com.linkedin.dagli.dag.DAG6x3 Maven / Gradle / Ivy
Go to download
Show more of this group Show more artifacts with this name
Show all versions of core Show documentation
Show all versions of core Show documentation
DAG-oriented machine learning framework for bug-resistant, readable, efficient, maintainable and trivially deployable models in Java and other JVM languages
// AUTOGENERATED CODE. DO NOT MODIFY DIRECTLY! Instead, please modify the dag/DAGXxY.ftl file.
// See the README in the module's src/template directory for details.
package com.linkedin.dagli.dag;
import java.util.Collection;
import java.util.Collections;
import com.linkedin.dagli.annotation.equality.IgnoredByValueEquality;
import com.linkedin.dagli.annotation.equality.ValueEquality;
import com.linkedin.dagli.objectio.ObjectIterator;
import com.linkedin.dagli.objectio.ObjectReader;
import com.linkedin.dagli.preparer.Preparer6;
import com.linkedin.dagli.preparer.AbstractBatchPreparer6;
import com.linkedin.dagli.preparer.PreparerContext;
import com.linkedin.dagli.preparer.PreparerResultMixed;
import com.linkedin.dagli.producer.Producer;
import com.linkedin.dagli.reducer.Reducer;
import com.linkedin.dagli.transformer.AbstractPreparableTransformer6;
import com.linkedin.dagli.transformer.AbstractPreparedTransformer6;
import com.linkedin.dagli.transformer.PreparedTransformer6;
import com.linkedin.dagli.util.array.AutoCloseableArray;
import com.linkedin.dagli.generator.Generator;
import com.linkedin.dagli.placeholder.Placeholder;
import com.linkedin.dagli.tuple.Tuple3;
import com.linkedin.dagli.tuple.Tuple6;
import com.linkedin.dagli.transformer.Value0FromTuple;
import com.linkedin.dagli.transformer.Value1FromTuple;
import com.linkedin.dagli.transformer.Value2FromTuple;
/**
* DAGs, directed acyclic graphs, contain root nodes ({@link com.linkedin.dagli.placeholder.Placeholder}s and {@link com.linkedin.dagli.generator.Generator}s) and child nodes
* ({@link com.linkedin.dagli.transformer.PreparedTransformer}s,
* {@link com.linkedin.dagli.transformer.PreparableTransformer}s, and {@link com.linkedin.dagli.view.TransformerView}s).
*
* DAG6x3 is a preparable DAG.
*
* Before it can be used to transform new examples, it must be prepared with training data (e.g. by calling the
* prepare() method). The prepareAndApply() method both trains the DAG and applies it to the training data, yielding
* both the prepared DAG and the results of applying the DAG to the training data (this can be useful in autoevaluation,
* e.g. how well a classification model performs when predicting labels for its own training data).
*
* DAG6x3 is also a {@link com.linkedin.dagli.transformer.PreparableTransformer} and can be used as
* a transformer within other DAGs.
*
* @param the type of the first input
* @param the type of the second input
* @param the type of the third input
* @param the type of the fourth input
* @param the type of the fifth input
* @param the type of the sixth input
* @param the type of the first result
* @param the type of the second result
* @param the type of the third result
*/
@ValueEquality
public class DAG6x3
extends
AbstractPreparableTransformer6, DAG6x3.Prepared, DAG6x3>
implements
PreparableDAGTransformer, DAG6x3.Prepared, DAG6x3> {
private static final long serialVersionUID = 1;
// the DAGStructure stores the actual graph
private DAGStructure> _dag;
// keep track of what level of reduction we used to construct this DAG; this will be carried over to the prepared DAG
@IgnoredByValueEquality
// irrelevant given _dag and not (directly) externally observable
private Reducer.Level _reductionLevel = null; // null -> no reductions applied
// remember what executor should be used to prepare the DAG; this will also be "inherited" by the prepared DAG and
// used for inference
@IgnoredByValueEquality
// not a factor in the semantics of the DAG and not (directly) externally observable
private DAGExecutor _executor;
/**
* Creates a new DAG instance from the specified {@link DAGStructure} graph with the specified inputs. DAG objects
* have inputs because they can also serve as transformers within other DAGs. In the common case where the DAG is
* used directly and not embedded within another DAG, these inputs don't matter (and can simply be left as
* {@link com.linkedin.dagli.producer.MissingInput}s).
*
* By default, this DAG will use a {@link LocalDAGExecutor}; to use a different executor, use the
* {@link #withExecutor(DAGExecutor)} method. When this DAG is prepared, the DAG executor will be "inherited" by the
* resultant prepared DAG.
*
* @param dag the {@link DAGStructure} representing the actual graph of nodes and directed (parent/child) edges
* @param input1 the {@link Producer} providing the first input to the transformer
* @param input2 the {@link Producer} providing the second input to the transformer
* @param input3 the {@link Producer} providing the third input to the transformer
* @param input4 the {@link Producer} providing the fourth input to the transformer
* @param input5 the {@link Producer} providing the fifth input to the transformer
* @param input6 the {@link Producer} providing the sixth input to the transformer
*/
DAG6x3(DAGStructure> dag, Producer input1, Producer input2,
Producer input3, Producer input4, Producer input5,
Producer input6) {
super(input1, input2, input3, input4, input5, input6);
_dag = dag;
_executor = new LocalDAGExecutor();
}
/**
* Returns a DAG derived from this one that has been reduced to at least the specified level.
*
* If level is less than this DAG's current reduction level, {@code this} DAG is returned.
*
* Otherwise, this DAG is reduced, and the new, reduced DAG is returned.
*
* @param level the level of reduction desired; all reducers at or above this level will be applied
* @return a DAG that has been reduced to at least the specified level (possibly this same DAG)
*/
@Override
public DAG6x3 withReduction(Reducer.Level level) {
if (Reducer.Level.compare(level, _reductionLevel) >= 0) {
return this;
}
DeduplicatedDAG reduced = DAGReducer.reduce(new DeduplicatedDAG(_dag), level);
return clone(c -> {
c._dag = new DAGStructure<>(reduced);
c._reductionLevel = level;
});
}
@Override
public InternalAPI internalAPI() {
return new InternalAPI();
}
public class InternalAPI
extends
AbstractPreparableTransformer6, DAG6x3.Prepared, DAG6x3>.InternalAPI
implements
PreparableDAGTransformer.InternalAPI, DAG6x3.Prepared, DAG6x3> {
@Override
public DAGStructure> getDAGStructure() {
return _dag;
}
@Override
public Reducer.Level getReductionLevel() {
return _reductionLevel;
}
@Override
public DAGExecutor getDAGExecutor() {
return _executor;
}
@Override
public DAG6x3 getInstance() {
return DAG6x3.this;
}
}
@Override
protected Collection>> getGraphReducers() {
return Collections.singletonList(DAGTransformerReducer.INSTANCE);
}
@Override
protected boolean hasAlwaysConstantResult() {
return _dag._isAlwaysConstant;
}
/**
* Returns a copy of this DAG that will use the given {@link DAGExecutor} that will be used to execute this DAG.
*
* This executor will also be "inherited" by the resultant prepared DAG.
*
* @param executor the {@link DAGExecutor} to use
* @return a copy of this instance that will use the provided executor
*/
public DAG6x3 withExecutor(DAGExecutor executor) {
return clone(r -> r._executor = executor);
}
@Override
protected boolean hasIdempotentPreparer() {
return _dag._hasIdempotentPreparer;
}
/**
* Sets the parents of this transformer. The results of the parent nodes will be the inputs to the transformer.
*
* @param input1 the {@link Producer} providing the first input to the transformer
* @param input2 the {@link Producer} providing the second input to the transformer
* @param input3 the {@link Producer} providing the third input to the transformer
* @param input4 the {@link Producer} providing the fourth input to the transformer
* @param input5 the {@link Producer} providing the fifth input to the transformer
* @param input6 the {@link Producer} providing the sixth input to the transformer
* @return a copy of this instance that will have the specified parents
*/
public DAG6x3 withInputs(Producer input1, Producer input2,
Producer input3, Producer input4, Producer input5,
Producer input6) {
return super.withAllInputs(input1, input2, input3, input4, input5, input6);
}
/**
* Creates a new DAG that invokes this transformer, except that the first input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the first input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the first input
* @return a new DAG that invokes this transformer with the same inputs, except the first, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput1(Generator generator) {
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3 dag =
this.withInputs(generator, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG
.withPlaceholders(nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput2(), getInput3(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the second input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the second input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the second input
* @return a new DAG that invokes this transformer with the same inputs, except the second, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput2(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3 dag =
this.withInputs(nestedPlaceholder1, generator, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput3(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the third input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the third input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the third input
* @return a new DAG that invokes this transformer with the same inputs, except the third, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput3(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3 dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, generator, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the fourth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the fourth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the fourth input
* @return a new DAG that invokes this transformer with the same inputs, except the fourth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput4(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3 dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, generator, nestedPlaceholder5,
nestedPlaceholder6);
return DAG
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the fifth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the fifth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the fifth input
* @return a new DAG that invokes this transformer with the same inputs, except the fifth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput5(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3 dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, generator,
nestedPlaceholder6);
return DAG
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput4(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the sixth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the sixth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the sixth input
* @return a new DAG that invokes this transformer with the same inputs, except the sixth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3 withGeneratorAsInput6(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
DAG6x3 dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder5, generator);
return DAG
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder5)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput4(), getInput5());
}
/**
* The {@link com.linkedin.dagli.preparer.Preparer} that prepares (trains) the DAG.
*
* @param the type of the first input
* @param the type of the second input
* @param the type of the third input
* @param the type of the fourth input
* @param the type of the fifth input
* @param the type of the sixth input
* @param the type of the first result
* @param the type of the second result
* @param the type of the third result
*/
private static class Preparer extends
AbstractBatchPreparer6, DAG6x3.Prepared> {
// the DAG being prepared
private final DAG6x3 _dag;
/**
* Creates a new instance.
*
* @param dag the DAG that will be prepared
*/
Preparer(DAG6x3 dag) {
_dag = dag;
}
@Override
@SuppressWarnings("unchecked")
// we know the returned type is correct by the semantics of executors
public PreparerResultMixed>, DAG6x3.Prepared> finish(
ObjectReader> inputs) {
return (PreparerResultMixed) _dag._executor.internalAPI().prepareUnsafe(_dag,
ObjectReader.split(6, inputs.lazyMap(tuple -> tuple.toArray())));
}
@Override
public void process(A value1, B value2, C value3, D value4, E value5, F value6) {
// noop
}
}
@Override
protected Preparer6, DAG6x3.Prepared> getPreparer(
PreparerContext context) {
return new Preparer<>(this);
}
/**
* Prepares (trains) the DAG on the provided input data (examples), and applies it to this same data. The returned
* {@link Result} instance contains both the prepared DAG as well as the results from applying the DAG on the
* preparation data.
*
* The inputs to this method take the form of parallel sequences of values. Each sequence will provide the values
* for a particular {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG (the ordering of the
* {@link com.linkedin.dagli.placeholder.Placeholder}s in a DAG is determined by their
* ordering in the call to "DAG.withPlaceholders(...)...." that created it). They are "parallel" because, e.g. all the
* values corresponding to the fourth example will be the fourth value in their respective input sequences (all
* sequences must, of course, be equal in size).
