deepdoctection 0.44.1__tar.gz → 0.46__tar.gz

{deepdoctection-0.44.1 → deepdoctection-0.46}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: deepdoctection
-Version: 0.44.1
+Version: 0.46
 Summary: Repository for Document AI
 Home-page: https://github.com/deepdoctection/deepdoctection
 Author: Dr. Janis Meyer
@@ -19,18 +19,15 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: catalogue==2.0.10
 Requires-Dist: huggingface_hub>=0.26.0
-Requires-Dist: importlib-metadata>=5.0.0
 Requires-Dist: jsonlines==3.1.0
 Requires-Dist: lazy-imports==0.3.1
 Requires-Dist: mock==4.0.3
-Requires-Dist: networkx>=2.7.1
-Requires-Dist: numpy<2.0,>=1.21
+Requires-Dist: numpy>2.0
 Requires-Dist: packaging>=20.0
 Requires-Dist: Pillow>=10.0.0
 Requires-Dist: pypdf>=6.0.0
 Requires-Dist: pypdfium2>=4.30.0
 Requires-Dist: pyyaml>=6.0.1
-Requires-Dist: pyzmq>=16
 Requires-Dist: scipy>=1.13.1
 Requires-Dist: termcolor>=1.1
 Requires-Dist: tabulate>=0.7.7
@@ -38,18 +35,15 @@ Requires-Dist: tqdm>=4.64.0
 Provides-Extra: tf
 Requires-Dist: catalogue==2.0.10; extra == "tf"
 Requires-Dist: huggingface_hub>=0.26.0; extra == "tf"
-Requires-Dist: importlib-metadata>=5.0.0; extra == "tf"
 Requires-Dist: jsonlines==3.1.0; extra == "tf"
 Requires-Dist: lazy-imports==0.3.1; extra == "tf"
 Requires-Dist: mock==4.0.3; extra == "tf"
-Requires-Dist: networkx>=2.7.1; extra == "tf"
-Requires-Dist: numpy<2.0,>=1.21; extra == "tf"
+Requires-Dist: numpy>2.0; extra == "tf"
 Requires-Dist: packaging>=20.0; extra == "tf"
 Requires-Dist: Pillow>=10.0.0; extra == "tf"
 Requires-Dist: pypdf>=6.0.0; extra == "tf"
 Requires-Dist: pypdfium2>=4.30.0; extra == "tf"
 Requires-Dist: pyyaml>=6.0.1; extra == "tf"
-Requires-Dist: pyzmq>=16; extra == "tf"
 Requires-Dist: scipy>=1.13.1; extra == "tf"
 Requires-Dist: termcolor>=1.1; extra == "tf"
 Requires-Dist: tabulate>=0.7.7; extra == "tf"
@@ -58,30 +52,28 @@ Requires-Dist: tensorpack==0.11; extra == "tf"
 Requires-Dist: protobuf==3.20.1; extra == "tf"
 Requires-Dist: tensorflow-addons>=0.17.1; extra == "tf"
 Requires-Dist: tf2onnx>=1.9.2; extra == "tf"
-Requires-Dist: python-doctr==0.9.0; extra == "tf"
+Requires-Dist: python-doctr==0.10.0; extra == "tf"
 Requires-Dist: pycocotools>=2.0.2; extra == "tf"
 Requires-Dist: boto3==1.34.102; extra == "tf"
 Requires-Dist: pdfplumber>=0.11.0; extra == "tf"
-Requires-Dist: fasttext-wheel; extra == "tf"
+Requires-Dist: pyzmq>=16; extra == "tf"
 Requires-Dist: jdeskew>=0.2.2; extra == "tf"
 Requires-Dist: apted==1.0.3; extra == "tf"
 Requires-Dist: distance==0.1.3; extra == "tf"
 Requires-Dist: lxml>=4.9.1; extra == "tf"
+Requires-Dist: networkx>=2.7.1; extra == "tf"
 Provides-Extra: pt
 Requires-Dist: catalogue==2.0.10; extra == "pt"
 Requires-Dist: huggingface_hub>=0.26.0; extra == "pt"
-Requires-Dist: importlib-metadata>=5.0.0; extra == "pt"
 Requires-Dist: jsonlines==3.1.0; extra == "pt"
 Requires-Dist: lazy-imports==0.3.1; extra == "pt"
 Requires-Dist: mock==4.0.3; extra == "pt"
-Requires-Dist: networkx>=2.7.1; extra == "pt"
-Requires-Dist: numpy<2.0,>=1.21; extra == "pt"
+Requires-Dist: numpy>2.0; extra == "pt"
 Requires-Dist: packaging>=20.0; extra == "pt"
 Requires-Dist: Pillow>=10.0.0; extra == "pt"
 Requires-Dist: pypdf>=6.0.0; extra == "pt"
 Requires-Dist: pypdfium2>=4.30.0; extra == "pt"
 Requires-Dist: pyyaml>=6.0.1; extra == "pt"
-Requires-Dist: pyzmq>=16; extra == "pt"
 Requires-Dist: scipy>=1.13.1; extra == "pt"
 Requires-Dist: termcolor>=1.1; extra == "pt"
 Requires-Dist: tabulate>=0.7.7; extra == "pt"
@@ -89,15 +81,16 @@ Requires-Dist: tqdm>=4.64.0; extra == "pt"
 Requires-Dist: timm>=0.9.16; extra == "pt"
 Requires-Dist: transformers>=4.48.0; extra == "pt"
 Requires-Dist: accelerate>=0.29.1; extra == "pt"
-Requires-Dist: python-doctr==0.9.0; extra == "pt"
+Requires-Dist: python-doctr==0.10.0; extra == "pt"
 Requires-Dist: pycocotools>=2.0.2; extra == "pt"
 Requires-Dist: boto3==1.34.102; extra == "pt"
 Requires-Dist: pdfplumber>=0.11.0; extra == "pt"
-Requires-Dist: fasttext-wheel; extra == "pt"
+Requires-Dist: pyzmq>=16; extra == "pt"
 Requires-Dist: jdeskew>=0.2.2; extra == "pt"
 Requires-Dist: apted==1.0.3; extra == "pt"
 Requires-Dist: distance==0.1.3; extra == "pt"
 Requires-Dist: lxml>=4.9.1; extra == "pt"
+Requires-Dist: networkx>=2.7.1; extra == "pt"
 Provides-Extra: docs
 Requires-Dist: tensorpack==0.11; extra == "docs"
 Requires-Dist: boto3==1.34.102; extra == "docs"
@@ -183,7 +176,8 @@ It also provides a framework for training, evaluating and inferencing Document A
   [**LiLT**](https://github.com/jpWang/LiLT) and selected
   [**Bert**](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)-style including features like sliding windows.
 - Text mining for native PDFs with [**pdfplumber**](https://github.com/jsvine/pdfplumber),
-- Language detection with [**fastText**](https://github.com/facebookresearch/fastText),
+- Language detection with `papluca/xlm-roberta-base-language-detection`. [**fastText**](https://github.com/facebookresearch/fastText) is still available but
+  but will be removed in a future version.
 - Deskewing and rotating images with [**jdeskew**](https://github.com/phamquiluan/jdeskew).
 - Fine-tuning and evaluation tools.
 - Lot's of [tutorials](https://github.com/deepdoctection/notebooks)
@@ -294,7 +288,7 @@ alt="text" width="40%">
 
