Only look for classes when we load them

kaldi_data_generator/main.py

@@ -172,13 +172,16 @@ class HTRDataGenerator:
                 with_classes=self.should_filter_by_class,
                 with_metadata=self.should_filter_by_metadatas,
             ):
-                elem_classes = [c for c in elt["classes"] if c["state"] != "rejected"]
 
                 should_accept = True
                 if self.should_filter_by_class:
                     # at first filter to only have elements with accepted classes
                     # if accepted classes list is empty then should accept all
                     # except for ignored classes
+                    elem_classes = [
+                        c for c in elt["classes"] if c["state"] != "rejected"
+                    ]
+
                     should_accept = len(self.accepted_classes) == 0
                     for classification in elem_classes:
                         class_name = classification["ml_class"]["name"]
@@ -193,6 +196,9 @@ class HTRDataGenerator:
                     continue
 
                 if self.should_filter_by_style:
+                    elem_classes = [
+                        c for c in elt["classes"] if c["state"] != "rejected"
+                    ]
                     style_counts = Counter()
                     for classification in elem_classes:
                         class_name = classification["ml_class"]["name"]