Gradients

public final class Gradients

Adds operations to compute the partial derivatives of sum of ys w.r.t xs, i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...

If Options.dx() values are set, they are as the initial symbolic partial derivatives of some loss function L w.r.t. y. Options.dx() must have the size of y.

If Options.dx() is not set, the implementation will use dx of OnesLike for all shapes in y.

The partial derivatives are returned in output dy, with the size of x.

Example of usage:

Gradients gradients = tf.gradients(loss, Arrays.asList(w, b));
 Constant<TFloat32> alpha = tf.constant(1.0f);
 tf.train.applyGradientDescent(w, alpha, gradients.<Float>dy(0));
 tf.train.applyGradientDescent(b, alpha, gradients.<Float>dy(1));
 

Nested Classes

class

Gradients.Options

Optional attributes for Gradients

Public Methods

static Gradients	create(Scope scope, Operand<?> y, Iterable<? extends Operand<?>> x, Options... options) Adds gradients computation ops to the graph according to scope.
static Gradients	create(Scope scope, Iterable<? extends Operand<?>> y, Iterable<? extends Operand<?>> x, Options... options) Adds gradients computation ops to the graph according to scope.
static Gradients.Options	dx(Iterable<? extends Operand<?>> dx)
<T extends TType> Output<T>	dy(int index) Returns a symbolic handle to one of the gradient operation output Warning: Does not check that the type of the tensor matches T.
List<Output<?>>	dy() Partial derivatives of ys w.r.t.
Iterator<Operand<?>>	iterator()

Inherited Methods

From class java.lang.Object

boolean	equals(Object arg0)
final Class<?>	getClass()
int	hashCode()
final void	notify()
final void	notifyAll()
String	toString()
final void	wait(long arg0, int arg1)
final void	wait(long arg0)
final void	wait()

From interface java.lang.Iterable

abstract void	forEach(Consumer<? super T> arg0)
abstract Iterator<Operand<?>>	iterator()
abstract Spliterator<Operand<?>>	spliterator()

Public Methods

public static Gradients create (Scope scope, Operand<?> y, Iterable<? extends Operand<?>> x, Options... options)

Adds gradients computation ops to the graph according to scope.

This is a simplified version of ERROR(/#create(Scope, Iterable, Iterable, Options...)) where y is a single output.

Parameters

scope	current graph scope
y	output of the function to derive
x	inputs of the function for which partial derivatives are computed
options	carries optional attributes values

Returns

• a new instance of Gradients

Throws

IllegalArgumentException	if execution environment is not a graph

public static Gradients create (Scope scope, Iterable<? extends Operand<?>> y, Iterable<? extends Operand<?>> x, Options... options)

Adds gradients computation ops to the graph according to scope.

Parameters

scope	current graph scope
y	outputs of the function to derive
x	inputs of the function for which partial derivatives are computed
options	carries optional attributes values

Returns

• a new instance of Gradients

Throws

IllegalArgumentException	if execution environment is not a graph

public static Gradients.Options dx (Iterable<? extends Operand<?>> dx)

Parameters

dx	partial derivatives of some loss function L w.r.t. y

Returns

• builder to add more options to this operation

public Output<T> dy (int index)

Returns a symbolic handle to one of the gradient operation output

Warning: Does not check that the type of the tensor matches T. It is recommended to call this method with an explicit type parameter rather than letting it be inferred, e.g. gradients.<Float>dy(0)

Parameters

index	The index of the output among the gradients added by this operation

public List<Output<?>> dy ()

Partial derivatives of ys w.r.t. xs, with the size of x

public Iterator<Operand<?>> iterator ()