 - Linux or macOS. Windows is not supported but there is a [Dockerfile](./docker/pytorch-cpu-jupyter/Dockerfile) available.
 - Python >= 3.9
-- 2.2 \<= PyTorch **or** 2.11 \<= Tensorflow < 2.16. (For lower Tensorflow versions the code will only run on a GPU).
+- 2.6 \<= PyTorch **or** 2.11 \<= Tensorflow < 2.16. (For lower Tensorflow versions the code will only run on a GPU).
   Tensorflow support will be stopped from Python 3.11 onwards.
 - To fine-tune models, a GPU is recommended.
 
@@ -321,7 +315,7 @@ For a simple setup which is enough to parse documents with the default setting,
 
 ```
 pip install transformers
-pip install python-doctr==0.9.0
+pip install python-doctr==0.10.0 # If you use Python 3.10 or higher you can use the latest version.
 pip install deepdoctection
 ```
 
@@ -329,8 +323,9 @@ pip install deepdoctection
 
 ```
 pip install tensorpack
-pip install python-doctr==0.9.0
 pip install deepdoctection
+pip install "numpy>=1.21,<2.0" --upgrade --force-reinstall  # because TF 2.11 does not support numpy 2.0 
+pip install "python-doctr==0.9.0"
 ```
 
 Both setups are sufficient to run the [**introduction notebook**](https://github.com/deepdoctection/notebooks/blob/main/Get_Started.ipynb).

