fix loss when minibatch size is different (intel-analytics#3021)

* fix loss * fix ut
dding3 · Jul 1, 2020 · dd96ee3 · dd96ee3
1 parent 89992a6
commit dd96ee3
Show file tree

Hide file tree

Showing 3 changed files with 30 additions and 6 deletions.
diff --git a/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ValidationMethod.scala b/scala/dllib/src/main/scala/com/intel/analytics/bigdl/dllib/optim/ValidationMethod.scala
@@ -1089,9 +1089,9 @@ class Loss[@specialized(Float, Double)T: ClassTag](
       output.toTensor[T]
     }
     val loss = ev.toType[Float](criterion.forward(_output, _target))
-    val count = 1
+    val count = _target.size().head
 
-    new LossResult(loss, count)
+    new LossResult(loss * count, count)
   }
 
   override def format(): String = "Loss"

diff --git a/spark/dl/src/test/scala/com/intel/analytics/bigdl/optim/EvaluatorSpec.scala b/spark/dl/src/test/scala/com/intel/analytics/bigdl/optim/EvaluatorSpec.scala
@@ -58,7 +58,7 @@ class EvaluatorSpec extends SparkContextLifeCycle with Matchers {
 
     result(0)._1 should be (new AccuracyResult(0, 100))
     result(1)._1 should be (new AccuracyResult(100, 100))
-    result(2)._1 should be (new LossResult(57.610695f, 25))
+    result(2)._1 should be (new LossResult(230.44278f, 100))
     result(0)._1.result()._1 should be (0f)
     result(1)._1.result()._1 should be (1f)
     result(2)._1.result()._1 should be (2.3044279f+-0.000001f)
@@ -82,7 +82,31 @@ class EvaluatorSpec extends SparkContextLifeCycle with Matchers {
 
     result(0)._1 should be (new AccuracyResult(0, 100))
     result(1)._1 should be (new AccuracyResult(100, 100))
-    result(2)._1 should be (new LossResult(57.610695f, 25))
+    result(2)._1 should be (new LossResult(230.44278f, 100))
+    result(0)._1.result()._1 should be (0f)
+    result(1)._1.result()._1 should be (1f)
+    result(2)._1.result()._1 should be (2.3044279f+-0.000001f)
+  }
+
+  "Evaluator different MiniBatch" should "be correct" in {
+    RNG.setSeed(100)
+    val tmp = new Array[MiniBatch[Float]](25)
+    var i = 1
+    while (i <= tmp.length) {
+      val input = Tensor[Float](i, 28, 28).fill(0.8f)
+      val label = Tensor[Float](i).fill(1.0f)
+      tmp(i - 1) = MiniBatch(input, label)
+      i += 1
+    }
+    val model = LeNet5(classNum = 10)
+    val dataSet = DataSet.array(tmp, sc).toDistributed().data(train = false)
+
+    val result = model.evaluate(dataSet, Array(new Top1Accuracy[Float](), new Top5Accuracy[Float](),
+      new Loss[Float](CrossEntropyCriterion[Float]())))
+
+    result(0)._1 should be (new AccuracyResult(0, 325))
+    result(1)._1 should be (new AccuracyResult(325, 325))
+    result(2)._1 should be (new LossResult(748.93896f, 325))
     result(0)._1.result()._1 should be (0f)
     result(1)._1.result()._1 should be (1f)
     result(2)._1.result()._1 should be (2.3044279f+-0.000001f)

diff --git a/spark/dl/src/test/scala/com/intel/analytics/bigdl/utils/intermediate/DnnPredictorSpec.scala b/spark/dl/src/test/scala/com/intel/analytics/bigdl/utils/intermediate/DnnPredictorSpec.scala
@@ -142,7 +142,7 @@ class DnnPredictorSpec extends FlatSpec with Matchers with BeforeAndAfter {
     dataSet.partitions.length should be(4)
     result(0)._1 should be (new AccuracyResult(0, 100))
     result(1)._1 should be (new AccuracyResult(100, 100))
-    result(2)._1 should be (new LossResult(16.130993f, 7))
+    result(2)._1 should be (new LossResult(230.44278f, 100))
     result(0)._1.result()._1 should be (0f)
     result(1)._1.result()._1 should be (1f)
     result(2)._1.result()._1 should be (2.3044279f+-0.000001f)
@@ -167,7 +167,7 @@ class DnnPredictorSpec extends FlatSpec with Matchers with BeforeAndAfter {
     dataSet.partitions.length should be(4)
     result(0)._1 should be (new AccuracyResult(0, 100))
     result(1)._1 should be (new AccuracyResult(100, 100))
-    result(2)._1 should be (new LossResult(57.610695f, 25))
+    result(2)._1 should be (new LossResult(230.44278f, 100))
     result(0)._1.result()._1 should be (0f)
     result(1)._1.result()._1 should be (1f)
     result(2)._1.result()._1 should be (2.3044279f+-0.000001f)