python-doctr 0.9.0__py3-none-any.whl → 0.10.0__py3-none-any.whl

doctr/datasets/cord.py

@@ -33,6 +33,7 @@ class CORD(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -53,6 +54,7 @@ class CORD(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         url, sha256, name = self.TRAIN if train else self.TEST
@@ -64,10 +66,15 @@ class CORD(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
 
         # List images
         tmp_root = os.path.join(self.root, "image")
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         self.train = train
         np_dtype = np.float32
         for img_path in tqdm(iterable=os.listdir(tmp_root), desc="Unpacking CORD", total=len(os.listdir(tmp_root))):
@@ -109,6 +116,8 @@ class CORD(VisionDataset):
                 )
                 for crop, label in zip(crops, list(text_targets)):
                     self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((img_path, np.asarray(box_targets, dtype=int).clip(min=0)))
             else:
                 self.data.append((
                     img_path,

doctr/datasets/funsd.py

@@ -33,6 +33,7 @@ class FUNSD(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -45,6 +46,7 @@ class FUNSD(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -55,6 +57,12 @@ class FUNSD(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
         np_dtype = np.float32
 
@@ -63,7 +71,7 @@ class FUNSD(VisionDataset):
 
         # # List images
         tmp_root = os.path.join(self.root, subfolder, "images")
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         for img_path in tqdm(iterable=os.listdir(tmp_root), desc="Unpacking FUNSD", total=len(os.listdir(tmp_root))):
             # File existence check
             if not os.path.exists(os.path.join(tmp_root, img_path)):
@@ -100,6 +108,8 @@ class FUNSD(VisionDataset):
                     # filter labels with unknown characters
                     if not any(char in label for char in ["☑", "☐", "\uf703", "\uf702"]):
                         self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((img_path, np.asarray(box_targets, dtype=np_dtype)))
             else:
                 self.data.append((
                     img_path,

doctr/datasets/ic03.py

@@ -32,6 +32,7 @@ class IC03(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -51,6 +52,7 @@ class IC03(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         url, sha256, file_name = self.TRAIN if train else self.TEST
@@ -62,8 +64,14 @@ class IC03(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         # Load xml data
@@ -117,6 +125,8 @@ class IC03(VisionDataset):
                     for crop, label in zip(crops, labels):
                         if crop.shape[0] > 0 and crop.shape[1] > 0 and len(label) > 0:
                             self.data.append((crop, label))
+                elif detection_task:
+                    self.data.append((name.text, boxes))
                 else:
                     self.data.append((name.text, dict(boxes=boxes, labels=labels)))
 

doctr/datasets/ic13.py

@@ -38,6 +38,7 @@ class IC13(AbstractDataset):
         label_folder: folder with all annotation files for the images
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `AbstractDataset`.
     """
 
@@ -47,11 +48,17 @@ class IC13(AbstractDataset):
         label_folder: str,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
             img_folder, pre_transforms=convert_target_to_relative if not recognition_task else None, **kwargs
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
 
         # File existence check
         if not os.path.exists(label_folder) or not os.path.exists(img_folder):
@@ -59,7 +66,7 @@ class IC13(AbstractDataset):
                 f"unable to locate {label_folder if not os.path.exists(label_folder) else img_folder}"
             )
 
-        self.data: List[Tuple[Union[Path, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[Path, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         img_names = os.listdir(img_folder)
@@ -95,5 +102,7 @@ class IC13(AbstractDataset):
                 crops = crop_bboxes_from_image(img_path=img_path, geoms=box_targets)
                 for crop, label in zip(crops, labels):
                     self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((img_path, box_targets))
             else:
                 self.data.append((img_path, dict(boxes=box_targets, labels=labels)))

doctr/datasets/iiit5k.py

@@ -34,6 +34,7 @@ class IIIT5K(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -45,6 +46,7 @@ class IIIT5K(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -55,6 +57,12 @@ class IIIT5K(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
 
         # Load mat data
@@ -62,7 +70,7 @@ class IIIT5K(VisionDataset):
         mat_file = "trainCharBound" if self.train else "testCharBound"
         mat_data = sio.loadmat(os.path.join(tmp_root, f"{mat_file}.mat"))[mat_file][0]
 
