Versions Compared

Key

  • This line was added.
  • This line was removed.
  • Formatting was changed.

...

JIRA: https://issues.apache.org/jira/browse/FLINK-22915

Released: 1.1415

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

...

The existing Pipeline API can not be used to compose such an Estimator with other Transformer/Estimator because it is expected to use the same inputs for the same stage in the Pipeline.

4) Compose a directed-acyclic-graph (i.e DAG) Estimator/Transformer into an Estimator/Transformer.

For example, the workflow may involve the join of 2 tables, where each table could be generated by a chain of Estimator/Transformer. The entire workflow is therefore a DAG of Estimator/Transformer.

5) Online learning where a long-running instance Transformer needs to be updated by the latest model data generated by another long-running instance of Estimator.

In this scenario, we need to allow the Estimator to be run on a different machine than the Transformer. So that Estimator could consume sufficient computation resource in a cluster while the Transformer could be deployed on edge devices.

65) Provide APIs to allow Estimator/Transformer to be efficiently saved/loaded even if state (e.g. model data) of Estimator/Transformer is more than 10s of GBs.

...

In addition to addressing the above use-cases, this FLIP also proposes a few more changes to simplify the class hierarchy and improve API usability. The existing Flink ML library has the following usability issues:
7
6) The Model interface (which currently simply extends the Transformer interface without adding any extra logic) does not provide any added value. The added class hierarchy complexity is not justified.

87) fit/transform API requires users to explicitly provide the TableEnvironment, where the TableEnvironment could be retrieved from the Table instance given to the fit/transform.

98) A Pipeline is both a Transformer and an Estimator. The experience of using Pipeline is therefore different from the experience of using Estimator (with the needFit API).

109) There is no API provided by the Estimator/Transformer interface to validate the schema consistency of a Pipeline. Users would have to instantiate Tables (with I/O logics) and run fit/transform to know whether the stages in the Pipeline are compatible with each other.

...

Therefore, this design doc proposes to add the Graph/graphTransformerGraphTransformer/GraphBuilder classes to provide the following capability:

...

This change addresses the use-cases described in the motivation section, e.g. a graph embedding Estimator needs to take 2 tables as inputs.

3) Added Graph, graphTransformer and GraphBuilder.

This change addresses the use-cases described in the motivation section, where we need to compose an Estimator from a DAG of Estimator/Transformer. Note that the Graph/GraphBuilder supports Estimator class whose input schemas are different from its fitted Transformer.

4) Added setStateStreams and getStateStreams to the Transformer setStateStreams and getStateStreams to the Transformer interface.

This change addresses the use-cases described in the motivation section, where a running Transformer needs to ingest the model state streams emitted by a Estimator, which could be running on a different machine.

54) Removed the methods PipelineStage::toJson and PipelineStage::loadJson. Add methods save(...) and load(...) to the Stage interface.


The following changes are relatively minor:

65) Removed TableEnvironment from the parameter list of fit/transform APIs.

This change simplifies the usage of fit/transform APIs.

76) Added pipelineTransformer and let Pipeline implement only the Estimator. Pipeline is no longer a Transformer.

This change makes the experience of using Pipeline consistent with the experience of using Estimator/Transformer, where a class is either an Estimator or a Transformer.

87) Removed Pipeline::appendStage from the Pipeline class.

This change makes the concept of Pipeline consistent with that of Graph/GraphBuilder. Neither Graph nor Pipeline provides the API to construct themselves.

98) Removed the Model interface. And renamed PipelineModel to PipelineTransformer.

This change simplifies the class hierarchy by removing a redundant class. It follows the philosophy of only adding complexity when we have explicit use-case for it.

109) Renamed PipelineStage to Stage and add the PublicEvolving tag to the Stage interface.

...

