python-doctr 0.8.0__py3-none-any.whl → 0.9.0__py3-none-any.whl

doctr/transforms/functional/pytorch.py

@@ -35,9 +35,9 @@ def invert_colors(img: torch.Tensor, min_val: float = 0.6) -> torch.Tensor:
     rgb_shift = min_val + (1 - min_val) * torch.rand(shift_shape)
     # Inverse the color
     if out.dtype == torch.uint8:
-        out = (out.to(dtype=rgb_shift.dtype) * rgb_shift).to(dtype=torch.uint8)  # type: ignore[attr-defined]
+        out = (out.to(dtype=rgb_shift.dtype) * rgb_shift).to(dtype=torch.uint8)
     else:
-        out = out * rgb_shift.to(dtype=out.dtype)  # type: ignore[attr-defined]
+        out = out * rgb_shift.to(dtype=out.dtype)
     # Inverse the color
     out = 255 - out if out.dtype == torch.uint8 else 1 - out
     return out
@@ -81,7 +81,7 @@ def rotate_sample(
     rotated_geoms: np.ndarray = rotate_abs_geoms(
         _geoms,
         angle,
-        img.shape[1:],
+        img.shape[1:],  # type: ignore[arg-type]
         expand,
     ).astype(np.float32)
 
@@ -132,7 +132,7 @@ def random_shadow(img: torch.Tensor, opacity_range: Tuple[float, float], **kwarg
     -------
         shaded image
     """
-    shadow_mask = create_shadow_mask(img.shape[1:], **kwargs)
+    shadow_mask = create_shadow_mask(img.shape[1:], **kwargs)  # type: ignore[arg-type]
 
     opacity = np.random.uniform(*opacity_range)
     shadow_tensor = 1 - torch.from_numpy(shadow_mask[None, ...])

doctr/transforms/modules/base.py

@@ -5,7 +5,7 @@
 
 import math
 import random
-from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+from typing import Any, Callable, List, Optional, Tuple, Union
 
 import numpy as np
 
@@ -168,11 +168,11 @@ class OneOf(NestedObject):
     def __init__(self, transforms: List[Callable[[Any], Any]]) -> None:
         self.transforms = transforms
 
-    def __call__(self, img: Any) -> Any:
+    def __call__(self, img: Any, target: Optional[np.ndarray] = None) -> Union[Any, Tuple[Any, np.ndarray]]:
         # Pick transformation
         transfo = self.transforms[int(random.random() * len(self.transforms))]
         # Apply
-        return transfo(img)
+        return transfo(img) if target is None else transfo(img, target)  # type: ignore[call-arg]
 
 
 class RandomApply(NestedObject):
@@ -261,17 +261,39 @@ class RandomCrop(NestedObject):
     def extra_repr(self) -> str:
         return f"scale={self.scale}, ratio={self.ratio}"
 
-    def __call__(self, img: Any, target: Dict[str, np.ndarray]) -> Tuple[Any, Dict[str, np.ndarray]]:
+    def __call__(self, img: Any, target: np.ndarray) -> Tuple[Any, np.ndarray]:
         scale = random.uniform(self.scale[0], self.scale[1])
         ratio = random.uniform(self.ratio[0], self.ratio[1])
-        # Those might overflow
-        crop_h = math.sqrt(scale * ratio)
-        crop_w = math.sqrt(scale / ratio)
-        xmin, ymin = random.uniform(0, 1 - crop_w), random.uniform(0, 1 - crop_h)
-        xmax, ymax = xmin + crop_w, ymin + crop_h
-        # Clip them
-        xmin, ymin = max(xmin, 0), max(ymin, 0)
-        xmax, ymax = min(xmax, 1), min(ymax, 1)
-
-        croped_img, crop_boxes = F.crop_detection(img, target["boxes"], (xmin, ymin, xmax, ymax))
-        return croped_img, dict(boxes=crop_boxes)
+
+        height, width = img.shape[:2]
+
+        # Calculate crop size
+        crop_area = scale * width * height
+        aspect_ratio = ratio * (width / height)
+        crop_width = int(round(math.sqrt(crop_area * aspect_ratio)))
+        crop_height = int(round(math.sqrt(crop_area / aspect_ratio)))
+
+        # Ensure crop size does not exceed image dimensions
+        crop_width = min(crop_width, width)
+        crop_height = min(crop_height, height)
+
+        # Randomly select crop position
+        x = random.randint(0, width - crop_width)
+        y = random.randint(0, height - crop_height)
+
+        # relative crop box
+        crop_box = (x / width, y / height, (x + crop_width) / width, (y + crop_height) / height)
+        if target.shape[1:] == (4, 2):
+            min_xy = np.min(target, axis=1)
+            max_xy = np.max(target, axis=1)
+            _target = np.concatenate((min_xy, max_xy), axis=1)
+        else:
+            _target = target
+
+        # Crop image and targets
+        croped_img, crop_boxes = F.crop_detection(img, _target, crop_box)
+        # hard fallback if no box is kept
+        if crop_boxes.shape[0] == 0:
+            return img, target
+        # clip boxes
+        return croped_img, np.clip(crop_boxes, 0, 1)

