Charset should only include training characters

Closes #190 (closed)

dan/datasets/extract/extract.py

@@ -277,7+277,7 @@
 
         if self.unknown_token in text:
             raise UnknownTokenInText(element_id=element.id)
 
         image_path = Path(self.output, IMAGES_DIR, split, element.id).with_suffix(
             self.image_extension
         )
@@ -293,7+293,7 @@
                 }
             )
 
-        self.data[split][str(image_path)] = self.format_text(
+        text = self.format_text(
             text,
             # Do not replace unknown characters in train split
             charset=self.charset if split != TRAIN_NAME else None,
         )
 
-        if split == TRAIN_NAME:
-            self.charset = self.charset.union(set(text))
+        self.data[split][str(image_path)] = text
+        self.charset = self.charset.union(set(text))
 
     def process_parent(
         self,