']
          bool_attributes : ['negation']  # default standoff to doc converter
        - '@factory': eds.sentences
          nlp: ${nlp}
        - '@factory': eds.split
          nlp: null
          max_length: 2000
          regex: '\n\n+'
    shuffle: dataset
    batch_size: 8 docs
    pipe_names: [ "ner" ]
  - data:
      '@readers': standoff
      path: ./dataset/
      converter:
        - '@factory': eds.standoff_dict2doc
          span_setter : 'gold_spans'
          span_attributes : ['sosy', 'unit', 'negation']
          bool_attributes : ['negation']  # default standoff to doc converter
    shuffle: dataset
    batch_size: 16 spans
    pipe_names: [ "qualifier" ]

val_data:
  '@readers': standoff
  path: ./dataset/
  converter:
    - '@factory': eds.standoff_dict2doc
      span_setter : 'gold_spans'
      span_attributes : ['sosy', 'unit', 'negation']
      bool_attributes : ['negation']  # default standoff to doc converter

# 🚀 TRAIN SCRIPT OPTIONS
train:
  nlp: ${ nlp }
  train_data: ${ train_data }
  val_data: ${ val_data }
  max_steps: 40
  validation_interval: 10
  max_grad_norm: 1.0
  scorer: ${ scorer }
  num_workers: 0
  optimizer: ${ optimizer }
  grad_dev_policy: "clip_threshold"