python-doctr 0.9.0__py3-none-any.whl → 0.11.0__py3-none-any.whl

doctr/models/recognition/predictor/_utils.py

@@ -1,9 +1,8 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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 List, Tuple, Union
 
 import numpy as np
 
@@ -13,16 +12,15 @@ __all__ = ["split_crops", "remap_preds"]
 
 
 def split_crops(
-    crops: List[np.ndarray],
+    crops: list[np.ndarray],
     max_ratio: float,
     target_ratio: int,
     dilation: float,
     channels_last: bool = True,
-) -> Tuple[List[np.ndarray], List[Union[int, Tuple[int, int]]], bool]:
+) -> tuple[list[np.ndarray], list[int | tuple[int, int]], bool]:
     """Chunk crops horizontally to match a given aspect ratio
 
     Args:
-    ----
         crops: list of numpy array of shape (H, W, 3) if channels_last or (3, H, W) otherwise
         max_ratio: the maximum aspect ratio that won't trigger the chunk
         target_ratio: when crops are chunked, they will be chunked to match this aspect ratio
@@ -30,12 +28,11 @@ def split_crops(
         channels_last: whether the numpy array has dimensions in channels last order
 
     Returns:
-    -------
         a tuple with the new crops, their mapping, and a boolean specifying whether any remap is required
     """
     _remap_required = False
-    crop_map: List[Union[int, Tuple[int, int]]] = []
-    new_crops: List[np.ndarray] = []
+    crop_map: list[int | tuple[int, int]] = []
+    new_crops: list[np.ndarray] = []
     for crop in crops:
         h, w = crop.shape[:2] if channels_last else crop.shape[-2:]
         aspect_ratio = w / h
@@ -71,8 +68,8 @@ def split_crops(
 
 
 def remap_preds(
-    preds: List[Tuple[str, float]], crop_map: List[Union[int, Tuple[int, int]]], dilation: float
-) -> List[Tuple[str, float]]:
+    preds: list[tuple[str, float]], crop_map: list[int | tuple[int, int]], dilation: float
+) -> list[tuple[str, float]]:
     remapped_out = []
     for _idx in crop_map:
         # Crop hasn't been split

doctr/models/recognition/predictor/pytorch.py

@@ -1,9 +1,10 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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, Sequence, Tuple, Union
+from collections.abc import Sequence
+from typing import Any
 
 import numpy as np
 import torch
@@ -21,7 +22,6 @@ class RecognitionPredictor(nn.Module):
     """Implements an object able to identify character sequences in images
 
     Args:
-    ----
         pre_processor: transform inputs for easier batched model inference
         model: core detection architecture
         split_wide_crops: wether to use crop splitting for high aspect ratio crops
@@ -44,9 +44,9 @@ class RecognitionPredictor(nn.Module):
     @torch.inference_mode()
     def forward(
         self,
-        crops: Sequence[Union[np.ndarray, torch.Tensor]],
+        crops: Sequence[np.ndarray | torch.Tensor],
         **kwargs: Any,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         if len(crops) == 0:
             return []
         # Dimension check
@@ -67,7 +67,7 @@ class RecognitionPredictor(nn.Module):
                 crops = new_crops
 
         # Resize & batch them
-        processed_batches = self.pre_processor(crops)
+        processed_batches = self.pre_processor(crops)  # type: ignore[arg-type]
 
         # Forward it
         _params = next(self.model.parameters())

doctr/models/recognition/predictor/tensorflow.py

@@ -1,9 +1,9 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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, Tuple, Union
+from typing import Any
 
 import numpy as np
 import tensorflow as tf
@@ -21,13 +21,12 @@ class RecognitionPredictor(NestedObject):
     """Implements an object able to identify character sequences in images
 
     Args:
-    ----
         pre_processor: transform inputs for easier batched model inference
         model: core detection architecture
         split_wide_crops: wether to use crop splitting for high aspect ratio crops
     """
 
-    _children_names: List[str] = ["pre_processor", "model"]
+    _children_names: list[str] = ["pre_processor", "model"]
 
     def __init__(
         self,
@@ -45,9 +44,9 @@ class RecognitionPredictor(NestedObject):
 
     def __call__(
         self,
-        crops: List[Union[np.ndarray, tf.Tensor]],
+        crops: list[np.ndarray | tf.Tensor],
         **kwargs: Any,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         if len(crops) == 0:
             return []
         # Dimension check

doctr/models/recognition/sar/__init__.py

@@ -1,6 +1,6 @@
 from doctr.file_utils import is_tf_available, is_torch_available
 
-if is_tf_available():
-    from .tensorflow import *
-elif is_torch_available():
-    from .pytorch import *  # type: ignore[assignment]
+if is_torch_available():
+    from .pytorch import *
+elif is_tf_available():
+    from .tensorflow import *  # type: ignore[assignment]

doctr/models/recognition/sar/pytorch.py

@@ -1,10 +1,11 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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 collections.abc import Callable
 from copy import deepcopy
-from typing import Any, Callable, Dict, List, Optional, Tuple
+from typing import Any
 
 import torch
 from torch import nn
@@ -19,7 +20,7 @@ from ..core import RecognitionModel, RecognitionPostProcessor
 
 __all__ = ["SAR", "sar_resnet31"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "sar_resnet31": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
@@ -80,7 +81,6 @@ class SARDecoder(nn.Module):
     """Implements decoder module of the SAR model
 
     Args:
-    ----
         rnn_units: number of hidden units in recurrent cells
         max_length: maximum length of a sequence
         vocab_size: number of classes in the model alphabet
@@ -114,12 +114,12 @@ class SARDecoder(nn.Module):
         self,
         features: torch.Tensor,  # (N, C, H, W)
         holistic: torch.Tensor,  # (N, C)
-        gt: Optional[torch.Tensor] = None,  # (N, L)
+        gt: torch.Tensor | None = None,  # (N, L)
     ) -> torch.Tensor:
         if gt is not None:
             gt_embedding = self.embed_tgt(gt)
 
-        logits_list: List[torch.Tensor] = []
+        logits_list: list[torch.Tensor] = []
 
         for t in range(self.max_length + 1):  # 32
             if t == 0:
@@ -166,7 +166,6 @@ class SAR(nn.Module, RecognitionModel):
     Irregular Text Recognition" <https://arxiv.org/pdf/1811.00751.pdf>`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         rnn_units: number of hidden units in both encoder and decoder LSTM
@@ -187,9 +186,9 @@ class SAR(nn.Module, RecognitionModel):
         attention_units: int = 512,
         max_length: int = 30,
         dropout_prob: float = 0.0,
-        input_shape: Tuple[int, int, int] = (3, 32, 128),
+        input_shape: tuple[int, int, int] = (3, 32, 128),
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -232,10 +231,10 @@ class SAR(nn.Module, RecognitionModel):
     def forward(
         self,
         x: torch.Tensor,
-        target: Optional[List[str]] = None,
+        target: list[str] | None = None,
         return_model_output: bool = False,
         return_preds: bool = False,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         features = self.feat_extractor(x)["features"]
         # NOTE: use max instead of functional max_pool2d which leads to ONNX incompatibility (kernel_size)
         # Vertical max pooling (N, C, H, W) --> (N, C, W)
@@ -254,7 +253,7 @@ class SAR(nn.Module, RecognitionModel):
 
         decoded_features = _bf16_to_float32(self.decoder(features, encoded, gt=None if target is None else gt))
 
-        out: Dict[str, Any] = {}
+        out: dict[str, Any] = {}
         if self.exportable:
             out["logits"] = decoded_features
             return out
@@ -263,8 +262,13 @@ class SAR(nn.Module, RecognitionModel):
             out["out_map"] = decoded_features
 
         if target is None or return_preds:
+            # Disable for torch.compile compatibility
+            @torch.compiler.disable  # type: ignore[attr-defined]
+            def _postprocess(decoded_features: torch.Tensor) -> list[tuple[str, float]]:
+                return self.postprocessor(decoded_features)
+
             # Post-process boxes
-            out["preds"] = self.postprocessor(decoded_features)
+            out["preds"] = _postprocess(decoded_features)
 
         if target is not None:
             out["loss"] = self.compute_loss(decoded_features, gt, seq_len)
@@ -281,19 +285,17 @@ class SAR(nn.Module, RecognitionModel):
         Sequences are masked after the EOS character.
 
         Args:
-        ----
             model_output: predicted logits of the model
             gt: the encoded tensor with gt labels
             seq_len: lengths of each gt word inside the batch
 
         Returns:
-        -------
             The loss of the model on the batch
         """
         # Input length : number of timesteps
         input_len = model_output.shape[1]
         # Add one for additional <eos> token
-        seq_len = seq_len + 1
+        seq_len = seq_len + 1  # type: ignore[assignment]
         # Compute loss
         # (N, L, vocab_size + 1)
         cce = F.cross_entropy(model_output.permute(0, 2, 1), gt, reduction="none")
@@ -308,14 +310,13 @@ class SARPostProcessor(RecognitionPostProcessor):
     """Post processor for SAR architectures
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 
     def __call__(
         self,
         logits: torch.Tensor,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         # compute pred with argmax for attention models
         out_idxs = logits.argmax(-1)
         # N x L
@@ -338,7 +339,7 @@ def _sar(
     backbone_fn: Callable[[bool], nn.Module],
     layer: str,
     pretrained_backbone: bool = True,
-    ignore_keys: Optional[List[str]] = None,
+    ignore_keys: list[str] | None = None,
     **kwargs: Any,
 ) -> SAR:
     pretrained_backbone = pretrained_backbone and not pretrained
@@ -379,12 +380,10 @@ def sar_resnet31(pretrained: bool = False, **kwargs: Any) -> SAR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the SAR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _sar(

doctr/models/recognition/sar/tensorflow.py

@@ -1,10 +1,10 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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 copy import deepcopy
-from typing import Any, Dict, List, Optional, Tuple
+from typing import Any
 
 import tensorflow as tf
 from tensorflow.keras import Model, Sequential, layers
@@ -13,18 +13,18 @@ from doctr.datasets import VOCABS
 from doctr.utils.repr import NestedObject
 
 from ...classification import resnet31
-from ...utils.tensorflow import _bf16_to_float32, load_pretrained_params
+from ...utils.tensorflow import _bf16_to_float32, _build_model, load_pretrained_params
 from ..core import RecognitionModel, RecognitionPostProcessor
 
 __all__ = ["SAR", "sar_resnet31"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "sar_resnet31": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
         "input_shape": (32, 128, 3),
         "vocab": VOCABS["french"],
-        "url": "https://doctr-static.mindee.com/models?id=v0.6.0/sar_resnet31-c41e32a5.zip&src=0",
+        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/sar_resnet31-5a58806c.weights.h5&src=0",
     },
 }
 
@@ -33,7 +33,6 @@ class SAREncoder(layers.Layer, NestedObject):
     """Implements encoder module of the SAR model
 
     Args:
-    ----
         rnn_units: number of hidden rnn units
         dropout_prob: dropout probability
     """
@@ -58,7 +57,6 @@ class AttentionModule(layers.Layer, NestedObject):
     """Implements attention module of the SAR model
 
     Args:
-    ----
         attention_units: number of hidden attention units
 
     """
@@ -120,7 +118,6 @@ class SARDecoder(layers.Layer, NestedObject):
     """Implements decoder module of the SAR model
 
     Args:
-    ----
         rnn_units: number of hidden units in recurrent cells
         max_length: maximum length of a sequence
         vocab_size: number of classes in the model alphabet
@@ -159,13 +156,13 @@ class SARDecoder(layers.Layer, NestedObject):
         self,
         features: tf.Tensor,
         holistic: tf.Tensor,
-        gt: Optional[tf.Tensor] = None,
+        gt: tf.Tensor | None = None,
         **kwargs: Any,
     ) -> tf.Tensor:
         if gt is not None:
             gt_embedding = self.embed_tgt(gt, **kwargs)
 
-        logits_list: List[tf.Tensor] = []
+        logits_list: list[tf.Tensor] = []
 
         for t in range(self.max_length + 1):  # 32
             if t == 0:
@@ -210,7 +207,6 @@ class SAR(Model, RecognitionModel):
     Irregular Text Recognition" <https://arxiv.org/pdf/1811.00751.pdf>`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         rnn_units: number of hidden units in both encoder and decoder LSTM
@@ -223,7 +219,7 @@ class SAR(Model, RecognitionModel):
         cfg: dictionary containing information about the model
     """
 
-    _children_names: List[str] = ["feat_extractor", "encoder", "decoder", "postprocessor"]
+    _children_names: list[str] = ["feat_extractor", "encoder", "decoder", "postprocessor"]
 
     def __init__(
         self,
@@ -236,7 +232,7 @@ class SAR(Model, RecognitionModel):
         num_decoder_cells: int = 2,
         dropout_prob: float = 0.0,
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -269,13 +265,11 @@ class SAR(Model, RecognitionModel):
         Sequences are masked after the EOS character.
 
         Args:
-        ----
             gt: the encoded tensor with gt labels
             model_output: predicted logits of the model
             seq_len: lengths of each gt word inside the batch
 
         Returns:
-        -------
             The loss of the model on the batch
         """
         # Input length : number of timesteps
@@ -296,11 +290,11 @@ class SAR(Model, RecognitionModel):
     def call(
         self,
         x: tf.Tensor,
-        target: Optional[List[str]] = None,
+        target: list[str] | None = None,
         return_model_output: bool = False,
         return_preds: bool = False,
         **kwargs: Any,
-    ) -> Dict[str, Any]:
+    ) -> dict[str, Any]:
         features = self.feat_extractor(x, **kwargs)
         # vertical max pooling --> (N, C, W)
         pooled_features = tf.reduce_max(features, axis=1)
@@ -318,7 +312,7 @@ class SAR(Model, RecognitionModel):
             self.decoder(features, encoded, gt=None if target is None else gt, **kwargs)
         )
 
-        out: Dict[str, tf.Tensor] = {}
+        out: dict[str, tf.Tensor] = {}
         if self.exportable:
             out["logits"] = decoded_features
             return out
@@ -340,14 +334,13 @@ class SARPostProcessor(RecognitionPostProcessor):
     """Post processor for SAR architectures
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """
 
     def __call__(
         self,
         logits: tf.Tensor,
-    ) -> List[Tuple[str, float]]:
+    ) -> list[tuple[str, float]]:
         # compute pred with argmax for attention models
         out_idxs = tf.math.argmax(logits, axis=2)
         # N x L
@@ -371,7 +364,7 @@ def _sar(
     pretrained: bool,
     backbone_fn,
     pretrained_backbone: bool = True,
-    input_shape: Optional[Tuple[int, int, int]] = None,
+    input_shape: tuple[int, int, int] | None = None,
     **kwargs: Any,
 ) -> SAR:
     pretrained_backbone = pretrained_backbone and not pretrained
@@ -392,9 +385,13 @@ def _sar(
 
     # Build the model
     model = SAR(feat_extractor, cfg=_cfg, **kwargs)
+    _build_model(model)
     # Load pretrained parameters
     if pretrained:
-        load_pretrained_params(model, default_cfgs[arch]["url"])
+        # The given vocab differs from the pretrained model => skip the mismatching layers for fine tuning
+        load_pretrained_params(
+            model, default_cfgs[arch]["url"], skip_mismatch=kwargs["vocab"] != default_cfgs[arch]["vocab"]
+        )
 
     return model
 
@@ -410,12 +407,10 @@ def sar_resnet31(pretrained: bool = False, **kwargs: Any) -> SAR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the SAR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _sar("sar_resnet31", pretrained, resnet31, **kwargs)

doctr/models/recognition/utils.py

@@ -1,9 +1,8 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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 List
 
 from rapidfuzz.distance import Levenshtein
 
@@ -14,18 +13,16 @@ def merge_strings(a: str, b: str, dil_factor: float) -> str:
     """Merges 2 character sequences in the best way to maximize the alignment of their overlapping characters.
 
