python-doctr 0.8.1__py3-none-any.whl → 0.9.0__py3-none-any.whl

doctr/models/predictor/tensorflow.py

@@ -9,10 +9,10 @@ import numpy as np
 import tensorflow as tf
 
 from doctr.io.elements import Document
-from doctr.models._utils import estimate_orientation, get_language
+from doctr.models._utils import get_language
 from doctr.models.detection.predictor import DetectionPredictor
 from doctr.models.recognition.predictor import RecognitionPredictor
-from doctr.utils.geometry import rotate_image
+from doctr.utils.geometry import detach_scores
 from doctr.utils.repr import NestedObject
 
 from .base import _OCRPredictor
@@ -56,7 +56,13 @@ class OCRPredictor(NestedObject, _OCRPredictor):
         self.det_predictor = det_predictor
         self.reco_predictor = reco_predictor
         _OCRPredictor.__init__(
-            self, assume_straight_pages, straighten_pages, preserve_aspect_ratio, symmetric_pad, **kwargs
+            self,
+            assume_straight_pages,
+            straighten_pages,
+            preserve_aspect_ratio,
+            symmetric_pad,
+            detect_orientation,
+            **kwargs,
         )
         self.detect_orientation = detect_orientation
         self.detect_language = detect_language
@@ -81,19 +87,16 @@ class OCRPredictor(NestedObject, _OCRPredictor):
             for out_map in out_maps
         ]
         if self.detect_orientation:
-            origin_page_orientations = [estimate_orientation(seq_map) for seq_map in seg_maps]
+            general_pages_orientations, origin_pages_orientations = self._get_orientations(pages, seg_maps)
             orientations = [
-                {"value": orientation_page, "confidence": None} for orientation_page in origin_page_orientations
+                {"value": orientation_page, "confidence": None} for orientation_page in origin_pages_orientations
             ]
         else:
             orientations = None
+            general_pages_orientations = None
+            origin_pages_orientations = None
         if self.straighten_pages:
-            origin_page_orientations = (
-                origin_page_orientations
-                if self.detect_orientation
-                else [estimate_orientation(seq_map) for seq_map in seg_maps]
-            )
-            pages = [rotate_image(page, -angle, expand=False) for page, angle in zip(pages, origin_page_orientations)]
+            pages = self._straighten_pages(pages, seg_maps, general_pages_orientations, origin_pages_orientations)
             # forward again to get predictions on straight pages
             loc_preds_dict = self.det_predictor(pages, **kwargs)  # type: ignore[assignment]
 
@@ -101,9 +104,8 @@ class OCRPredictor(NestedObject, _OCRPredictor):
             len(loc_pred) == 1 for loc_pred in loc_preds_dict
         ), "Detection Model in ocr_predictor should output only one class"
         loc_preds: List[np.ndarray] = [list(loc_pred.values())[0] for loc_pred in loc_preds_dict]  # type: ignore[union-attr]
-
-        # Rectify crops if aspect ratio
-        loc_preds = self._remove_padding(pages, loc_preds)
+        # Detach objectness scores from loc_preds
+        loc_preds, objectness_scores = detach_scores(loc_preds)
 
         # Apply hooks to loc_preds if any
         for hook in self.hooks:
@@ -113,14 +115,20 @@ class OCRPredictor(NestedObject, _OCRPredictor):
         crops, loc_preds = self._prepare_crops(
             pages, loc_preds, channels_last=True, assume_straight_pages=self.assume_straight_pages
         )
-        # Rectify crop orientation
+        # Rectify crop orientation and get crop orientation predictions
+        crop_orientations: Any = []
         if not self.assume_straight_pages:
-            crops, loc_preds = self._rectify_crops(crops, loc_preds)
+            crops, loc_preds, _crop_orientations = self._rectify_crops(crops, loc_preds)
+            crop_orientations = [
+                {"value": orientation[0], "confidence": orientation[1]} for orientation in _crop_orientations
+            ]
 
         # Identify character sequences
         word_preds = self.reco_predictor([crop for page_crops in crops for crop in page_crops], **kwargs)
+        if not crop_orientations:
+            crop_orientations = [{"value": 0, "confidence": None} for _ in word_preds]
 