{deepdoctection-0.44.1 → deepdoctection-0.46}/README.md

@@ -40,7 +40,8 @@ It also provides a framework for training, evaluating and inferencing Document A
   [**LiLT**](https://github.com/jpWang/LiLT) and selected
   [**Bert**](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)-style including features like sliding windows.
 - Text mining for native PDFs with [**pdfplumber**](https://github.com/jsvine/pdfplumber),
-- Language detection with [**fastText**](https://github.com/facebookresearch/fastText),
+- Language detection with `papluca/xlm-roberta-base-language-detection`. [**fastText**](https://github.com/facebookresearch/fastText) is still available but
+  but will be removed in a future version.
 - Deskewing and rotating images with [**jdeskew**](https://github.com/phamquiluan/jdeskew).
 - Fine-tuning and evaluation tools.
 - Lot's of [tutorials](https://github.com/deepdoctection/notebooks)
@@ -151,7 +152,7 @@ alt="text" width="40%">
 
 - Linux or macOS. Windows is not supported but there is a [Dockerfile](./docker/pytorch-cpu-jupyter/Dockerfile) available.
 - Python >= 3.9
-- 2.2 \<= PyTorch **or** 2.11 \<= Tensorflow < 2.16. (For lower Tensorflow versions the code will only run on a GPU).
+- 2.6 \<= PyTorch **or** 2.11 \<= Tensorflow < 2.16. (For lower Tensorflow versions the code will only run on a GPU).
   Tensorflow support will be stopped from Python 3.11 onwards.
 - To fine-tune models, a GPU is recommended.
 
@@ -178,7 +179,7 @@ For a simple setup which is enough to parse documents with the default setting,
 
 ```
 pip install transformers
-pip install python-doctr==0.9.0
+pip install python-doctr==0.10.0 # If you use Python 3.10 or higher you can use the latest version.
 pip install deepdoctection
 ```
 
@@ -186,8 +187,9 @@ pip install deepdoctection
 
 ```
 pip install tensorpack
-pip install python-doctr==0.9.0
 pip install deepdoctection
+pip install "numpy>=1.21,<2.0" --upgrade --force-reinstall  # because TF 2.11 does not support numpy 2.0 
+pip install "python-doctr==0.9.0"
 ```
 
 Both setups are sufficient to run the [**introduction notebook**](https://github.com/deepdoctection/notebooks/blob/main/Get_Started.ipynb).

{deepdoctection-0.44.1 → deepdoctection-0.46}/deepdoctection/__init__.py

@@ -25,11 +25,10 @@ from .utils.logger import LoggingRecord, logger
 
 # pylint: enable=wrong-import-position
 
-__version__ = "0.44.1"
+__version__ = "0.46"
 
 _IMPORT_STRUCTURE = {
-    "analyzer": ["config_sanity_checks", "get_dd_analyzer", "ServiceFactory"],
-    "configs": ["update_cfg_from_defaults"],
+    "analyzer": ["config_sanity_checks", "get_dd_analyzer", "ServiceFactory", "update_cfg_from_defaults"],
     "dataflow": [
         "DataFlowTerminated",
         "DataFlowResetStateNotCalled",
@@ -186,7 +185,9 @@ _IMPORT_STRUCTURE = {
         "HFLayoutLmv3SequenceClassifier",
         "HFLiltTokenClassifier",
         "HFLiltSequenceClassifier",
+        "HFLmTokenClassifier",
         "HFLmSequenceClassifier",
+        "HFLmLanguageDetector",
         "ModelProfile",
         "ModelCatalog",
         "print_model_infos",
@@ -270,6 +271,7 @@ _IMPORT_STRUCTURE = {
         "MultiThreadPipelineComponent",
         "DoctectionPipe",
         "LanguageDetectionService",
+        "skip_if_category_or_service_extracted",
         "ImageLayoutService",
         "LMTokenClassifierService",
         "LMSequenceClassifierService",
@@ -309,12 +311,14 @@ _IMPORT_STRUCTURE = {
         "get_tensorpack_requirement",
         "pytorch_available",
         "get_pytorch_requirement",
+        "pyzmq_available",
         "lxml_available",
         "get_lxml_requirement",
         "apted_available",
         "get_apted_requirement",
         "distance_available",
         "get_distance_requirement",
+        "networkx_available",
         "numpy_v1_available",
         "get_numpy_v1_requirement",
         "transformers_available",

{deepdoctection-0.44.1 → deepdoctection-0.46}/deepdoctection/analyzer/config.py

@@ -520,6 +520,16 @@ cfg.USE_LAYOUT_LINK = False
 # (e.g., by grouping orphan text containers). Only applicable if list items were previously grouped.
 cfg.USE_LINE_MATCHER = False
 