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         for img_path, label, box_targets in tqdm(iterable=mat_data, desc="Unpacking IIIT5K", total=len(mat_data)):
@@ -73,24 +81,26 @@ class IIIT5K(VisionDataset):
             if not os.path.exists(os.path.join(tmp_root, _raw_path)):
                 raise FileNotFoundError(f"unable to locate {os.path.join(tmp_root, _raw_path)}")
 
+            if use_polygons:
+                # (x, y) coordinates of top left, top right, bottom right, bottom left corners
+                box_targets = [
+                    [
+                        [box[0], box[1]],
+                        [box[0] + box[2], box[1]],
+                        [box[0] + box[2], box[1] + box[3]],
+                        [box[0], box[1] + box[3]],
+                    ]
+                    for box in box_targets
+                ]
+            else:
+                # xmin, ymin, xmax, ymax
+                box_targets = [[box[0], box[1], box[0] + box[2], box[1] + box[3]] for box in box_targets]
+
             if recognition_task:
                 self.data.append((_raw_path, _raw_label))
+            elif detection_task:
+                self.data.append((_raw_path, np.asarray(box_targets, dtype=np_dtype)))
             else:
-                if use_polygons:
-                    # (x, y) coordinates of top left, top right, bottom right, bottom left corners
-                    box_targets = [
-                        [
-                            [box[0], box[1]],
-                            [box[0] + box[2], box[1]],
-                            [box[0] + box[2], box[1] + box[3]],
-                            [box[0], box[1] + box[3]],
-                        ]
-                        for box in box_targets
-                    ]
-                else:
-                    # xmin, ymin, xmax, ymax
-                    box_targets = [[box[0], box[1], box[0] + box[2], box[1] + box[3]] for box in box_targets]
-
                 # label are casted to list where each char corresponds to the character's bounding box
                 self.data.append((
                     _raw_path,

doctr/datasets/imgur5k.py

@@ -46,6 +46,7 @@ class IMGUR5K(AbstractDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `AbstractDataset`.
     """
 
@@ -56,17 +57,23 @@ class IMGUR5K(AbstractDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
             img_folder, pre_transforms=convert_target_to_relative if not recognition_task else None, **kwargs
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
 
         # File existence check
         if not os.path.exists(label_path) or not os.path.exists(img_folder):
             raise FileNotFoundError(f"unable to locate {label_path if not os.path.exists(label_path) else img_folder}")
 
-        self.data: List[Tuple[Union[str, Path, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, Path, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         self.train = train
         np_dtype = np.float32
 
@@ -132,6 +139,8 @@ class IMGUR5K(AbstractDataset):
                                 tmp_img = Image.fromarray(crop)
                                 tmp_img.save(os.path.join(reco_folder_path, f"{reco_images_counter}.png"))
                                 reco_images_counter += 1
+                elif detection_task:
+                    self.data.append((img_path, np.asarray(box_targets, dtype=np_dtype)))
                 else:
                     self.data.append((img_path, dict(boxes=np.asarray(box_targets, dtype=np_dtype), labels=labels)))
 

doctr/datasets/sroie.py

@@ -33,6 +33,7 @@ class SROIE(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -52,6 +53,7 @@ class SROIE(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         url, sha256, name = self.TRAIN if train else self.TEST
@@ -63,10 +65,16 @@ class SROIE(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
 
         tmp_root = os.path.join(self.root, "images")
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         for img_path in tqdm(iterable=os.listdir(tmp_root), desc="Unpacking SROIE", total=len(os.listdir(tmp_root))):
@@ -94,6 +102,8 @@ class SROIE(VisionDataset):
                 for crop, label in zip(crops, labels):
                     if crop.shape[0] > 0 and crop.shape[1] > 0 and len(label) > 0:
                         self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((img_path, coords))
             else:
                 self.data.append((img_path, dict(boxes=coords, labels=labels)))
 

doctr/datasets/svhn.py

@@ -32,6 +32,7 @@ class SVHN(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -52,6 +53,7 @@ class SVHN(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         url, sha256, name = self.TRAIN if train else self.TEST
@@ -63,8 +65,14 @@ class SVHN(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         tmp_root = os.path.join(self.root, "train" if train else "test")
@@ -122,6 +130,8 @@ class SVHN(VisionDataset):
                     for crop, label in zip(crops, label_targets):
                         if crop.shape[0] > 0 and crop.shape[1] > 0 and len(label) > 0:
                             self.data.append((crop, label))
+                elif detection_task:
+                    self.data.append((img_name, box_targets))
                 else:
                     self.data.append((img_name, dict(boxes=box_targets, labels=label_targets)))
 

