Skip to content

Latest commit

 

History

History
28 lines (27 loc) · 2.36 KB

v15.0.0-beta3.md

File metadata and controls

28 lines (27 loc) · 2.36 KB

Version 15.0.0-beta3 Changelog

  • Input configurators such as DenseFeatureVectorInput now provide a standardized, convenient way to specify inputs.

    • Correspondingly, the API for many transformers and neural network layers have been greatly simplified and standardized; e.g. new NNDenseLayer().withInputFromDensifiedVectors(...) is now new NNDenseLayer().withInput().fromVectors(...).
    • The AbstractInput and vector configurators can be found here.

  • MermaidVisualization in the visualization-mermaid module now provides visualizations of Dagli graphs as Mermaid markup. For example, here's a visualization of the neural network bag-of-ngrams example with the values for a single training example displayed along the edges: A visualized DAG

  • New visualization API: AbstractVisualization in the visualization module can be extended to create new visualizers in addition to the existing Mermaid and ASCII (visualization-ascii) alternatives.

  • Dagli's Kryo 4.* dependencies have been updated to Kryo 5.

  • All Dagli nodes (and neural network layers) now implement the Named interface and can be given arbitrary names via their withName(...) methods (this is primarily useful for distinguishing nodes when visualizing the graph).

  • LinkedStack (a stack implemented as a singly-linked list) is now immutable and replaces LinkedNode.

  • Three transformers have been renamed for consistency with the naming of other transformers:

    • VectorAsDoubleArray is renamed to DoubleArrayFromVector
    • SparseVectorizedDistribution is renamed to SparseVectorFromDistribution
    • DenseVectorizedDistribution is renamed to DenseVectorFromDistribution

  • LiblinearClassification now uses DenseVector features, allowing the work of "densifying" the feature vector to be done discretely in the DAG rather than in the model transformer itself (all other models that require dense features already do this).

  • Added xgboost-core module which includes the XGBoost transformers without the XGBoost4J dependency, allowing the client to use their own XGBoost4J dependency (for Windows or multi-threaded OSX support). The xgboost module continues to include the official XGBoost4J release (currently supporting Linux and single-threaded OSX).