Fix in TrainerPerfectClassification.

Author: otiliastr

neural_structured_learning/research/gam/trainer/trainer_classification.py

@@ -763,4 +763,8 @@ def train(self, unused_data, unused_session=None, **unused_kwargs):
     return 1.0, 1.0
 
   def predict(self, unused_session, indices_unlabeled):
-    return self.data.get_original_labels(indices_unlabeled)
+    labels = self.data.get_original_labels(indices_unlabeled)
+    num_samples = len(indices_unlabeled)
+    predictions = np.zeros((num_samples, self.data.num_classes))
+    predictions[np.arange(num_samples), labels] = 1.0
+    return predictions