-        boxes, text_preds = self._process_predictions(loc_preds, word_preds)
+        boxes, text_preds, crop_orientations = self._process_predictions(loc_preds, word_preds, crop_orientations)
 
         if self.detect_language:
             languages = [get_language(" ".join([item[0] for item in text_pred])) for text_pred in text_preds]
@@ -131,8 +139,10 @@ class OCRPredictor(NestedObject, _OCRPredictor):
         out = self.doc_builder(
             pages,
             boxes,
+            objectness_scores,
             text_preds,
             origin_page_shapes,  # type: ignore[arg-type]
+            crop_orientations,
             orientations,
             languages_dict,
         )

doctr/models/preprocessor/pytorch.py

@@ -79,7 +79,7 @@ class PreProcessor(nn.Module):
         else:
             x = x.to(dtype=torch.float32)  # type: ignore[union-attr]
 
-        return x  # type: ignore[return-value]
+        return x
 
     def __call__(self, x: Union[torch.Tensor, np.ndarray, List[Union[torch.Tensor, np.ndarray]]]) -> List[torch.Tensor]:
         """Prepare document data for model forwarding
@@ -103,7 +103,7 @@ class PreProcessor(nn.Module):
             elif x.dtype not in (torch.uint8, torch.float16, torch.float32):
                 raise TypeError("unsupported data type for torch.Tensor")
             # Resizing
-            if x.shape[-2] != self.resize.size[0] or x.shape[-1] != self.resize.size[1]:  # type: ignore[union-attr]
+            if x.shape[-2] != self.resize.size[0] or x.shape[-1] != self.resize.size[1]:
                 x = F.resize(
                     x, self.resize.size, interpolation=self.resize.interpolation, antialias=self.resize.antialias
                 )
@@ -118,11 +118,11 @@ class PreProcessor(nn.Module):
             # Sample transform (to tensor, resize)
             samples = list(multithread_exec(self.sample_transforms, x))
             # Batching
-            batches = self.batch_inputs(samples)  # type: ignore[assignment]
+            batches = self.batch_inputs(samples)
         else:
             raise TypeError(f"invalid input type: {type(x)}")
 
         # Batch transforms (normalize)
         batches = list(multithread_exec(self.normalize, batches))
 
-        return batches  # type: ignore[return-value]
+        return batches

doctr/models/preprocessor/tensorflow.py

@@ -41,6 +41,7 @@ class PreProcessor(NestedObject):
         self.resize = Resize(output_size, **kwargs)
         # Perform the division by 255 at the same time
         self.normalize = Normalize(mean, std)
+        self._runs_on_cuda = tf.test.is_gpu_available()
 
     def batch_inputs(self, samples: List[tf.Tensor]) -> List[tf.Tensor]:
         """Gather samples into batches for inference purposes
@@ -113,13 +114,13 @@ class PreProcessor(NestedObject):
 
         elif isinstance(x, list) and all(isinstance(sample, (np.ndarray, tf.Tensor)) for sample in x):
             # Sample transform (to tensor, resize)
-            samples = list(multithread_exec(self.sample_transforms, x))
+            samples = list(multithread_exec(self.sample_transforms, x, threads=1 if self._runs_on_cuda else None))
             # Batching
             batches = self.batch_inputs(samples)
         else:
             raise TypeError(f"invalid input type: {type(x)}")
 
         # Batch transforms (normalize)
-        batches = list(multithread_exec(self.normalize, batches))
+        batches = list(multithread_exec(self.normalize, batches, threads=1 if self._runs_on_cuda else None))
 
         return batches

doctr/models/recognition/master/pytorch.py

@@ -107,7 +107,7 @@ class MASTER(_MASTER, nn.Module):
         # NOTE: nn.TransformerDecoder takes the inverse from this implementation
         # [True, True, True, ..., False, False, False] -> False is masked
         # (N, 1, 1, max_length)
-        target_pad_mask = (target != self.vocab_size + 2).unsqueeze(1).unsqueeze(1)  # type: ignore[attr-defined]
+        target_pad_mask = (target != self.vocab_size + 2).unsqueeze(1).unsqueeze(1)
         target_length = target.size(1)
         # sub mask filled diagonal with True = see and False = masked (max_length, max_length)
         # NOTE: onnxruntime tril/triu works only with float currently (onnxruntime 1.11.1 - opset 14)
@@ -142,7 +142,7 @@ class MASTER(_MASTER, nn.Module):
         # Input length : number of timesteps
         input_len = model_output.shape[1]
         # Add one for additional <eos> token (sos disappear in shift!)
-        seq_len = seq_len + 1  # type: ignore[assignment]
+        seq_len = seq_len + 1
         # Compute loss: don't forget to shift gt! Otherwise the model learns to output the gt[t-1]!
         # The "masked" first gt char is <sos>. Delete last logit of the model output.
         cce = F.cross_entropy(model_output[:, :-1, :].permute(0, 2, 1), gt[:, 1:], reduction="none")

doctr/models/recognition/parseq/pytorch.py

@@ -212,7 +212,7 @@ class PARSeq(_PARSeq, nn.Module):
 