doctr/datasets/svt.py

@@ -32,6 +32,7 @@ class SVT(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -43,6 +44,7 @@ class SVT(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -53,8 +55,14 @@ class SVT(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         # Load xml data
@@ -108,6 +116,8 @@ class SVT(VisionDataset):
                 for crop, label in zip(crops, labels):
                     if crop.shape[0] > 0 and crop.shape[1] > 0 and len(label) > 0:
                         self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((name.text, boxes))
             else:
                 self.data.append((name.text, dict(boxes=boxes, labels=labels)))
 

doctr/datasets/synthtext.py

@@ -35,6 +35,7 @@ class SynthText(VisionDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `VisionDataset`.
     """
 
@@ -46,6 +47,7 @@ class SynthText(VisionDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
@@ -56,8 +58,14 @@ class SynthText(VisionDataset):
             pre_transforms=convert_target_to_relative if not recognition_task else None,
             **kwargs,
         )
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         self.train = train
-        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
         np_dtype = np.float32
 
         # Load mat data
@@ -111,6 +119,8 @@ class SynthText(VisionDataset):
                             tmp_img = Image.fromarray(crop)
                             tmp_img.save(os.path.join(reco_folder_path, f"{reco_images_counter}.png"))
                             reco_images_counter += 1
+            elif detection_task:
+                self.data.append((img_path[0], np.asarray(word_boxes, dtype=np_dtype)))
             else:
                 self.data.append((img_path[0], dict(boxes=np.asarray(word_boxes, dtype=np_dtype), labels=labels)))
 

doctr/datasets/utils.py

@@ -169,8 +169,13 @@ def encode_sequences(
     return encoded_data
 
 
-def convert_target_to_relative(img: ImageTensor, target: Dict[str, Any]) -> Tuple[ImageTensor, Dict[str, Any]]:
-    target["boxes"] = convert_to_relative_coords(target["boxes"], get_img_shape(img))
+def convert_target_to_relative(
+    img: ImageTensor, target: Union[np.ndarray, Dict[str, Any]]
+) -> Tuple[ImageTensor, Union[Dict[str, Any], np.ndarray]]:
+    if isinstance(target, np.ndarray):
+        target = convert_to_relative_coords(target, get_img_shape(img))
+    else:
+        target["boxes"] = convert_to_relative_coords(target["boxes"], get_img_shape(img))
     return img, target
 
 

doctr/datasets/vocabs.py

@@ -25,6 +25,7 @@ VOCABS: Dict[str, str] = {
     "hindi_punctuation": "।,?!:्ॐ॰॥॰",
     "bangla_letters": "অআইঈউঊঋএঐওঔকখগঘঙচছজঝঞটঠডঢণতথদধনপফবভমযরলশষসহ়ঽািীুূৃেৈোৌ্ৎংঃঁ",
     "bangla_digits": "০১২৩৪৫৬৭৮৯",
+    "generic_cyrillic_letters": "абвгдежзийклмнопрстуфхцчшщьюяАБВГДЕЖЗИЙКЛМНОПРСТУФХЦЧШЩЬЮЯ",
 }
 