+# Enables a sequence classification pipeline component, e.g. a LayoutLM or a Bert-like model.
+cfg.USE_LM_SEQUENCE_CLASS = False
+
+# Enables a token classification pipeline component, e.g. a LayoutLM or Bert-like model
+cfg.USE_LM_TOKEN_CLASS = False
+
+# Specifies the selection of the rotation model. There are two models available: A rotation estimator
+# based on Tesseract ('tesseract'), and a rotation estimator based on DocTr ('doctr').
+cfg.ROTATOR.MODEL = "tesseract"
+
 # Relevant when LIB = TF. Specifies the layout detection model.
 # This model should detect multiple or single objects across an entire page.
 # Currently, only one default model is supported.
@@ -899,6 +909,40 @@ cfg.LAYOUT_LINK.PARENTAL_CATEGORIES = [LayoutType.FIGURE, LayoutType.TABLE]
 # These are typically smaller or subordinate elements (e.g., captions).
 cfg.LAYOUT_LINK.CHILD_CATEGORIES = [LayoutType.CAPTION]
 
+
+# Weights configuration for sequence classifier. This will be a fine-tuned version of a LayoutLM, LayoutLMv2,
+# LayoutXLM, LayoutLMv3, LiLT or Roberta base model for sequence classification.
+cfg.LM_SEQUENCE_CLASS.WEIGHTS = None
+
+# When predicting document classes, it might be possible that some pages are empty or do not contain any text, in
+# which case the model will be unable to predict anything. If set to `True` it will
+# assign images with no features the category `TokenClasses.OTHER`.
+cfg.LM_SEQUENCE_CLASS.USE_OTHER_AS_DEFAULT_CATEGORY = False
+
+# Weights configuration for sequence classifier. This will be a fine-tuned version of a LayoutLM, LayoutLMv2,
+# LayoutXLM, LayoutLMv3, LiLT or Roberta base model for token classification.
+cfg.LM_TOKEN_CLASS.WEIGHTS = None
+
+# When predicting token classes, it might be possible that some words might not get sent to the model because they are
+# categorized as not eligible token (e.g. empty string). If set to `True` it will assign all words without token
+# as `TokenClasses.OTHER`.
+cfg.LM_TOKEN_CLASS.USE_OTHER_AS_DEFAULT_CATEGORY = False
+
+# Using bounding boxes of segments instead of words might improve model accuracy
+# for models that have been trained on segments rather than words (e.g. LiLT, LayoutLMv3).
+# Choose a single or a sequence of layout segments to use their bounding boxes. Note,
+# that the layout segments need to have a child-relationship with words. If a word
+# does not appear as child, it will use the word bounding box.
+cfg.LM_TOKEN_CLASS.SEGMENT_POSITIONS = None
+
+# If the output of the `tokenizer` exceeds the `max_length` sequence length, a
+# sliding window will be created with each window having `max_length` sequence
+# input. When using `SLIDING_WINDOW_STRIDE=0` no strides will be created,
+# otherwise it will create slides with windows shifted `SLIDING_WINDOW_STRIDE` to
+# the right.
+cfg.LM_TOKEN_CLASS.SLIDING_WINDOW_STRIDE = 0
+
+
 # Freezes the configuration to make it immutable.
 # This prevents accidental modification at runtime.
 cfg.freeze()

{deepdoctection-0.44.1 → deepdoctection-0.46}/deepdoctection/analyzer/factory.py

@@ -19,16 +19,29 @@
 `ServiceFactory` for building analyzers
 """
 
+from __future__ import annotations
 
 from os import environ
-from typing import Union
+from typing import TYPE_CHECKING, Literal, Union
 
 from lazy_imports import try_import
 
 from ..extern.base import ImageTransformer, ObjectDetector, PdfMiner
 from ..extern.d2detect import D2FrcnnDetector, D2FrcnnTracingDetector
-from ..extern.doctrocr import DoctrTextlineDetector, DoctrTextRecognizer
+from ..extern.doctrocr import DocTrRotationTransformer, DoctrTextlineDetector, DoctrTextRecognizer
 from ..extern.hfdetr import HFDetrDerivedDetector
+from ..extern.hflayoutlm import (
+    HFLayoutLmSequenceClassifier,
+    HFLayoutLmTokenClassifier,
+    HFLayoutLmv2SequenceClassifier,
+    HFLayoutLmv2TokenClassifier,
+    HFLayoutLmv3SequenceClassifier,
+    HFLayoutLmv3TokenClassifier,
+    HFLiltSequenceClassifier,
+    HFLiltTokenClassifier,
+    get_tokenizer_from_model_class,
+)
+from ..extern.hflm import HFLmSequenceClassifier, HFLmTokenClassifier
 from ..extern.model import ModelCatalog, ModelDownloadManager
 from ..extern.pdftext import PdfPlumberTextDetector
 from ..extern.tessocr import TesseractOcrDetector, TesseractRotationTransformer
@@ -45,6 +58,7 @@ from ..pipe.common import (
 )
 from ..pipe.doctectionpipe import DoctectionPipe
 from ..pipe.layout import ImageLayoutService, skip_if_category_or_service_extracted
+from ..pipe.lm import LMSequenceClassifierService, LMTokenClassifierService
 from ..pipe.order import TextOrderService
 from ..pipe.refine import TableSegmentationRefinementService
 from ..pipe.segment import PubtablesSegmentationService, TableSegmentationService
@@ -60,6 +74,11 @@ from ..utils.transform import PadTransform
 with try_import() as image_guard:
     from botocore.config import Config  # type: ignore
 
