All Downloads are FREE. Search and download functionalities are using the official Maven repository.

com.intel.analytics.bigdl.dllib.optim.DistriOptimizer.scala Maven / Gradle / Ivy

The newest version!
/*
 * Copyright 2016 The BigDL 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.bigdl.dllib.optim

import com.intel.analytics.bigdl.dllib.feature.dataset.{DistributedDataSet, MiniBatch, PaddingParam, Sample}
import com.intel.analytics.bigdl.dllib.models.utils.{CachedModels, ModelBroadcast}
import com.intel.analytics.bigdl.dllib.nn.mkldnn.Phase.TrainingPhase
import com.intel.analytics.bigdl.dllib.nn.mkldnn.{DnnGraph, MklDnnContainer, MklDnnModule}
import com.intel.analytics.bigdl.dllib.nn.{Container, Graph, Module, Utils}
import com.intel.analytics.bigdl.dllib.optim.parameters.{AllReduceParameter, ParameterProcessor}
import com.intel.analytics.bigdl.dllib.tensor.Tensor
import com.intel.analytics.bigdl.dllib.tensor.TensorNumericMath.TensorNumeric
import com.intel.analytics.bigdl.dllib.utils._
import com.intel.analytics.bigdl.dllib.utils.intermediate.{ConversionUtils, IRGraph}
import com.intel.analytics.bigdl.dllib.visualization.{TrainSummary, ValidationSummary}
import com.intel.analytics.bigdl.dllib.utils._
import com.intel.analytics.bigdl.{Module, _}
import java.io.{File, FilenameFilter}
import java.text.SimpleDateFormat
import java.util.Calendar

import org.apache.logging.log4j.{LogManager, Logger}
import org.apache.spark.TaskContext
import org.apache.spark.rdd.RDD
import org.apache.spark.util.AccumulatorV2

import scala.collection.mutable
import scala.collection.mutable.ArrayBuffer
import scala.concurrent.Future
import scala.reflect.ClassTag

object DistriOptimizer extends AbstractOptimizer {

  import Optimizer._

  val logger: Logger = LogManager.getLogger(getClass)

  /**
   * Optimizer cache some metadata on each executor
   *
   * @param localModels cached models
   * @param modelWeights weights of the cached models
   * @param modelGradients gradients of the cached models
   * @param localCriterions cached criterion
   * @param localStates cached state
   * @param moduleTimeList module running time
   * @param localMethods cached validation methods
   * @param optimMethods cached optim methods
   * @param parameterSynchronizer cached parameter synchronizer
   * @tparam T Tensor element type
   */

  abstract class Cache[T] {
    def localModels: Array[Module[T]]
    def modelWeights: Array[Tensor[T]]
    def modelGradients: Array[Tensor[T]]
    def localCriterions: Array[Criterion[T]]
    def localMethods: Array[Option[Array[ValidationMethod[T]]]]
    def optimMethods: Map[String, OptimMethod[T]]
    def moduleTimeList: Array[Long]
    def parameterSynchronizer: DistriParameterSynchronizer[T]
  }

  case class CacheV1[T](
    localModels: Array[Module[T]],
    modelWeights: Array[Tensor[T]],
    modelGradients: Array[Tensor[T]],
    localCriterions: Array[Criterion[T]],
    localStates: Array[Table],
    var moduleTimeList: Array[Long] = null,
    localMethods: Array[Option[Array[ValidationMethod[T]]]],
    var optimMethods: Map[String, OptimMethod[T]],
    parameterSynchronizer: DistriParameterSynchronizer[T] = null
  ) extends Cache[T]

  /**
   * Train the model.
   *
   * @param dataset train dataset
   * @param coresPerNode cores per node
   * @param state state table
   * @param endWhen trigger to stop training
   * @param metrics metrics
   * @param models cached models
   * @param optimMethods optimization methods
   * @param parameters [[AllReduceParameter]]
   * @param parameterSplits the segments of parameters (offset, length)
   * @param validationTrigger validation trigger
   * @param validationDataSet validation dataset
   * @param validationMethods validation methods
   * @param cacheTrigger cache trigger
   * @param cachePath cache path
   * @param trainSummary train summary
   * @param validationSummary validation summary
   * @param isOverWrite if overwrite the checkpoint
   * @param parameterProcessers a list of ParameterProcessor used to process parameters
   */
  private[optim] def optimize[T: ClassTag](
    trainingModel: Module[T],
    dataset: DistributedDataSet[MiniBatch[T]],
    coresPerNode: Int,
    state: Table,
    endWhen: Trigger,
    metrics: Metrics,
    models: RDD[CacheV1[T]],
    optimMethods: Map[String, OptimMethod[T]],
    parameters: AllReduceParameter[T],
    parameterSplits: Map[String, (Int, Int)],
    validationTrigger: Option[Trigger],
    validationDataSet: Option[DataSet[MiniBatch[T]]],
    validationMethods: Option[Array[ValidationMethod[T]]],
    cacheTrigger: Option[Trigger],
    cachePath: Option[String],
    trainSummary: Option[TrainSummary],
    validationSummary: Option[ValidationSummary],
    isOverWrite: Boolean,
    parameterProcessers: Array[ParameterProcessor]
  )(implicit ev: TensorNumeric[T]): Unit = {
    val sc = dataset.originRDD().sparkContext
    val partitionNum = dataset.originRDD().partitions.length
    var wallClockTime = 0L
    var lastEpochTime = 0L

