Remove unnecessary preprocessing steps of VI-LayoutXLM to improve performance.

Bourn3z · Bourn3z · commit 07806991dff4 · 2024-01-22T11:23:07.000+08:00
diff --git a/configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yaml b/configs/kie/vi_layoutxlm/re_vi_layoutxlm_xfund_zh.yaml
@@ -59,12 +59,9 @@ train:
     label_file: XFUND/zh_train/train.json
     sample_ratio: 1.0
     transform_pipeline:
-      - DecodeImage:
-          img_mode: RGB
-          to_float32: False
       - VQATokenLabelEncode:
           contains_re: True
-          algorithm: &algorithm LayoutXLM
+          algorithm: &algorithm VI-LayoutXLM
           class_path: *class_path
           order_method: tb-yx
       - VQATokenPad:
@@ -75,30 +72,21 @@ train:
           max_seq_len: *max_seq_len
       - TensorizeEntitiesRelations:
           max_relation_len: 5000
-      - LayoutResize:
-          size: [224, 224]
-      - NormalizeImage:
-          bgr_to_rgb: False
-          is_hwc: True
-          mean: imagenet
-          std: imagenet
-      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visualize
     output_columns:
       [
         "input_ids",
         "bbox",
         "attention_mask",
         "token_type_ids",
-        "image",
         "question",
         "question_label",
         "answer",
         "answer_label",
         "relation_label",
       ]
-    net_input_column_index: [0, 1, 2, 3, 4, 5, 6, 7, 8] # input indices for network forward func in output_columns
-    label_column_index: [9] # input indices marked as label
+    net_input_column_index: [0, 1, 2, 3, 4, 5, 6, 7] # input indices for network forward func in output_columns
+    label_column_index: [8] # input indices marked as label
 
   loader:
     shuffle: True
@@ -117,9 +105,6 @@ eval:
     sample_ratio: 1.0
     shuffle: False
     transform_pipeline:
-      - DecodeImage:
-          img_mode: RGB
-          to_float32: False
       - VQATokenLabelEncode:
           contains_re: True
           algorithm: *algorithm
@@ -133,30 +118,21 @@ eval:
           max_seq_len: *max_seq_len
       - TensorizeEntitiesRelations:
           max_relation_len: 5000
-      - LayoutResize:
-          size: [224, 224]
-      - NormalizeImage:
-          bgr_to_rgb: False
-          is_hwc: True
-          mean: imagenet
-          std: imagenet
-      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visualize
     output_columns:
       [
         "input_ids",
         "bbox",
         "attention_mask",
         "token_type_ids",
-        "image",
         "question",
         "question_label",
         "answer",
         "answer_label",
         "relation_label",
       ]
-    net_input_column_index: [0, 1, 2, 3, 4, 5, 6, 7, 8] # input indices for network forward func in output_columns
-    label_column_index: [9] # input indices marked as label
+    net_input_column_index: [0, 1, 2, 3, 4, 5, 6, 7] # input indices for network forward func in output_columns
+    label_column_index: [8] # input indices marked as label
 
   loader:
     shuffle: False
diff --git a/configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yaml b/configs/kie/vi_layoutxlm/ser_vi_layoutxlm_xfund_zh.yaml
@@ -57,31 +57,20 @@ train:
     label_file: XFUND/zh_train/train.json
     sample_ratio: 1.0
     transform_pipeline:
-      - DecodeImage:
-          img_mode: RGB
-          to_float32: False
       - VQATokenLabelEncode:
           contains_re: False
-          algorithm: &algorithm LayoutXLM
+          algorithm: &algorithm VI-LayoutXLM
           class_path: *class_path
           order_method: tb-yx
       - VQATokenPad:
           max_seq_len: &max_seq_len 512
           return_attention_mask: True
       - VQASerTokenChunk:
           max_seq_len: *max_seq_len
-      - LayoutResize:
-          size: [ 224, 224 ]
-      - NormalizeImage:
-          bgr_to_rgb: False
-          is_hwc: True
-          mean: imagenet
-          std: imagenet
-      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the input labels for the loss function, and optional data for debug/visualize
-    output_columns: [ 'input_ids', 'bbox','attention_mask','token_type_ids', 'image', 'labels' ]
-    net_input_column_index: [ 0, 1, 2, 3, 4 ] # input indices for network forward func in output_columns
-    label_column_index: [ 2, 5 ] # input indices marked as label
+    output_columns: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'labels' ]
+    net_input_column_index: [ 0, 1, 2, 3 ] # input indices for network forward func in output_columns
+    label_column_index: [ 2, 4 ] # input indices marked as label
 
