org.tensorflow

Class Session.Runner

java.lang.Object

org.tensorflow.Session.Runner

Enclosing class:

Session

public final class Session.Runner
extends Object

Run Operations and evaluate Tensors.

A Runner runs the necessary graph fragments to execute every Operation required to evaluate the Tensors to fetch. The feed(String,int,Tensor) call allows callers to override the value of Tensors in the graph by substituting the provided Tensors for the outputs of the operations provided to feed(String,int,Tensor).

Constructor Summary

Constructors

Constructor and Description
Runner()

Method Summary

Modifier and Type	Method and Description
Session.Runner	addTarget(Operand<?> operand) Make run() execute operand, but not return any evaluated Tensors.
Session.Runner	addTarget(Operation operation) Make run() execute operation, but not return any evaluated Tensors.
Session.Runner	addTarget(String operation) Make run() execute operation, but not return any evaluated Tensors.
Session.Runner	feed(Output<?> o, Tensor<?> t) Use t instead of the Tensor referred to by executing the operation referred to by output.
Session.Runner	feed(String operation, int index, Tensor<?> t) Avoid evaluating the index-th output of operation by substituting t for the value it produces.
Session.Runner	feed(String operation, Tensor<?> t) Avoid evaluating operation and substitute t for the value it produces.
Session.Runner	fetch(Operand<?> operand) Makes run() return the Tensor referred to by the output of operand.
Session.Runner	fetch(Output<?> output) Makes run() return the Tensor referred to by output.
Session.Runner	fetch(String operation) Make run() return the output of operation.
Session.Runner	fetch(String operation, int index) Make run() return the index-th output of operation.
List<Tensor<?>>	run() Execute the graph fragments necessary to compute all requested fetches.
Session.Run	runAndFetchMetadata() Execute graph fragments to compute requested fetches and return metadata about the run.
Session.Runner	setOptions(byte[] options) (Experimental method): set options (typically for debugging) for this run.

Methods inherited from class java.lang.Object

clone, equals, finalize, getClass, hashCode, notify, notifyAll, toString, wait, wait, wait

Constructor Detail

Runner

public Runner()

Method Detail

feed

public Session.Runner feed(String operation,
                           Tensor<?> t)

Avoid evaluating operation and substitute t for the value it produces.

Parameters:

operation - Is either the string name of the operation, in which case this method is a shorthand for feed(operation, 0), or it is a string of the form operation_name:output_index , in which case this method acts like feed(operation_name, output_index). These colon-separated names are commonly used in the SignatureDef protocol buffer messages that are included in SavedModelBundle.metaGraphDef().

feed

public Session.Runner feed(String operation,
                           int index,
                           Tensor<?> t)

Avoid evaluating the index-th output of operation by substituting t for the value it produces.

Operations in a Graph can have multiple outputs, index identifies which one t is being provided for.

feed

public Session.Runner feed(Output<?> o,
                           Tensor<?> t)

Use t instead of the Tensor referred to by executing the operation referred to by output.

fetch

public Session.Runner fetch(String operation)

Make run() return the output of operation.

Parameters:

operation - Is either the string name of the operation, in which case this method is a shorthand for fetch(operation, 0), or it is a string of the form operation_name:output_index , in which case this method acts like fetch(operation_name, output_index). These colon-separated names are commonly used in the SignatureDef protocol buffer messages that are included in SavedModelBundle.metaGraphDef().

fetch

public Session.Runner fetch(String operation,
                            int index)

Make run() return the index-th output of operation.

Operations in a Graph can have multiple outputs, index identifies which one to return.

fetch

public Session.Runner fetch(Output<?> output)

Makes run() return the Tensor referred to by output.

fetch

public Session.Runner fetch(Operand<?> operand)

Makes run() return the Tensor referred to by the output of operand.

addTarget

public Session.Runner addTarget(String operation)

Make run() execute operation, but not return any evaluated Tensors.

addTarget

public Session.Runner addTarget(Operation operation)

Make run() execute operation, but not return any evaluated Tensors.

addTarget

public Session.Runner addTarget(Operand<?> operand)

Make run() execute operand, but not return any evaluated Tensors.

setOptions

public Session.Runner setOptions(byte[] options)

(Experimental method): set options (typically for debugging) for this run.

The options are presented as a serialized RunOptions protocol buffer.

The org.tensorflow package is free of any protocol buffer dependencies in order to remain friendly to resource constrained systems (where something like nanoproto may be more appropriate). A cost of that is this lack of type-safety in this API function. This choice is under review and this function may be replaced by more type-safe equivalents at any time.

run

public List<Tensor<?>> run()

Execute the graph fragments necessary to compute all requested fetches.

WARNING: The caller assumes ownership of all returned Tensors, i.e., the caller must call Tensor.close() on all elements of the returned list to free up resources.

TODO(ashankar): Reconsider the return type here. Two things in particular: (a) Make it easier for the caller to cleanup (perhaps returning something like AutoCloseableList in SessionTest.java), and (b) Evaluate whether the return value should be a list, or maybe a Map<Output, Tensor>?

TODO(andrewmyers): It would also be good if whatever is returned here made it easier to extract output tensors in a type-safe way.

runAndFetchMetadata

public Session.Run runAndFetchMetadata()

Execute graph fragments to compute requested fetches and return metadata about the run.

This is exactly like run(), but in addition to the requested Tensors, also returns metadata about the graph execution in the form of a serialized RunMetadata protocol buffer.