*
* Note that the results from applying the DAG to the preparation data with this method are not necessarily the same
* results that would be obtained from calling the apply(...) method on prepared DAG with the same data. This is
* because certain transformers (e.g. KFoldCrossTrained and PreparedByGroup) intentionally transform training data in
* a different way than inference data. Please see the documentation of these transformers for details.
*
* @param values1 the sequence of values for the first
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values2 the sequence of values for the second
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values3 the sequence of values for the third
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values4 the sequence of values for the fourth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values5 the sequence of values for the fifth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values6 the sequence of values for the sixth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @return a {@link Result} instance containing both the prepared DAG and the result of applying the DAG on the
* preparation data
*/
public Result prepareAndApply(Iterable values1,
Iterable values2, Iterable values3, Iterable values4,
Iterable values5, Iterable values6) {
DAGExecutionResult, DAG6x3.Prepared> res =
_executor
.internalAPI()
.prepareAndApplyUnsafe(
this,
new ObjectReader[] { ObjectReader.wrap(values1), ObjectReader.wrap(values2), ObjectReader.wrap(values3), ObjectReader
.wrap(values4), ObjectReader.wrap(values5), ObjectReader.wrap(values6) });
return new Result<>((DAG6x3.Prepared) res.getPreparerResult()
.getPreparedTransformerForNewData(), res.getOutputs());
}
/**
* Prepares (trains) the DAG on the provided input data (examples). The resultant prepared (trained) DAG can then be
* applied to new examples (inference).
*
* The inputs to this method take the form of parallel sequences of values. Each sequence will provide the values
* for a particular {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG (the ordering of the
* {@link com.linkedin.dagli.placeholder.Placeholder}s in a DAG is determined by their
* ordering in the call to "DAG.withPlaceholders(...)...." that created it). They are "parallel" because, e.g. all the
* values corresponding to the fourth example will be the fourth value in their respective input sequences (all
* sequences must, of course, be equal in size).
*
* @param values1 the sequence of values for the first
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values2 the sequence of values for the second
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values3 the sequence of values for the third
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values4 the sequence of values for the fourth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values5 the sequence of values for the fifth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values6 the sequence of values for the sixth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @return a prepared (trained) DAG that can be applied to new examples (inference)
*/
public DAG6x3.Prepared prepare(Iterable values1,
Iterable values2, Iterable values3, Iterable values4,
Iterable values5, Iterable values6) {
return (DAG6x3.Prepared) _executor
.internalAPI()
.prepareUnsafe(
this,
new ObjectReader[] { ObjectReader.wrap(values1), ObjectReader.wrap(values2), ObjectReader.wrap(values3), ObjectReader
.wrap(values4), ObjectReader.wrap(values5), ObjectReader.wrap(values6) })
.getPreparedTransformerForNewData();
}
/**
* {@link Result}s are returned by the DAG's prepareAndApply(...) methods and contain both the prepared (trained) DAG
* and the results of applying that DAG to the preparation (training) data.
*
* @param the type of the first input
* @param the type of the second input
* @param the type of the third input
* @param the type of the fourth input
* @param the type of the fifth input
* @param the type of the sixth input
* @param the type of the first result
* @param the type of the second result
* @param the type of the third result
*/
public static final class Result extends AbstractDAGResult3 {
// the prepared DAG
private final DAG6x3.Prepared _preparedDAG;
/**
* @return the {@link Prepared} DAG that was prepared (trained) on the provided preparation (training) data.
*/
public DAG6x3.Prepared getPreparedDAG() {
return _preparedDAG;
}
/**
* Creates a new instance.