doctr/transforms/modules/pytorch.py

@@ -4,7 +4,7 @@
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
 import math
-from typing import Any, Dict, Optional, Tuple, Union
+from typing import Optional, Tuple, Union
 
 import numpy as np
 import torch
@@ -15,7 +15,7 @@ from torchvision.transforms import transforms as T
 
 from ..functional.pytorch import random_shadow
 
-__all__ = ["Resize", "GaussianNoise", "ChannelShuffle", "RandomHorizontalFlip", "RandomShadow"]
+__all__ = ["Resize", "GaussianNoise", "ChannelShuffle", "RandomHorizontalFlip", "RandomShadow", "RandomResize"]
 
 
 class Resize(T.Resize):
@@ -135,9 +135,9 @@ class GaussianNoise(torch.nn.Module):
         # Reshape the distribution
         noise = self.mean + 2 * self.std * torch.rand(x.shape, device=x.device) - self.std
         if x.dtype == torch.uint8:
-            return (x + 255 * noise).round().clamp(0, 255).to(dtype=torch.uint8)  # type: ignore[attr-defined]
+            return (x + 255 * noise).round().clamp(0, 255).to(dtype=torch.uint8)
         else:
-            return (x + noise.to(dtype=x.dtype)).clamp(0, 1)  # type: ignore[attr-defined]
+            return (x + noise.to(dtype=x.dtype)).clamp(0, 1)
 
     def extra_repr(self) -> str:
         return f"mean={self.mean}, std={self.std}"
@@ -159,13 +159,16 @@ class RandomHorizontalFlip(T.RandomHorizontalFlip):
     """Randomly flip the input image horizontally"""
 
     def forward(
-        self, img: Union[torch.Tensor, Image], target: Dict[str, Any]
-    ) -> Tuple[Union[torch.Tensor, Image], Dict[str, Any]]:
+        self, img: Union[torch.Tensor, Image], target: np.ndarray
+    ) -> Tuple[Union[torch.Tensor, Image], np.ndarray]:
         if torch.rand(1) < self.p:
             _img = F.hflip(img)
             _target = target.copy()
             # Changing the relative bbox coordinates
-            _target["boxes"][:, ::2] = 1 - target["boxes"][:, [2, 0]]
+            if target.shape[1:] == (4,):
+                _target[:, ::2] = 1 - target[:, [2, 0]]
+            else:
+                _target[..., 0] = 1 - target[..., 0]
             return _img, _target
         return img, target
 
@@ -199,7 +202,7 @@ class RandomShadow(torch.nn.Module):
                             self.opacity_range,
                         )
                     )
-                    .round()  # type: ignore[attr-defined]
+                    .round()
                     .clip(0, 255)
                     .to(dtype=torch.uint8)
                 )
@@ -210,3 +213,58 @@ class RandomShadow(torch.nn.Module):
 
