Package ai.djl.nn

Class Parameter

java.lang.Object

ai.djl.nn.Parameter

All Implemented Interfaces:

java.lang.AutoCloseable

public class Parameter
extends java.lang.Object
implements java.lang.AutoCloseable

Parameter is a container class that holds a learnable parameter of a model.

Every Parameter is associated with a Block. The output of the block's forward function depends on the values in the Parameter. During training, the values in the Parameter are updated to reflect the training data. This process forms the crux of learning.

See Also:

The D2L chapter on parameter management

Nested Class Summary

Nested Classes

Modifier and Type	Class	Description
static class	Parameter.Builder	A Builder to construct a Parameter.
static class	Parameter.Type	Enumerates the types of Parameter.

Method Summary

Modifier and Type	Method	Description
static Parameter.Builder	builder()	Creates a builder to build a Parameter.
void	close()
void	freeze(boolean freeze)	Freezes or unfreezes the parameter for training.
NDArray	getArray()	Gets the values of this Parameter as an NDArray.
java.lang.String	getId()	Gets the ID of this Parameter.
Initializer	getInitializer()	Returns the Initializer for this Parameter, if not already set.
java.lang.String	getName()	Gets the name of this Parameter.
Shape	getShape()	Gets the shape of this Parameter.
Parameter.Type	getType()	Gets the type of this Parameter.
void	initialize(NDManager manager, DataType dataType)	Initializes the parameter with the given NDManager, with given DataType for the given expected input shapes.
boolean	isInitialized()	Checks if this Parameter is initialized.
void	load(NDManager manager, java.io.DataInputStream dis)	Loads parameter NDArrays from InputStream.
boolean	requiresGradient()	Returns whether this parameter needs gradients to be computed.
void	save(java.io.DataOutputStream dos)	Writes the parameter NDArrays to the given output stream.
void	setArray(NDArray array)	Sets the values of this Parameter.
void	setInitializer(Initializer initializer)	Sets the Initializer for this Parameter, if not already set.
void	setShape(Shape shape)	Sets the shape of this Parameter.

Methods inherited from class java.lang.Object

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

Method Detail

getId

public java.lang.String getId()

Gets the ID of this Parameter.

Returns:

the ID of this Parameter

getName

public java.lang.String getName()

Gets the name of this Parameter.

Returns:

the name of this Parameter

getType

public Parameter.Type getType()

Gets the type of this Parameter.

Returns:

the type of this Parameter

setArray

public void setArray(NDArray array)

Sets the values of this Parameter.

Parameters:

array - the NDArray that contains values of this Parameter

setShape

public void setShape(Shape shape)

Sets the shape of this Parameter.

Parameters:

shape - the shape of this Parameter

getShape

public Shape getShape()

Gets the shape of this Parameter.

Returns:

the shape of this Parameter

getArray

public NDArray getArray()

Gets the values of this Parameter as an NDArray.

Returns:

an NDArray that contains values of this Parameter

requiresGradient

public boolean requiresGradient()

Returns whether this parameter needs gradients to be computed.

Returns:

whether this parameter needs gradients to be computed

freeze

public void freeze(boolean freeze)

Freezes or unfreezes the parameter for training.

Sometimes during training, especially during transfer learning, it is typical to train only part of the model. For this, the freeze can be used to prevent certain parts from being trained.

This modifies the requiresGradient() of the parameter.

Parameters:

freeze - true if the parameter should be frozen (freeze == !requiresGradient())

isInitialized

public boolean isInitialized()

Checks if this Parameter is initialized.

Returns:

true if this Parameter is initialized

setInitializer

public void setInitializer(Initializer initializer)

Sets the Initializer for this Parameter, if not already set. If overwrite flag is true, sets the initializer regardless.

Parameters:

initializer - the initializer to be set

getInitializer

public Initializer getInitializer()

Returns the Initializer for this Parameter, if not already set. If overwrite flag is true, sets the initializer regardless.

Returns:

the initializer of this Parameter

initialize

public void initialize(NDManager manager,
                       DataType dataType)

Initializes the parameter with the given NDManager, with given DataType for the given expected input shapes.

Parameters:

manager - an NDManager to create the arrays

dataType - the datatype of the Parameter

save

public void save(java.io.DataOutputStream dos)
          throws java.io.IOException

Writes the parameter NDArrays to the given output stream.

Parameters:

dos - the output stream to write to

Throws:

java.io.IOException - if the write operation fails

load

public void load(NDManager manager,
                 java.io.DataInputStream dis)
          throws java.io.IOException,
                 MalformedModelException

Loads parameter NDArrays from InputStream.

Currently, we cannot deserialize into the exact subclass of NDArray. The SparseNDArray will be loaded as NDArray only.

Parameters:

manager - the NDManager

dis - the InputStream

Throws:

java.io.IOException - if failed to read

MalformedModelException - Exception thrown when model is not in expected format (parameters).

close

public void close()

Specified by:

close in interface java.lang.AutoCloseable

builder

public static Parameter.Builder builder()

Creates a builder to build a Parameter.

The methods start with set are required fields, and opt for optional fields.

Returns:

a new builder