# Objectosphere classifier (Dhamija et al., NeurIPS 2018).
#
# Extends the Entropic Open-Set loss with a logit-norm objective:
#   - Known samples: CE + hinge pushing ||logits|| >= xi
#   - Background samples: entropy maximisation + magnitude suppression (zeta * ||logits||^2)
#
# At inference time, flag unknown inputs with a norm threshold:
#   is_unknown = logit_norm < threshold
# Choose the threshold from the validation set (e.g. 5th percentile of
# known-class norms).

model_type: ecd

input_features:
  - name: feature_1
    type: number
  - name: feature_2
    type: number

output_features:
  - name: label
    type: category
    loss:
      type: objectosphere
      background_class: 4   # adjust to match your vocabulary index
      xi: 10.0              # minimum logit norm for known-class samples
      zeta: 0.1             # weight for unknown-class magnitude suppression

trainer:
  epochs: 50
  learning_rate: 0.001