     def extra_repr(self) -> str:
         return f"opacity_range={self.opacity_range}"
+
+
+class RandomResize(torch.nn.Module):
+    """Randomly resize the input image and align corresponding targets
+
+    >>> import torch
+    >>> from doctr.transforms import RandomResize
+    >>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
+    >>> out = transfo(torch.rand((3, 64, 64)))
+
+    Args:
+    ----
+        scale_range: range of the resizing factor for width and height (independently)
+        preserve_aspect_ratio: whether to preserve the aspect ratio of the image,
+            given a float value, the aspect ratio will be preserved with this probability
+        symmetric_pad: whether to symmetrically pad the image,
+            given a float value, the symmetric padding will be applied with this probability
+        p: probability to apply the transformation
+    """
+
+    def __init__(
+        self,
+        scale_range: Tuple[float, float] = (0.3, 0.9),
+        preserve_aspect_ratio: Union[bool, float] = False,
+        symmetric_pad: Union[bool, float] = False,
+        p: float = 0.5,
+    ) -> None:
+        super().__init__()
+        self.scale_range = scale_range
+        self.preserve_aspect_ratio = preserve_aspect_ratio
+        self.symmetric_pad = symmetric_pad
+        self.p = p
+        self._resize = Resize
+
+    def forward(self, img: torch.Tensor, target: np.ndarray) -> Tuple[torch.Tensor, np.ndarray]:
+        if torch.rand(1) < self.p:
+            scale_h = np.random.uniform(*self.scale_range)
+            scale_w = np.random.uniform(*self.scale_range)
+            new_size = (int(img.shape[-2] * scale_h), int(img.shape[-1] * scale_w))
+
+            _img, _target = self._resize(
+                new_size,
+                preserve_aspect_ratio=self.preserve_aspect_ratio
+                if isinstance(self.preserve_aspect_ratio, bool)
+                else bool(torch.rand(1) <= self.symmetric_pad),
+                symmetric_pad=self.symmetric_pad
+                if isinstance(self.symmetric_pad, bool)
+                else bool(torch.rand(1) <= self.symmetric_pad),
+            )(img, target)
+
+            return _img, _target
+        return img, target
+
+    def extra_repr(self) -> str:
+        return f"scale_range={self.scale_range}, preserve_aspect_ratio={self.preserve_aspect_ratio}, symmetric_pad={self.symmetric_pad}, p={self.p}"  # noqa: E501

doctr/transforms/modules/tensorflow.py

@@ -4,7 +4,7 @@
 # See LICENSE or go to <https://opensource.org/licenses/Apache-2.0> for full license details.
 
 import random
-from typing import Any, Callable, Dict, Iterable, List, Optional, Tuple, Union
+from typing import Any, Callable, Iterable, List, Optional, Tuple, Union
 
 import numpy as np
 import tensorflow as tf
@@ -30,6 +30,7 @@ __all__ = [
     "GaussianNoise",
     "RandomHorizontalFlip",
     "RandomShadow",
+    "RandomResize",
 ]
 
 
@@ -457,10 +458,7 @@ class RandomHorizontalFlip(NestedObject):
     >>> from doctr.transforms import RandomHorizontalFlip
     >>> transfo = RandomHorizontalFlip(p=0.5)
     >>> image = tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1)
-    >>> target = {
-    >>> "boxes": np.array([[0.1, 0.1, 0.4, 0.5] ], dtype= np.float32),
-    >>> "labels": np.ones(1, dtype= np.int64)
-    >>> }
+    >>> target = np.array([[0.1, 0.1, 0.4, 0.5] ], dtype= np.float32)
     >>> out = transfo(image, target)
 
     Args:
@@ -472,12 +470,15 @@ class RandomHorizontalFlip(NestedObject):
         super().__init__()
         self.p = p
 
-    def __call__(self, img: Union[tf.Tensor, np.ndarray], target: Dict[str, Any]) -> Tuple[tf.Tensor, Dict[str, Any]]:
+    def __call__(self, img: Union[tf.Tensor, np.ndarray], target: np.ndarray) -> Tuple[tf.Tensor, np.ndarray]:
         if np.random.rand(1) <= self.p:
             _img = tf.image.flip_left_right(img)
             _target = target.copy()
             # Changing the relative bbox coordinates
-            _target["boxes"][:, ::2] = 1 - target["boxes"][:, [2, 0]]
+            if target.shape[1:] == (4,):
+                _target[:, ::2] = 1 - target[:, [2, 0]]
+            else:
+                _target[..., 0] = 1 - target[..., 0]
             return _img, _target
         return img, target
 
