First draft of ElasticAverageCollideBinder

Author: Kevin Peng

src/main/scala/BIDMach/allreduce/binder/ElasticAverageCollideBinder.scala

@@ -0,0 +1,273 @@
+package BIDMach.allreduce.binder
+
+import java.util.ArrayDeque
+import java.util.concurrent.atomic.AtomicInteger
+import java.util.logging.Logger
+import scala.util.Random
+
+import BIDMach.allreduce.binder.AllreduceBinder.{DataSink, DataSource}
+//import BIDMach.models.Model
+import BIDMach.updaters.Grad
+import BIDMat.{Mat, FMat, GMat}
+
+
+/**
+  * Linearize input model mats, and elastic-average update to the same model.
+  * Perform momentum exchange among several nodes in a cluster, preserving total energy of the nodes.
+  *
+  * @param model
+  * @param alphaFromIter
+  */
+// FIXME: should get rndseed, node num and # nodes from worker
+class ElasticAverageCollideBinder(updater: Grad, alphaFromIter: Int => Float, hardness: Float, rndseed: Long, inode: Int,
+                                  nnodes: Int, logger: Logger) extends AllreduceBinder {
+
+  val model = updater.model
+  // Keeping track of elastic updates
+  var tic = System.currentTimeMillis()
+  val reduceCount = new AtomicInteger()
+  
+  val random = new Random(rndseed)
+  // TODO: make these GMats when applicable
+  val rawRandVecs = new Array[Array[FMat]](nnodes)
+  val randVecs = new Array[Array[FMat]](nnodes)
+  val randVecSqNorms = new Array[Array[Float]](nnodes)
+  var rvOffset = 0
+  // TODO: think about GMats too
+  val aelem = FMat(1, 1)
+  
+  // TODO: make this more efficient by making use of functionality in SciFunctions etc.
+  def genRandomVector(out: FMat) = {
+    var i = 0
+    val len = out.length
+    while (i < len) {
+      out.data(i) = random.nextGaussian().toFloat
+    }
+  }
+  
+  def dotprod(a:Mat, b:Mat):Float = {
+    aelem ~ a.contents dot b.contents
+    aelem.dv.toFloat;
+  }
+  
+  // TODO: is synchronization necessary to get updater momentum lengths
+  def initRandVecs = {
+    if (rawRandVecs(0) eq null) {
+      for (i <- 0 until nnodes) {
+        rawRandVecs(i) = new Array(updater.momentum.length)
+        
+        for ((pm, j) <- updater.momentum.iterator.zipWithIndex) {
+          val fmat = FMat.make(pm.dims)
+          genRandomVector(fmat.contents())
+          pm match {
+            case _: GMat => rawRandVecs(i)(j) = GMat(fmat)
+            case _: FMat => rawRandVecs(i)(j) = fmat
+          }
+        }
+      }
+      
+      for (i <- 0 until nnodes) {
+        randVecs(i) = new Array(updater.momentum.length)
+        randVecSqNorms(i) = new Array(updater.momentum.length)
+        for (j <- 0 until updater.momentum.length) {
+          randVecs(i)(j) = rawRandVecs(i)(j) - rawRandVecs((i + 1) % nnodes)(j)
+          randVecSqNorms(i)(j) = dotprod(randVecs(i)(j), randVecs(i)(j))
+        }
+      }
+    }
+  }
+  
+  def rotateRndVecs = {
+    val prevOffset = (rvOffset + nnodes - 1) % nnodes
+    
+    for (randMat <- rawRandVecs(rvOffset)) {
+      randMat match {
+        case gmat: GMat =>
+          val fmat = FMat.make(randMat.dims)
+          genRandomVector(fmat)
+          gmat <-- fmat
+        case fmat: FMat => genRandomVector(fmat)
+      }
+    }
+    
+    for (offset <- Array(prevOffset, rvOffset)) {
+      val nextOffset = (offset + 1) % nnodes
+      for ((v1, v2) <- randVecs(offset) zip randVecs(nextOffset)) {
+        v1 ~ v1 - v2
+      }
+      for ((v, i) <- randVecs(offset).iterator.zipWithIndex) {
+        randVecSqNorms(offset)(i) = dotprod(v, v)
+      }
+    }
+    
+    rvOffset += 1
+    if (rvOffset == nnodes) rvOffset = 0
+  }
+
+  override lazy val totalDataSize: Int = {
+    var ret = 0
+    updater.momentum.synchronized {
+      // Momentum mats
+      for (p <- updater.momentum) ret += p.length
+      // Squared magnitudes of momentum mats
+      ret += updater.momentum.length
+      // Dot product of momentum mats and random mats
+      ret += updater.momentum.length
+    }
+    // Model mats
+    model.modelmats.synchronized {
+      for (mat <- model.modelmats) ret += mat.length
+    }
+    ret
+  }
+
+  override def dataSource: DataSource = inputRequest => {
+    initRandVecs
+
+    val ret: Array[Float] = new Array[Float](totalDataSize)
+    var current = totalDataSize
+    val myRandVecs = randVecs((rvOffset + inode) % nnodes)
+    
+    // TODO: do we need to lock on the model and updater mats
+
+    // backward traversing model mats, assuming forward traversal by the training model
+    for (mm <- model.modelmats.reverseIterator) {
+      current -= mm.length
+      mm match {
+        case gmat: GMat => GMat.GPUtoCPUarraycopy(gmat.pdata, 0, ret, current, gmat.length, "ElasticAverageBinder dataSource")
+        case fmat: FMat => System.arraycopy(fmat.contents().data, 0, ret, current, fmat.length)
+      }
+    }
+
+    // dot product of momentum and random vectors
+    // backward traversing update mats, assuming forward traversal by updater
+    for ((pm, r) <- updater.momentum.reverseIterator zip myRandVecs.reverseIterator) {
+      current -= 1
+      ret(current) = dotprod(pm, r)
+    }
+    
+    // squared norm of momentums
+    for (pm <- updater.momentum.reverseIterator) {
+      current -= 1
+      ret(current) = dotprod(pm, pm)
+    }
+    
+    // backward traversing update mats, assuming forward traversal by updater
+    for (pm <- updater.momentum.reverseIterator) {
+      current -= pm.length
+      pm match {
+        case gmat: GMat => GMat.GPUtoCPUarraycopy(gmat.pdata, 0, ret, current, gmat.length, "ElasticAverageBinder dataSource")
+        case fmat: FMat => System.arraycopy(fmat.contents().data, 0, ret, current, fmat.length)
+      }
+    }
+    
+    assert(current == 0, "current should be zero after iteration")
+
+    AllReduceInput(ret)
+
+  }
+
+
+
+  override def dataSink: DataSink = reducedOutput => {
+
+    reduceCount.synchronized {
+      val currentCount: Int = reduceCount.getAndIncrement()
+      val updateCounts = 10
+      if (currentCount % updateCounts == 0) {
+        val toc = System.currentTimeMillis()
+        if (currentCount > 0) {
+          logger.info(f"elastic_updates/s=${updateCounts/((toc - tic) / 1.0e3)}%2.2f, total_updates=$currentCount")
+        }
+        tic = toc
+      }
+    }
+    val reducedData = reducedOutput.data
+
+    assert(reducedData.length == totalDataSize, "Reduced output should be same length as input")
+
+    // backward traversing model mats, assuming forward traversal by the training model
+    // using while instead of for loop due to performance
+    var current = totalDataSize
+    val alpha = alphaFromIter(reducedOutput.iteration)
+
+    for (mm <- model.modelmats.reverseIterator) {
+      current -= mm.length
+      mm.synchronized {
+        mm match {
+          case gmat: GMat =>
+            val gReduced = GMat.make(gmat.dims)
+            GMat.CPUtoGPUarraycopy(reducedData, current, gReduced.pdata, 0, gmat.length, "ElasticAverageCollideBinder dataSink")
+            gReduced ~ gReduced / aelem.set(nnodes)
+            gmat ~ gmat * aelem.set(1 - alpha)
+            gReduced ~ gReduced * aelem.set(alpha)
+            gmat ~ gReduced + gmat
+            gReduced.free()
+          case fmat: FMat =>
+            val fReduced = FMat.make(fmat.dims)
+            System.arraycopy(reducedData, current, fReduced.contents().data, 0, fmat.length)
+            fReduced ~ fReduced / aelem.set(nnodes)
+            fmat ~ fmat * aelem.set(1 - alpha)
+            fReduced ~ fReduced * aelem.set(alpha)
+            fmat ~ fReduced + fmat
+        }
+      }
+    }
+    
+    val sumPmR = new Array[Float](updater.modelmats.length)
+    current -= updater.modelmats.length
+    System.arraycopy(reducedData, current, sumPmR, 0, updater.modelmats.length)
+    
+    val sumPmPm = new Array[Float](updater.modelmats.length)
+    current -= updater.modelmats.length
+    System.arraycopy(reducedData, current, sumPmPm, 0, updater.modelmats.length)
+    
+    val meanP = new Array[Mat](updater.modelmats.length)
+    for (i <- updater.modelmats.length - 1 to 0 by -1) {
+      current -= updater.modelmats(i).length
+      val pbar = updater.modelmats(i) match {
+        case _: GMat =>
+          val pbar = GMat.make(updater.modelmats(i).dims)
+          GMat.CPUtoGPUarraycopy(reducedData, current, pbar.pdata, 0, updater.modelmats(i).length, "ElasticAverageCollideBinder dataSink")
+          pbar
+        case _: FMat =>
+          val pbar = FMat.make(updater.modelmats(i).dims)
+          System.arraycopy(reducedData, current, pbar.contents().data, 0, updater.modelmats(i).length)
+          pbar
+      }
+      pbar ~ pbar / aelem.set(nnodes)
+      meanP(i) = pbar
+    }
+    
+    assert(current == 0, "current should be zero after iteration")
+    
+    for (j <- updater.modelmats.length - 1 to 0 by -1) {
+      // TODO: not hold the lock for 1293579813753 years, but also avoid data races
+      updater.modelmats(j) synchronized {
+        val x = meanP(j) - updater.modelmats(j)
+        x ~ x * aelem.set(hardness)
+        x ~ x + updater.modelmats(j)
+        
+        val sumXR = (1 - hardness) * sumPmR(j)
+        val sumXXminusPmPm = hardness * (hardness - 2) * (sumPmPm(j) - nnodes * dotprod(meanP(j), meanP(j)))
+        
+        val twoSumXR = 2 * sumXR
+        val sumRR = randVecSqNorms.map(_(j)).reduce(_ + _)
+        // Discriminant should always be positive for any hardness in [0, 1]
+        val discr = twoSumXR*twoSumXR - 4*sumRR*sumXXminusPmPm
+        val epsilon = 1e-36f
+        val beta = if (Mat.myrand.nextFloat() < 0.5f) {
+          (-twoSumXR + math.sqrt(discr).toFloat) / (2 * sumRR + epsilon)
+        } else {
+          (-twoSumXR - math.sqrt(discr).toFloat) / (2 * sumRR + epsilon)
+        }
+        
+        updater.modelmats(j) ~ x - aelem.set(beta) * randVecs((rvOffset + inode) % nnodes)(j)
+      }
+    }
+    
+    rotateRndVecs
+  }
+
+}
+