Throw on requesting gradient for targets

Author: FlorentinD

ml/ml-core/src/main/java/org/neo4j/gds/ml/core/functions/CrossEntropyLoss.java

@@ -78,14 +78,12 @@ public Tensor<?> gradient(Variable<?> parent, ComputationContext ctx) {
 
                 // Compare to a threshold value rather than `0`, very small probability can result in setting infinite gradient values.
                 if (predictedProbabilityForTrueClass > PREDICTED_PROBABILITY_THRESHOLD) {
-                    gradient.setDataAt(row, trueClass, multiplier / predictedProbabilityForTrueClass
-                    );
+                    gradient.setDataAt(row, trueClass, multiplier / predictedProbabilityForTrueClass);
                 }
             }
             return gradient;
         } else {
-            // targets should never require a gradient
-            return ctx.data(parent).createWithSameDimensions();
+            throw new IllegalStateException("The gradient should not be necessary for the targets. But got: " + targets.render());
         }
     }
 }