   loader:
     shuffle: True
@@ -100,9 +89,6 @@ eval:
     sample_ratio: 1.0
     shuffle: False
     transform_pipeline:
-      - DecodeImage:
-          img_mode: RGB
-          to_float32: False
       - VQATokenLabelEncode:
           contains_re: False
           algorithm: *algorithm
@@ -113,18 +99,10 @@ eval:
           return_attention_mask: True
       - VQASerTokenChunk:
           max_seq_len: *max_seq_len
-      - LayoutResize:
-          size: [ 224, 224 ]
-      - NormalizeImage:
-          bgr_to_rgb: False
-          is_hwc: True
-          mean: imagenet
-          std: imagenet
-      - ToCHWImage:
     #  the order of the dataloader list, matching the network input and the labels for evaluation
-    output_columns: [ 'input_ids', 'bbox', 'attention_mask','token_type_ids','image', 'labels' ]
-    net_input_column_index: [ 0, 1, 2, 3, 4 ] # input indices for network forward func in output_columns
-    label_column_index: [ 2, 5 ] # input indices marked as label
+    output_columns: [ 'input_ids', 'bbox', 'attention_mask', 'token_type_ids', 'labels' ]
+    net_input_column_index: [ 0, 1, 2, 3 ] # input indices for network forward func in output_columns
+    label_column_index: [ 2, 4 ] # input indices marked as label
 
   loader:
     shuffle: False
diff --git a/mindocr/data/transforms/layoutlm_transforms.py b/mindocr/data/transforms/layoutlm_transforms.py
@@ -3,6 +3,7 @@
 from collections import defaultdict
 
 import cv2
+from PIL import Image
 import numpy as np
 
 from mindspore import nn
@@ -65,8 +66,10 @@ def __init__(
         super(VQATokenLabelEncode, self).__init__()
         tokenizer_dict = {
             "LayoutXLM": {"class": LayoutXLMTokenizer, "pretrained_model": "layoutxlm-base-uncased"},
+            "VI-LayoutXLM": {"class": LayoutXLMTokenizer, "pretrained_model": "layoutxlm-base-uncased"},
         }
         self.contains_re = contains_re
+        self.algorithm = algorithm
         tokenizer_config = tokenizer_dict[algorithm]
         self.tokenizer = tokenizer_config["class"].from_pretrained(tokenizer_config["pretrained_model"])  # to replace
         self.label2id_map, id2label_map = load_vqa_bio_label_maps(class_path)
@@ -141,8 +144,10 @@ def __call__(self, data):
         train_re = self.contains_re and not self.infer_mode
         if train_re:
             ocr_info = self.filter_empty_contents(ocr_info)
-
-        height, width, _ = data["image"].shape
+        if self.algorithm == "VI-LayoutXLM":
+            width, height = Image.open(data["img_path"]).size
+        elif self.algorithm == "LayoutXLM":
+            height, width, _ = data["image"].shape
 
         words_list = []
         bbox_list = []
@@ -258,10 +263,23 @@ def trans_poly_to_bbox(poly):
         return [x1, y1, x2, y2]
 
     def _load_ocr_info(self, data):
-        """read text info from 'label' data"""
-        info = data["label"]
-        info_dict = json.loads(info)
-        return info_dict
+        if self.infer_mode:
+            ocr_result = self.ocr_engine.ocr(data["image"], cls=False)[0]
+            ocr_info = []
+            for res in ocr_result:
+                ocr_info.append(
+                    {
+                        "transcription": res[1][0],
+                        "bbox": self.trans_poly_to_bbox(res[0]),
+                        "points": res[0],
+                    }
+                )
+            return ocr_info
+        else:
+            info = data["label"]
+            # read text info
+            info_dict = json.loads(info)
+            return info_dict
 
     @staticmethod
     def _smooth_box(bboxes, height, width):
