org.deeplearning4j.nn.modelimport.keras

Class KerasModel

java.lang.Object

org.deeplearning4j.nn.modelimport.keras.KerasModel

Direct Known Subclasses:

KerasSequentialModel

public class KerasModel
extends Object

Build ComputationGraph from Keras (Functional API) Model or Sequential model configuration.

Author:

[email protected]

Field Summary

Fields

Modifier and Type	Field and Description
protected String	className
protected boolean	enforceTrainingConfig
static String	HDF5_MODEL_CONFIG_ATTRIBUTE
static String	HDF5_MODEL_WEIGHTS_ROOT
static String	HDF5_TRAINING_CONFIG_ATTRIBUTE
protected ArrayList<String>	inputLayerNames
protected Map<String,KerasLayer>	layers
protected List<KerasLayer>	layersOrdered
static String	MODEL_CLASS_NAME_MODEL
static String	MODEL_CLASS_NAME_SEQUENTIAL
static String	MODEL_CONFIG_FIELD_INPUT_LAYERS
static String	MODEL_CONFIG_FIELD_LAYERS
static String	MODEL_CONFIG_FIELD_OUTPUT_LAYERS
static String	MODEL_FIELD_CLASS_NAME
static String	MODEL_FIELD_CONFIG
protected ArrayList<String>	outputLayerNames
protected Map<String,InputType>	outputTypes
static String	TRAINING_CONFIG_FIELD_LOSS
protected int	truncatedBPTT
protected boolean	useTruncatedBPTT

Constructor Summary

Constructors

Modifier	Constructor and Description
protected	KerasModel()
	KerasModel(org.deeplearning4j.nn.modelimport.keras.KerasModel.ModelBuilder modelBuilder) (Recommended) Builder-pattern constructor for (Functional API) Model.
protected	KerasModel(String modelJson, String modelYaml, Hdf5Archive weightsArchive, String weightsRoot, String trainingJson, Hdf5Archive trainingArchive, boolean enforceTrainingConfig) (Not recommended) Constructor for (Functional API) Model from model configuration (JSON or YAML), training configuration (JSON), weights, and "training mode" boolean indicator.

Method Summary

Modifier and Type	Method and Description
ComputationGraph	getComputationGraph() Build a ComputationGraph from this Keras Model configuration and import weights.
ComputationGraph	getComputationGraph(boolean importWeights) Build a ComputationGraph from this Keras Model configuration and (optionally) import weights.
ComputationGraphConfiguration	getComputationGraphConfiguration() Configure a ComputationGraph from this Keras Model configuration.
protected Model	helperCopyWeightsToModel(Model model) Helper function to import weights from nested Map into existing model.
protected void	helperImportTrainingConfiguration(String trainingConfigJson) Helper method called from constructor.
protected void	helperImportWeights(Hdf5Archive weightsArchive, String weightsRoot) Store weights to import with each associated Keras layer.
protected void	helperInferOutputTypes() Helper method called from constructor.
protected void	helperPrepareLayers(List<Object> layerConfigs) Helper method called from constructor.
protected List<String>	helperRecurseWeightsArchive(Hdf5Archive weightsArchive, String weightsRoot, String layerName)
static Map<String,Object>	parseJsonString(String json) Convenience function for parsing JSON strings.
static Map<String,Object>	parseYamlString(String json) Convenience function for parsing JSON strings.

Methods inherited from class java.lang.Object

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

Field Detail

MODEL_FIELD_CLASS_NAME

public static final String MODEL_FIELD_CLASS_NAME

See Also:

Constant Field Values

MODEL_CLASS_NAME_SEQUENTIAL

public static final String MODEL_CLASS_NAME_SEQUENTIAL

See Also:

Constant Field Values

MODEL_CLASS_NAME_MODEL

public static final String MODEL_CLASS_NAME_MODEL

See Also:

Constant Field Values

MODEL_FIELD_CONFIG

public static final String MODEL_FIELD_CONFIG

See Also:

Constant Field Values

MODEL_CONFIG_FIELD_LAYERS

public static final String MODEL_CONFIG_FIELD_LAYERS

See Also:

Constant Field Values

MODEL_CONFIG_FIELD_INPUT_LAYERS

public static final String MODEL_CONFIG_FIELD_INPUT_LAYERS

See Also:

Constant Field Values

MODEL_CONFIG_FIELD_OUTPUT_LAYERS

public static final String MODEL_CONFIG_FIELD_OUTPUT_LAYERS

See Also:

Constant Field Values

TRAINING_CONFIG_FIELD_LOSS

public static final String TRAINING_CONFIG_FIELD_LOSS

See Also:

Constant Field Values

HDF5_MODEL_WEIGHTS_ROOT

public static final String HDF5_MODEL_WEIGHTS_ROOT

See Also:

Constant Field Values

HDF5_MODEL_CONFIG_ATTRIBUTE

public static final String HDF5_MODEL_CONFIG_ATTRIBUTE

See Also:

Constant Field Values

HDF5_TRAINING_CONFIG_ATTRIBUTE

public static final String HDF5_TRAINING_CONFIG_ATTRIBUTE

See Also:

