Developer Guide and Reference

  • 2021.3
  • 06/28/2021
  • Public Content
Contents

Algorithms

All Algorithms classes are derived from the base class
AlgorithmIface
. It provides interfaces for computations covering a variety of usage scenarios. Basic methods that you typically call are
compute()
and
finalizeCompute()
. In a very generic form algorithms accept one or several numeric tables or models as an input and return one or several numeric tables and models as an output. Algorithms may also require algorithm-specific parameters that you can modify by accessing the
parameter
field of the algorithm. Because most of algorithm parameters are preset with default values, you can often omit initialization of the parameter.

Algorithm Input

An algorithm can accept one or several numeric tables or models as an input. In computation modes that permit multiple calls to the
compute()
method, ensure that the structure of the input data, that is, the number of features, their order, and type, is the same for all the calls. The following methods are available to provide input to an algorithm:
input.set(Input ID, InputData)
Use to set a pointer to the input argument with the
Input ID
identifier. This method overwrites the previous input pointer stored in the algorithm.
input.add(Input ID, InputData)
Use in the distributed computation mode to add the pointers with the
Input ID
identifier. Unlike the
input.set()
method,
input.add()
does not overwrite the previously set input pointers, but stores all the input pointers until the
compute()
method is called.
input.get(Input ID)
Use to get a reference to the pointer to the input data with the
Input ID
identifier.
For the input that each specific algorithm accepts, refer to the description of this algorithm.

Algorithm Output

Output of an algorithm can be one or several models or numeric tables. To retrieve the results of the algorithm computation, call the
getResult()
method. To access specific results, use the
get(Result ID)
method with the appropriate
Result ID
identifier. In the distributed processing mode, to get access to partial results of the algorithm computation, call the
getPartialResult()
method on each computation node. For a full list of algorithm computation results available, refer to the description of an appropriate algorithm.
By default, all algorithms allocate required memory to store partial and final results. Follow these steps to provide user allocated memory for partial or final results to the algorithm:
  1. Create an object of an appropriate class for the results. For the classes supported, refer to the description of a specific algorithm.
  2. Provide a pointer to that object to the algorithm by calling the
    setPartialResult()
    or
    setResult()
    method as appropriate.
  3. Call the
    compute()
    method. After the call, the object created contains partial or final results.

Algorithm Parameters

Most of algorithms in oneDAL have a set of algorithm-specific parameters. Because most of the parameters are optional and preset with default values, you can often omit parameter modification. Provide required parameters to the algorithm using the constructor during algorithm initialization. If you need to change the parameters, you can do it by accessing the public field parameter of the algorithm. Some algorithms have an initialization procedure that sets or precomputes specific parameters needed to compute the algorithm. You can use the InitializationProcedureIface interface class to implement your own initialization procedure when the default implementation does not meet your specific needs.
Each algorithm also has generic parameters, such as the floating-point type, computation method, and computation step for the distributed processing mode.
  • In C++, these parameters are defined as template parameters, and in most cases they are preset with default values. You can change the template parameters while declaring the algorithm.
  • In Java, the generic parameters have no default values, and you need to define them in the constructor during algorithm initialization.
For a list of algorithm parameters, refer to the description of an appropriate algorithm.

Product and Performance Information

1

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex.