Bump Python requirement torchvision to 0.21.0

dan/ocr/evaluate.py

@@ -31,6 +31,8 @@ logger = logging.getLogger(__name__)
 NERVAL_THRESHOLD = 0.30
 NB_WORST_PREDICTIONS = 5
 
+EMPTY_STRING = "∅"
+
 
 def parse_threshold(value: str) -> float:
     """
@@ -87,14 +89,24 @@ def print_worst_predictions(all_inferences: Dict[str, List[Inference]]):
         reverse=True,
     )[:NB_WORST_PREDICTIONS]
     for inference in worst_inferences:
+        if not inference.ground_truth:
+            logger.warning(
+                f"Ground truth is empty for {inference.image}. `{EMPTY_STRING}` will be displayed"
+            )
+
+        if not inference.prediction:
+            logger.warning(
+                f"Prediction is empty for {inference.image}. `{EMPTY_STRING}` will be displayed"
+            )
+
         alignment = getNiceAlignment(
             align(
-                inference.ground_truth,
-                inference.prediction,
+                inference.ground_truth or EMPTY_STRING,
+                inference.prediction or EMPTY_STRING,
                 task="path",
             ),
-            inference.ground_truth,
-            inference.prediction,
+            inference.ground_truth or EMPTY_STRING,
+            inference.prediction or EMPTY_STRING,
         )
         alignment_str = f'{alignment["query_aligned"]}\n{alignment["matched_aligned"]}\n{alignment["target_aligned"]}'
         table.add_row([inference.image, round(inference.wer * 100, 2), alignment_str])