fix: Fixing flaky unit tests (#1215)

microsoft · Oct 19, 2021 · e23041f · e23041f
1 parent 5d31e3e
commit e23041f
Showing 1 changed file with 46 additions and 4 deletions.
diff --git a/...est/scala/com/microsoft/azure/synapse/ml/explainers/split1/VectorLIMEExplainerSuite.scala b/...est/scala/com/microsoft/azure/synapse/ml/explainers/split1/VectorLIMEExplainerSuite.scala
@@ -10,10 +10,12 @@ import com.microsoft.azure.synapse.ml.core.test.fuzzing.{TestObject, Transformer
 import com.microsoft.azure.synapse.ml.core.utils.BreezeUtils._
 import com.microsoft.azure.synapse.ml.explainers.LocalExplainer.LIME
 import com.microsoft.azure.synapse.ml.explainers.VectorLIME
-import org.apache.spark.ml.linalg.{Vector, Vectors}
+import org.apache.spark.ml.linalg.{DenseVector, SQLDataTypes, Vector, Vectors}
 import org.apache.spark.ml.regression.{LinearRegression, LinearRegressionModel}
 import org.apache.spark.ml.util.MLReadable
 import org.apache.spark.sql.DataFrame
+import org.apache.spark.sql.functions.col
+import org.apache.spark.sql.types.{ArrayType, DataType}
 import org.scalactic.Equality
 
 class VectorLIMEExplainerSuite extends TestBase
@@ -22,9 +24,50 @@ class VectorLIMEExplainerSuite extends TestBase
   import spark.implicits._
 
   implicit val vectorEquality: Equality[BDV[Double]] = breezeVectorEq(1E-6)
-  implicit val matrixEquality: Equality[BDM[Double]] = breezeMatrixEq(1E-6)
 
-  override val sortInDataframeEquality = true
+  // The equality comparison for the weights matrices are quite complicated,
+  // creating an override here to handle the logic.
+  override def assertDFEq(df1: DataFrame, df2: DataFrame)(implicit eq: Equality[DataFrame]): Unit = {
+    assert(DataType.equalsStructurally(df1.schema, df2.schema))
+
+    val arrayOfVectorFields = df1.schema.fields.filter {
+      f => DataType.equalsStructurally(f.dataType, ArrayType(SQLDataTypes.VectorType))
+    }.map(_.name)
+
+    // Non-matrix columns, defer to super.assertDFEq
+    val otherFields = df1.schema.fields.map(_.name).filterNot(arrayOfVectorFields.contains)
+
+    super.assertDFEq(
+      df1.select(otherFields.map(col): _*),
+      df2.select(otherFields.map(col): _*)
+    )(eq)
+
+    val (a, b) = (df1.select(arrayOfVectorFields.map(col): _*), df2.select(arrayOfVectorFields.map(col): _*)) match {
+      case (x, y) =>
+        val sorted = if (sortInDataframeEquality) {
+          (x.sort(arrayOfVectorFields.map(col): _*), y.sort(arrayOfVectorFields.map(col): _*))
+        } else {
+          (x, y)
+        }
+
+        (sorted._1.collect, sorted._2.collect)
+    }
+
+    assert(a.length === b.length)
+    a.indices foreach { // For each row
+      i =>
+        arrayOfVectorFields.indices foreach {
+          j => // For each matrix type column
+            val x: Seq[DenseVector] = a(i).getAs[Seq[DenseVector]](j)
+            val y: Seq[DenseVector] = b(i).getAs[Seq[DenseVector]](j)
+            assert(x.length === y.length)
+            x.indices foreach {
+              k =>
+                x(k) === y(k)
+            }
+        }
+    }
+  }
 
   val d1 = 3
   val d2 = 1
@@ -54,7 +97,6 @@ class VectorLIMEExplainerSuite extends TestBase
     .setNumSamples(1000)
 
   test("VectorLIME can explain a model locally") {
-
     val predicted = model.transform(df)
     val weights = lime
       .transform(predicted)