Code Block
languagejava
/**
 * Base class for a stage in a Pipeline or Graph. The interface is only a concept, and does not have any actual
 * functionality. Its subclasses could be Estimator, Transformer or AlgoOperator. No other classes should inherit this
 * interface directly.
 *
 * <p>Each stage is with parameters, and requires a public empty constructor for restoration.
 *
 * @param <T> The class type of the Stage implementation itself.
 * @see WithParams
 */
@PublicEvolving
interface Stage<T extends Stage<T>> extends WithParams<T>, Serializable {
    /**
     * Saves this stage to the given path.
     */
    void save(String path);

    /**
     * Loads this stage from the given path.
     */
    void load(String path);
}

/**
 * A AlgoOperator is a Stage that takes a list of tables as inputs and produces a list of
 * tables as results. It can be used to encode a generic multi-input multi-output machine learning function.
 *
 * @param <T> The class type of the AlgoOperator implementation itself.
 */
@PublicEvolving
public interface AlgoOperator<T extends AlgoOperator<T>> extends Stage<T> {
    /**
     * Applies the AlgoOperator on the given input tables, and returns the result tables.
     *
     * @param inputs a list of tables
     * @return a list of tables
     */
    Table[] transform(Table... inputs);
}

/**
 * A Transformer is a AlgoOperator with additional support for state streams, which could be set by the Estimator that fitted
 * this Transformer. Unlike AlgoOperator, a Transformer is typically associated with an Estimator.
 *
 * @param <T> The class type of the Transformer implementation itself.
 */
@PublicEvolving
public interface Transformer<T extends Transformer<T>> extends AlgoOperator<T> {
    /**
     * Uses the given list of tables to update internal states. This can be useful for e.g. online
     * learning where an Estimator fits an infinite stream of training samples and streams the model
     * diff data to this Transformer.
     *
     * <p>This method may be called at most once.
     *
     * @param inputs a list of tables
     */
    default void setStateStreams(Table... inputs) {
        throw new UnsupportedOperationException("this method is not implemented");
    }

    /**
     * Gets a list of tables representing changes of internal states of this Transformer. These
     * tables might come from the Estimator that instantiated this Transformer.
     *
     * @return a list of tables
     */
    default Table[] getStateStreams() {
        throw new UnsupportedOperationException("this method is not implemented");
    }
}

/**
 * An Estimator is a Stage that takes a list of tables as inputs and produces a Transformer.
 *
 * @param <E> class type of the Estimator implementation itself.
 * @param <M> class type of the Transformer this Estimator produces.
 */
@PublicEvolving
public interface Estimator<E extends Estimator<E, M>, M extends Transformer<M>> extends Stage<E> {
    /**
     * Trains on the given inputs and produces a Transformer.
     *
     * @param inputs a list of tables
     * @return a Transformer
     */
    M fit(Table... inputs);
}

/**
 * A Pipeline acts as an Estimator. It consists of an ordered list of stages, each of which could be
 * an Estimator, Transformer or AlgoOperator.
 */
@PublicEvolving
public final class Pipeline implements Estimator<Pipeline, pipelineTransformer> {

    public Pipeline(List<Stage<?>> stages) {...}

    @Override
    public pipelineTransformer fit(Table... inputs) {...}

    /** Skipped a few methods, including the implementations of the Estimator APIs. */
}

/**
 * A pipelineTransformer acts as a Transformer. It consists of an ordered list of Transformers or AlgoOperators.
 */
@PublicEvolving
public final class pipelineTransformer implements Transformer<pipelineTransformer> {

    public pipelineTransformer(List<Transformer<?>> transformers) {...}

    /** Skipped a few methods, including the implementations of the Transformer APIs. */
}

The following code block shows the interface of Graph, graphTransformer and GraphBuilder that we propose to add.

Code Block
languagejava
/**
 * A Graph acts as an Estimator. A Graph consists of a DAG of stages, each of which could be
 * an Estimator, Transformer or AlgoOperator. When `Graph::fit` is called, the stages are executed in a
 * topologically-sorted order. If a stage is an Estimator, its `Estimator::fit` method will be
 * called on the input tables (from the input edges) to fit a model. Then the model, which is a
 * Transformer, will be used to transform the input tables to produce output tables to the output
 * edges. If a stage is a Transformer or AlgoOperator, its `AlgoOperator::transform` method will be called on the
 * input tables to produce output tables to the output edges. The fitted model from a Graph is a
 * graphTransformer, which consists of fitted models and transformers, corresponding to the Graph's
 * stages.
 */
@PublicEvolving
public final class Graph implements Estimator<Graph, graphTransformer> {
    public Graph(...) {...}

    @Override
    public graphTransformer fit(Table... inputs) {...}

    /** Skipped a few methods, including the implementations of some Estimator APIs. */
}