+if TYPE_CHECKING:
+    from ..extern.hflayoutlm import LayoutSequenceModels, LayoutTokenModels
+    from ..extern.hflm import LmSequenceModels, LmTokenModels
+
+    RotationTransformer = Union[TesseractRotationTransformer, DocTrRotationTransformer]
 
 __all__ = [
     "ServiceFactory",
@@ -172,24 +191,32 @@ class ServiceFactory:
         return ServiceFactory._build_layout_detector(config, mode)
 
     @staticmethod
-    def _build_rotation_detector() -> TesseractRotationTransformer:
+    def _build_rotation_detector(rotator_name: Literal["tesseract", "doctr"]) -> RotationTransformer:
         """
         Building a rotation detector.
 
         Returns:
             TesseractRotationTransformer: Rotation detector instance.
         """
-        return TesseractRotationTransformer()
+
+        if rotator_name == "tesseract":
+            return TesseractRotationTransformer()
+        if rotator_name == "doctr":
+            return DocTrRotationTransformer()
+        raise ValueError(
+            f"You have chosen rotator_name: {rotator_name} which is not allowed. Only tesseract or "
+            f"doctr are allowed."
+        )
 
     @staticmethod
-    def build_rotation_detector() -> TesseractRotationTransformer:
+    def build_rotation_detector(rotator_name: Literal["tesseract", "doctr"]) -> RotationTransformer:
         """
         Building a rotation detector.
 
         Returns:
             TesseractRotationTransformer: Rotation detector instance.
         """
-        return ServiceFactory._build_rotation_detector()
+        return ServiceFactory._build_rotation_detector(rotator_name)
 
     @staticmethod
     def _build_transform_service(transform_predictor: ImageTransformer) -> SimpleTransformService:
@@ -841,6 +868,226 @@ class ServiceFactory:
         """
         return ServiceFactory._build_text_order_service(config)
 
+    @staticmethod
+    def _build_sequence_classifier(config: AttrDict) -> Union[LayoutSequenceModels, LmSequenceModels]:
+        """
+        Builds and returns a sequence classifier instance.
+
+        Args:
+            config: Configuration object that determines the type of sequence classifier to construct.
+
+        Returns:
+            A sequence classifier instance constructed according to the specified configuration.
+        """
+        config_path = ModelCatalog.get_full_path_configs(config.LM_SEQUENCE_CLASS.WEIGHTS)
+        weights_path = ModelDownloadManager.maybe_download_weights_and_configs(config.LM_SEQUENCE_CLASS.WEIGHTS)
+        profile = ModelCatalog.get_profile(config.LM_SEQUENCE_CLASS.WEIGHTS)
+        categories = profile.categories if profile.categories is not None else {}
+        use_xlm_tokenizer = "xlm_tokenizer" == profile.architecture
+
+        if profile.model_wrapper in ("HFLayoutLmSequenceClassifier",):
+            return HFLayoutLmSequenceClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        if profile.model_wrapper in ("HFLayoutLmv2SequenceClassifier",):
+            return HFLayoutLmv2SequenceClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        if profile.model_wrapper in ("HFLayoutLmv3SequenceClassifier",):
+            return HFLayoutLmv3SequenceClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        if profile.model_wrapper in ("HFLiltSequenceClassifier",):
+            return HFLiltSequenceClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        if profile.model_wrapper in ("HFLmSequenceClassifier",):
+            return HFLmSequenceClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        raise ValueError(f"Unsupported model wrapper: {profile.model_wrapper}")
+
+    @staticmethod
+    def build_sequence_classifier(config: AttrDict) -> Union[LayoutSequenceModels, LmSequenceModels]:
+        """
+        Builds and returns a sequence classifier instance.
+
+        Args:
+            config: Configuration object that determines the type of sequence classifier to construct.
+
+        Returns:
+            A sequence classifier instance constructed according to the specified configuration.
+        """
+        return ServiceFactory._build_sequence_classifier(config)
+
+    @staticmethod
+    def _build_sequence_classifier_service(
+        config: AttrDict, sequence_classifier: Union[LayoutSequenceModels, LmSequenceModels]
+    ) -> LMSequenceClassifierService:
+        """
+        Building a sequence classifier service.
+
+        Args:
+            config: Configuration object.
+            sequence_classifier: Sequence classifier instance.
+
+        Returns:
+            LMSequenceClassifierService: Text order service instance.
+        """
+        tokenizer_fast = get_tokenizer_from_model_class(
+            sequence_classifier.model.__class__.__name__, sequence_classifier.use_xlm_tokenizer
+        )
+
+        return LMSequenceClassifierService(
+            tokenizer=tokenizer_fast,
+            language_model=sequence_classifier,
+            use_other_as_default_category=config.LM_SEQUENCE_CLASS.USE_OTHER_AS_DEFAULT_CATEGORY,
+        )
+
+    @staticmethod
+    def build_sequence_classifier_service(
+        config: AttrDict, sequence_classifier: Union[LayoutSequenceModels, LmSequenceModels]
+    ) -> LMSequenceClassifierService:
+        """
+        Building a sequence classifier service.
+
+        Args:
+            config: Configuration object.
+            sequence_classifier: Sequence classifier instance.
+
+        Returns:
+            LMSequenceClassifierService: Text order service instance.
+        """
+        return ServiceFactory._build_sequence_classifier_service(config, sequence_classifier)
+
+    @staticmethod
+    def _build_token_classifier(config: AttrDict) -> Union[LayoutTokenModels, LmTokenModels]:
+        """
+        Builds and returns a token classifier model.
+
+        Args:
+            config: Configuration object.
+
+        Returns:
+            The instantiated token classifier model.
+        """
+        config_path = ModelCatalog.get_full_path_configs(config.LM_TOKEN_CLASS.WEIGHTS)
+        weights_path = ModelDownloadManager.maybe_download_weights_and_configs(config.LM_TOKEN_CLASS.WEIGHTS)
+        profile = ModelCatalog.get_profile(config.LM_TOKEN_CLASS.WEIGHTS)
+        categories = profile.categories if profile.categories is not None else {}
+        use_xlm_tokenizer = "xlm_tokenizer" == profile.architecture
+        if profile.model_wrapper in ("HFLayoutLmTokenClassifier",):
+            return HFLayoutLmTokenClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+                use_xlm_tokenizer=use_xlm_tokenizer,
+            )
+        if profile.model_wrapper in ("HFLayoutLmv2TokenClassifier",):
+            return HFLayoutLmv2TokenClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+            )
+        if profile.model_wrapper in ("HFLayoutLmv3TokenClassifier",):
+            return HFLayoutLmv3TokenClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+            )
+        if profile.model_wrapper in ("HFLiltTokenClassifier",):
+            return HFLiltTokenClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+                device=config.DEVICE,
+            )
+        if profile.model_wrapper in ("HFLmTokenClassifier",):
+            return HFLmTokenClassifier(
+                path_config_json=config_path,
+                path_weights=weights_path,
+                categories=categories,
+            )
+        raise ValueError(f"Unsupported model wrapper: {profile.model_wrapper}")
+
+    @staticmethod
+    def build_token_classifier(config: AttrDict) -> Union[LayoutTokenModels, LmTokenModels]:
+        """
+        Builds and returns a token classifier model.
+
+        Args:
+            config: Configuration object.
+
+        Returns:
+            The instantiated token classifier model.
+        """
+        return ServiceFactory._build_token_classifier(config)
+
+    @staticmethod
+    def _build_token_classifier_service(
+        config: AttrDict, token_classifier: Union[LayoutTokenModels, LmTokenModels]
+    ) -> LMTokenClassifierService:
+        """
+        Building a token classifier service.
+
+        Args:
+            config: Configuration object.
+            token_classifier: Token classifier instance.
+
+        Returns:
+             A LMTokenClassifierService instance.
+        """
+        tokenizer_fast = get_tokenizer_from_model_class(
+            token_classifier.model.__class__.__name__, token_classifier.use_xlm_tokenizer
+        )
+
+        return LMTokenClassifierService(
+            tokenizer=tokenizer_fast,
+            language_model=token_classifier,
+            use_other_as_default_category=config.LM_TOKEN_CLASS.USE_OTHER_AS_DEFAULT_CATEGORY,
+            segment_positions=config.LM_TOKEN_CLASS.SEGMENT_POSITIONS,
+            sliding_window_stride=config.LM_TOKEN_CLASS.SLIDING_WINDOW_STRIDE,
+        )
+
+    @staticmethod
+    def build_token_classifier_service(
+        config: AttrDict, token_classifier: Union[LayoutTokenModels, LmTokenModels]
+    ) -> LMTokenClassifierService:
+        """
+        Building a token classifier service.
+
+        Args:
+            config: Configuration object.
+            token_classifier: Token classifier instance.
+
+        Returns:
+             A LMTokenClassifierService instance.
+        """
+        return ServiceFactory._build_token_classifier_service(config, token_classifier)
+
     @staticmethod
     def _build_page_parsing_service(config: AttrDict) -> PageParsingService:
         """
@@ -885,7 +1132,7 @@ class ServiceFactory:
         pipe_component_list: list[PipelineComponent] = []
 
