com.kotlinnlp.simplednn.deeplearning.mergelayers.biaffine.BiaffineLayerStructure.kt Maven / Gradle / Ivy

Go to download

Show more of this group Show more artifacts with this name
Show all versions of simplednn Show documentation

SimpleDNN is a machine learning lightweight open-source library written in Kotlin whose purpose is to support the development of feed-forward and recurrent Artificial Neural Networks.

There is a newer version: 0.14.0

Show newest version

/* Copyright 2016-present The KotlinNLP Authors. All Rights Reserved.
 *
 * This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, you can obtain one at http://mozilla.org/MPL/2.0/.
 * ------------------------------------------------------------------*/

package com.kotlinnlp.simplednn.deeplearning.mergelayers.biaffine

import com.kotlinnlp.simplednn.core.functionalities.activations.ActivationFunction
import com.kotlinnlp.simplednn.core.arrays.AugmentedArray
import com.kotlinnlp.simplednn.deeplearning.mergelayers.MergeLayer
import com.kotlinnlp.simplednn.simplemath.ndarray.NDArray
import com.kotlinnlp.simplednn.simplemath.ndarray.Shape
import com.kotlinnlp.simplednn.simplemath.ndarray.dense.DenseNDArray
import com.kotlinnlp.simplednn.simplemath.ndarray.dense.DenseNDArrayFactory
import com.kotlinnlp.simplednn.utils.ItemsPool

/**
 * The Biaffine Layer Structure.
 * WARNING: actually the forward and backward operations are NOT OPTIMIZED for sparse inputs.
 *
 * @property inputArray the first input array of the layer
 * @property inputArray2 the second input array of the layer
 * @property outputArray the output array of the layer
 * @property params the parameters which connect the input to the output
 * @property activationFunction the activation function of the layer
 * @property dropout the probability of dropout (default 0.0).
 *                   If applying it, the usual value is 0.5 (better 0.25 if it's the first layer).
 * @property id a unique id for this item (default = 0)
 *
 */
class BiaffineLayerStructure>(
  inputArray1: AugmentedArray,
  inputArray2: AugmentedArray,
  outputArray: AugmentedArray,
  override val params: BiaffineLayerParameters,
  activationFunction: ActivationFunction? = null,
  dropout: Double = 0.0,
  override val id: Int = 0
) :
  ItemsPool.IDItem,
  MergeLayer(
    inputArray1 = inputArray1,
    inputArray2 = inputArray2,
    outputArray = outputArray,
    params = params,
    activationFunction = activationFunction,
    dropout = dropout) {

  /**
   * Constructor by params.
   *
   * @property params the parameters which connect the input to the output
   * @property activationFunction the activation function of the layer
   * @property dropout the probability of dropout (default 0.0).
   *                   If applying it, the usual value is 0.5 (better 0.25 if it's the first layer).
   * @property id a unique id for this item (default = 0)
   */
  constructor(params: BiaffineLayerParameters,
              activationFunction: ActivationFunction? = null,
              dropout: Double = 0.0,
              id: Int = 0): this(
    inputArray1 = AugmentedArray(size = params.inputSize1),
    inputArray2 = AugmentedArray(size = params.inputSize2),
    outputArray = AugmentedArray(size = params.outputSize),
    params = params,
    activationFunction = activationFunction,
    dropout = dropout,
    id = id
  )

  /**
   * A support structure used for calculations. Each array wx1i is obtained by: wi (dot) x1.
   */
  val wx1Arrays: Array = Array(
    size = this.params.outputSize,
    init = { DenseNDArrayFactory.emptyArray(Shape(this.params.inputSize2)) }
  )

  /**
   * The helper which execute the forward.
   */
  override val forwardHelper = BiaffineForwardHelper(layer = this)

  /**
   * The helper which execute the backward.
   */
  override val backwardHelper = BiaffineBackwardHelper(layer = this)

  /**
   * The helper which calculates the relevance.
   */
  override val relevanceHelper = BiaffineRelevanceHelper(layer = this)

  /**
   * Initialization: set the activation function of the outputArray.
   */
  init {
    if (activationFunction != null) {
      outputArray.setActivation(activationFunction)
    }
  }

  /**
   * @return the [BiaffineLayerParameters] used to store errors
   */
  override fun parametersErrorsFactory() = BiaffineLayerParameters(
    inputSize1 = this.params.inputSize1,
    inputSize2 = this.params.inputSize2,
    outputSize = this.params.outputSize,
    sparseInput = this.params.sparseInput,
    weightsInitializer = null,
    biasesInitializer = null
  )
}