*
* @param preparedDAG the prepared DAG to be stored in this instance
* @param results an array of {@link ObjectReader}s that contain each resultant output value of applying the DAG to
* the preparation (training) data
*/
Result(DAG6x3.Prepared preparedDAG, ObjectReader[] results) {
super(results);
_preparedDAG = preparedDAG;
}
}
/**
* Prepared DAGs (directed acyclic graphs) contain root nodes ({@link com.linkedin.dagli.placeholder.Placeholder}s and
* {@link com.linkedin.dagli.generator.Generator}s) and child nodes
* ({@link com.linkedin.dagli.transformer.PreparedTransformer}s) Unlike preparable DAGs, prepared DAGs do not contain
* {@link com.linkedin.dagli.transformer.PreparableTransformer}s nor {@link com.linkedin.dagli.view.TransformerView}s
* and can be applied to examples to transform them to some desired result (often an inference, such as a
* classification or regression) using the apply(...) (for single examples) or applyAll(...) (for multiple examples)
* methods.
*
* Prepared DAGs are also themselves {@link com.linkedin.dagli.transformer.PreparedTransformer}s, and can be used as
* such within another DAG.
*
* @param the type of the first input
* @param the type of the second input
* @param the type of the third input
* @param the type of the fourth input
* @param the type of the fifth input
* @param the type of the sixth input
* @param the type of the first result
* @param the type of the second result
* @param the type of the third result
*/
@ValueEquality
public static class Prepared extends
AbstractPreparedTransformer6, DAG6x3.Prepared>
implements PreparedDAGTransformer, DAG6x3.Prepared> {
private static final long serialVersionUID = 1;
// the DAGStructure stores the actual graph
private DAGStructure> _dag;
// keep track of what level of reduction we used to construct this DAG
@IgnoredByValueEquality
// irrelevant given _dag and not (directly) externally observable
private Reducer.Level _reductionLevel = null; // null -> no reductions applied
// the executor that will be used to apply the DAG to new examples
@IgnoredByValueEquality
// not a factor in the semantics of the DAG and not (directly) externally observable
private PreparedDAGExecutor _executor;
/**
* Creates a new DAG instance from the specified {@link DAGStructure} graph with the specified inputs. DAG objects
* have inputs because they can also serve as transformers within other DAGs. In the common case where the DAG is
* used directly and not embedded within another DAG, these inputs don't matter (and can simply be left as
* {@link com.linkedin.dagli.producer.MissingInput}s).
*
* By default, this DAG will use a {@link LocalDAGExecutor}; to use a different executor, use the
* {@link #withExecutor(DAGExecutor)} method.
*
* @param dag the {@link DAGStructure} representing the actual graph of nodes and directed (parent/child) edges
* @param input1 the {@link Producer} providing the first input to the transformer
* @param input2 the {@link Producer} providing the second input to the transformer
* @param input3 the {@link Producer} providing the third input to the transformer
* @param input4 the {@link Producer} providing the fourth input to the transformer
* @param input5 the {@link Producer} providing the fifth input to the transformer
* @param input6 the {@link Producer} providing the sixth input to the transformer
*/
Prepared(DAGStructure> dag, Producer input1, Producer input2,
Producer input3, Producer input4, Producer input5,
Producer input6) {
super(input1, input2, input3, input4, input5, input6);
_dag = dag;
_executor = new LocalDAGExecutor();
}
/**
* Returns a DAG derived from this one that has been reduced to at least the specified level.
*
* If level is less than this DAG's current reduction level, {@code this} DAG is returned.
*
* Otherwise, this DAG is reduced, and the new, reduced DAG is returned.
