com.intel.analytics.zoo.models.seq2seq.RNNEncoder.scala Maven / Gradle / Ivy

Go to download
Show more of this group Show more artifacts with this name
Show all versions of analytics-zoo-bigdl_0.9.1-spark_2.2.1 Show documentation
The newest version!
/*
 * Copyright 2018 Analytics Zoo Authors.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

package com.intel.analytics.zoo.models.seq2seq

import com.intel.analytics.bigdl.nn.abstractnn.{AbstractModule, Activity}
import com.intel.analytics.bigdl.nn.keras.{KerasLayer, KerasLayerSerializable}
import com.intel.analytics.bigdl.serialization.Bigdl.{AttrValue, BigDLModule}
import com.intel.analytics.bigdl.tensor.Tensor
import com.intel.analytics.bigdl.tensor.TensorNumericMath.TensorNumeric
import com.intel.analytics.bigdl.utils.{Shape, _}
import com.intel.analytics.bigdl.utils.serializer.{DeserializeContext, ModuleSerializable, ModuleSerializer, SerializeContext}
import com.intel.analytics.bigdl.utils.serializer.converters.DataConverter
import com.intel.analytics.bigdl.utils.serializer.converters.DataConverter.ArrayConverter
import com.intel.analytics.zoo.pipeline.api.Net
import com.intel.analytics.zoo.pipeline.api.keras.layers.utils.KerasUtils
import com.intel.analytics.zoo.pipeline.api.keras.layers._
import com.intel.analytics.zoo.pipeline.api.keras.models.Sequential

import scala.collection.mutable.ArrayBuffer
import scala.reflect.ClassTag
import scala.reflect.runtime._

/**
 * [[RNNEncoder]] A generic recurrent neural network encoder
 *
 * @param rnns rnn layers used for encoder, support stacked rnn layers
 * @param embedding embedding layer in encoder
 * @param inputShape shape of input
 */
class RNNEncoder[T: ClassTag](val rnns: Array[Recurrent[T]],
  val embedding: KerasLayer[Tensor[T], Tensor[T], T],
  var inputShape: Shape = null)(implicit ev: TensorNumeric[T])
  extends Encoder[T](inputShape) {

  override def doBuild(inputShape: Shape): AbstractModule[Tensor[T], Activity, T] = {
    val layer = Sequential()
    layer.add(InputLayer(KerasUtils.removeBatch(inputShape)))
    if (embedding != null) {
      layer.add(embedding)
    }
    rnns.foreach(layer.add(_))
    layer.asInstanceOf[AbstractModule[Tensor[T], Activity, T]]
  }

  override def computeOutputShape(inputShape: Shape): Shape = {
    // batch x seq x hidden
    val rnnShape = labor.getOutputShape()
    val sizes = rnnShape.toSingle().toArray

    val statesShape = rnns.map { rnn =>
      MultiShape(rnn.getHiddenShape().map(x => Shape(Array(sizes(0), x) ++ sizes.drop(3))).toList)
    }

    Shape(List(rnnShape, MultiShape(statesShape.toList)))
  }

  // ouput is T(rnnOutput, T(layer1states, layer2states, ...))
  override def updateOutput(input: Tensor[T]): Activity = {
    val laborOutput = labor.updateOutput(input)
    val states = rnns.map(_.getHiddenState())

    output = T(laborOutput, T.array(states))
    output
  }

  override def updateGradInput(input: Tensor[T], gradOutput: Activity): Tensor[T] = {
    val rnnGradOutput = gradOutput.toTable[Tensor[T]](1)
    val gradStates = gradOutput.toTable[Table](2)

    var i = 0
    while (i < rnns.size) {
      rnns(i).setGradHiddenState(gradStates(i + 1))
      i += 1
    }
    gradInput = labor.updateGradInput(input, rnnGradOutput)
    gradInput
  }

  override def accGradParameters(input: Tensor[T], gradOutput: Activity): Unit = {
    val rnnGradOutput = gradOutput.toTable[Tensor[T]](1)
    val gradStates = gradOutput.toTable[Table](2)

    var i = 0
    while (i < rnns.size) {
      rnns(i).setGradHiddenState(gradStates(i + 1))
      i += 1
    }
    labor.accGradParameters(input, rnnGradOutput)
  }
}

object RNNEncoder extends KerasLayerSerializable {
  ModuleSerializer.registerModule(
    "com.intel.analytics.zoo.models.seq2seq.RNNEncoder",
    RNNEncoder)

  /**
   * [[RNNEncoder]] A generic recurrent neural network encoder
   *
   * @param rnns rnn layers used for encoder, support stacked rnn layers
   * @param embedding embedding layer in encoder
   * @param inputShape shape of input
   */
  def apply[@specialized(Float, Double) T: ClassTag](rnns: Array[Recurrent[T]],
    embedding: KerasLayer[Tensor[T], Tensor[T], T],
    inputShape: Shape)(implicit ev: TensorNumeric[T]): RNNEncoder[T] = {
    new RNNEncoder[T](rnns, embedding, inputShape)
  }

  /**
   * [[RNNEncoder]] A generic recurrent neural network encoder
   *
   * @param rnnType style of recurrent unit, one of [SimpleRNN, LSTM, GRU]
   * @param numLayers number of layers used in encoder
   * @param hiddenSize hidden size of encoder
   * @param embedding embedding layer in encoder
   * @param inputShape shape of input
   */
  def apply[@specialized(Float, Double) T: ClassTag](rnnType: String,
    numLayers: Int,
    hiddenSize: Int,
    embedding: KerasLayer[Tensor[T], Tensor[T], T] = null,
    inputShape: Shape = null)(implicit ev: TensorNumeric[T]): RNNEncoder[T] = {
    val rnn = new ArrayBuffer[Recurrent[T]]()
    rnnType.toLowerCase() match {
      case "lstm" =>
        for (i <- 1 to numLayers) rnn.append(LSTM(hiddenSize, returnSequences = true))
      case "gru" =>
        for (i <- 1 to numLayers) rnn.append(GRU(hiddenSize, returnSequences = true))
      case "simplernn" =>
        for (i <- 1 to numLayers) rnn.append(SimpleRNN(hiddenSize, returnSequences = true))
      case _ => throw new IllegalArgumentException(s"Please use " +
        s"RNNEncoder(rnn: Array[Recurrent[T]], embedding: KerasLayer[Activity, Activity, T])" +
        s"to create a encoder")
    }
    RNNEncoder[T](rnn.toArray, embedding, inputShape)
  }

  override def doLoadModule[T: ClassTag](context : DeserializeContext)
    (implicit ev: TensorNumeric[T]) : AbstractModule[Activity, Activity, T] = {

    val attrMap = context.bigdlModule.getAttrMap

    val rnns = DataConverter.getAttributeValue(context, attrMap.get("rnns")).
      asInstanceOf[Array[AbstractModule[_, _, T]]].map(_.asInstanceOf[Recurrent[T]])
    rnns.map(_.labor = null)
    rnns.map(_.modules.clear())

    val embeddingAttr = attrMap.get("embedding")
    val embedding = DataConverter.getAttributeValue(context, embeddingAttr).
      asInstanceOf[KerasLayer[Tensor[T], Tensor[T], T]]
    if (embedding != null) {
      embedding.labor = null
      embedding.modules.clear()
    }

    val shapeAttr = attrMap.get("shape")
    val shape = DataConverter.getAttributeValue(context, shapeAttr).asInstanceOf[Shape]
    val encoder = RNNEncoder(rnns, embedding, shape)

    encoder.build(KerasUtils.addBatch(shape))
    encoder.asInstanceOf[AbstractModule[Activity, Activity, T]]
  }

  override def doSerializeModule[T: ClassTag](context: SerializeContext[T],
    encoderBuilder : BigDLModule.Builder)
    (implicit ev: TensorNumeric[T]) : Unit = {

    val encoder = context.moduleData.module.asInstanceOf[RNNEncoder[T]]
    require(encoder.inputShape != null, "Encoder serialize cannot work without inputShape")

    val rnnsBuilder = AttrValue.newBuilder
    ArrayConverter.setAttributeValue(context, rnnsBuilder,
      context.moduleData.module.asInstanceOf[RNNEncoder[T]].rnns,
      scala.reflect.runtime.universe.typeOf[Array[_ <:
        AbstractModule[_ <: Activity, _ <:  Activity, _ <: Any]]])
    encoderBuilder.putAttr("rnns", rnnsBuilder.build)

    val embeddingBuilder = AttrValue.newBuilder
    DataConverter.setAttributeValue(context, embeddingBuilder,
      encoder.embedding, ModuleSerializer.abstractModuleType)
    encoderBuilder.putAttr("embedding", embeddingBuilder.build)

    val shapeBuilder = AttrValue.newBuilder
    DataConverter.setAttributeValue(context, shapeBuilder,
      encoder.inputShape, universe.typeOf[Shape])
    encoderBuilder.putAttr("shape", shapeBuilder.build)

    appendKerasLabel(context, encoderBuilder)
  }
}