syntax = "proto3";

package sarus_data_spec;

message Scalar {
  // A Scalar represents data that does not fulfill the promise of a Dataset.
  // A Dataset promises to have a schema and the possibility to iterate on
  // pyarrow.RecordBatches

  // A Scalar does not ensure this possibility. As a consequence, oprations
  // from standard libraries are allowed (pandas.mean, numpy.std,...) but
  // operations implemented for Datasets by Sarus like computing marginals or
  // fitting a Keras model cannot be performed on a Scalar.

  // Scalars are generated by transforms that explicitly require a specific
  // format (e.g. as_pandas, as_numpy,...) or as byproducts of transforms
  // (model weights, training history,...).

  string uuid = 1; // e.g. RFC 4122 id used to refer to the dataset (content linked?)
  string name = 2;
  string doc = 3;
  Spec spec = 4;
  map<string, string> properties = 5; // Other properties

  // Definitions
  message Spec { // How to obtain the dataset
    oneof spec {
      Transformed transformed = 1;
      Model model = 2;
      PrivacyParameters privacy_params=3;
      RandomSeed random_seed = 4;
      SyntheticModel synthetic_model=5;
    }
  }

  message Transformed {
    string transform = 1; // Transform id
    repeated string arguments = 2; // Dataset or other object ids
    map<string,string> named_arguments = 3;
  }

  message Model {
    bytes arguments = 1;
    bytes named_arguments = 2;
    ModelClass model_class = 3;

    enum ModelClass {
      TF_KERAS = 0;
      SK_SVC = 1;
      SK_ONEHOT = 2;
      SK_PCA = 3;

      // Cluster
      SK_AFFINITY_PROPAGATION = 4;
      SK_AGGLOMERATIVE_CLUSTERING = 5;
      SK_BIRCH = 6;
      SK_DBSCAN = 7;
      SK_FEATURE_AGGLOMERATION = 8;
      SK_KMEANS = 9;
      SK_MINIBATCH_KMEANS = 10;
      SK_MEAN_SHIFT = 11;
      SK_OPTICS = 12;
      SK_SPECTRAL_CLUSTERING = 13;
      SK_SPECTRAL_BICLUSTERING = 14;
      SK_SPECTRAL_COCLUSTERING = 15;

      // ensemble
      SK_ADABOOST_CLASSIFIER = 60;
      SK_ADABOOST_REGRESSOR = 61;
      SK_BAGGING_CLASSIFIER = 62;
      SK_BAGGING_REGRESSOR = 63;
      SK_EXTRA_TREES_REGRESSOR = 64;
      SK_EXTRA_TREES_CLASSIFIER = 65;
      SK_GRADIENT_BOOSTING_CLASSIFIER = 66;
      SK_GRADIENT_BOOSTING_REGRESSOR = 67;
      SK_ISOLATION_FOREST = 68;
      SK_RANDOM_FOREST_CLASSIFIER = 69;
      SK_RANDOM_FOREST_REGRESSOR = 70;
      SK_RANDOM_TREES_EMBEDDING = 71;
      SK_STACKING_CLASSIFIER = 72;
      SK_STACKING_REGRESSOR = 73;
      SK_VOTING_CLASSIFIER = 74;
      SK_VOTING_REGRESSOR = 75;
      SK_HIST_GRADIENT_BOOSTING_REGRESSOR = 76;
      SK_HIST_GRADIENT_BOOSTING_CLASSIFIER = 77;

      // Model selection
      SK_REPEATED_STRATIFIED_KFOLD = 80;

      // XGB
      XGB_CLASSIFIER = 92;

      SK_LABEL_ENCODER = 98;
      SK_KFOLD = 99;
    };
  }

  message PrivacyParameters {
    repeated Point points = 1;

    message Point {
      double epsilon = 1;
      double delta = 2;
    }
  }

  message RandomSeed {
    int32 value=1;
  }
  message SyntheticModel{}
}