         sos_idx = torch.zeros(len(final_perms), 1, device=seqlen.device)
         eos_idx = torch.full((len(final_perms), 1), max_num_chars + 1, device=seqlen.device)
-        combined = torch.cat([sos_idx, final_perms + 1, eos_idx], dim=1).int()  # type: ignore
+        combined = torch.cat([sos_idx, final_perms + 1, eos_idx], dim=1).int()
         if len(combined) > 1:
             combined[1, 1:] = max_num_chars + 1 - torch.arange(max_num_chars + 1, device=seqlen.device)
         return combined
@@ -282,7 +282,8 @@ class PARSeq(_PARSeq, nn.Module):
                 ys[:, i + 1] = pos_prob.squeeze().argmax(-1)
 
                 # Stop decoding if all sequences have reached the EOS token
-                if max_len is None and (ys == self.vocab_size).any(dim=-1).all():  # type: ignore[attr-defined]
+                # NOTE: `break` isn't correctly translated to Onnx so we don't break here if we want to export
+                if not self.exportable and max_len is None and (ys == self.vocab_size).any(dim=-1).all():
                     break
 
         logits = torch.cat(pos_logits, dim=1)  # (N, max_length, vocab_size + 1)
@@ -297,7 +298,7 @@ class PARSeq(_PARSeq, nn.Module):
 
         # Create padding mask for refined target input maskes all behind EOS token as False
         # (N, 1, 1, max_length)
-        target_pad_mask = ~((ys == self.vocab_size).int().cumsum(-1) > 0).unsqueeze(1).unsqueeze(1)  # type: ignore[attr-defined]
+        target_pad_mask = ~((ys == self.vocab_size).int().cumsum(-1) > 0).unsqueeze(1).unsqueeze(1)
         mask = (target_pad_mask.bool() & query_mask[:, : ys.shape[1]].bool()).int()
         logits = self.head(self.decode(ys, features, mask, target_query=pos_queries))
 

doctr/models/recognition/parseq/tensorflow.py

@@ -288,10 +288,11 @@ class PARSeq(_PARSeq, Model):
                 )
 
                 # Stop decoding if all sequences have reached the EOS token
-                # We need to check it on True to be compatible with ONNX
+                # NOTE: `break` isn't correctly translated to Onnx so we don't break here if we want to export
                 if (
-                    max_len is None
-                    and tf.reduce_any(tf.reduce_all(tf.equal(ys, tf.constant(self.vocab_size)), axis=-1)) is True
+                    not self.exportable
+                    and max_len is None
+                    and tf.reduce_any(tf.reduce_all(tf.equal(ys, tf.constant(self.vocab_size)), axis=-1))
                 ):
                     break
 

doctr/models/recognition/sar/pytorch.py

@@ -125,25 +125,26 @@ class SARDecoder(nn.Module):
             if t == 0:
                 # step to init the first states of the LSTMCell
                 hidden_state_init = cell_state_init = torch.zeros(
-                    features.size(0), features.size(1), device=features.device
+                    features.size(0), features.size(1), device=features.device, dtype=features.dtype
                 )
                 hidden_state, cell_state = hidden_state_init, cell_state_init
                 prev_symbol = holistic
             elif t == 1:
                 # step to init a 'blank' sequence of length vocab_size + 1 filled with zeros
                 # (N, vocab_size + 1) --> (N, embedding_units)
-                prev_symbol = torch.zeros(features.size(0), self.vocab_size + 1, device=features.device)
+                prev_symbol = torch.zeros(
+                    features.size(0), self.vocab_size + 1, device=features.device, dtype=features.dtype
+                )
                 prev_symbol = self.embed(prev_symbol)
             else:
-                if gt is not None:
+                if gt is not None and self.training:
                     # (N, embedding_units) -2 because of <bos> and <eos> (same)
                     prev_symbol = self.embed(gt_embedding[:, t - 2])
                 else:
                     # -1 to start at timestep where prev_symbol was initialized
                     index = logits_list[t - 1].argmax(-1)
                     # update prev_symbol with ones at the index of the previous logit vector
-                    # (N, embedding_units)
-                    prev_symbol = prev_symbol.scatter_(1, index.unsqueeze(1), 1)
+                    prev_symbol = self.embed(self.embed_tgt(index))
 