    // driverState is needed to prevent serializing the whole optimizer
    optimMethods.values.foreach { optimMethod =>
      if (!optimMethod.state.contains("epoch")) optimMethod.state.update("epoch", 1)
      if (!optimMethod.state.contains("neval")) optimMethod.state.update("neval", 1)
      if (!optimMethod.state.contains("Loss")) {
        optimMethod.state.update("Loss", Float.PositiveInfinity)
      }
      if (!optimMethod.state.contains("score")) optimMethod.state.update("score", 0f)
      if (!optimMethod.state.contains("recordsProcessedThisEpoch")) {
        optimMethod.state.update("recordsProcessedThisEpoch", 0)
      }
    }

    val _subModelNumber = Engine.getEngineType() match {
      case MklBlas => coresPerNode
      case MklDnn => 1
    }
    val driverState = T(
      "epoch" -> optimMethods.values.head.state("epoch"),
      "neval" -> optimMethods.values.head.state("neval"),
      "Loss" -> optimMethods.values.head.state("Loss"),
      "score" -> optimMethods.values.head.state("score"),
      "parallelism" -> _subModelNumber
    )

    logger.info("Count dataset")
    val countBefore = System.nanoTime()
    val numSamples = dataset.data(train = false).map(_.size()).reduce(_ + _)
    val countAfter = System.nanoTime()
    logger.info(s"Count dataset complete. Time elapsed: ${(countAfter - countBefore) / 1e9}s")
    if (numSamples != dataset.size()) {
      logger.warn("If the dataset is built directly from RDD[Minibatch], the data in each " +
        "minibatch is fixed, and a single minibatch is randomly selected in each partition. If " +
        "the dataset is transformed from RDD[Sample], each minibatch will be constructed on the " +
        "fly from random samples, which is better for convergence.")
    }

    logger.info(s"config $state")
    var recordsProcessedThisEpoch = optimMethods.values.head.state[Int]("recordsProcessedThisEpoch")
    if (recordsProcessedThisEpoch == 0) {
      val shuffleBefore = System.nanoTime()
      logger.info("Shuffle data")
      dataset.shuffle()
      val shuffleEnd = System.nanoTime()
      logger.info(s"Shuffle data complete. Takes ${(shuffleEnd - shuffleBefore) / 1e9}s")
    }

    var tasks: ArrayBuffer[Future[_]] = new ArrayBuffer()
    var threshold = Long.MaxValue
    var timeout = Long.MaxValue
    var iteration = 0
    val dropPercentage = state.get[Double]("dropPercentage").get
    val warmupIterationNum = state.get[Int]("warmupIterationNum").get
    val computeThresholdbatchSize = state.get[Int]("computeThresholdbatchSize").get
    val maxDropPercentage = state.get[Double]("maxDropPercentage").get
    val driverSubModelNum = partitionNum * _subModelNumber
    var dropModelNumBatch = 0
    var lossArray = new Array[Double](_subModelNumber)

    var epochStart = System.nanoTime()
    var dataRDD = dataset.data(train = true)

    while (!endWhen(driverState)) {
      val lossSum = sc.doubleAccumulator("loss sum")
      val recordsNum = sc.doubleAccumulator("record number")
      metrics.set("computing time for each node", mutable.ArrayBuffer[Double](), sc)
      metrics.set("get weights for each node", mutable.ArrayBuffer[Double](), sc)
      metrics.set("computing time average", 0.0, sc, partitionNum)
      metrics.set("aggregate gradient time", 0.0, sc, partitionNum)
      metrics.set("get weights average", 0.0, sc, partitionNum)
      metrics.set("put gradient", 0.0, sc, Engine.nodeNumber())
      metrics.set("aggregrateGradientParition average executor", 0.0, sc, Engine.nodeNumber())
      metrics.set("compute weight average", 0.0, sc, Engine.nodeNumber())
      metrics.set("send weights average", 0.0, sc, Engine.nodeNumber())

      val driverMetrics = metrics
      val start = System.nanoTime()
      /*
        Run the forwards/backwards pass using multiple threads in each partition, and track the
        number of model updates that finished before the thread timeout mechanism.
       */
      val numFinishedModelUpdates: Int = dataRDD
        .zipPartitions(models, preservesPartitioning = true) { (data, modelIter) => {
          val cached = modelIter.next()
          val syWStart = System.nanoTime()
          /*
            Note: All models in `cached` share the same storage for weights, so we only need to
            copy the weights from parameter server into the first model's weights.
           */
          val weightsResults = parameters.getWeights(cached.modelWeights.head.narrow(1,
            parameters.paramOffset, parameters.size))
          val miniBatchBuffer = new Array[MiniBatch[T]](_subModelNumber)
          val batch = data.next()
          val stackSize = batch.size() / _subModelNumber
          tasks += Engine.default.invoke(() => {
            Log4Error.invalidOperationError((batch.size() >= _subModelNumber) &&
              (batch.size() % _subModelNumber == 0), "total batch size: " +
              s"${batch.size()} should be divided by total core number: ${_subModelNumber}")
            if (batch.size() < _subModelNumber * 2) {
              logger.warn("Warning: for better training speed, " +
                "total batch size is recommended to be at least two times of core number" +
                s"${_subModelNumber}, please tune your batch size accordingly")
            }
            var b = 0
            while (b < _subModelNumber) {
              miniBatchBuffer(b) = batch.slice(b * stackSize + 1, stackSize)
              b += 1
            }
          })
          Engine.default.sync(tasks)
          weightsResults.waitResult()
          val weightSyncTime = System.nanoTime() - syWStart
          driverMetrics.add("get weights average", weightSyncTime)
          driverMetrics.add("get weights for each node", weightSyncTime)
          tasks.clear()

