deeplearning4j / arbiter

A tool for hyperparameter optimization of machine learning models

Github

Arbiter

A tool dedicated to tuning (hyperparameter optimization) of machine learning models. Part of the DL4J Suite of Machine Learning / Deep Learning tools for the enterprise.

Modules

Arbiter contains the following modules:

  • arbiter-core: Defines the API and core functionality, and also contains functionality for the Arbiter UI
  • arbiter-deeplearning4j: For hyperparameter optimization of DL4J models (MultiLayerNetwork and ComputationGraph networks)

Hyperparameter Optimization Functionality

The open-source version of Arbiter currently defines two methods of hyperparameter optimization:

  • Grid search
  • Random search

For optimization of complex models such as neural networks (those with more than a few hyperparameters), random search is superior to grid search, though Bayesian hyperparameter optimization schemes For a comparison of random and grid search methods, see Random Search for Hyper-parameter Optimization (Bergstra and Bengio, 2012).

Core Concepts and Classes in Arbiter for Hyperparameter Optimization

In order to conduct hyperparameter optimization in Arbiter, it is necessary for the user to understand and define the following:

  • Parameter Space: A ParameterSpace<P> specifies the type and allowable values of hyperparameters for a model configuration of type P. For example, P could be a MultiLayerConfiguration for DL4J
  • Candidate Generator: A CandidateGenerator<C> is used to generate candidate models configurations of some type C. The following implementations are defined in arbiter-core:
    • RandomSearchCandidateGenerator
    • GridSearchCandidateGenerator
  • Score Function: A ScoreFunction<M,D> is used to score a model of type M given data of type D. For example, in DL4J a score function might be used to calculate the classification accuracy from a DataSetIterator
    • A key concept here is that they score is a single numerical (double precision) value that we either want to minimize or maximize - this is the goal of hyperparameter optimization
  • Termination Conditions: One or more TerminationCondition instances must be provided to the OptimizationConfiguration. TerminationCondition instances are used to control when hyperparameter optimization should be stopped. Some built-in termination conditions:
    • MaxCandidatesCondition: Terminate if more than the specified number of candidate hyperparameter configurations have been executed
    • MaxTimeCondition: Terminate after a specified amount of time has elapsed since starting the optimization
  • Result Saver: The ResultSaver<C,M,A> interface is used to specify how the results of each hyperparameter optimization run should be saved. For example, whether saving should be done to local disk, to a database, to HDFS, or simply stored in memory.
    • Note that ResultSaver.saveModel method returns a ResultReference object, which provides a mechanism for re-loading both the model and score from wherever it may be saved.
  • Optimization Configuration: An OptimizationConfiguration<C,M,D,A> ties together the above configuration options in a fluent (builder) pattern.
  • Candidate Executor: The CandidateExecutor<C,M,D,A> interface provides a layer of abstraction between the configuration and execution of each instance of learning. Currently, the only option is the LocalCandidateExecutor, which is used to execute learning on a single machine (in the current JVM). In principle, other execution methods (for example, on Spark or cloud computing machines) could be implemented.
  • Optimization Runner: The OptimizationRunner uses an OptimizationConfiguration and a CandidateExecutor to actually run the optimization, and save the results.

Optimization of DeepLearning4J Models

(This section: forthcoming)