@@ -515,3 +516,58 @@ class RandomShadow(NestedObject):
 
     def extra_repr(self) -> str:
         return f"opacity_range={self.opacity_range}"
+
+
+class RandomResize(NestedObject):
+    """Randomly resize the input image and align corresponding targets
+
+    >>> import tensorflow as tf
+    >>> from doctr.transforms import RandomResize
+    >>> transfo = RandomResize((0.3, 0.9), preserve_aspect_ratio=True, symmetric_pad=True, p=0.5)
+    >>> out = transfo(tf.random.uniform(shape=[64, 64, 3], minval=0, maxval=1))
+
+    Args:
+    ----
+        scale_range: range of the resizing factor for width and height (independently)
+        preserve_aspect_ratio: whether to preserve the aspect ratio of the image,
+            given a float value, the aspect ratio will be preserved with this probability
+        symmetric_pad: whether to symmetrically pad the image,
+            given a float value, the symmetric padding will be applied with this probability
+        p: probability to apply the transformation
+    """
+
+    def __init__(
+        self,
+        scale_range: Tuple[float, float] = (0.3, 0.9),
+        preserve_aspect_ratio: Union[bool, float] = False,
+        symmetric_pad: Union[bool, float] = False,
+        p: float = 0.5,
+    ):
+        super().__init__()
+        self.scale_range = scale_range
+        self.preserve_aspect_ratio = preserve_aspect_ratio
+        self.symmetric_pad = symmetric_pad
+        self.p = p
+        self._resize = Resize
+
+    def __call__(self, img: tf.Tensor, target: np.ndarray) -> Tuple[tf.Tensor, np.ndarray]:
+        if np.random.rand(1) <= self.p:
+            scale_h = random.uniform(*self.scale_range)
+            scale_w = random.uniform(*self.scale_range)
+            new_size = (int(img.shape[-3] * scale_h), int(img.shape[-2] * scale_w))
+
+            _img, _target = self._resize(
+                new_size,
+                preserve_aspect_ratio=self.preserve_aspect_ratio
+                if isinstance(self.preserve_aspect_ratio, bool)
+                else bool(np.random.rand(1) <= self.symmetric_pad),
+                symmetric_pad=self.symmetric_pad
+                if isinstance(self.symmetric_pad, bool)
+                else bool(np.random.rand(1) <= self.symmetric_pad),
+            )(img, target)
+
+            return _img, _target
+        return img, target
+
+    def extra_repr(self) -> str:
+        return f"scale_range={self.scale_range}, preserve_aspect_ratio={self.preserve_aspect_ratio}, symmetric_pad={self.symmetric_pad}, p={self.p}"  # noqa: E501

doctr/utils/fonts.py

@@ -5,14 +5,16 @@
 
 import logging
 import platform
-from typing import Optional
+from typing import Optional, Union
 
 from PIL import ImageFont
 
 __all__ = ["get_font"]
 
 
-def get_font(font_family: Optional[str] = None, font_size: int = 13) -> ImageFont.ImageFont:
+def get_font(
+    font_family: Optional[str] = None, font_size: int = 13
+) -> Union[ImageFont.FreeTypeFont, ImageFont.ImageFont]:
     """Resolves a compatible ImageFont for the system
 
     Args:
@@ -28,14 +30,14 @@ def get_font(font_family: Optional[str] = None, font_size: int = 13) -> ImageFon
     if font_family is None:
         try:
             font = ImageFont.truetype("FreeMono.ttf" if platform.system() == "Linux" else "Arial.ttf", font_size)
-        except OSError:
-            font = ImageFont.load_default()
+        except OSError:  # pragma: no cover
+            font = ImageFont.load_default()  # type: ignore[assignment]
             logging.warning(
                 "unable to load recommended font family. Loading default PIL font,"
                 "font size issues may be expected."
                 "To prevent this, it is recommended to specify the value of 'font_family'."
             )
-    else:
+    else:  # pragma: no cover
         font = ImageFont.truetype(font_family, font_size)
 
     return font

doctr/utils/geometry.py

@@ -25,6 +25,7 @@ __all__ = [
     "rotate_abs_geoms",
     "extract_crops",
     "extract_rcrops",
+    "detach_scores",
 ]
 
 
@@ -57,6 +58,28 @@ def polygon_to_bbox(polygon: Polygon4P) -> BoundingBox:
     return (min(x), min(y)), (max(x), max(y))
 
 
+def detach_scores(boxes: List[np.ndarray]) -> Tuple[List[np.ndarray], List[np.ndarray]]:
+    """Detach the objectness scores from box predictions
+
+    Args:
+    ----
+        boxes: list of arrays with boxes of shape (N, 5) or (N, 5, 2)
+
+    Returns:
+    -------
+        a tuple of two lists: the first one contains the boxes without the objectness scores,
+        the second one contains the objectness scores
+    """
+
+    def _detach(boxes: np.ndarray) -> Tuple[np.ndarray, np.ndarray]:
+        if boxes.ndim == 2:
+            return boxes[:, :-1], boxes[:, -1]
+        return boxes[:, :-1], boxes[:, -1, -1]
+
+    loc_preds, obj_scores = zip(*(_detach(box) for box in boxes))
+    return list(loc_preds), list(obj_scores)
+
+
 def resolve_enclosing_bbox(bboxes: Union[List[BoundingBox], np.ndarray]) -> Union[BoundingBox, np.ndarray]:
     """Compute enclosing bbox either from:
 
@@ -64,18 +87,18 @@ def resolve_enclosing_bbox(bboxes: Union[List[BoundingBox], np.ndarray]) -> Unio
     ----
         bboxes: boxes in one of the following formats:
 
-            - an array of boxes: (*, 5), where boxes have this shape:
-            (xmin, ymin, xmax, ymax, score)
+            - an array of boxes: (*, 4), where boxes have this shape:
+            (xmin, ymin, xmax, ymax)
 
             - a list of BoundingBox
 
     Returns:
     -------
-        a (1, 5) array (enclosing boxarray), or a BoundingBox
+        a (1, 4) array (enclosing boxarray), or a BoundingBox
     """
     if isinstance(bboxes, np.ndarray):
-        xmin, ymin, xmax, ymax, score = np.split(bboxes, 5, axis=1)
-        return np.array([xmin.min(), ymin.min(), xmax.max(), ymax.max(), score.mean()])
+        xmin, ymin, xmax, ymax = np.split(bboxes, 4, axis=1)
+        return np.array([xmin.min(), ymin.min(), xmax.max(), ymax.max()])
     else:
         x, y = zip(*[point for box in bboxes for point in box])
         return (min(x), min(y)), (max(x), max(y))
@@ -88,21 +111,21 @@ def resolve_enclosing_rbbox(rbboxes: List[np.ndarray], intermed_size: int = 1024
     ----
         rbboxes: boxes in one of the following formats:
 
-            - an array of boxes: (*, 5), where boxes have this shape:
-            (xmin, ymin, xmax, ymax, score)
+            - an array of boxes: (*, 4, 2), where boxes have this shape:
+            (x1, y1), (x2, y2), (x3, y3), (x4, y4)
 
             - a list of BoundingBox
         intermed_size: size of the intermediate image
 
     Returns:
     -------
-        a (1, 5) array (enclosing boxarray), or a BoundingBox
+        a (4, 2) array (enclosing rotated box)
     """
     cloud: np.ndarray = np.concatenate(rbboxes, axis=0)
     # Convert to absolute for minAreaRect
     cloud *= intermed_size
     rect = cv2.minAreaRect(cloud.astype(np.int32))
-    return cv2.boxPoints(rect) / intermed_size  # type: ignore[operator]
+    return cv2.boxPoints(rect) / intermed_size  # type: ignore[return-value]
 
 
 def rotate_abs_points(points: np.ndarray, angle: float = 0.0) -> np.ndarray:
@@ -232,7 +255,7 @@ def rotate_boxes(
 
     Args:
     ----
-        loc_preds: (N, 5) or (N, 4, 2) array of RELATIVE boxes
+        loc_preds: (N, 4) or (N, 4, 2) array of RELATIVE boxes
         angle: angle between -90 and +90 degrees
         orig_shape: shape of the origin image
         min_angle: minimum angle to rotate boxes
@@ -320,7 +343,7 @@ def rotate_image(
             # Pad height
             else:
                 h_pad, w_pad = int(rot_img.shape[1] * image.shape[0] / image.shape[1] - rot_img.shape[0]), 0
-            rot_img = np.pad(rot_img, ((h_pad // 2, h_pad - h_pad // 2), (w_pad // 2, w_pad - w_pad // 2), (0, 0)))
+            rot_img = np.pad(rot_img, ((h_pad // 2, h_pad - h_pad // 2), (w_pad // 2, w_pad - w_pad // 2), (0, 0)))  # type: ignore[assignment]
         if preserve_origin_shape:
             # rescale
             rot_img = cv2.resize(rot_img, image.shape[:-1][::-1], interpolation=cv2.INTER_LINEAR)
@@ -453,4 +476,4 @@ def extract_rcrops(
         )
         for idx in range(_boxes.shape[0])
     ]
-    return crops
+    return crops  # type: ignore[return-value]