python-doctr 0.8.1__py3-none-any.whl → 0.10.0__py3-none-any.whl

doctr/__init__.py

@@ -1,3 +1,3 @@
-from . import io, models, datasets, transforms, utils
+from . import io, models, datasets, contrib, transforms, utils
 from .file_utils import is_tf_available, is_torch_available
 from .version import __version__  # noqa: F401

doctr/contrib/artefacts.py

@@ -0,0 +1,131 @@
+# Copyright (C) 2021-2024, Mindee.
+
+# This program is licensed under the Apache License 2.0.
+# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
+
+from typing import Any, Dict, List, Optional, Tuple
+
+import cv2
+import numpy as np
+
+from doctr.file_utils import requires_package
+
+from .base import _BasePredictor
+
+__all__ = ["ArtefactDetector"]
+
+default_cfgs: Dict[str, Dict[str, Any]] = {
+    "yolov8_artefact": {
+        "input_shape": (3, 1024, 1024),
+        "labels": ["bar_code", "qr_code", "logo", "photo"],
+        "url": "https://doctr-static.mindee.com/models?id=v0.8.1/yolo_artefact-f9d66f14.onnx&src=0",
+    },
+}
+
+
+class ArtefactDetector(_BasePredictor):
+    """
+    A class to detect artefacts in images
+
+    >>> from doctr.io import DocumentFile
+    >>> from doctr.contrib.artefacts import ArtefactDetector
+    >>> doc = DocumentFile.from_images(["path/to/image.jpg"])
+    >>> detector = ArtefactDetector()
+    >>> results = detector(doc)
+
+    Args:
+    ----
+        arch: the architecture to use
+        batch_size: the batch size to use
+        model_path: the path to the model to use
+        labels: the labels to use
+        input_shape: the input shape to use
+        mask_labels: the mask labels to use
+        conf_threshold: the confidence threshold to use
+        iou_threshold: the intersection over union threshold to use
+        **kwargs: additional arguments to be passed to `download_from_url`
+    """
+
+    def __init__(
+        self,
+        arch: str = "yolov8_artefact",
+        batch_size: int = 2,
+        model_path: Optional[str] = None,
+        labels: Optional[List[str]] = None,
+        input_shape: Optional[Tuple[int, int, int]] = None,
+        conf_threshold: float = 0.5,
+        iou_threshold: float = 0.5,
+        **kwargs: Any,
+    ) -> None:
+        super().__init__(batch_size=batch_size, url=default_cfgs[arch]["url"], model_path=model_path, **kwargs)
+        self.labels = labels or default_cfgs[arch]["labels"]
+        self.input_shape = input_shape or default_cfgs[arch]["input_shape"]
+        self.conf_threshold = conf_threshold
+        self.iou_threshold = iou_threshold
+
+    def preprocess(self, img: np.ndarray) -> np.ndarray:
+        return np.transpose(cv2.resize(img, (self.input_shape[2], self.input_shape[1])), (2, 0, 1)) / np.array(255.0)
+
+    def postprocess(self, output: List[np.ndarray], input_images: List[List[np.ndarray]]) -> List[List[Dict[str, Any]]]:
+        results = []
+
+        for batch in zip(output, input_images):
+            for out, img in zip(batch[0], batch[1]):
+                org_height, org_width = img.shape[:2]
+                width_scale, height_scale = org_width / self.input_shape[2], org_height / self.input_shape[1]
+                for res in out:
+                    sample_results = []
+                    for row in np.transpose(np.squeeze(res)):
+                        classes_scores = row[4:]
+                        max_score = np.amax(classes_scores)
+                        if max_score >= self.conf_threshold:
+                            class_id = np.argmax(classes_scores)
+                            x, y, w, h = row[0], row[1], row[2], row[3]
+                            # to rescaled xmin, ymin, xmax, ymax
+                            xmin = int((x - w / 2) * width_scale)
+                            ymin = int((y - h / 2) * height_scale)
+                            xmax = int((x + w / 2) * width_scale)
+                            ymax = int((y + h / 2) * height_scale)
+
+                            sample_results.append({
+                                "label": self.labels[class_id],
+                                "confidence": float(max_score),
+                                "box": [xmin, ymin, xmax, ymax],
+                            })
+
+                    # Filter out overlapping boxes
+                    boxes = [res["box"] for res in sample_results]
+                    scores = [res["confidence"] for res in sample_results]
+                    keep_indices = cv2.dnn.NMSBoxes(boxes, scores, self.conf_threshold, self.iou_threshold)  # type: ignore[arg-type]
+                    sample_results = [sample_results[i] for i in keep_indices]
+
+                    results.append(sample_results)
+
+        self._results = results
+        return results
+
+    def show(self, **kwargs: Any) -> None:
+        """
+        Display the results
+
+        Args:
+        ----
+            **kwargs: additional keyword arguments to be passed to `plt.show`
+        """
+        requires_package("matplotlib", "`.show()` requires matplotlib installed")
+        import matplotlib.pyplot as plt
+        from matplotlib.patches import Rectangle
+
+        # visualize the results with matplotlib
+        if self._results and self._inputs:
+            for img, res in zip(self._inputs, self._results):
+                plt.figure(figsize=(10, 10))
+                plt.imshow(img)
+                for obj in res:
+                    xmin, ymin, xmax, ymax = obj["box"]
+                    label = obj["label"]
+                    plt.text(xmin, ymin, f"{label} {obj['confidence']:.2f}", color="red")
+                    plt.gca().add_patch(
+                        Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, fill=False, edgecolor="red", linewidth=2)
+                    )
+                plt.show(**kwargs)

