Evaluate predictions with nerval

Based on !340 (merged)

Closes #231 (closed)

dan/ocr/manager/metrics.py

@@ -26,8 +26,8 @@ METRICS_KEYWORD = {"cer": "chars", "wer": "words", "ner": "tokens"}
 
 @dataclass
 class Inference:
-    ground_truth: List[str]
-    prediction: List[str]
+    ground_truth: str
+    prediction: str
 
 
 class MetricManager:
@@ -47,9 +47,6 @@ class MetricManager:
         self.metric_names: List[str] = metric_names
         self.epoch_metrics = defaultdict(list)
 
-        # List of inferences (prediction with their ground truth)
-        self.inferences = []
-
     def format_string_for_cer(self, text: str, remove_token: bool = False):
         """
         Format string for CER computation: remove layout tokens and extra spaces
@@ -165,7 +162,6 @@ class MetricManager:
             metrics["time"] = [values["time"]]
 
         gt, prediction = values["str_y"], values["str_x"]
-        self.inferences.append(Inference(ground_truth=gt, prediction=prediction))
 
         for metric_name in metric_names:
             match metric_name: