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

doctr/models/recognition/vitstr/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 .base import _ViTSTR, _ViTSTRPostProcessor
 
 __all__ = ["ViTSTR", "vitstr_small", "vitstr_base"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "vitstr_small": {
         "mean": (0.694, 0.695, 0.693),
         "std": (0.299, 0.296, 0.301),
@@ -42,7 +43,6 @@ class ViTSTR(_ViTSTR, nn.Module):
     Efficient Scene Text Recognition" <https://arxiv.org/pdf/2105.08582.pdf>`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         embedding_units: number of embedding units
@@ -59,9 +59,9 @@ class ViTSTR(_ViTSTR, nn.Module):
         vocab: str,
         embedding_units: int,
         max_length: int = 32,  # different from paper
-        input_shape: Tuple[int, int, int] = (3, 32, 128),  # different from paper
+        input_shape: tuple[int, int, int] = (3, 32, 128),  # different from paper
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -77,10 +77,10 @@ class ViTSTR(_ViTSTR, nn.Module):
     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"]  # (batch_size, patches_seqlen, d_model)
 
         if target is not None:
@@ -98,7 +98,7 @@ class ViTSTR(_ViTSTR, nn.Module):
         logits = self.head(features).view(B, N, len(self.vocab) + 1)  # (batch_size, max_length, vocab + 1)
         decoded_features = _bf16_to_float32(logits[:, 1:])  # remove cls_token
 
-        out: Dict[str, Any] = {}
+        out: dict[str, Any] = {}
         if self.exportable:
             out["logits"] = decoded_features
             return out
@@ -107,8 +107,13 @@ class ViTSTR(_ViTSTR, nn.Module):
             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)
@@ -125,19 +130,17 @@ class ViTSTR(_ViTSTR, nn.Module):
         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 steps
         input_len = model_output.shape[1]
         # Add one for additional <eos> token (sos disappear in shift!)
-        seq_len = seq_len + 1
+        seq_len = seq_len + 1  # type: ignore[assignment]
         # 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")
@@ -153,14 +156,13 @@ class ViTSTRPostProcessor(_ViTSTRPostProcessor):
     """Post processor for ViTSTR architecture
 
     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)
         preds_prob = torch.softmax(logits, -1).max(dim=-1)[0]
@@ -183,7 +185,7 @@ def _vitstr(
     pretrained: bool,
     backbone_fn: Callable[[bool], nn.Module],
     layer: str,
-    ignore_keys: Optional[List[str]] = None,
+    ignore_keys: list[str] | None = None,
     **kwargs: Any,
 ) -> ViTSTR:
     # Patch the config
@@ -228,12 +230,10 @@ def vitstr_small(pretrained: bool = False, **kwargs: Any) -> ViTSTR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _vitstr(
@@ -259,12 +259,10 @@ def vitstr_base(pretrained: bool = False, **kwargs: Any) -> ViTSTR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _vitstr(

doctr/models/recognition/vitstr/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, layers
@@ -12,25 +12,25 @@ from tensorflow.keras import Model, layers
 from doctr.datasets import VOCABS
 
 from ...classification import vit_b, vit_s
-from ...utils.tensorflow import _bf16_to_float32, load_pretrained_params
+from ...utils.tensorflow import _bf16_to_float32, _build_model, load_pretrained_params
 from .base import _ViTSTR, _ViTSTRPostProcessor
 
 __all__ = ["ViTSTR", "vitstr_small", "vitstr_base"]
 
-default_cfgs: Dict[str, Dict[str, Any]] = {
+default_cfgs: dict[str, dict[str, Any]] = {
     "vitstr_small": {
         "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/vitstr_small-358fab2e.zip&src=0",
+        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/vitstr_small-d28b8d92.weights.h5&src=0",
     },
     "vitstr_base": {
         "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/vitstr_base-2889159a.zip&src=0",
+        "url": "https://doctr-static.mindee.com/models?id=v0.9.0/vitstr_base-9ad6eb84.weights.h5&src=0",
     },
 }
 
@@ -40,7 +40,6 @@ class ViTSTR(_ViTSTR, Model):
     Efficient Scene Text Recognition" <https://arxiv.org/pdf/2105.08582.pdf>`_.
 