 VOCABS["latin"] = VOCABS["digits"] + VOCABS["ascii_letters"] + VOCABS["punctuation"]
@@ -53,12 +54,15 @@ VOCABS["finnish"] = VOCABS["english"] + "äöÄÖ"
 VOCABS["swedish"] = VOCABS["english"] + "åäöÅÄÖ"
 VOCABS["vietnamese"] = (
     VOCABS["english"]
-    + "áàảạãăắằẳẵặâấầẩẫậéèẻẽẹêếềểễệóòỏõọôốồổộỗơớờởợỡúùủũụưứừửữựiíìỉĩịýỳỷỹỵ"
-    + "ÁÀẢẠÃĂẮẰẲẴẶÂẤẦẨẪẬÉÈẺẼẸÊẾỀỂỄỆÓÒỎÕỌÔỐỒỔỘỖƠỚỜỞỢỠÚÙỦŨỤƯỨỪỬỮỰIÍÌỈĨỊÝỲỶỸỴ"
+    + "áàảạãăắằẳẵặâấầẩẫậđéèẻẽẹêếềểễệóòỏõọôốồổộỗơớờởợỡúùủũụưứừửữựiíìỉĩịýỳỷỹỵ"
+    + "ÁÀẢẠÃĂẮẰẲẴẶÂẤẦẨẪẬĐÉÈẺẼẸÊẾỀỂỄỆÓÒỎÕỌÔỐỒỔỘỖƠỚỜỞỢỠÚÙỦŨỤƯỨỪỬỮỰIÍÌỈĨỊÝỲỶỸỴ"
 )
 VOCABS["hebrew"] = VOCABS["english"] + "אבגדהוזחטיכלמנסעפצקרשת" + "₪"
 VOCABS["hindi"] = VOCABS["hindi_letters"] + VOCABS["hindi_digits"] + VOCABS["hindi_punctuation"]
 VOCABS["bangla"] = VOCABS["bangla_letters"] + VOCABS["bangla_digits"]
+VOCABS["ukrainian"] = (
+    VOCABS["generic_cyrillic_letters"] + VOCABS["digits"] + VOCABS["punctuation"] + VOCABS["currency"] + "ґіїєҐІЇЄ₴"
+)
 VOCABS["multilingual"] = "".join(
     dict.fromkeys(
         VOCABS["french"]

doctr/datasets/wildreceipt.py

@@ -40,6 +40,7 @@ class WILDRECEIPT(AbstractDataset):
         train: whether the subset should be the training one
         use_polygons: whether polygons should be considered as rotated bounding box (instead of straight ones)
         recognition_task: whether the dataset should be used for recognition task
+        detection_task: whether the dataset should be used for detection task
         **kwargs: keyword arguments from `AbstractDataset`.
     """
 
@@ -50,11 +51,19 @@ class WILDRECEIPT(AbstractDataset):
         train: bool = True,
         use_polygons: bool = False,
         recognition_task: bool = False,
+        detection_task: bool = False,
         **kwargs: Any,
     ) -> None:
         super().__init__(
             img_folder, pre_transforms=convert_target_to_relative if not recognition_task else None, **kwargs
         )
+        # Task check
+        if recognition_task and detection_task:
+            raise ValueError(
+                "`recognition_task` and `detection_task` cannot be set to True simultaneously. "
+                + "To get the whole dataset with boxes and labels leave both parameters to False."
+            )
+
         # File existence check
         if not os.path.exists(label_path) or not os.path.exists(img_folder):
             raise FileNotFoundError(f"unable to locate {label_path if not os.path.exists(label_path) else img_folder}")
@@ -62,7 +71,7 @@ class WILDRECEIPT(AbstractDataset):
         tmp_root = img_folder
         self.train = train
         np_dtype = np.float32
-        self.data: List[Tuple[Union[str, Path, np.ndarray], Union[str, Dict[str, Any]]]] = []
+        self.data: List[Tuple[Union[str, Path, np.ndarray], Union[str, Dict[str, Any], np.ndarray]]] = []
 
         with open(label_path, "r") as file:
             data = file.read()
@@ -100,6 +109,8 @@ class WILDRECEIPT(AbstractDataset):
                 for crop, label in zip(crops, list(text_targets)):
                     if label and " " not in label:
                         self.data.append((crop, label))
+            elif detection_task:
+                self.data.append((img_path, np.asarray(box_targets, dtype=int).clip(min=0)))
             else:
                 self.data.append((
                     img_path,

doctr/file_utils.py

@@ -35,6 +35,20 @@ else:  # pragma: no cover
     logging.info("Disabling PyTorch because USE_TF is set")
     _torch_available = False
 
+# Compatibility fix to make sure tensorflow.keras stays at Keras 2
+if "TF_USE_LEGACY_KERAS" not in os.environ:
+    os.environ["TF_USE_LEGACY_KERAS"] = "1"
+
+elif os.environ["TF_USE_LEGACY_KERAS"] != "1":
+    raise ValueError(
+        "docTR is only compatible with Keras 2, but you have explicitly set `TF_USE_LEGACY_KERAS` to `0`. "
+    )
+
+
+def ensure_keras_v2() -> None:  # pragma: no cover
+    if not os.environ.get("TF_USE_LEGACY_KERAS") == "1":
+        os.environ["TF_USE_LEGACY_KERAS"] = "1"
+
 
 if USE_TF in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TORCH not in ENV_VARS_TRUE_VALUES:
     _tf_available = importlib.util.find_spec("tensorflow") is not None
@@ -65,6 +79,11 @@ if USE_TF in ENV_VARS_TRUE_AND_AUTO_VALUES and USE_TORCH not in ENV_VARS_TRUE_VA
             _tf_available = False
         else:
             logging.info(f"TensorFlow version {_tf_version} available.")
+            ensure_keras_v2()
+            import tensorflow as tf
+
+            # Enable eager execution - this is required for some models to work properly
+            tf.config.run_functions_eagerly(True)
 else:  # pragma: no cover
     logging.info("Disabling Tensorflow because USE_TORCH is set")
     _tf_available = False

doctr/io/elements.py

@@ -168,7 +168,7 @@ class Line(Element):
         if geometry is None:
             # Check whether this is a rotated or straight box
             box_resolution_fn = resolve_enclosing_rbbox if len(words[0].geometry) == 4 else resolve_enclosing_bbox
-            geometry = box_resolution_fn([w.geometry for w in words])  # type: ignore[operator]
+            geometry = box_resolution_fn([w.geometry for w in words])  # type: ignore[misc]
 
         super().__init__(words=words)
         self.geometry = geometry
@@ -232,7 +232,7 @@ class Block(Element):
             box_resolution_fn = (
                 resolve_enclosing_rbbox if isinstance(lines[0].geometry, np.ndarray) else resolve_enclosing_bbox
             )
-            geometry = box_resolution_fn(line_boxes + artefact_boxes)  # type: ignore[operator]
+            geometry = box_resolution_fn(line_boxes + artefact_boxes)  # type: ignore
 
         super().__init__(lines=lines, artefacts=artefacts)
         self.geometry = geometry
@@ -310,6 +310,10 @@ class Page(Element):
     def synthesize(self, **kwargs) -> np.ndarray:
         """Synthesize the page from the predictions
 
+        Args:
+        ----
+            **kwargs: keyword arguments passed to the `synthesize_page` method
+
         Returns
         -------
             synthesized page
@@ -493,7 +497,7 @@ class KIEPage(Element):
 
         Args:
         ----
-            **kwargs: keyword arguments passed to the matplotlib.pyplot.show method
+            **kwargs: keyword arguments passed to the `synthesize_kie_page` method
 
         Returns:
         -------
@@ -603,11 +607,15 @@ class Document(Element):
     def synthesize(self, **kwargs) -> List[np.ndarray]:
         """Synthesize all pages from their predictions
 
+        Args:
+        ----
+            **kwargs: keyword arguments passed to the `Page.synthesize` method
+
         Returns
         -------
             list of synthesized pages
         """
-        return [page.synthesize() for page in self.pages]
+        return [page.synthesize(**kwargs) for page in self.pages]
 
     def export_as_xml(self, **kwargs) -> List[Tuple[bytes, ET.ElementTree]]:
         """Export the document as XML (hOCR-format)

doctr/models/builder.py

@@ -266,7 +266,7 @@ class DocumentBuilder(NestedObject):
                 Line([
                     Word(
                         *word_preds[idx],
-                        tuple([tuple(pt) for pt in boxes[idx].tolist()]),  # type: ignore[arg-type]
+                        tuple(tuple(pt) for pt in boxes[idx].tolist()),  # type: ignore[arg-type]
                         float(objectness_scores[idx]),
                         crop_orientations[idx],
                     )
@@ -500,7 +500,7 @@ class KIEDocumentBuilder(DocumentBuilder):
             Prediction(
                 value=word_preds[idx][0],
                 confidence=word_preds[idx][1],
-                geometry=tuple([tuple(pt) for pt in boxes[idx].tolist()]),  # type: ignore[arg-type]
+                geometry=tuple(tuple(pt) for pt in boxes[idx].tolist()),  # type: ignore[arg-type]
                 objectness_score=float(objectness_scores[idx]),
                 crop_orientation=crop_orientations[idx],
             )