          // ======================Start train models===================================
          var time = System.nanoTime()
          if (dropPercentage > 0.0 && iteration > warmupIterationNum +
            computeThresholdbatchSize - 1) {
            timeout = threshold - weightSyncTime
          }
          val pre = (iteration % computeThresholdbatchSize) * _subModelNumber
          val trainingThreads = Engine.default.invokeAndWait2((0 until _subModelNumber).map(i =>
            () => {
              val trainStart = System.nanoTime()
              val localModel = cached.localModels(i)
              localModel.training()
              val localCriterion = cached.localCriterions(i)
              val input = miniBatchBuffer(i).getInput()
              val target = miniBatchBuffer(i).getTarget()

              if (Engine.getEngineType() == MklBlas) {
                val output = localModel.forward(input)
                lossArray(i) = ev.toType[Double](localCriterion.forward(output, target))
                val errors = localCriterion.backward(output, target)
                localModel.backward(input, errors)
              } else if (localModel.isInstanceOf[IRGraph[T]]) {
                val output = localModel.forward(input)
                Engine.dnnComputing.invokeAndWait2(Array(0).map(_ => () => {
                  lossArray(i) = ev.toType[Double](localCriterion.forward(output, target))
                  localCriterion.backward(output, target)
                }))
                localModel.backward(input, localCriterion.gradInput)
              } else {
                Engine.dnnComputing.invokeAndWait2(Array(0).map(_ => () => {
                  val output = localModel.forward(input)
                  lossArray(i) = ev.toType[Double](localCriterion.forward(output, target))
                  val errors = localCriterion.backward(output, target)
                  localModel.backward(input, errors)
                }))
              }
              cached.moduleTimeList(i + pre) = System.nanoTime() - trainStart + weightSyncTime
              i
            }
          ), timeout)
          val computingTime = System.nanoTime() - time
          driverMetrics.add("computing time average", computingTime)
          driverMetrics.add("computing time for each node", computingTime)

          val finishedThreads = trainingThreads.filter(!_.isCancelled).map(_.get())
          recordsNum.add(finishedThreads.size * stackSize)
          var i = 0
          while (i < finishedThreads.size) {
            lossSum.add(lossArray(finishedThreads(i)))
            i += 1
          }

          if (finishedThreads.nonEmpty) {
            val finishedGradients = finishedThreads.map(cached.modelGradients(_))

            time = System.nanoTime()
            val pOffset = parameters.paramOffset
            val pLength = parameters.size
            val taskSize = pLength / _subModelNumber
            val extraTask = pLength % _subModelNumber

            // Aggregate multi-model's gradient to the first model's gradient
            val parallelNum = if (taskSize == 0) extraTask else _subModelNumber
            if (parallelNum != 1) {
              Engine.default.invokeAndWait((0 until parallelNum).map(tid => () => {
                val offset = pOffset + tid * taskSize + math.min(tid, extraTask)
                val length = taskSize + (if (tid < extraTask) 1 else 0)
                var i = 1
                while (i < finishedGradients.length) {
                  finishedGradients(0).narrow(1, offset, length)
                    .add(finishedGradients(i).narrow(1, offset, length))
                  i += 1
                }
              }))
              driverMetrics.add("aggregate gradient time", System.nanoTime() - time)
            }
            val putG = System.nanoTime()
            // Put first finished model's gradient who aggregated
            // all other models' gradient to AllReduceParameter
            parameters.putGradients(finishedGradients(0).narrow(1, pOffset, pLength))
            driverMetrics.add("put gradient", System.nanoTime() - putG)

          } else {
            val putG = System.nanoTime()
            // zero gradient in BlockManager when no thread finished.
            cached.modelGradients(0).zero()
            parameters.putGradients(cached.modelGradients(0).narrow(1, parameters.paramOffset,
              parameters.size))
            driverMetrics.add("put gradient", System.nanoTime() - putG)
          }

          tasks ++= Engine.default.invoke {
            (0 until _subModelNumber).map { i =>
              () => {
                cached.localModels(i).training()
                cached.localModels(i).zeroGradParameters()
              }
            }
          }
          Iterator.single(finishedThreads.size)
        }
        }.reduce(_ + _)

      dropModelNumBatch += (driverSubModelNum - numFinishedModelUpdates)
      if (dropPercentage == 0.0 ||
        numFinishedModelUpdates >= driverSubModelNum * (1.0 - maxDropPercentage)) {
        // enough records were processed for this batch, so update the model
        val value = lossSum.value / numFinishedModelUpdates

