Merge branch 'layoutlm' of https://github.com/KevinNuNu/mmocr into layoutlm

Author: KevinNuNu

projects/LayoutLMv3/models/ser_postprocessor.py

@@ -70,6 +70,40 @@ def __call__(self, outputs: torch.Tensor,
         pred_label_ids = max_idx.numpy().tolist()
         pred_label_scores = max_value.numpy().tolist()
 
+        # inference process do not have item in gt_label,
+        # so select valid token with word_ids rather than
+        # with gt_label_ids like official code.
+        pred_words_biolabels = []
+        word_biolabels = []
+        pre_word_id = None
+        for idx, cur_word_id in enumerate(word_ids):
+            if cur_word_id is not None:
+                if cur_word_id != pre_word_id:
+                    if word_biolabels:
+                        pred_words_biolabels.append(word_biolabels)
+                        word_biolabels = []
+                word_biolabels.append((self.id2biolabel[pred_label_ids[idx]],
+                                       pred_label_scores[idx]))
+            else:
+                pred_words_biolabels.append(word_biolabels)
+                break
+            pre_word_id = cur_word_id
+        # record pred_label
+        if self.only_label_first_subword:
+            pred_label = LabelData()
+            pred_label.item = [
+                pred_word_biolabels[0][0]
+                for pred_word_biolabels in pred_words_biolabels
+            ]
+            pred_label.score = [
+                pred_word_biolabels[0][1]
+                for pred_word_biolabels in pred_words_biolabels
+            ]
+            merged_data_sample.pred_label = pred_label
+        else:
+            raise NotImplementedError(
+                'The `only_label_first_subword=False` is not support yet.')
+
         # inference process do not have item in gt_label,
         # so select valid token with word_ids rather than
         # with gt_label_ids like official code.