Constant Field Values

className

protected String className

enforceTrainingConfig

protected boolean enforceTrainingConfig

layersOrdered

protected List<KerasLayer> layersOrdered

layers

protected Map<String,KerasLayer> layers

outputTypes

protected Map<String,InputType> outputTypes

inputLayerNames

protected ArrayList<String> inputLayerNames

outputLayerNames

protected ArrayList<String> outputLayerNames

useTruncatedBPTT

protected boolean useTruncatedBPTT

truncatedBPTT

protected int truncatedBPTT

Constructor Detail

KerasModel

public KerasModel(org.deeplearning4j.nn.modelimport.keras.KerasModel.ModelBuilder modelBuilder)
           throws UnsupportedKerasConfigurationException,
                  IOException,
                  InvalidKerasConfigurationException

(Recommended) Builder-pattern constructor for (Functional API) Model.

Parameters:

modelBuilder - builder object

Throws:

IOException

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

KerasModel

protected KerasModel(String modelJson,
                     String modelYaml,
                     Hdf5Archive weightsArchive,
                     String weightsRoot,
                     String trainingJson,
                     Hdf5Archive trainingArchive,
                     boolean enforceTrainingConfig)
              throws IOException,
                     InvalidKerasConfigurationException,
                     UnsupportedKerasConfigurationException

(Not recommended) Constructor for (Functional API) Model from model configuration (JSON or YAML), training configuration (JSON), weights, and "training mode" boolean indicator. When built in training mode, certain unsupported configurations (e.g., unknown regularizers) will throw Exceptions. When enforceTrainingConfig=false, these will generate warnings but will be otherwise ignored.

Parameters:

modelJson - model configuration JSON string

modelYaml - model configuration YAML string

enforceTrainingConfig - whether to enforce training-related configurations

Throws:

IOException

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

KerasModel

protected KerasModel()

Method Detail

helperPrepareLayers

protected void helperPrepareLayers(List<Object> layerConfigs)
                            throws InvalidKerasConfigurationException,
                                   UnsupportedKerasConfigurationException

Helper method called from constructor. Converts layer configuration JSON into KerasLayer objects.

Parameters:

layerConfigs - List of Keras layer configurations

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

helperImportTrainingConfiguration

protected void helperImportTrainingConfiguration(String trainingConfigJson)
                                          throws IOException,
                                                 InvalidKerasConfigurationException,
                                                 UnsupportedKerasConfigurationException

Helper method called from constructor. Incorporate training configuration details into model. Includes loss function, optimization details, etc.

Parameters:

trainingConfigJson - JSON containing Keras training configuration

Throws:

IOException

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

helperInferOutputTypes

protected void helperInferOutputTypes()
                               throws InvalidKerasConfigurationException,
                                      UnsupportedKerasConfigurationException

Helper method called from constructor. Infers and records output type for every layer.

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

helperRecurseWeightsArchive

protected List<String> helperRecurseWeightsArchive(Hdf5Archive weightsArchive,
                                                   String weightsRoot,
                                                   String layerName)

helperImportWeights

protected void helperImportWeights(Hdf5Archive weightsArchive,
                                   String weightsRoot)
                            throws InvalidKerasConfigurationException,
                                   UnsupportedKerasConfigurationException

Store weights to import with each associated Keras layer.

Parameters:

weightsArchive - Hdf5Archive

weightsRoot -

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

getComputationGraphConfiguration

public ComputationGraphConfiguration getComputationGraphConfiguration()
                                                               throws InvalidKerasConfigurationException,
                                                                      UnsupportedKerasConfigurationException

Configure a ComputationGraph from this Keras Model configuration.

Returns:

ComputationGraph

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

getComputationGraph

public ComputationGraph getComputationGraph()
                                     throws InvalidKerasConfigurationException,
                                            UnsupportedKerasConfigurationException

Build a ComputationGraph from this Keras Model configuration and import weights.

Returns:

ComputationGraph

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

getComputationGraph

public ComputationGraph getComputationGraph(boolean importWeights)
                                     throws InvalidKerasConfigurationException,
                                            UnsupportedKerasConfigurationException

Build a ComputationGraph from this Keras Model configuration and (optionally) import weights.

Parameters:

importWeights - whether to import weights

Returns:

ComputationGraph

Throws:

InvalidKerasConfigurationException

UnsupportedKerasConfigurationException

parseJsonString

public static Map<String,Object> parseJsonString(String json)
                                          throws IOException

Convenience function for parsing JSON strings.

Parameters:

json - String containing valid JSON

Returns:

Nested (key,value) map of arbitrary depth

Throws:

IOException

parseYamlString

public static Map<String,Object> parseYamlString(String json)
                                          throws IOException

Convenience function for parsing JSON strings.

Parameters:

json - String containing valid JSON

Returns:

Nested (key,value) map of arbitrary depth

Throws:

IOException

helperCopyWeightsToModel

protected Model helperCopyWeightsToModel(Model model)
                                  throws InvalidKerasConfigurationException

Helper function to import weights from nested Map into existing model. Depends critically on matched layer and parameter names. In general this seems to be straightforward for most Keras models and layersOrdered, but there may be edge cases.

Parameters:

model - DL4J Model interface

Returns:

DL4J Model interface

Throws:

InvalidKerasConfigurationException