        driverState("numFinishedModel") = numFinishedModelUpdates
        // isGradientUpdated is flag to mark whether gradient is updated. May changed in the future.
        driverState("isGradientUpdated") = false
        // parameterProcesser like L2NormClippingProcessor may aggregate gradient,
        // and change the value of isGradientUpdated in driverState.
        parameterProcessers.foreach(_.collectGlobalData(models.asInstanceOf[RDD[Cache[T]]],
          parameters, metrics, driverState))

        val isGradientUpdated = driverState[Boolean]("isGradientUpdated")
        val stateBroadcast = sc.broadcast(driverState)

        models.mapPartitions { modelIter =>
          val (paramLocalStart, paramLocalLen) = parameters.localPartitionRange
          val modelCache = modelIter.next()
          // if parameterProcesser has aggregated gradient, we can skip this aggregation.
          if (!isGradientUpdated) {
            val getG = System.nanoTime()
            parameters.aggregateGradientPartition(numFinishedModelUpdates)
            driverMetrics.add("aggregrateGradientParition average executor",
              System.nanoTime() - getG)
          }
          parameterProcessers.foreach(_.processParameters(parameters, modelCache, driverState))

          modelCache.optimMethods.foreach { case (name, optimMethod) =>

            optimMethod.state.update("epoch", driverState[Int]("epoch"))
            optimMethod.state.update("neval", driverState[Int]("neval"))
            optimMethod.state.update("Loss", driverState[Float]("Loss"))
            if (validationMethods.isDefined) {
              optimMethod.state.update("score", driverState[Float]("score"))
            }
            val p = parameterSplits(name)
            val startIdx = Math.max(paramLocalStart, p._1)
            val endIdx = Math.min(paramLocalStart + paramLocalLen, p._1 + p._2)
            if (endIdx > startIdx) {

              optimMethod.optimize(_ => (ev.fromType(value), parameters.gradientPartition.narrow(1,
                startIdx - paramLocalStart + 1, endIdx - startIdx)),
                parameters.weightPartition.narrow(1,
                  startIdx - paramLocalStart + 1, endIdx - startIdx))
            }

          }
          var time = System.nanoTime()
          driverMetrics.add("compute weight average", System.nanoTime() - time)
          parameters.sendWeightPartition()
          time = System.nanoTime()
          driverMetrics.add("send weights average", System.nanoTime() - time)
          Iterator.empty
        }.count()

        stateBroadcast.destroy()
        recordsProcessedThisEpoch += (recordsNum.value).toInt
        val end = System.nanoTime()
        wallClockTime += end - start
        driverState("isGradientUpdated") = true
        driverState("Loss") = lossSum.value.toFloat / numFinishedModelUpdates
        optimMethods.foreach { v =>
          v._2.updateHyperParameter()
        }
        // TODO: Support show learningrate for multiOptimMethod
        driverState(s"LearningRate") = optimMethods.head._2.getLearningRate().toFloat

        driverState("Throughput") = recordsNum.value.toFloat / ((end - start) / 1e9f)
        val _header = header(driverState[Int]("epoch"), recordsProcessedThisEpoch, numSamples,
          driverState[Int]("neval"), wallClockTime)
        logger.info(s"${_header} Trained ${recordsNum.value} records in ${(end - start) / 1e9} " +
          s"seconds. Throughput is ${driverState("Throughput")} records/second. Loss is ${
            driverState("Loss")
          }. ${getHyperParameterLog(optimMethods)}")
        logger.debug("\n" + metrics.summary())
        logger.debug("Dropped modules: " + (driverSubModelNum - numFinishedModelUpdates))
        lossArray = new Array[Double](_subModelNumber)

        // compute threshold
        iteration += 1
        if (dropPercentage > 0.0 && iteration > warmupIterationNum &&
          iteration % computeThresholdbatchSize == 0) {
          val moduleTimeList = models.mapPartitions { iter =>
            iter.next().moduleTimeList.iterator
          }.collect()

          val k = (dropPercentage * computeThresholdbatchSize * driverSubModelNum).toInt
          if (k > dropModelNumBatch) {
            threshold = Util.kthLargest(moduleTimeList, 0, moduleTimeList.length - 1,
              k - dropModelNumBatch)
          } else {
            threshold = (threshold * 1.01).toLong
          }
          logger.info("threshold: " + threshold)

          // clear moduleTimeList in each node
          models.mapPartitions { iter =>
            val timeList = iter.next.moduleTimeList
            var i = 0
            while (i < timeList.length) {
              timeList(i) = 0
              i += 1
            }
            Iterator.empty
          }.count()
          dropModelNumBatch = 0
        }

        driverState("neval") = driverState[Int]("neval") + 1
        if (recordsProcessedThisEpoch >= numSamples) {
          // Epoch is finished
          val epochEnd = System.nanoTime()
          wallClockTime = lastEpochTime + epochEnd - epochStart
          lastEpochTime = wallClockTime
          epochStart = System.nanoTime()
          logger.info(s"${_header} Epoch finished. Wall clock time is ${wallClockTime / 1e6} ms")

          driverState("epoch") = driverState[Int]("epoch") + 1
          dataset.shuffle()
          dataRDD = dataset.data(train = true)
          recordsProcessedThisEpoch = 0
        }

        optimMethods.map { case (moduleName, optimMethod) =>
          optimMethod.state.update("recordsProcessedThisEpoch", recordsProcessedThisEpoch)
          optimMethod.state.update("epoch", driverState[Int]("epoch"))
          optimMethod.state.update("neval", driverState[Int]("neval"))
          optimMethod.state.update("Loss", driverState[Float]("Loss"))
          if (validationMethods.isDefined) {
            optimMethod.state.update("score", driverState[Float]("score"))
          }
        }

        validate(
          validationTrigger,
          validationDataSet,
          validationMethods,
          coresPerNode,
          models.asInstanceOf[RDD[Cache[T]]],
          driverState,
          validationSummary,
          _header,
          parameters
        )

        trainSummary.foreach { summary =>
          saveSummary(
            summary,
            models.asInstanceOf[RDD[Cache[T]]],
            driverState,
            parameters,
            trainingModel
          )
        }

        checkpoint(
          cacheTrigger,
          cachePath,
          isOverWrite,
          wallClockTime,
          models.asInstanceOf[RDD[Cache[T]]],
          driverState,
          parameters,
          optimMethods,
          trainingModel
        )

      } else {
        logger.info(s"Warning! Not enough training samples were successfully processed in this " +
          s"iteration due to some slow tasks. The gradients computed in this iteration will be " +
          s"discarded. Only $numFinishedModelUpdates/$driverSubModelNum threads successfully " +
          s"completed training.")
      }
    }
  }

  /**
   * Init engine and cache models, weights, gradients, criterions, state tables
   * and validation methods on worker nodes.
   *
   * @param model train model
   * @param dataset train dataset
   * @param criterion loss function
   * @param state state table
   * @param nodeNumber node number
   * @param coresPerNode cores per node
   * @param checkSingleton if checkSingleton
   * @param allReduceParameter all reduce parameter instance
   * @param parameterSplits the mapping from module names to the parameter segments (offset,
   * length)
   * @param validationMethods validation methods
   * @param optimMethod optimization method
   * @param parameterProcessors a list of ParameterProcessor used to process parameters
   * @return cached models
   */
  private[bigdl] def initThreadModels[T: ClassTag](
    model: Module[T],
    dataset: DistributedDataSet[MiniBatch[T]],
    criterion: Criterion[T],
    state: Table,
    nodeNumber: Int,
    coresPerNode: Int,
    checkSingleton: Boolean,
    allReduceParameter: AllReduceParameter[T],
    parameterSplits: Map[String, (Int, Int)],
    validationMethods: Option[Array[ValidationMethod[T]]],
    optimMethod: Map[String, OptimMethod[T]],
    parameterProcessors: ArrayBuffer[ParameterProcessor]
  )(implicit ev: TensorNumeric[T]): (RDD[DistriOptimizer.CacheV1[T]], ModelBroadcast[T]) = {
    val sc = dataset.originRDD().sparkContext
    val broadcast = sc.broadcast((criterion, state, validationMethods, optimMethod))
    val convertedModel = ConversionUtils.convert(model)
    // ensure model's parameter is compacted for getting a better performance when broadcasting
    convertedModel.getParameters()
    // As cloneModel is using Serialization to implement deep copy, and will throw OOMError
    // when model's size is bigger than SerializationUtils' buffer size. So we can use
    // ModelBroadcast to clone model here.
    // Notes: All models returned by modelBroadcast.value() share the same weight&bias, while
    // gradWeight&gradBias is unshared.
    val modelBroadcast = ModelBroadcast[T]().broadcast(sc, convertedModel)
    val _subModelNumber = Engine.getEngineType match {
      case MklBlas => coresPerNode
      case MklDnn => 1
      case _ =>
        Log4Error.invalidInputError(false, s"unexpected engine type ${Engine.getEngineType}",
          "only support MklBlas and MklDnn")
        0
    }

    Log4Error.invalidOperationError(dataset.originRDD().partitions.length == nodeNumber,
      s"Passed in rdd partition number ${dataset.originRDD().partitions.length}" +
        s" is not equal to configured node number ${nodeNumber}")


    val computeThresholdbatchSize = state.get[Int]("computeThresholdbatchSize").get
    val nExecutor = Engine.nodeNumber()
    val executorCores = Engine.coreNumber()

    val models = dataset.originRDD().mapPartitions(_ => {
      val partitionId = TaskContext.getPartitionId
      val (broadcastCriterion, broadcastState, broadcastMethod,
      broadcastOptim) = broadcast.value
      if (!Engine.checkSingleton()) {
        if (checkSingleton) {
          Log4Error.invalidOperationError(Engine.checkSingleton(),
            "Partitions of the training data are not evenly" +
            "distributed across the executors in the Spark cluster; are there sufficient " +
            "training" +
            "data to be distributed? Set property \"bigdl.check.singleton\" to false to skip " +
            "this check")
        } else {
          logger.warn("Partitions of the training data are not evenly" +
            "distributed across the executors in the Spark cluster; are there sufficient " +
            "training" +
            "data to be distributed?")
        }
      }
      Engine.setNodeAndCore(nExecutor, executorCores)
      val cached = (0 until _subModelNumber).map { _ =>
        val localModel = modelBroadcast.value(true)
        if (Engine.getEngineType() == MklDnn && !localModel.isInstanceOf[IRGraph[T]]) {
          Engine.dnnComputing.invokeAndWait2((0 until _subModelNumber).map(i =>
            () => {
              localModel match {
                case container: MklDnnContainer => container.compile(TrainingPhase)
                case graph: DnnGraph => graph.compile(TrainingPhase)
                case _ =>
              }
            }))
        }
        setModelId(localModel, partitionId)
        val localCriterion = broadcastCriterion.cloneCriterion()
        val localState = broadcastState.clone()
        val localMethod =
          if (broadcastMethod.isDefined) Some(broadcastMethod.get.map(_.clone())) else None
        val (weights, grads) = localModel.getParameters()
        (localModel, weights, grads, localCriterion, localState, localMethod)
      }.toArray

      logger.info("model thread pool size is " + Engine.model.getPoolSize)
      val weights = cached.head._2
      allReduceParameter.init(weights.narrow(1, allReduceParameter.paramOffset,
        allReduceParameter.size))

      Iterator.single(CacheV1(
        cached.map(_._1), // models
        cached.map(_._2), // weights
        cached.map(_._3), // gradients
        cached.map(_._4), // criterions
        cached.map(_._5), // states
        new Array[Long](_subModelNumber * computeThresholdbatchSize),
        cached.map(_._6),
        broadcastOptim.map(v => (v._1, v._2.clone()))
      ))
    }).persist()
    models.setName("Thread Model RDD")
    logger.info("Cache thread models...")
    models.count()
    logger.info("Cache thread models... done")
    (models, modelBroadcast)
  }

  private def setModelId[T: ClassTag](model: Module[T], partitionId: Int): Unit = {
    model.setId(partitionId)
    if (model.isInstanceOf[Container[_, _, T]]) {
      model.asInstanceOf[Container[_, _, T]].modules.
        foreach(sub => setModelId(sub, partitionId))
    }
  }

  /**
   * Fetch current model parameters to driver, and copy to trainingModel.
   *
   * @param models cached models
   * @param parameters [[AllReduceParameter]]
   * @param trainingModel the model is trained by optimizer
   * @return trained model
   */
  override protected def getModel[T: ClassTag](
    models: RDD[Cache[T]],
    parameters: AllReduceParameter[T],
    trainingModel: Module[T])(implicit
    ev: TensorNumeric[T])
  : Module[T] = {
    val partitionNum = models.partitions.length

    models.mapPartitions(iter => {
      iter.next().localModels.head.beforeGetModel()
      Iterator.single(1)
    }).reduce(_ + _)

    Util.setExtraParametersFromModelRDD(models, trainingModel, maxSize = 500000000)

    // make sure gradient is as the same length as weight
    val parameterArray = trainingModel.parameters()
    (0 until parameterArray._2.length).foreach(i =>
      parameterArray._2(i).resizeAs(parameterArray._1(i))
    )

    val (parameter, gradientParameter) = trainingModel.getParameters()


    val (weights, gradients) = models.mapPartitions(iter => {
      val cached = iter.next()
      val curPartitionId = TaskContext.getPartitionId()
      Iterator.single((Map(curPartitionId -> parameters.weightPartition),
        Map(curPartitionId -> parameters.gradientPartition)))
    }).reduce((a, b) => (a._1 ++ b._1, a._2 ++ b._2))

    val taskSize = parameters.size / partitionNum
    Log4Error.invalidOperationError(taskSize != 0,
      "parameter length should not less than partition number")
    val extraSize = parameters.size % partitionNum

    (0 until partitionNum).map(pid => {
      val start = parameters.paramOffset + pid * taskSize + math.min(pid, extraSize)
      val length = taskSize + (if (pid < extraSize) 1 else 0)
      parameter.narrow(1, start, length).copy(weights(pid))
      gradientParameter.narrow(1, start, length).copy(gradients(pid))
    })


    trainingModel
  }

}

/**
 * The optimizer run on a distributed cluster.
 *
 * @param _model train model
 * @param _dataset train dataset
 * @param _criterion loss function
 */
class DistriOptimizer[T: ClassTag](
  _model: Module[T],
  _dataset: DistributedDataSet[MiniBatch[T]],
  _criterion: Criterion[T]
)(implicit ev: TensorNumeric[T])
  extends Optimizer[T, MiniBatch[T]](
    _model, _dataset, _criterion) {


  var compress: String = "fp16"

  def setCompressType(compressType: String): this.type = {
    compress = compressType
    this
  }

  val metrics = new Metrics

  private var models: RDD[DistriOptimizer.CacheV1[T]] = null
  // this variable is used to check the models cloned when broadcast, if there're native resources,
  // it will be deleted at the end of Optimizer.
  private var modelBroadcast: ModelBroadcast[T] = null

  /**
   * Clean some internal states, so this or other optimizers can run optimize again
   *
   * This method will be called at the end of optimize. You need not call it if optimize succeed.
   * If the optimize fails, you may call it before next optimize.
   */
  def clearState(): Unit = {
    DistriOptimizer.clearState(models.asInstanceOf[RDD[DistriOptimizer.Cache[T]]])
  }


  // By default, optimMethod internal state for each worker will not be reserved and reuse.
  private var reserveOptimMethod = false
  private[bigdl] var previousOptim: RDD[Map[String, OptimMethod[T]]] = null
  /**
   * If you want to reserve optimMethod for each worker, and reuse those methods in
   * next training task, you can call it.
   */

  /**
   * If you want to reserve optimMethod for each worker and reuse those methods in
   * next training task, please set reserve = true
   * Otherwise, if just using optimMethod you set in optimizer, please set reserve = false
   *
   * @param reserve whether to reserve optim method for each worker
   * @return
   */
  override def reserveOptim(reserve: Boolean): this.type = {
    reserveOptimMethod = reserve
    this
  }

  // replace optim methods with previous
  private def resetOptimMethods[T: ClassTag](
    models: RDD[DistriOptimizer.CacheV1[T]],
    previousOptimMethods: RDD[Map[String, OptimMethod[T]]]): RDD[DistriOptimizer.CacheV1[T]] = {
    models.zipPartitions(previousOptimMethods) { (m1, m2) => {
      val cache = m1.next()
      cache.optimMethods = m2.next()
      Iterator(cache)
    }
    }
  }

  private[bigdl] def endEpoch(): Unit = {
    DistriOptimizer.endEpoch(optimMethods)
  }

  override def setTrainData(sampleRDD: RDD[Sample[T]],
    batchSize: Int,
    miniBatch: MiniBatch[T]): this.type = {
    this.dataset = DistriOptimizer.setTrainData(sampleRDD, batchSize, miniBatch)
    // if current epoch is not finished, we will end the
    // current epoch and start a new epoch when optimize is called
    endEpoch()
    this
  }

  override def setTrainData(sampleRDD: RDD[Sample[T]],
    batchSize: Int,
    featurePaddingParam: PaddingParam[T] = null,
    labelPaddingParam: PaddingParam[T] = null): this.type = {
    val _featurePaddingParam = if (featurePaddingParam != null) Some(featurePaddingParam) else None
    val _labelPaddingParam = if (labelPaddingParam != null) Some(labelPaddingParam) else None
    this.dataset = DistriOptimizer.setTrainData(sampleRDD, batchSize,
      featurePaddingParam, labelPaddingParam)
    // if current epoch is not finished, we will end the
    // current epoch and start a new epoch when optimize is called
    endEpoch()
    this
  }

  override def prepareInput(): Unit = {
    if (!dataset.toDistributed().isCached) {
      DistriOptimizer.logger.info("caching training rdd ...")
      DistriOptimizer.prepareInput(this.dataset, this.validationDataSet)
    }
  }

  override def optimize(): Module[T] = {

    val distDataset = dataset.toDistributed()
    val trainingModel = if (Engine.getEngineType() == MklDnn && !model.isInstanceOf[MklDnnModule]
      && !model.isInstanceOf[IRGraph[T]] && !model.isInstanceOf[Graph[T]]) {
      model.toGraph().setName(model.getName())
    } else model

    optimMethods.values.foreach { optimMethod =>
      optimMethod.clearHistory()
    }

    // To be compatible with the old usage that user define hyperparameters in a table.
    if (optimMethods.size == 1) {
      optimMethods.head._2.loadFromTable(state)
    }

    state("dropPercentage") = dropPercentage
    state("warmupIterationNum") = warmupIterationNum
    state("computeThresholdbatchSize") = computeThresholdbatchSize
    state("maxDropPercentage") = maxDropPercentage
    state("isLayerwiseScaled") = Utils.isLayerwiseScaled(_model)

    val nodeNumber = Engine.nodeNumber()
    val coresPerNode = Engine.coreNumber()

    val partitionNum = distDataset.originRDD().partitions.length
    val modelParameters = trainingModel.getParameters()
    val allReduceParameter = AllReduceParameter.newParameter[T](partitionNum,
      modelParameters._1.nElement(), compress = this.compress)
    // subModuleName -> (storageOffset, length, AllReduceParameter)
    val parameterSplits = if (optimMethods.size != 1) {
      val p = optimMethods.map { case (subModuleName, optimMethod) =>
        val subModule = trainingModel(subModuleName)
        Log4Error.invalidOperationError(subModule.isDefined,
          s"Optimizer couldn't find $subModuleName in $model")
        val subModuleWeights = subModule.get.getParameters()._1
        (subModuleName, subModuleWeights)
      }
      val sortedWeights = p.values.toArray.sortWith((a, b) => a.storageOffset() < b.storageOffset())
      val compactWeights = Module.isCompact(sortedWeights)
      Log4Error.invalidOperationError(modelParameters._1 == compactWeights,
        s"DistriOptimizer: All subModules should have an OptimMethod.")
      p.map { case (subModuleName, weights) =>
        (subModuleName, (weights.storageOffset(), weights.nElement()))
      }
    } else if (optimMethods.contains(trainingModel.getName())) {
      Map(trainingModel.getName() -> (1, modelParameters._1.nElement()))
    } else {
      Log4Error.invalidOperationError(false,
        s"${trainingModel.getName()} doesn't " +
        s"have corresponding OptimMethod")
      null
    }

    LarsSGD.containsLarsSGD(optimMethods).foreach(weightDecay =>
      parameterProcessors.append(new LarsProcessor(parameterSplits, weightDecay))
    )

    prepareInput()

    val modelsAndBroadcast = DistriOptimizer.initThreadModels(trainingModel, distDataset, criterion,
      state, nodeNumber, coresPerNode, checkSingleton, allReduceParameter, parameterSplits,
      validationMethods,
      optimMethods, parameterProcessors)

    models = if (reserveOptimMethod && previousOptim != null) {
      // replace optimMethods with previous ones
      resetOptimMethods(modelsAndBroadcast._1, previousOptim)
    } else {
      modelsAndBroadcast._1
    }
    modelBroadcast = modelsAndBroadcast._2

    if (checkpointPath.isDefined) {
      val file = checkpointPath.get + "/" +
        new SimpleDateFormat("yyyyMMdd_HHmmss").format(Calendar.getInstance().getTime())
      new File(file).mkdir()
      checkpointPath = Some(file)
    }

    var retryNum = 0
    val maxRetry = System.getProperty("bigdl.failure.retryTimes", "5").toInt
    val retryTimeInterval = System.getProperty("bigdl.failure.retryTimeInterval", "120").toInt
    var lastFailureTimestamp = System.nanoTime()

    while (retryNum < maxRetry) {
      try {
        DistriOptimizer.optimize(
          trainingModel,
          distDataset,
          coresPerNode,
          state,
          endWhen,
          metrics,
          models,
          optimMethods,
          allReduceParameter,
          parameterSplits,
          validationTrigger,
          validationDataSet,
          validationMethods,
          checkpointTrigger,
          checkpointPath,
          trainSummary,
          validationSummary,
          isOverWrite,
          parameterProcessors.toArray
        )
        retryNum = Int.MaxValue
      } catch {
        case e: IllegalArgumentException =>
          Log4Error.invalidOperationError(false, e.getMessage, cause = e)
        case t: Throwable =>
//          DistriOptimizer.logger.error("Error: " + ExceptionUtils.getStackTrace(t))
          if (checkpointPath.isDefined) {
            /* To avoid retry number is used up by first few exceptions, we count time here.
             * If exception exceeds maxRetry times in maxRetry*retryTimeInterval seconds,
             * we will give up retry Or we will reset retryNum
             */
            if (System.nanoTime() - lastFailureTimestamp < maxRetry * retryTimeInterval * 1e9) {
              retryNum += 1
              if (retryNum == maxRetry) {
                Log4Error.invalidOperationError(false, t.getMessage, cause = t)
              }
            } else {
              retryNum = 1
            }
            DistriOptimizer.logger.info(s"Retrying $retryNum times")
            lastFailureTimestamp = System.nanoTime()

            val modelFile = getLatestFile(checkpointPath.get, "model")
            clearState()
            models.unpersist()
            val newModel = if (modelFile != null) {
              DistriOptimizer.logger.info("Model recover from last snapshot")
              Module.load[T](modelFile)
            } else {
              DistriOptimizer.logger.info("Model recover from origin model")
              trainingModel
            }
            optimMethods = optimMethods.map { case (moduleName, optimMethod) =>
              val methodFile = getLatestFile(checkpointPath.get, s"optimMethod-$moduleName")

              val newOptimMethod = if (methodFile != null) {
                DistriOptimizer.logger.info(s"$moduleName's OptimMethod recover from last snapshot")
                OptimMethod.load[T](methodFile)
              } else {
                DistriOptimizer.logger.info(s"$moduleName's OptimMethod recover from origin model")
                optimMethod
              }
              newOptimMethod.clearHistory()
              (moduleName, newOptimMethod)
            }
            val modelsAndBroadcast = DistriOptimizer.initThreadModels(newModel, distDataset,
              criterion, state, nodeNumber, coresPerNode, checkSingleton, allReduceParameter,
              parameterSplits, validationMethods, optimMethods, parameterProcessors)
            models = modelsAndBroadcast._1
            modelBroadcast = modelsAndBroadcast._2
          } else {
            Log4Error.invalidOperationError(false, t.getMessage, cause = t)
          }
      }
    }

    DistriOptimizer.getModel(models.asInstanceOf[RDD[DistriOptimizer.Cache[T]]],
      allReduceParameter, trainingModel)

    // Reset some internal states, so this or other optimizers can run optimize again
    clearState()

    // unpersist the model because the next time optimize is called, new `models` will be
    // created
    shutdown()

    // reserve optimMethod internal state for each worker if need
    if (reserveOptimMethod) {
      previousOptim = models.map(m => m.optimMethods).cache()
      previousOptim.count()
    } else {
      if (previousOptim != null) previousOptim.unpersist()
    }
    models.unpersist()

    trainingModel
  }

  private def getLatestFile(path: String, fileName: String): String = {
    val fl = new java.io.File(path)
    val files = fl.listFiles(new FilenameFilter {
      override def accept(dir: File, name: String): Boolean = {
        name.startsWith(fileName)
      }
    })

    var lastMod = Long.MinValue
    var choice: String = null
    files.map { file =>
      if (file.lastModified() > lastMod) {
        choice = file.getPath;
        lastMod = file.lastModified();
      }
    }
    return choice;
  }

  // this shutdown should not be called out of this scope.
  private[optim] override def shutdown(): Unit = {
    models.mapPartitions { iter =>
      iter.foreach { arrayModels =>
        arrayModels.localModels.foreach(_.release())
      }

      iter
    }.count()
    CachedModels.deleteKey(modelBroadcast.uuid)
  }
}




© 2015 - 2024 Weber Informatics LLC | Privacy Policy