ParaDime SpecificationsΒΆ

In addition to the object-oriented way of specifying ParaDime routines, ParaDime also defines a grammar for YAML specifications. This grammar is explained in detail in our arXiv preprint.

A specification includes the top-level fields relations, losses, training phases, and derived data. Of these two, losses and training phases are required. The following is a simple example of such a YAML specification, which specifies a parametric version of metric MDS:

  - name: dists hd
    level: global
    type: pairwise
      metric: euclidean
  - name: dists ld
    level: batch
    type: pairwise
      metric: euclidean
  - name: mds
    type: relation
    func: mse
      - dists hd
      - dists ld
training phases:
    - loss:
      components: mds

ParaDime routines can be constructed from such specifications using the from_spec() class method:

dr = paradime.dr.ParametricDR.from_spec(
    <file or dict>,

This method accepts either a string with the name of a YAML file or a dictionary (e.g., an already parsed specification). Along with the specification you also need to pass a model. Data is added later, either with the train() call or using add_data().

The full schema of required fields and allowed values in ParaDime specifications is defined in the parsing module, using the Cerberus syntax for validation schemata.