*
* @param level the level of reduction desired; all reducers at or above this level will be applied
* @return a DAG that has been reduced to at least the specified level (possibly this same DAG)
*/
@Override
public DAG6x3.Prepared withReduction(Reducer.Level level) {
if (Reducer.Level.compare(level, _reductionLevel) >= 0) {
return this;
}
DeduplicatedDAG reduced = DAGReducer.reduce(new DeduplicatedDAG(_dag), level);
return clone(c -> {
c._dag = new DAGStructure<>(reduced);
c._reductionLevel = level;
});
}
@Override
public InternalAPI internalAPI() {
return new InternalAPI();
}
public class InternalAPI
extends
AbstractPreparedTransformer6, DAG6x3.Prepared>.InternalAPI
implements
PreparedDAGTransformer.InternalAPI, DAG6x3.Prepared> {
@Override
public DAGStructure> getDAGStructure() {
return _dag;
}
@Override
public Reducer.Level getReductionLevel() {
return _reductionLevel;
}
@Override
public PreparedDAGExecutor getDAGExecutor() {
return _executor;
}
@Override
public DAG6x3.Prepared getInstance() {
return DAG6x3.Prepared.this;
}
}
@Override
protected Collection>> getGraphReducers() {
return Collections.singletonList(DAGTransformerReducer.INSTANCE);
}
@Override
protected boolean hasAlwaysConstantResult() {
return _dag._isAlwaysConstant;
}
/**
* Returns a copy of this DAG that will use the given {@link PreparedDAGExecutor} that will be used to execute this DAG.
*
* @param executor the {@link DAGExecutor} to use
* @return a copy of this instance that will use the provided executor
*/
public DAG6x3.Prepared withExecutor(PreparedDAGExecutor executor) {
return clone(r -> r._executor = executor);
}
/**
* Sets the parents of this transformer. The results of the parent nodes will be the inputs to the transformer.
*
* @param input1 the {@link Producer} providing the first input to the transformer
* @param input2 the {@link Producer} providing the second input to the transformer
* @param input3 the {@link Producer} providing the third input to the transformer
* @param input4 the {@link Producer} providing the fourth input to the transformer
* @param input5 the {@link Producer} providing the fifth input to the transformer
* @param input6 the {@link Producer} providing the sixth input to the transformer
* @return a copy of this instance that will have the specified parents
*/
public DAG6x3.Prepared withInputs(Producer input1,
Producer input2, Producer input3, Producer input4,
Producer input5, Producer input6) {
return super.withAllInputs(input1, input2, input3, input4, input5, input6);
}
/**
* Creates a new DAG that invokes this transformer, except that the first input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the first input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the first input
* @return a new DAG that invokes this transformer with the same inputs, except the first, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput1(Generator generator) {
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3.Prepared dag =
this.withInputs(generator, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput2(), getInput3(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the second input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the second input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the second input
* @return a new DAG that invokes this transformer with the same inputs, except the second, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput2(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3.Prepared dag =
this.withInputs(nestedPlaceholder1, generator, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder3, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput3(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the third input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the third input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the third input
* @return a new DAG that invokes this transformer with the same inputs, except the third, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput3(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3.Prepared dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, generator, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder4, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput4(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the fourth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the fourth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the fourth input
* @return a new DAG that invokes this transformer with the same inputs, except the fourth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput4(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3.Prepared dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, generator, nestedPlaceholder5,
nestedPlaceholder6);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder5,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput5(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the fifth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the fifth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the fifth input
* @return a new DAG that invokes this transformer with the same inputs, except the fifth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput5(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder6 = new Placeholder<>("Original Input 6");
DAG6x3.Prepared dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4, generator,
nestedPlaceholder6);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder6)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput4(), getInput6());
}
/**
* Creates a new DAG that invokes this transformer, except that the sixth input is replaced by the specified {@link Generator}.
* The resultant DAG will thus have an arity of 5 (one less than this transformer), and values that were formerly
* provided as the sixth input to the transformer will now instead be generated by the {@link Generator}.
*
* This method is useful for eliding inputs to a transformer. For example, a prepared classifier might have a
* "vestigial" label input (inherited from a corresponding preparable classifier from which it was obtained) that can
* be replaced by Constant.nullValue(). Since the label input isn't used by the
* prepared classifier (instead, it predicts a label from its other inputs and produces this as its result) replacing
* it with a null has no effect and eliminates an unnecessary input.