doctr/contrib/base.py

@@ -0,0 +1,105 @@
+# Copyright (C) 2021-2024, Mindee.
+
+# This program is licensed under the Apache License 2.0.
+# See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
+
+from typing import Any, List, Optional
+
+import numpy as np
+
+from doctr.file_utils import requires_package
+from doctr.utils.data import download_from_url
+
+
+class _BasePredictor:
+    """
+    Base class for all predictors
+
+    Args:
+    ----
+        batch_size: the batch size to use
+        url: the url to use to download a model if needed
+        model_path: the path to the model to use
+        **kwargs: additional arguments to be passed to `download_from_url`
+    """
+
+    def __init__(self, batch_size: int, url: Optional[str] = None, model_path: Optional[str] = None, **kwargs) -> None:
+        self.batch_size = batch_size
+        self.session = self._init_model(url, model_path, **kwargs)
+
+        self._inputs: List[np.ndarray] = []
+        self._results: List[Any] = []
+
+    def _init_model(self, url: Optional[str] = None, model_path: Optional[str] = None, **kwargs: Any) -> Any:
+        """
+        Download the model from the given url if needed
+
+        Args:
+        ----
+            url: the url to use
+            model_path: the path to the model to use
+            **kwargs: additional arguments to be passed to `download_from_url`
+
+        Returns:
+        -------
+            Any: the ONNX loaded model
+        """
+        requires_package("onnxruntime", "`.contrib` module requires `onnxruntime` to be installed.")
+        import onnxruntime as ort
+
+        if not url and not model_path:
+            raise ValueError("You must provide either a url or a model_path")
+        onnx_model_path = model_path if model_path else str(download_from_url(url, cache_subdir="models", **kwargs))  # type: ignore[arg-type]
+        return ort.InferenceSession(onnx_model_path, providers=["CUDAExecutionProvider", "CPUExecutionProvider"])
+
+    def preprocess(self, img: np.ndarray) -> np.ndarray:
+        """
+        Preprocess the input image
+
+        Args:
+        ----
+            img: the input image to preprocess
+
+        Returns:
+        -------
+            np.ndarray: the preprocessed image
+        """
+        raise NotImplementedError
+
+    def postprocess(self, output: List[np.ndarray], input_images: List[List[np.ndarray]]) -> Any:
+        """
+        Postprocess the model output
+
+        Args:
+        ----
+            output: the model output to postprocess
+            input_images: the input images used to generate the output
+
+        Returns:
+        -------
+            Any: the postprocessed output
+        """
+        raise NotImplementedError
+
+    def __call__(self, inputs: List[np.ndarray]) -> Any:
+        """
+        Call the model on the given inputs
+
+        Args:
+        ----
+            inputs: the inputs to use
+
+        Returns:
+        -------
+            Any: the postprocessed output
+        """
+        self._inputs = inputs
+        model_inputs = self.session.get_inputs()
+
+        batched_inputs = [inputs[i : i + self.batch_size] for i in range(0, len(inputs), self.batch_size)]
+        processed_batches = [
+            np.array([self.preprocess(img) for img in batch], dtype=np.float32) for batch in batched_inputs
+        ]
+
+        outputs = [self.session.run(None, {model_inputs[0].name: batch}) for batch in processed_batches]
+        return self.postprocess(outputs, batched_inputs)

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/datasets/pytorch.py

@@ -50,9 +50,9 @@ class AbstractDataset(_AbstractDataset):
     @staticmethod
     def collate_fn(samples: List[Tuple[torch.Tensor, Any]]) -> Tuple[torch.Tensor, List[Any]]:
         images, targets = zip(*samples)
-        images = torch.stack(images, dim=0)
+        images = torch.stack(images, dim=0)  # type: ignore[assignment]
 