     Args:
-    ----
         a: first char seq, suffix should be similar to b's prefix.
         b: second char seq, prefix should be similar to a's suffix.
         dil_factor: dilation factor of the boxes to overlap, should be > 1. This parameter is
             only used when the mother sequence is splitted on a character repetition
 
     Returns:
-    -------
         A merged character sequence.
 
     Example::
-        >>> from doctr.model.recognition.utils import merge_sequences
+        >>> from doctr.models.recognition.utils import merge_sequences
         >>> merge_sequences('abcd', 'cdefgh', 1.4)
         'abcdefgh'
         >>> merge_sequences('abcdi', 'cdefgh', 1.4)
@@ -61,26 +58,24 @@ def merge_strings(a: str, b: str, dil_factor: float) -> str:
     return a[:-1] + b[index - 1 :]
 
 
-def merge_multi_strings(seq_list: List[str], dil_factor: float) -> str:
+def merge_multi_strings(seq_list: list[str], dil_factor: float) -> str:
     """Recursively merges consecutive string sequences with overlapping characters.
 
     Args:
-    ----
         seq_list: list of sequences to merge. Sequences need to be ordered from left to right.
         dil_factor: dilation factor of the boxes to overlap, should be > 1. This parameter is
             only used when the mother sequence is splitted on a character repetition
 
     Returns:
-    -------
         A merged character sequence
 
     Example::
-        >>> from doctr.model.recognition.utils import merge_multi_sequences
+        >>> from doctr.models.recognition.utils import merge_multi_sequences
         >>> merge_multi_sequences(['abc', 'bcdef', 'difghi', 'aijkl'], 1.4)
         'abcdefghijkl'
     """
 
-    def _recursive_merge(a: str, seq_list: List[str], dil_factor: float) -> str:
+    def _recursive_merge(a: str, seq_list: list[str], dil_factor: float) -> str:
         # Recursive version of compute_overlap
         if len(seq_list) == 1:
             return merge_strings(a, seq_list[0], dil_factor)

doctr/models/recognition/vitstr/__init__.py

@@ -1,6 +1,6 @@
 from doctr.file_utils import is_tf_available, is_torch_available
 
-if is_tf_available():
-    from .tensorflow import *
-elif is_torch_available():
-    from .pytorch import *  # type: ignore[assignment]
+if is_torch_available():
+    from .pytorch import *
+elif is_tf_available():
+    from .tensorflow import *  # type: ignore[assignment]

doctr/models/recognition/vitstr/base.py

@@ -1,9 +1,8 @@
-# Copyright (C) 2021-2024, Mindee.
+# Copyright (C) 2021-2025, 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 List, Tuple
 
 import numpy as np
 
@@ -17,17 +16,15 @@ class _ViTSTR:
 
     def build_target(
         self,
-        gts: List[str],
-    ) -> Tuple[np.ndarray, List[int]]:
+        gts: list[str],
+    ) -> tuple[np.ndarray, list[int]]:
         """Encode a list of gts sequences into a np array and gives the corresponding*
         sequence lengths.
 
         Args:
-        ----
             gts: list of ground-truth labels
 
         Returns:
-        -------
             A tuple of 2 tensors: Encoded labels and sequence lengths (for each entry of the batch)
         """
         encoded = encode_sequences(
@@ -45,7 +42,6 @@ class _ViTSTRPostProcessor(RecognitionPostProcessor):
     """Abstract class to postprocess the raw output of the model
 
     Args:
-    ----
         vocab: string containing the ordered sequence of supported characters
     """