fix Bug in BertScore calculation: pred target misalignment

src/torchmetrics/functional/text/bert.py

@@ -436,18 +436,15 @@ def bert_score(
  preds_loader, preds_dataset.max_length, model, device, num_layers, all_layers, idf, verbose, user_forward_fn
  )
 
+ preds_embeddings = preds_embeddings[preds_loader.dataset.sorting_indices]
+ target_embeddings = target_embeddings[target_loader.dataset.sorting_indices]
+
+ preds_idf_scale = preds_idf_scale[preds_loader.dataset.sorting_indices]
+ target_idf_scale = target_idf_scale[target_loader.dataset.sorting_indices]
+
  precision, recall, f1_score = _get_precision_recall_f1(
  preds_embeddings, target_embeddings, preds_idf_scale, target_idf_scale
  )
- # Sort predictions
- if len(precision.shape) == 1: # i.e. when all_layers = False
- precision = precision[preds_loader.dataset.sorting_indices]
- recall = recall[preds_loader.dataset.sorting_indices]
- f1_score = f1_score[preds_loader.dataset.sorting_indices]
- elif len(precision.shape) == 2: # i.e. when all_layers = True
- precision = precision[:, preds_loader.dataset.sorting_indices]
- recall = recall[:, preds_loader.dataset.sorting_indices]
- f1_score = f1_score[:, preds_loader.dataset.sorting_indices]
 
  if baseline is not None:
  precision, recall, f1_score = _rescale_metrics_with_baseline(