-        return images, list(targets)
+        return images, list(targets)  # type: ignore[return-value]
 
 
 class VisionDataset(AbstractDataset, _VisionDataset):  # noqa: D101

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/generator/base.py

@@ -20,7 +20,7 @@ def synthesize_text_img(
     font_family: Optional[str] = None,
     background_color: Optional[Tuple[int, int, int]] = None,
     text_color: Optional[Tuple[int, int, int]] = None,
-) -> Image:
+) -> Image.Image:
     """Generate a synthetic text image
 
     Args:
@@ -81,7 +81,7 @@ class _CharacterGenerator(AbstractDataset):
         self._data: List[Image.Image] = []
         if cache_samples:
             self._data = [
-                (synthesize_text_img(char, font_family=font), idx)
+                (synthesize_text_img(char, font_family=font), idx)  # type: ignore[misc]
                 for idx, char in enumerate(self.vocab)
                 for font in self.font_family
             ]
@@ -93,7 +93,7 @@ class _CharacterGenerator(AbstractDataset):
         # Samples are already cached
         if len(self._data) > 0:
             idx = index % len(self._data)
-            pil_img, target = self._data[idx]
+            pil_img, target = self._data[idx]  # type: ignore[misc]
         else:
             target = index % len(self.vocab)
             pil_img = synthesize_text_img(self.vocab[target], font_family=random.choice(self.font_family))
@@ -132,7 +132,8 @@ class _WordGenerator(AbstractDataset):
         if cache_samples:
             _words = [self._generate_string(*self.wordlen_range) for _ in range(num_samples)]
             self._data = [
-                (synthesize_text_img(text, font_family=random.choice(self.font_family)), text) for text in _words
+                (synthesize_text_img(text, font_family=random.choice(self.font_family)), text)  # type: ignore[misc]
+                for text in _words
             ]
 
     def _generate_string(self, min_chars: int, max_chars: int) -> str:
@@ -145,7 +146,7 @@ class _WordGenerator(AbstractDataset):
     def _read_sample(self, index: int) -> Tuple[Any, str]:
         # Samples are already cached
         if len(self._data) > 0:
-            pil_img, target = self._data[index]
+            pil_img, target = self._data[index]  # type: ignore[misc]
         else:
             target = self._generate_string(*self.wordlen_range)
             pil_img = synthesize_text_img(target, font_family=random.choice(self.font_family))

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
 
@@ -112,7 +119,7 @@ class IMGUR5K(AbstractDataset):
                 if ann["word"] != "."
             ]
             # (x, y) coordinates of top left, top right, bottom right, bottom left corners
-            box_targets = [cv2.boxPoints(((box[0], box[1]), (box[2], box[3]), box[4])) for box in _boxes]  # type: ignore[arg-type]
+            box_targets = [cv2.boxPoints(((box[0], box[1]), (box[2], box[3]), box[4])) for box in _boxes]
 
             if not use_polygons:
                 # xmin, ymin, xmax, ymax
@@ -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/loader.py

@@ -9,8 +9,6 @@ from typing import Callable, Optional
 import numpy as np
 import tensorflow as tf
 
-from doctr.utils.multithreading import multithread_exec
-
 __all__ = ["DataLoader"]
 
 
@@ -47,7 +45,6 @@ class DataLoader:
         shuffle: whether the samples should be shuffled before passing it to the iterator
         batch_size: number of elements in each batch
         drop_last: if `True`, drops the last batch if it isn't full
-        num_workers: number of workers to use for data loading
         collate_fn: function to merge samples into a batch
     """
 
@@ -57,7 +54,6 @@ class DataLoader:
         shuffle: bool = True,
         batch_size: int = 1,
         drop_last: bool = False,
-        num_workers: Optional[int] = None,
         collate_fn: Optional[Callable] = None,
     ) -> None:
         self.dataset = dataset
@@ -69,7 +65,6 @@ class DataLoader:
             self.collate_fn = self.dataset.collate_fn if hasattr(self.dataset, "collate_fn") else default_collate
         else:
             self.collate_fn = collate_fn
-        self.num_workers = num_workers
         self.reset()
 
     def __len__(self) -> int:
@@ -92,7 +87,7 @@ class DataLoader:
             idx = self._num_yielded * self.batch_size
             indices = self.indices[idx : min(len(self.dataset), idx + self.batch_size)]
 
-            samples = list(multithread_exec(self.dataset.__getitem__, indices, threads=self.num_workers))
+            samples = list(map(self.dataset.__getitem__, indices))
 
             batch_data = self.collate_fn(samples)