/**
 * A graphTransformer acts as a Transformer. A graphTransformer consists of a DAG of Transformers or AlgoOperators. When
 * `graphTransformer::transform` is called, the stages are executed in a topologically-sorted order. When
 * a stage is executed, its `AlgoOperator::transform` method will be called on the input tables (from
 * the input edges) to produce output tables to the output edges.
 */
public final class graphTransformer implements Transformer<graphTransformer> {
    /** Skipped a few methods, including the implementations of the Transformer APIs. */
}

/**
 * A GraphBuilder provides APIs to build Graph and graphTransformer from a DAG of Estimator, Transformer and AlgoOperator instances.
 */
@PublicEvolving
public final class GraphBuilder {
    /**
     * Specifies the upper bound (could be loose) of the number of output tables that can be
     * returned by the Transformer::getStateStreams and Transformer::transform methods, for any
     * stage involved in this Graph.
     *
     * <p>The default upper bound is 20.
     */
    public GraphBuilder setMaxOutputLength(int maxOutputLength) {...}

    /**
     * Creates a TableId associated with this GraphBuilder. It can be used to specify the passing of
     * tables between stages, as well as the input/output tables of the Graph/graphTransformer generated
     * by this builder.
     */
    public TableId createTableId() {...}

    /**
     * If the stage is an Estimator, both its fit method and the transform method of its fitted Transformer would be
     * invoked with the given inputs when the graph runs.
     *
     * <p>If this stage is a Transformer or AlgoOperator, its transform method would be invoked with the given
     * inputs when the graph runs.
     *
     * Returns a list of TableIds, which represents outputs of the Transformer::transform invocation.
     */
    public TableId[] getOutputs(Stage<?> stage, TableId... inputs) {...}

    /**
     * If this stage is an Estimator, its fit method would be invoked with estimatorInputs, and the transform method
     * of its fitted Transformer would be invoked with transformerInputs, when the graph runs.
     *
     * <p>This method throws Exception if the stage is a Transformer or AlgoOperator.
     *
     * This method is useful when the state is an Estimator AND the Estimator::fit needs to take a different list of
     * Tables from the Transformer::transform of the fitted Transformer.
     *
     * Returns a list of TableIds, which represents outputs of the Transformer::transform invocation.
     */
    public TableId[] getOutputs(Stage<?> stage, TableId[] estimatorInputs, TableId[] transformerInputs) {...}

    /**
     * The graphTransformer::setStateStreams should invoke the setStateStreams of the corresponding stage
     * with the corresponding inputs.
     */
    void setStateStreams(Stage<?> stage, TableId... inputs) {...}

    /**
     * The graphTransformer::getStateStreams should invoke the getStateStreams of the corresponding stage.
     *
     * <p>Returns a list of TableIds, which represents outputs of the getStateStreams invocation.
     */
    TableId[] getStateStreams(Stage<?> stage) {...}

    /**
     * Returns a Graph instance which the following API specification:
     *
     * 1) Graph::fit should take inputs and returns a graphTransformer with the following specification.
     *
     * 2) graphTransformer::transform should take inputs and return outputs.
     *
     * 3) graphTransformer::setStateStreams should take inputStates.
     *
     * 4) graphTransformer::getStateStreams should return outputStates.
     *
     * The fit/transform/setStateStreams/getStateStreams should invoke the APIs of the internal stages in the order specified by the DAG of stages.
     */
    Graph build(TableId[] inputs, TableId[] outputs, TableId[] inputStates, TableId[] outputStates) {...}

    /**
     * Returns a Graph instance which the following API specification:
     *
     * 1) Graph::fit should take estimatorInputs and returns a graphTransformer with the following specification.
     *
     * 2) graphTransformer::transform should take transformerInputs and return outputs.
     *
     * 3) graphTransformer::setStateStreams should take inputStates.
     *
     * 4) graphTransformer::getStateStreams should return outputStates.
     *
     * The fit/transform/setStateStreams/getStateStreams should invoke the APIs of the internal stages in the order specified by the DAG of stages.
     *
     * This method is useful when the Graph::fit needs to take a different list of Tables from the graphTransformer::transform of the fitted graphTransformer.
     */
    Graph build(TableId[] estimatorInputs, TableId[] transformerInputs, TableId[] outputs, TableId[] inputStates, TableId[] outputStates) {...}

