GraphTrainableModel

abstract class GraphTrainableModel(layers: Layer) : TrainableModel

GraphTrainableModel model groups a linear stack of layers into a graph-based TensorFlow Model. Also, it provides training and inference features on this model.

Constructors

GraphTrainableModel

fun GraphTrainableModel(vararg layers: Layer)

Creates a GraphTrainableModel model via sequence of layers.

Functions

open override fun close()

Closes internal resources: session and kGraph.

compile

open override fun compile(optimizer: Optimizer, loss: LossFunction, metrics: List<Metric>)

open override fun compile(optimizer: Optimizer, loss: LossFunction, metric: Metric)

open override fun compile(optimizer: Optimizer, loss: LossFunction, metric: Metrics)

open override fun compile(optimizer: Optimizer, loss: Losses, metric: Metric)

open override fun compile(optimizer: Optimizer, loss: Losses, metric: Metrics)

Configures the model for training.

copy

open override fun copy(copiedModelName: String?, saveOptimizerState: Boolean, copyWeights: Boolean): TensorFlowInferenceModel

Creates a copy.

evaluate

fun evaluate(dataset: Dataset, metric: Metrics): Double

Evaluates dataset via metric.

open override fun evaluate(dataset: Dataset, batchSize: Int, callbacks: List<Callback>): EvaluationResult

fun evaluate(dataset: Dataset, batchSize: Int = 256, callback: Callback): EvaluationResult

Returns the metrics and loss values for the model in test (evaluation) mode.

fit

open override fun fit(dataset: Dataset, epochs: Int, batchSize: Int, callbacks: List<Callback>): TrainingHistory

open override fun fit(trainingDataset: Dataset, validationDataset: Dataset, epochs: Int, trainBatchSize: Int, validationBatchSize: Int, callbacks: List<Callback>): TrainingHistory

Trains the model for a fixed number of epochs (iterations over a dataset).

fun fit(dataset: Dataset, epochs: Int = 5, batchSize: Int = 32, callback: Callback): TrainingHistory

fun fit(trainingDataset: Dataset, validationDataset: Dataset, epochs: Int = 5, trainBatchSize: Int = 32, validationBatchSize: Int = 256, callback: Callback): TrainingHistory

Trains the model for a fixed number of epochs (iterations over a dataset).

fun fit(dataset: OnHeapDataset, validationRate: Double, epochs: Int, trainBatchSize: Int, validationBatchSize: Int, callbacks: List<Callback> = listOf()): TrainingHistory

fun fit(dataset: OnHeapDataset, validationRate: Double, epochs: Int, trainBatchSize: Int, validationBatchSize: Int, callback: Callback): TrainingHistory

Trains the model for a fixed number of epochs (iterations on a dataset).

getLayer

infix fun getLayer(layerName: String): Layer

Return layer by layerName.

graphToString

fun graphToString(): String

Forms the graph description in string format.

init

fun init()

Initializes kGraph variables.

input

fun input(inputOp: Input)

Chain-like setter to set up inputOp.

kGraph

fun kGraph(): KGraph

Returns KGraph.

loadWeights

open override fun loadWeights(modelDirectory: File, loadOptimizerState: Boolean)

Loads variable data from .txt files.

output

fun output(outputOp: Output)

Chain-like setter to set up outputOp.

predict

open override fun predict(inputData: FloatArray): Int

Generates output prediction for the input sample.

fun predict(dataset: Dataset): List<Int>

Predicts labels for all observation in dataset.

open override fun predict(inputData: FloatArray, predictionTensorName: String): Int

Generates output prediction for the input sample using output of the predictionTensorName tensor.

open override fun predict(dataset: Dataset, batchSize: Int, callbacks: List<Callback>): IntArray

fun predict(dataset: Dataset, batchSize: Int, callback: Callback): IntArray

Generates output predictions for the input samples.

predictAndGetActivations

open override fun predictAndGetActivations(inputData: FloatArray, predictionTensorName: String): Pair<Int, List<*>>

Predicts and returns not only prediction but list of activations values from intermediate model layers (for visualisation or debugging purposes).

predictSoftly

open override fun predictSoftly(inputData: FloatArray, predictionTensorName: String): FloatArray

Predicts vector of probabilities instead of specific class in predict method.

open override fun predictSoftly(dataset: Dataset, batchSize: Int, callbacks: List<Callback>): Array<FloatArray>

Generates output predictions for the input samples. Each prediction is a vector of probabilities instead of specific class in predict method.

fun predictSoftly(dataset: Dataset, batchSize: Int, callback: Callback): Array<FloatArray>

Generates output predictions for the input samples. Each prediction is a vector of probabilities instead of specific class in predict method.

reset

fun reset()

It ignores that model is initialized already and call initializers under the hood to re-initialize kGraph variables.

reshape

open override fun reshape(vararg dims: Long)

Chain-like setter to set up input shape.

save

open override fun save(modelDirectory: File, savingFormat: SavingFormat, saveOptimizerState: Boolean, writingMode: WritingMode)

Saves the model as graph and weights.

summary

open override fun summary(): ModelSummary

Returns model summary.

toString

open override fun toString(): String

Properties

inputDimensions

open override val inputDimensions: LongArray

Returns input dimensions in order HWC (height, width, channels)

inputLayer

val inputLayer: Input

First layer that is responsible for the input shape of the Neural Network.

isBuiltForForwardMode

var isBuiltForForwardMode: Boolean = false

Is true when model is ready for forward mode.

isModelCompiled

var isModelCompiled: Boolean = false

Is true when model is compiled.

isModelInitialized

var isModelInitialized: Boolean = false

Is true when model is initialized.

isOptimizerVariableInitialized

var isOptimizerVariableInitialized: Boolean = false

Is true when model optimizer variables are initialized.

kGraph

lateinit var kGraph: KGraph

TensorFlow wrapped computational graph.

layers

var layers: List<Layer>

The layers to describe the model design. Main part of the internal state of the model.

logger

val logger: KLogger

Logger for the model.

loss

var loss: LossFunction

Loss function.

name

var name: String? = null

Model name.

numberOfClasses

var numberOfClasses: Long

Number of classes for classification tasks. -1 is a default value for regression tasks.

shape

lateinit var shape: LongArray

Data shape for prediction.

stopTraining

var stopTraining: Boolean = false

Special flag for callbacks.

Inheritors

Functional

Sequential

Extensions

freeze

fun GraphTrainableModel.freeze()

Freezes weights in all layers in this model, so they won't be changed during training.

loadWeights

fun GraphTrainableModel.loadWeights(hdfFile: HdfFile)

Loads weights from hdf5 file created in Keras TensorFlow framework.

fun GraphTrainableModel.loadWeights(hdfFile: HdfFile, layerList: List<Layer>)

Loads weights from hdf5 file created in Keras TensorFlow framework for pre-defined list of layers.

loadWeightsByPaths

fun GraphTrainableModel.loadWeightsByPaths(hdfFile: HdfFile, weightPaths: List<LayerPaths>, missedWeights: MissedWeightsStrategy = MissedWeightsStrategy.INITIALIZE, forFrozenLayersOnly: Boolean = false)

Loads weights from hdf5 file created in Keras TensorFlow framework.

fun GraphTrainableModel.loadWeightsByPaths(hdfFile: HdfFile, layerList: List<Layer>, kernelDataPathTemplate: String = KERNEL_DATA_PATH_TEMPLATE, biasDataPathTemplate: String = BIAS_DATA_PATH_TEMPLATE)

Loads weights from hdf5 file created in Keras TensorFlow framework for pre-defined list of layers.

loadWeightsByPathTemplates

fun GraphTrainableModel.loadWeightsByPathTemplates(hdfFile: HdfFile, kernelDataPathTemplate: String = KERNEL_DATA_PATH_TEMPLATE, biasDataPathTemplate: String = BIAS_DATA_PATH_TEMPLATE)

Loads weights from hdf5 file created in Keras TensorFlow framework.

fun GraphTrainableModel.loadWeightsByPathTemplates(hdfFile: HdfFile, layerList: List<Layer>, kernelDataPathTemplate: String = KERNEL_DATA_PATH_TEMPLATE, biasDataPathTemplate: String = BIAS_DATA_PATH_TEMPLATE)

Loads weights from hdf5 file created in Keras TensorFlow framework for pre-defined list of layers.

loadWeightsForFrozenLayers

fun GraphTrainableModel.loadWeightsForFrozenLayers(hdfFile: HdfFile)

Loads weights from hdf5 file created in Keras TensorFlow framework for non-trainable (or frozen) layers only.

loadWeightsForFrozenLayersByPathTemplates

fun GraphTrainableModel.loadWeightsForFrozenLayersByPathTemplates(hdfFile: HdfFile, kernelDataPathTemplate: String = KERNEL_DATA_PATH_TEMPLATE, biasDataPathTemplate: String = BIAS_DATA_PATH_TEMPLATE)

Loads weights from hdf5 file created in Keras TensorFlow framework for non-trainable (or frozen) layers only.

saveModelConfiguration

fun GraphTrainableModel.saveModelConfiguration(jsonConfigFile: File, isKerasFullyCompatible: Boolean = false)

Saves model description as json configuration file fully compatible with the Keras TensorFlow framework.