*
* @param generator the generator that will provide a value for the sixth input
* @return a new DAG that invokes this transformer with the same inputs, except the sixth, whose value is replaced by
* the produced by the provided generator
*/
public DAG5x3.Prepared withGeneratorAsInput6(Generator generator) {
Placeholder nestedPlaceholder1 = new Placeholder<>("Original Input 1");
Placeholder nestedPlaceholder2 = new Placeholder<>("Original Input 2");
Placeholder nestedPlaceholder3 = new Placeholder<>("Original Input 3");
Placeholder nestedPlaceholder4 = new Placeholder<>("Original Input 4");
Placeholder nestedPlaceholder5 = new Placeholder<>("Original Input 5");
DAG6x3.Prepared dag =
this.withInputs(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder5, generator);
return DAG.Prepared
.withPlaceholders(nestedPlaceholder1, nestedPlaceholder2, nestedPlaceholder3, nestedPlaceholder4,
nestedPlaceholder5)
.withOutputs(new Value0FromTuple<>(dag), new Value1FromTuple<>(dag), new Value2FromTuple<>(dag))
.withExecutor(_executor).withInputs(getInput1(), getInput2(), getInput3(), getInput4(), getInput5());
}
/**
* Applies the DAG to the provided input data (examples), running the model on the data and producing the results as a
* {@link Result} instance.
*
* The inputs to this method take the form of parallel sequences of values. Each sequence will provide the values
* for a particular {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG (the ordering of the
* {@link com.linkedin.dagli.placeholder.Placeholder}s in a DAG is determined by their
* ordering in the call to "DAG.withPlaceholders(...)...." that created it). They are "parallel" because, e.g. all the
* values corresponding to the fourth example will be the fourth value in their respective input sequences (all
* sequences must, of course, be equal in size).
*
* @param values1 the sequence of values for the first
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values2 the sequence of values for the second
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values3 the sequence of values for the third
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values4 the sequence of values for the fourth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values5 the sequence of values for the fifth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @param values6 the sequence of values for the sixth
* {@link com.linkedin.dagli.placeholder.Placeholder} in the DAG.
* @return a {@link Result} instance containing the results of applying the DAG on the provided data
*/
public Result applyAll(Iterable values1, Iterable values2,
Iterable values3, Iterable values4, Iterable values5,
Iterable values6) {
return new Result(
_executor.internalAPI()
.applyUnsafe(
this,
new ObjectReader[] { ObjectReader.wrap(values1), ObjectReader.wrap(values2), ObjectReader
.wrap(values3), ObjectReader.wrap(values4), ObjectReader.wrap(values5), ObjectReader
.wrap(values6) }));
}
@Override
public Tuple3 apply(A value1, B value2, C value3, D value4, E value5, F value6) {
try (
AutoCloseableArray> res =
new AutoCloseableArray<>(_executor.internalAPI().applyUnsafe(
this,
new ObjectReader[] { ObjectReader.singleton(value1), ObjectReader.singleton(value2), ObjectReader
.singleton(value3), ObjectReader.singleton(value4), ObjectReader.singleton(value5), ObjectReader
.singleton(value6) }));
ObjectIterator resultIterator0 = res.get(0).iterator();
ObjectIterator resultIterator1 = res.get(1).iterator();
ObjectIterator resultIterator2 = res.get(2).iterator()) {
return (Tuple3) Tuple3.of(resultIterator0.next(), resultIterator1.next(), resultIterator2.next());
}
}
/**
* Contains the results of applying the DAG to one or more examples, as produced by a call to applyAll(...).
*
* @param the type of the first result
* @param the type of the second result
* @param the type of the third result
*/
public static final class Result extends AbstractDAGResult3 {
/**
* Creates a new instance.
*
* @param results an array of {@link ObjectReader}s that contain each resultant output value of applying the DAG to
* the provided examples
*/
Result(ObjectReader[] results) {
super(results);
}
}
}
}
© 2015 - 2024 Weber Informatics LLC | Privacy Policy