             # (N, C), (N, C)  take the last hidden state and cell state from current timestep
             hidden_state_init, cell_state_init = self.lstm_cell(prev_symbol, (hidden_state_init, cell_state_init))
@@ -292,7 +293,7 @@ class SAR(nn.Module, RecognitionModel):
         # Input length : number of timesteps
         input_len = model_output.shape[1]
         # Add one for additional <eos> token
-        seq_len = seq_len + 1  # type: ignore[assignment]
+        seq_len = seq_len + 1
         # Compute loss
         # (N, L, vocab_size + 1)
         cce = F.cross_entropy(model_output.permute(0, 2, 1), gt, reduction="none")

doctr/models/recognition/sar/tensorflow.py

@@ -177,23 +177,17 @@ class SARDecoder(layers.Layer, NestedObject):
             elif t == 1:
                 # step to init a 'blank' sequence of length vocab_size + 1 filled with zeros
                 # (N, vocab_size + 1) --> (N, embedding_units)
-                prev_symbol = tf.zeros([features.shape[0], self.vocab_size + 1])
+                prev_symbol = tf.zeros([features.shape[0], self.vocab_size + 1], dtype=features.dtype)
                 prev_symbol = self.embed(prev_symbol, **kwargs)
             else:
-                if gt is not None:
+                if gt is not None and kwargs.get("training", False):
                     # (N, embedding_units) -2 because of <bos> and <eos> (same)
                     prev_symbol = self.embed(gt_embedding[:, t - 2], **kwargs)
                 else:
                     # -1 to start at timestep where prev_symbol was initialized
                     index = tf.argmax(logits_list[t - 1], axis=-1)
                     # update prev_symbol with ones at the index of the previous logit vector
-                    # (N, embedding_units)
-                    index = tf.ones_like(index)
-                    prev_symbol = tf.scatter_nd(
-                        tf.expand_dims(index, axis=1),
-                        prev_symbol,
-                        tf.constant([features.shape[0], features.shape[-1]], dtype=tf.int64),
-                    )
+                    prev_symbol = self.embed(self.embed_tgt(index, **kwargs), **kwargs)
 
             # (N, C), (N, C)  take the last hidden state and cell state from current timestep
             _, states = self.lstm_cells(prev_symbol, states, **kwargs)

doctr/models/recognition/vitstr/pytorch.py

@@ -137,7 +137,7 @@ class ViTSTR(_ViTSTR, nn.Module):
         # Input length : number of steps
         input_len = model_output.shape[1]
         # Add one for additional <eos> token (sos disappear in shift!)
-        seq_len = seq_len + 1  # type: ignore[assignment]
+        seq_len = seq_len + 1
         # Compute loss: don't forget to shift gt! Otherwise the model learns to output the gt[t-1]!
         # The "masked" first gt char is <sos>.
         cce = F.cross_entropy(model_output.permute(0, 2, 1), gt[:, 1:], reduction="none")

doctr/models/recognition/zoo.py

@@ -45,7 +45,7 @@ def _predictor(arch: Any, pretrained: bool, **kwargs: Any) -> RecognitionPredict
 
     kwargs["mean"] = kwargs.get("mean", _model.cfg["mean"])
     kwargs["std"] = kwargs.get("std", _model.cfg["std"])
-    kwargs["batch_size"] = kwargs.get("batch_size", 32)
+    kwargs["batch_size"] = kwargs.get("batch_size", 128)
     input_shape = _model.cfg["input_shape"][:2] if is_tf_available() else _model.cfg["input_shape"][-2:]
     predictor = RecognitionPredictor(PreProcessor(input_shape, preserve_aspect_ratio=True, **kwargs), _model)
 

doctr/models/zoo.py

@@ -61,7 +61,7 @@ def _predictor(
 
 
 def ocr_predictor(
-    det_arch: Any = "db_resnet50",
+    det_arch: Any = "fast_base",
     reco_arch: Any = "crnn_vgg16_bn",
     pretrained: bool = False,
     pretrained_backbone: bool = True,
@@ -175,7 +175,7 @@ def _kie_predictor(
 
 
 def kie_predictor(
-    det_arch: Any = "db_resnet50",
+    det_arch: Any = "fast_base",
     reco_arch: Any = "crnn_vgg16_bn",
     pretrained: bool = False,
     pretrained_backbone: bool = True,

doctr/transforms/functional/base.py

@@ -200,4 +200,4 @@ def create_shadow_mask(
     mask: np.ndarray = np.zeros((*target_shape, 1), dtype=np.uint8)
     mask = cv2.fillPoly(mask, [final_contour], (255,), lineType=cv2.LINE_AA)[..., 0]
 
-    return (mask / 255).astype(np.float32).clip(0, 1) * intensity_mask.astype(np.float32)  # type: ignore[operator]
+    return (mask / 255).astype(np.float32).clip(0, 1) * intensity_mask.astype(np.float32)

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