    /**
     * Returns a graphTransformer instance which the following API specification:
     *
     * 1) graphTransformer::transform should take inputs and returns outputs.
     *
     * 2) graphTransformer::setStateStreams should take inputStates.
     *
     * 3) graphTransformer::getStateStreams should return outputStates.
     *
     * The transform/setStateStreams/getStateStreams should invoke the APIs of the internal  stages in the order specified by the DAG of stages.
     *
     * This method throws Exception if any stage of this graph is an Estimator.
     */
    graphTransformer buildTransformer(TableId[] inputs, TableId[] outputs, TableId[] inputStates, TableId[] outputStates) {...}

    // The TableId is necessary to pass the inputs/outputs of various API calls across the
    // Graph/graphTransformer stags.
    static class TableId {}

}

Example Usage

In this section we provide examples code snippets to demonstrate how we can use the APIs proposed in this FLIP to address the use-cases in the motivation section.

Composing an Estimator from a DAG of Estimator/Transformer

Suppose we have the following Transformer and Estimator classes:

  • TransformerA whose transform(...) takes 1 input table and has 1 output table.
  • TransformerB whose transform(...) takes 2 input tables and has 1 output table.
  • EstimatorB whose fit(...) takes 2 input tables and returns an instance of TransformerB.

And we want to compose an Estimator (e.g. Graph) from the following DAG of Transformer/Estimator.

Image Removed

The resulting Graph::fit is expected to have the following behavior:

  • The method takes 2 input tables. The 1st input table is given to a TransformerA instance. And the 2nd input table is given to another TransformerA instance.
  • An EstimatorB instance fits the output tables of these two TransformerA instances and generates a new TransformerB instance.
  • Returns a graphTransformer instance which contains 2 TransformerA instance and 1 TransformerB instance, connected using the same DAG as shown above.

...

languagejava


Example Usage

In this section we provide examples code snippets to demonstrate how we can use the APIs proposed in this FLIP to address the use-cases in the motivation section.

...

Online learning by running Transformer and Estimator concurrently on different machines

...

  • The method takes 2 input tables. Both tables are given to EstimatorA::fit.
  • EstimatorA fits the input tables and generates a TransformerA instance. The TransformerA instance takes 1 table input, which is different from the 2 tables given to the EstimatorA.
  • Returns a graphTransformer GraphTransformer instance which contains a TransformerA instance and a TransformerB instance, which are connected as a chain.

The fitted graphTransformer GraphTransformer is represented by the following DAG:

Notes:

  • The fitted graphTransformer GraphTransformer takes only 1 table as input whereas the Graph takes 2 tables as inputs.
  • The proposed APIs also support composing an Estimator from a DAG of Estimator/Transformer whose input schemas are different from its fitted Transformer. 

...

Code Block
languagejava
GraphBuilder builder = new GraphBuilder();

// Creates nodes
Stage<?> stage1 = new EstimatorA();
Stage<?> stage2 = new TransformerB();
// Creates inputs
TableId estimatorInput1 = builder.createTableId();
TableId estimatorInput2 = builder.createTableId();
TableId transformerInput1 = builder.createTableId();

// Feeds inputs to nodes and gets outputs.
TableId output1 = builder.getOutputs(stage1, new TableId[] {estimatorInput1, estimatorInput2}, new TableId[] {transformerInput1})[0];
TableId output2 = builder.getOutputs(stage2, output1)[0];

// Specifies the ordered lists of estimator inputs, transformer inputs, outputs, input states and output states
// that will be used as the inputs/outputs of the corresponding Graph and graphTransformerGraphTransformer APIs.
TableId[] estimatorInputs = new TableId[] {estimatorInput1, estimatorInput2};
TableId[] transformerInputs = new TableId[] {transformerInput1};
TableId[] outputs = new TableId[] {output2};
TableId[] inputStates = new TableId[] {};
TableId[] outputStates = new TableId[] {};

// Generates the Graph instance.
Graph graph = builder.build(estimatorInputs, transformerInputs, outputs, inputStates, outputStates);
// The fit method takes 2 tables which are mapped to estimatorInput1 and estimatorInput2.
graphTransformerGraphTransformer transformer = graph.fit(...);
// The transform method takes 1 table which is mapped to transformerInput1.
Table[] results = transformer.transform(...);

...