         if config.USE_ROTATOR:
-            rotation_detector = ServiceFactory.build_rotation_detector()
+            rotation_detector = ServiceFactory.build_rotation_detector(config.ROTATOR.MODEL)
             transform_service = ServiceFactory.build_transform_service(transform_predictor=rotation_detector)
             pipe_component_list.append(transform_service)
 
@@ -955,6 +1202,16 @@ class ServiceFactory:
             line_list_matching_service = ServiceFactory.build_line_matching_service(config)
             pipe_component_list.append(line_list_matching_service)
 
+        if config.USE_LM_SEQUENCE_CLASS:
+            sequence_classifier = ServiceFactory.build_sequence_classifier(config)
+            sequence_classifier_service = ServiceFactory.build_sequence_classifier_service(config, sequence_classifier)
+            pipe_component_list.append(sequence_classifier_service)
+
+        if config.USE_LM_TOKEN_CLASS:
+            token_classifier = ServiceFactory.build_token_classifier(config)
+            token_classifier_service = ServiceFactory.build_token_classifier_service(config, token_classifier)
+            pipe_component_list.append(token_classifier_service)
+
         page_parsing_service = ServiceFactory.build_page_parsing_service(config)
 
         return DoctectionPipe(pipeline_component_list=pipe_component_list, page_parsing_service=page_parsing_service)

{deepdoctection-0.44.1 → deepdoctection-0.46}/deepdoctection/configs/profiles.jsonl

@@ -30,4 +30,5 @@
 {"name": "Felix92/doctr-torch-parseq-multilingual-v1/pytorch_model.bin", "description": "", "size": [63286381], "tp_model": false, "config": "Felix92/doctr-torch-parseq-multilingual-v1/config.json", "preprocessor_config": null, "hf_repo_id": "Felix92/doctr-torch-parseq-multilingual-v1", "hf_model_name": "pytorch_model.bin", "hf_config_file": ["config.json"], "urls": null, "categories": {}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "DoctrTextRecognizer", "architecture": "parseq", "padding": null}
 {"name": "doctr/crnn_vgg16_bn/pt/master-fde31e4a.pt", "description": "MASTER", "size": [63286381], "tp_model": false, "config": null, "preprocessor_config": null, "hf_repo_id": null, "hf_model_name": null, "hf_config_file": null, "urls": ["https://doctr-static.mindee.com/models?id=v0.7.0/master-fde31e4a.pt&src=0"], "categories": {}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "DoctrTextRecognizer", "architecture": "master", "padding": null}
 {"name": "Aryn/deformable-detr-DocLayNet/model.safetensors", "description": "Deformable DEtection TRansformer (DETR), trained on DocLayNet (including 80k annotated pages in 11 classes).", "size": [115511753], "tp_model": false, "config": "Aryn/deformable-detr-DocLayNet/config.json", "preprocessor_config": "Aryn/deformable-detr-DocLayNet/preprocessor_config.json", "hf_repo_id": "Aryn/deformable-detr-DocLayNet", "hf_model_name": "model.safetensors", "hf_config_file": ["config.json", "preprocessor_config.json"], "urls": null, "categories": {"1": "default_type", "2": "caption", "11": "text", "12": "title", "3": "footnote", "4": "formula", "5": "list_item", "6": "page_footer", "7": "page_header", "8": "figure", "9": "section_header", "10": "table"}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "HFDetrDerivedDetector", "architecture": null, "padding": null}
-{"name": "deepdoctection/tatr_tab_struct_v2/model.safetensors", "description": "Table Transformer (DETR) model trained on PubTables1M. It was introduced in the paper Aligning benchmark datasets for table structure recognition by Smock et al. This model is devoted to table structure recognition and assumes to receive a slightly croppedtable as input. It will predict rows, column and spanning cells. Use a padding of around 5 pixels. This artefact has been converted from deepdoctection/tatr_tab_struct_v2/pytorch_model.bin and should be used to reduce security issues", "size": [115511753], "tp_model": false, "config": "deepdoctection/tatr_tab_struct_v2/config.json", "preprocessor_config": "deepdoctection/tatr_tab_struct_v2/preprocessor_config.json", "hf_repo_id": "deepdoctection/tatr_tab_struct_v2", "hf_model_name": "model.safetensors", "hf_config_file": ["config.json", "preprocessor_config.json"], "urls": null, "categories": {"1": "table", "2": "column", "3": "row", "4": "column_header", "5": "projected_row_header", "6": "spanning"}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "HFDetrDerivedDetector", "architecture": null, "padding": null}
+{"name": "deepdoctection/tatr_tab_struct_v2/model.safetensors", "description": "Table Transformer (DETR) model trained on PubTables1M. It was introduced in the paper Aligning benchmark datasets for table structure recognition by Smock et al. This model is devoted to table structure recognition and assumes to receive a slightly croppedtable as input. It will predict rows, column and spanning cells. Use a padding of around 5 pixels. This artefact has been converted from deepdoctection/tatr_tab_struct_v2/pytorch_model.bin and should be used to reduce security issues", "size": [115511753], "tp_model": false, "config": "deepdoctection/tatr_tab_struct_v2/config.json", "preprocessor_config": "deepdoctection/tatr_tab_struct_v2/preprocessor_config.json", "hf_repo_id": "deepdoctection/tatr_tab_struct_v2", "hf_model_name": "model.safetensors", "hf_config_file": ["config.json", "preprocessor_config.json"], "urls": null, "categories": {"1": "table", "2": "column", "3": "row", "4": "column_header", "5": "projected_row_header", "6": "spanning"}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "HFDetrDerivedDetector", "architecture": null, "padding": null}
+{"name": "papluca/xlm-roberta-base-language-detection/model.safetensors", "description": "This model is an XLM-RoBERTa transformer model with a classification head on top (i.e. a linear layer on top of the pooled output). For additional information please refer to the xlm-roberta-base model card or to the paper Unsupervised Cross-lingual Representation Learning at Scale by Conneau et al.", "size": [101971449], "tp_model": false, "config": "papluca/xlm-roberta-base-language-detection/config.json", "preprocessor_config": null, "hf_repo_id": "papluca/xlm-roberta-base-language-detection", "hf_model_name": "model.safetensors", "hf_config_file": ["config.json"], "urls": null, "categories": {"1": "jpn", "2": "dut", "3": "ara", "4": "pol", "5": "deu", "6": "ita", "7": "por", "8": "tur", "9": "spa", "10": "hin", "11": "gre", "12": "urd", "13": "bul", "14": "eng", "15": "fre", "16": "chi", "17": "rus", "18": "tha", "19": "swa", "20": "vie"}, "categories_orig": null, "dl_library": "PT", "model_wrapper": "HFLmLanguageDetector", "architecture": null, "padding": null}

{deepdoctection-0.44.1 → deepdoctection-0.46}/deepdoctection/dataflow/parallel_map.py

@@ -24,15 +24,19 @@ from abc import ABC, abstractmethod
 from contextlib import contextmanager
 from typing import Any, Callable, Iterator, no_type_check
 
-import zmq
+from lazy_imports import try_import
 
 from ..utils.concurrency import StoppableThread, enable_death_signal, start_proc_mask_signal
 from ..utils.error import DataFlowTerminatedError
+from ..utils.file_utils import pyzmq_available
 from ..utils.logger import LoggingRecord, logger
 from .base import DataFlow, DataFlowReentrantGuard, ProxyDataFlow
 from .common import RepeatedData
 from .serialize import PickleSerializer
 
+with try_import() as import_guard:
+    import zmq
+
 
 @no_type_check
 def del_weakref(x):
@@ -77,6 +81,8 @@ def _get_pipe_name(name):
 
 class _ParallelMapData(ProxyDataFlow, ABC):
     def __init__(self, df: DataFlow, buffer_size: int, strict: bool = False) -> None:
+        if not pyzmq_available():
+            raise ModuleNotFoundError("pyzmq is required for running parallel dataflows (multiprocess/multithread).")
         super().__init__(df)
         if buffer_size <= 0:
             raise ValueError(f"buffer_size must be a positive number, got {buffer_size}")