     Args:
-    ----
         feature_extractor: the backbone serving as feature extractor
         vocab: vocabulary used for encoding
         embedding_units: number of embedding units
@@ -51,7 +50,7 @@ class ViTSTR(_ViTSTR, Model):
         cfg: dictionary containing information about the model
     """
 
-    _children_names: List[str] = ["feat_extractor", "postprocessor"]
+    _children_names: list[str] = ["feat_extractor", "postprocessor"]
 
     def __init__(
         self,
@@ -60,9 +59,9 @@ class ViTSTR(_ViTSTR, Model):
         embedding_units: int,
         max_length: int = 32,
         dropout_prob: float = 0.0,
-        input_shape: Tuple[int, int, int] = (32, 128, 3),  # different from paper
+        input_shape: tuple[int, int, int] = (32, 128, 3),  # different from paper
         exportable: bool = False,
-        cfg: Optional[Dict[str, Any]] = None,
+        cfg: dict[str, Any] | None = None,
     ) -> None:
         super().__init__()
         self.vocab = vocab
@@ -79,19 +78,17 @@ class ViTSTR(_ViTSTR, Model):
     def compute_loss(
         model_output: tf.Tensor,
         gt: tf.Tensor,
-        seq_len: List[int],
+        seq_len: list[int],
     ) -> tf.Tensor:
         """Compute categorical cross-entropy loss for the model.
         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 steps
@@ -114,11 +111,11 @@ class ViTSTR(_ViTSTR, Model):
     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)  # (batch_size, patches_seqlen, d_model)
 
         if target is not None:
@@ -136,7 +133,7 @@ class ViTSTR(_ViTSTR, Model):
         )  # (batch_size, max_length, vocab + 1)
         decoded_features = _bf16_to_float32(logits[:, 1:])  # remove cls_token
 
-        out: Dict[str, tf.Tensor] = {}
+        out: dict[str, tf.Tensor] = {}
         if self.exportable:
             out["logits"] = decoded_features
             return out
@@ -158,14 +155,13 @@ class ViTSTRPostProcessor(_ViTSTRPostProcessor):
     """Post processor for ViTSTR architecture
 
     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)
         preds_prob = tf.math.reduce_max(tf.nn.softmax(logits, axis=-1), axis=-1)
@@ -191,7 +187,7 @@ def _vitstr(
     arch: str,
     pretrained: bool,
     backbone_fn,
-    input_shape: Optional[Tuple[int, int, int]] = None,
+    input_shape: tuple[int, int, int] | None = None,
     **kwargs: Any,
 ) -> ViTSTR:
     # Patch the config
@@ -216,9 +212,14 @@ def _vitstr(
 
     # Build the model
     model = ViTSTR(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
 
@@ -234,12 +235,10 @@ def vitstr_small(pretrained: bool = False, **kwargs: Any) -> ViTSTR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _vitstr(
@@ -263,12 +262,10 @@ def vitstr_base(pretrained: bool = False, **kwargs: Any) -> ViTSTR:
     >>> out = model(input_tensor)
 
     Args:
-    ----
         pretrained (bool): If True, returns a model pre-trained on our text recognition dataset
         **kwargs: keyword arguments of the ViTSTR architecture
 
     Returns:
-    -------
         text recognition architecture
     """
     return _vitstr(

doctr/models/recognition/zoo.py

@@ -1,11 +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 typing import Any, List
+from typing import Any
 
-from doctr.file_utils import is_tf_available
+from doctr.file_utils import is_tf_available, is_torch_available
 from doctr.models.preprocessor import PreProcessor
 
 from .. import recognition
@@ -14,7 +14,7 @@ from .predictor import RecognitionPredictor
 __all__ = ["recognition_predictor"]
 
 
-ARCHS: List[str] = [
+ARCHS: list[str] = [
     "crnn_vgg16_bn",
     "crnn_mobilenet_v3_small",
     "crnn_mobilenet_v3_large",
@@ -35,9 +35,14 @@ def _predictor(arch: Any, pretrained: bool, **kwargs: Any) -> RecognitionPredict
             pretrained=pretrained, pretrained_backbone=kwargs.get("pretrained_backbone", True)
         )
     else:
-        if not isinstance(
-            arch, (recognition.CRNN, recognition.SAR, recognition.MASTER, recognition.ViTSTR, recognition.PARSeq)
-        ):
+        allowed_archs = [recognition.CRNN, recognition.SAR, recognition.MASTER, recognition.ViTSTR, recognition.PARSeq]
+        if is_torch_available():
+            # Adding the type for torch compiled models to the allowed architectures
+            from doctr.models.utils import _CompiledModule
+
+            allowed_archs.append(_CompiledModule)
+
+        if not isinstance(arch, tuple(allowed_archs)):
             raise ValueError(f"unknown architecture: {type(arch)}")
         _model = arch
 
@@ -52,7 +57,13 @@ def _predictor(arch: Any, pretrained: bool, **kwargs: Any) -> RecognitionPredict
     return predictor
 
 
-def recognition_predictor(arch: Any = "crnn_vgg16_bn", pretrained: bool = False, **kwargs: Any) -> RecognitionPredictor:
+def recognition_predictor(
+    arch: Any = "crnn_vgg16_bn",
+    pretrained: bool = False,
+    symmetric_pad: bool = False,
+    batch_size: int = 128,
+    **kwargs: Any,
+) -> RecognitionPredictor:
     """Text recognition architecture.
 
     Example::
@@ -63,13 +74,13 @@ def recognition_predictor(arch: Any = "crnn_vgg16_bn", pretrained: bool = False,
         >>> out = model([input_page])
 
     Args:
-    ----
         arch: name of the architecture or model itself to use (e.g. 'crnn_vgg16_bn')
         pretrained: If True, returns a model pre-trained on our text recognition dataset
+        symmetric_pad: if True, pad the image symmetrically instead of padding at the bottom-right
+        batch_size: number of samples the model processes in parallel
         **kwargs: optional parameters to be passed to the architecture
 
     Returns:
-    -------
         Recognition predictor
     """
-    return _predictor(arch, pretrained, **kwargs)
+    return _predictor(arch=arch, pretrained=pretrained, symmetric_pad=symmetric_pad, batch_size=batch_size, **kwargs)

doctr/models/utils/__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/utils/pytorch.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.
 
 import logging
-from typing import Any, List, Optional, Tuple, Union
+from typing import Any
 
 import torch
 from torch import nn
@@ -18,8 +18,12 @@ __all__ = [
     "export_model_to_onnx",
     "_copy_tensor",
     "_bf16_to_float32",
+    "_CompiledModule",
 ]
 
+# torch compiled model type
+_CompiledModule = torch._dynamo.eval_frame.OptimizedModule
+
 
 def _copy_tensor(x: torch.Tensor) -> torch.Tensor:
     return x.clone().detach()
@@ -32,9 +36,9 @@ def _bf16_to_float32(x: torch.Tensor) -> torch.Tensor:
 
 def load_pretrained_params(
     model: nn.Module,
-    url: Optional[str] = None,
-    hash_prefix: Optional[str] = None,
-    ignore_keys: Optional[List[str]] = None,
+    url: str | None = None,
+    hash_prefix: str | None = None,
+    ignore_keys: list[str] | None = None,
     **kwargs: Any,
 ) -> None:
     """Load a set of parameters onto a model
@@ -43,7 +47,6 @@ def load_pretrained_params(
     >>> load_pretrained_params(model, "https://yoursource.com/yourcheckpoint-yourhash.zip")
 
     Args:
-    ----
         model: the PyTorch model to be loaded
         url: URL of the zipped set of parameters
         hash_prefix: first characters of SHA256 expected hash
@@ -76,7 +79,7 @@ def conv_sequence_pt(
     relu: bool = False,
     bn: bool = False,
     **kwargs: Any,
-) -> List[nn.Module]:
+) -> list[nn.Module]:
     """Builds a convolutional-based layer sequence
 
     >>> from torch.nn import Sequential
@@ -84,7 +87,6 @@ def conv_sequence_pt(
     >>> module = Sequential(conv_sequence(3, 32, True, True, kernel_size=3))
 
     Args:
-    ----
         in_channels: number of input channels
         out_channels: number of output channels
         relu: whether ReLU should be used
@@ -92,13 +94,12 @@ def conv_sequence_pt(
         **kwargs: additional arguments to be passed to the convolutional layer
 
     Returns:
-    -------
         list of layers
     """
     # No bias before Batch norm
     kwargs["bias"] = kwargs.get("bias", not bn)
     # Add activation directly to the conv if there is no BN
-    conv_seq: List[nn.Module] = [nn.Conv2d(in_channels, out_channels, **kwargs)]
+    conv_seq: list[nn.Module] = [nn.Conv2d(in_channels, out_channels, **kwargs)]
 
     if bn:
         conv_seq.append(nn.BatchNorm2d(out_channels))
@@ -110,8 +111,8 @@ def conv_sequence_pt(
 
 
 def set_device_and_dtype(
-    model: Any, batches: List[torch.Tensor], device: Union[str, torch.device], dtype: torch.dtype
-) -> Tuple[Any, List[torch.Tensor]]:
+    model: Any, batches: list[torch.Tensor], device: str | torch.device, dtype: torch.dtype
+) -> tuple[Any, list[torch.Tensor]]:
     """Set the device and dtype of a model and its batches
 
     >>> import torch
@@ -122,14 +123,12 @@ def set_device_and_dtype(
     >>> model, batches = set_device_and_dtype(model, batches, device="cuda", dtype=torch.float16)
 
     Args:
-    ----
         model: the model to be set
         batches: the batches to be set
         device: the device to be used
         dtype: the dtype to be used
 
     Returns:
-    -------
         the model and batches set
     """
     return model.to(device=device, dtype=dtype), [batch.to(device=device, dtype=dtype) for batch in batches]
@@ -145,14 +144,12 @@ def export_model_to_onnx(model: nn.Module, model_name: str, dummy_input: torch.T
     >>> export_model_to_onnx(model, "my_model", dummy_input=torch.randn(1, 3, 32, 32))
 
     Args:
-    ----
         model: the PyTorch model to be exported
         model_name: the name for the exported model
         dummy_input: the dummy input to the model
         kwargs: additional arguments to be passed to torch.onnx.export
 
     Returns:
-    -------
         the path to the exported model
     """
     torch.onnx.export(