Upload invoice_layoutlmv3.py

invoice_layoutlmv3.py

@@ -41,7 +41,7 @@ def quad_to_box(quad):
         box = tuple(bbox)
     return box
 
-def augment_image(image, bboxes):
+def augment_image(file_path, file, bboxes):
     aug = iaa.OneOf([
         iaa.ElasticTransformation(alpha=(0, 2.0), sigma=0.25),
         # iaa.imgcorruptlike.GaussianBlur(severity=1),
@@ -50,7 +50,8 @@ def augment_image(image, bboxes):
         iaa.PerspectiveTransform(scale=(0.01, 0.15)),
         iaa.imgcorruptlike.Brightness(severity=2),
     ])
-    bbs = BoundingBoxesOnImage.from_xyxy_array(bboxes, shape=image.size)
+    image = imageio.imread(os.path.join(file_path, file))
+    bbs = BoundingBoxesOnImage.from_xyxy_array(bboxes, shape=image.shape)
     image_aug, bbs_aug = aug(image=image, bounding_boxes=bbs)
     bbs_aug = bbs_aug.remove_out_of_image()
     bbs_aug = bbs_aug.clip_out_of_image()
@@ -59,7 +60,7 @@ def augment_image(image, bboxes):
         return None, None
     else:
         aug_bboxes = bbs_aug.to_xyxy_array()
-        return image_aug, aug_bboxes
+        return Image.fromarray(image_aug, 'L'), aug_bboxes
 
 _URLS = {
 "image_files": _get_drive_url("https://drive.google.com/file/d/1bVc6xIAYO22RpehEkmuihaGsZ_VoOH2E"),
@@ -157,7 +158,6 @@ class IntellectAI(datasets.GeneratorBasedBuilder):
             ner_tags = []
             image_path = os.path.join(filepath, file)
             image, size = load_image(image_path)
-            print(size)
             data = [obj for obj in metadata if obj["documentName"]==file][0]
             for item in data["annotation"]:
                 cur_line_bboxes = []
@@ -183,7 +183,7 @@ class IntellectAI(datasets.GeneratorBasedBuilder):
                 bboxes.extend(cur_line_bboxes)
             image_variants = [(image, bboxes)]
             for _ in range(4):
-                aug_image, aug_bboxes = augment_image(image, bboxes)
+                aug_image, aug_bboxes = augment_image(filepath, file, bboxes)
                 if aug_image is not None and aug_bboxes is not None:
                     image_variants.append((aug_image, aug_bboxes))
             for img, bbs in image_variants: