deepdoctection 0.44.0__py3-none-any.whl → 0.45.0__py3-none-any.whl

deepdoctection/extern/hflm.py

@@ -20,16 +20,28 @@ Wrapper for the HF Language Model for sequence and token classification
 """
 from __future__ import annotations
 
+import os
 from abc import ABC
+from collections import defaultdict
 from pathlib import Path
-from typing import Literal, Mapping, Optional, Union
+from typing import TYPE_CHECKING, Any, Literal, Mapping, Optional, Sequence, Union
 
 from lazy_imports import try_import
+from typing_extensions import TypeAlias
 
 from ..utils.file_utils import get_pytorch_requirement, get_transformers_requirement
 from ..utils.settings import TypeOrStr
 from ..utils.types import JsonDict, PathLikeOrStr, Requirement
-from .base import LMSequenceClassifier, ModelCategories, SequenceClassResult
+from .base import (
+    DetectionResult,
+    LanguageDetector,
+    LMSequenceClassifier,
+    LMTokenClassifier,
+    ModelCategories,
+    NerModelCategories,
+    SequenceClassResult,
+    TokenClassResult,
+)
 from .hflayoutlm import get_tokenizer_from_model_class
 from .pt.ptutils import get_torch_device
 
@@ -38,14 +50,63 @@ with try_import() as pt_import_guard:
     import torch.nn.functional as F
 
 with try_import() as tr_import_guard:
-    from transformers import PretrainedConfig, XLMRobertaForSequenceClassification
+    from transformers import (
+        PretrainedConfig,
+        XLMRobertaForSequenceClassification,
+        XLMRobertaForTokenClassification,
+        XLMRobertaTokenizerFast,
+    )
 
 
-def predict_sequence_classes(
+def predict_token_classes_from_lm(
+    uuids: list[list[str]],
     input_ids: torch.Tensor,
     attention_mask: torch.Tensor,
     token_type_ids: torch.Tensor,
-    model: Union[XLMRobertaForSequenceClassification],
+    tokens: list[list[str]],
+    model: XLMRobertaForTokenClassification,
+) -> list[TokenClassResult]:
+    """
+    Args:
+        uuids: A list of uuids that correspond to a word that induces the resulting token
+        input_ids: Token converted to ids to be taken from `LayoutLMTokenizer`
+        attention_mask: The associated attention masks from padded sequences taken from `LayoutLMTokenizer`
+        token_type_ids: Torch tensor of token type ids taken from `LayoutLMTokenizer`
+        tokens: List of original tokens taken from `LayoutLMTokenizer`
+        model: layoutlm model for token classification
+
+    Returns:
+        A list of `TokenClassResult`s
+    """
+
+    outputs = model(input_ids=input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)
+
+    soft_max = F.softmax(outputs.logits, dim=2)
+    score = torch.max(soft_max, dim=2)[0].tolist()
+    token_class_predictions_ = outputs.logits.argmax(-1).tolist()
+    input_ids_list = input_ids.tolist()
+
+    all_results = defaultdict(list)
+    for idx, uuid_list in enumerate(uuids):
+        for pos, token in enumerate(uuid_list):
+            all_results[token].append(
+                (input_ids_list[idx][pos], token_class_predictions_[idx][pos], tokens[idx][pos], score[idx][pos])
+            )
+    all_token_classes = []
+    for uuid, res in all_results.items():
+        res.sort(key=lambda x: x[3], reverse=True)
+        output = res[0]
+        all_token_classes.append(
+            TokenClassResult(uuid=uuid, token_id=output[0], class_id=output[1], token=output[2], score=output[3])
+        )
+    return all_token_classes
+
+
+def predict_sequence_classes_from_lm(
+    input_ids: torch.Tensor,
+    attention_mask: torch.Tensor,
+    token_type_ids: torch.Tensor,
+    model: XLMRobertaForSequenceClassification,
 ) -> SequenceClassResult:
     """
     Args:
@@ -66,6 +127,250 @@ def predict_sequence_classes(
     return SequenceClassResult(class_id=sequence_class_predictions, score=float(score))  # type: ignore
 
 
+class HFLmTokenClassifierBase(LMTokenClassifier, ABC):
+    """
+    Abstract base class for wrapping Bert-like models for token classification into the framework.
+    """
+
+    def __init__(
+        self,
+        path_config_json: PathLikeOrStr,
+        path_weights: PathLikeOrStr,
+        categories_semantics: Optional[Sequence[TypeOrStr]] = None,
+        categories_bio: Optional[Sequence[TypeOrStr]] = None,
+        categories: Optional[Mapping[int, TypeOrStr]] = None,
+        device: Optional[Union[Literal["cpu", "cuda"], torch.device]] = None,
+        use_xlm_tokenizer: bool = False,
+    ):
+        """
+        Args:
+            path_config_json: path to `.json` config file
+            path_weights: path to model artifact
+            categories_semantics: A dict with key (indices) and values (category names) for `NER` semantics, i.e. the
+                                 entities self. To be consistent with detectors use only values `>0`. Conversion will
+                                 be done internally.
+            categories_bio: A dict with key (indices) and values (category names) for `NER` tags (i.e. `BIO`). To be
+                           consistent with detectors use only `values>0`. Conversion will be done internally.
+            categories: If you have a pre-trained model you can pass a complete dict of NER categories
+            device: The device (cpu,"cuda"), where to place the model.
+            use_xlm_tokenizer: True if one uses the `LayoutXLM` or a lilt model built with a xlm language model, e.g.
+                              `info-xlm` or `roberta-xlm`. (`LayoutXLM` cannot be distinguished from LayoutLMv2).
+        """
+
+        if categories is None:
+            if categories_semantics is None:
+                raise ValueError("If categories is None then categories_semantics cannot be None")
+            if categories_bio is None:
+                raise ValueError("If categories is None then categories_bio cannot be None")
+
+        self.path_config = Path(path_config_json)
+        self.path_weights = Path(path_weights)
+        self.categories = NerModelCategories(
+            init_categories=categories, categories_semantics=categories_semantics, categories_bio=categories_bio
+        )
+        self.device = get_torch_device(device)
+        self.use_xlm_tokenizer = use_xlm_tokenizer
+
+    @classmethod
+    def get_requirements(cls) -> list[Requirement]:
+        return [get_pytorch_requirement(), get_transformers_requirement()]
+
+    def _map_category_names(self, token_results: list[TokenClassResult]) -> list[TokenClassResult]:
+        for result in token_results:
+            result.class_name = self.categories.categories[result.class_id + 1]
+            output = self.categories.disentangle_token_class_and_tag(result.class_name)
+            if output is not None:
+                token_class, tag = output
+                result.semantic_name = token_class
+                result.bio_tag = tag
+            else:
+                result.semantic_name = result.class_name
+            result.class_id += 1
+        return token_results
+
+    def _validate_encodings(
+        self, **encodings: Any
+    ) -> tuple[list[list[str]], list[str], torch.Tensor, torch.Tensor, torch.Tensor, list[list[str]]]:
+        image_ids = encodings.get("image_ids", [])
+        ann_ids = encodings.get("ann_ids")
+        input_ids = encodings.get("input_ids")
+        attention_mask = encodings.get("attention_mask")
+        token_type_ids = encodings.get("token_type_ids")
+        tokens = encodings.get("tokens")
+
+        assert isinstance(ann_ids, list), type(ann_ids)
+        if len(set(image_ids)) > 1:
+            raise ValueError("HFLmTokenClassifier accepts for inference only one image.")
+        if isinstance(input_ids, torch.Tensor):
+            input_ids = input_ids.to(self.device)
+        else:
+            raise ValueError(f"input_ids must be list but is {type(input_ids)}")
+        if isinstance(attention_mask, torch.Tensor):
+            attention_mask = attention_mask.to(self.device)
+        else:
+            raise ValueError(f"attention_mask must be list but is {type(attention_mask)}")
+        if isinstance(token_type_ids, torch.Tensor):
+            token_type_ids = token_type_ids.to(self.device)
+        else:
+            raise ValueError(f"token_type_ids must be list but is {type(token_type_ids)}")
+        if not isinstance(tokens, list):
+            raise ValueError(f"tokens must be list but is {type(tokens)}")
+
+        return ann_ids, image_ids, input_ids, attention_mask, token_type_ids, tokens
+
+    def clone(self) -> HFLmTokenClassifierBase:
+        return self.__class__(
+            self.path_config,
+            self.path_weights,
+            self.categories.categories_semantics,
+            self.categories.categories_bio,
+            self.categories.get_categories(),
+            self.device,
+            self.use_xlm_tokenizer,
+        )
+
+    @staticmethod
+    def get_name(path_weights: PathLikeOrStr, architecture: str) -> str:
+        """Returns the name of the model"""
+        return f"Transformers_{architecture}_" + "_".join(Path(path_weights).parts[-2:])
+
+    @staticmethod
+    def get_tokenizer_class_name(model_class_name: str, use_xlm_tokenizer: bool) -> str:
+        """
+        A refinement for adding the tokenizer class name to the model configs.
+
+        Args:
+            model_class_name: The model name, e.g. `model.__class__.__name__`
+            use_xlm_tokenizer: Whether to use a `XLM` tokenizer.
+
+        Returns:
+            The name of the tokenizer class.
+        """
+        tokenizer = get_tokenizer_from_model_class(model_class_name, use_xlm_tokenizer)
+        return tokenizer.__class__.__name__
+
+    @staticmethod
+    def image_to_raw_features_mapping() -> str:
+        """Returns the mapping function to convert images into raw features."""
+        return "image_to_raw_lm_features"
+
+    @staticmethod
+    def image_to_features_mapping() -> str:
+        """Returns the mapping function to convert images into features."""
+        return "image_to_lm_features"
+
+
+class HFLmTokenClassifier(HFLmTokenClassifierBase):
+    """
+    A wrapper class for `transformers.XLMRobertaForTokenClassification` and similar models to use within a pipeline
+    component. Check <https://huggingface.co/docs/transformers/model_doc/xlm-roberta> for documentation of the
+    model itself.
+    Note that this model is equipped with a head that is only useful for classifying the tokens. For sequence
+    classification and other things please use another model of the family.
+
+    Example:
+        ```python
+        # setting up compulsory ocr service
+        tesseract_config_path = ModelCatalog.get_full_path_configs("/dd/conf_tesseract.yaml")
+        tess = TesseractOcrDetector(tesseract_config_path)
+        ocr_service = TextExtractionService(tess)
+
+        # hf tokenizer and token classifier
+        tokenizer = XLMRobertaTokenizerFast.from_pretrained("FacebookAI/xlm-roberta-base")
+        roberta = XLMRobertaForTokenClassification("path/to/config.json","path/to/model.bin",
+                                              categories=["first_name", "surname", "street"])
+
+        # token classification service
+        roberta_service = LMTokenClassifierService(tokenizer,roberta)
+
+        pipe = DoctectionPipe(pipeline_component_list=[ocr_service,roberta_service])
+
+        path = "path/to/some/form"
+        df = pipe.analyze(path=path)
+
+        for dp in df:
+            ...
+        ```
+    """
+
+    def __init__(
+        self,
+        path_config_json: PathLikeOrStr,
+        path_weights: PathLikeOrStr,
+        categories_semantics: Optional[Sequence[TypeOrStr]] = None,
+        categories_bio: Optional[Sequence[TypeOrStr]] = None,
+        categories: Optional[Mapping[int, TypeOrStr]] = None,
+        device: Optional[Union[Literal["cpu", "cuda"], torch.device]] = None,
+        use_xlm_tokenizer: bool = True,
+    ):
+        """
+        Args:
+            path_config_json: path to `.json` config file
+            path_weights: path to model artifact
+            categories_semantics: A dict with key (indices) and values (category names) for NER semantics, i.e. the
+                                 entities self. To be consistent with detectors use only values `>0`. Conversion will
+                                 be done internally.
+            categories_bio: A dict with key (indices) and values (category names) for `NER` tags (i.e. BIO). To be
+                           consistent with detectors use only values>0. Conversion will be done internally.
+            categories: If you have a pre-trained model you can pass a complete dict of NER categories
+            device: The device (cpu,"cuda"), where to place the model.
+            use_xlm_tokenizer: Do not change this value unless you pre-trained a bert-like model with a different
+                              Tokenizer.
+        """
+        super().__init__(
+            path_config_json, path_weights, categories_semantics, categories_bio, categories, device, use_xlm_tokenizer
+        )
+        self.name = self.get_name(path_weights, "bert-like-token-classification")
+        self.model_id = self.get_model_id()
+        self.model = self.get_wrapped_model(path_config_json, path_weights)
+        self.model.to(self.device)
+        self.model.config.tokenizer_class = self.get_tokenizer_class_name(
+            self.model.__class__.__name__, self.use_xlm_tokenizer
+        )
+
+    def predict(self, **encodings: Union[list[list[str]], torch.Tensor]) -> list[TokenClassResult]:
+        """
+        Launch inference on bert-like models for token classification. Pass the following arguments
+
+        Args:
+            encodings: input_ids: Token converted to ids to be taken from `LayoutLMTokenizer`
+                       attention_mask: The associated attention masks from padded sequences taken from
+                                       `LayoutLMTokenizer`
+                       token_type_ids: Torch tensor of token type ids taken from `LayoutLMTokenizer`
+                       boxes: Torch tensor of bounding boxes of type `xyxy`
+                       tokens: List of original tokens taken from `LayoutLMTokenizer`
+
+        Returns:
+            A list of `TokenClassResult`s
+        """
+
+        ann_ids, _, input_ids, attention_mask, token_type_ids, tokens = self._validate_encodings(**encodings)
+        results = predict_token_classes_from_lm(ann_ids, input_ids, attention_mask, token_type_ids, tokens, self.model)
+        return self._map_category_names(results)
+
+    @staticmethod
+    def get_wrapped_model(
+        path_config_json: PathLikeOrStr, path_weights: PathLikeOrStr
+    ) -> XLMRobertaForTokenClassification:
+        """
+        Get the inner (wrapped) model.
+
+        Args:
+            path_config_json: path to .json config file
+            path_weights: path to model artifact
+
+        Returns:
+            `nn.Module`
+        """
+        config = PretrainedConfig.from_pretrained(pretrained_model_name_or_path=os.fspath(path_config_json))
+        return XLMRobertaForTokenClassification.from_pretrained(
+            pretrained_model_name_or_path=os.fspath(path_weights), config=config
+        )
+
+    def clear_model(self) -> None:
+        self.model = None
+
+
 class HFLmSequenceClassifierBase(LMSequenceClassifier, ABC):
     """
     Abstract base class for wrapping Bert-type models for sequence classification into the deepdoctection framework.
@@ -208,10 +513,11 @@ class HFLmSequenceClassifier(HFLmSequenceClassifierBase):
         use_xlm_tokenizer: bool = True,
     ):
         super().__init__(path_config_json, path_weights, categories, device)
-        self.name = self.get_name(path_weights, "bert-like")
+        self.name = self.get_name(path_weights, "bert-like-sequence-classification")
         self.model_id = self.get_model_id()
         self.model = self.get_wrapped_model(path_config_json, path_weights)
         self.model.to(self.device)
+        self.use_xlm_tokenizer = use_xlm_tokenizer
         self.model.config.tokenizer_class = self.get_tokenizer_class_name(
             self.model.__class__.__name__, use_xlm_tokenizer
         )
@@ -219,7 +525,7 @@ class HFLmSequenceClassifier(HFLmSequenceClassifierBase):
     def predict(self, **encodings: Union[list[list[str]], torch.Tensor]) -> SequenceClassResult:
         input_ids, attention_mask, token_type_ids = self._validate_encodings(**encodings)
 
-        result = predict_sequence_classes(
+        result = predict_sequence_classes_from_lm(
             input_ids,
             attention_mask,
             token_type_ids,
@@ -262,3 +568,122 @@ class HFLmSequenceClassifier(HFLmSequenceClassifierBase):
 
     def clear_model(self) -> None:
         self.model = None
+
+
+class HFLmLanguageDetector(LanguageDetector):
+    """
+    Language detector using HuggingFace's `XLMRobertaForSequenceClassification`.
+
+    This class wraps a multilingual sequence classification model (XLMRobertaForSequenceClassification)
+    for language detection tasks. Input text is tokenized and truncated/padded to a maximum length of 512 tokens.
+    The prediction returns a `DetectionResult` containing the detected language code and its confidence score.
+    """
+
+    def __init__(
+        self,
+        path_config_json: PathLikeOrStr,
+        path_weights: PathLikeOrStr,
+        categories: Mapping[int, TypeOrStr],
+        device: Optional[Union[Literal["cpu", "cuda"], torch.device]] = None,
+        use_xlm_tokenizer: bool = True,
+    ):
+        super().__init__()
+        self.path_config = Path(path_config_json)
+        self.path_weights = Path(path_weights)
+        self.categories = ModelCategories(init_categories=categories)
+        self.device = get_torch_device(device)
+        self.use_xlm_tokenizer = use_xlm_tokenizer
+        self.model = self.get_wrapped_model(path_config_json, path_weights)
+        self.model.to(self.device)
+        self.tokenizer = XLMRobertaTokenizerFast.from_pretrained("xlm-roberta-base")
+        self.name = self.get_name(path_weights, "bert-like-language-detection")
+        self.model_id = self.get_model_id()
+
+    def predict(self, text_string: str) -> DetectionResult:
+        """
+        Predict the language of the input sequence.
+
+        Args:
+            text_string: The input text sequence to classify.
+
+        Returns:
+            DetectionResult: The detected language and its confidence score.
+        """
+        encoding = self.tokenizer(
+            text_string,
+            return_tensors="pt",
+            padding=True,
+            truncation=True,
+            max_length=512,
+        )
+        input_ids = encoding["input_ids"].to(self.device)
+        attention_mask = encoding["attention_mask"].to(self.device)
+        token_type_ids = encoding.get("token_type_ids")
+        if token_type_ids is not None:
+            token_type_ids = token_type_ids.to(self.device)
+        else:
+            token_type_ids = torch.zeros_like(input_ids)
+
+        self.model.eval()
+        with torch.no_grad():
+            outputs = self.model(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+            )
+            probs = torch.softmax(outputs.logits, dim=-1)
+            score, class_id_tensor = torch.max(probs, dim=-1)
+            class_id = int(class_id_tensor.item() + 1)
+            lang = self.categories.categories[class_id]
+
+        return DetectionResult(class_name=lang, score=float(score.item()))
+
+    def clear_model(self) -> None:
+        self.model = None
+
+    @classmethod
+    def get_requirements(cls) -> list[Requirement]:
+        return [get_pytorch_requirement(), get_transformers_requirement()]
+
+    @staticmethod
+    def get_wrapped_model(
+        path_config_json: PathLikeOrStr, path_weights: PathLikeOrStr
+    ) -> XLMRobertaForSequenceClassification:
+        """
+        Get the inner (wrapped) model.
+
+        Args:
+            path_config_json: path to .json config file
+            path_weights: path to model artifact
+
+        Returns:
+            `XLMRobertaForSequenceClassification`
+        """
+        config = PretrainedConfig.from_pretrained(pretrained_model_name_or_path=path_config_json)
+        return XLMRobertaForSequenceClassification.from_pretrained(
+            pretrained_model_name_or_path=path_weights, config=config
+        )
+
+    def clone(self) -> HFLmLanguageDetector:
+        return self.__class__(
+            self.path_config, self.path_weights, self.categories.get_categories(), self.device, self.use_xlm_tokenizer
+        )
+
+    @staticmethod
+    def get_name(path_weights: PathLikeOrStr, architecture: str) -> str:
+        """
+        Returns the name of the model
+
+        Args:
+            path_weights: Path to model weights
+            architecture: Architecture name
+
+        Returns:
+            str: Model name
+        """
+        return f"Transformers_{architecture}_" + "_".join(Path(path_weights).parts[-2:])
+
+
+if TYPE_CHECKING:
+    LmTokenModels: TypeAlias = Union[HFLmTokenClassifier,]
+    LmSequenceModels: TypeAlias = Union[HFLmSequenceClassifier,]

deepdoctection/mapper/laylmstruct.py

@@ -806,17 +806,17 @@ def image_to_raw_lm_features(
     raw_features["image_id"] = page.image_id
     raw_features["width"] = page.width
     raw_features["height"] = page.height
-    raw_features["ann_ids"] = text_["ann_ids"]
-    raw_features["words"] = text_["words"]
+    raw_features["ann_ids"] = text_.ann_ids
+    raw_features["words"] = text_.words
     # We use a dummy bounding box for all bounding boxes so that we can pass the raw features to
     # raw_features_to_layoutlm_features
-    raw_features["bbox"] = [_CLS_BOX] * len(text_["words"])
+    raw_features["bbox"] = [_CLS_BOX] * len(text_.words)
     raw_features["dataset_type"] = dataset_type
 
-    if use_token_tag and text_["token_tags"]:
-        raw_features["labels"] = text_["token_tags"]
-    elif text_["token_classes"]:
-        raw_features["labels"] = text_["token_classes"]
+    if use_token_tag and text_.token_tags:
+        raw_features["labels"] = text_.token_tags
+    elif text_.token_classes:
+        raw_features["labels"] = text_.token_classes
     elif page.document_type is not None:
         document_type_id = page.image_orig.summary.get_sub_category(PageType.DOCUMENT_TYPE).category_id - 1
         raw_features["labels"] = [document_type_id]

deepdoctection/pipe/language.py

@@ -21,7 +21,7 @@ Module for language detection pipeline component
 from typing import Optional, Sequence
 
 from ..datapoint.image import Image, MetaAnnotation
-from ..datapoint.view import ImageDefaults, Page
+from ..datapoint.view import IMAGE_DEFAULTS, Page
 from ..extern.base import LanguageDetector, ObjectDetector
 from ..utils.error import ImageError
 from ..utils.settings import PageType, TypeOrStr, get_type
@@ -75,11 +75,11 @@ class LanguageDetectionService(PipelineComponent):
 
         self.predictor = language_detector
         self.text_detector = text_detector
-        self.text_container = get_type(text_container) if text_container is not None else ImageDefaults.TEXT_CONTAINER
+        self.text_container = get_type(text_container) if text_container is not None else IMAGE_DEFAULTS.TEXT_CONTAINER
         self.floating_text_block_categories = (
             tuple(get_type(text_block) for text_block in floating_text_block_categories)
             if (floating_text_block_categories is not None)
-            else ()
+            else IMAGE_DEFAULTS.FLOATING_TEXT_BLOCK_CATEGORIES
         )
 
         super().__init__(self._get_name(self.predictor.name))
@@ -109,7 +109,7 @@ class LanguageDetectionService(PipelineComponent):
             text = " ".join((result.text for result in detect_result_list if result.text is not None))
         predict_result = self.predictor.predict(text)
         self.dp_manager.set_summary_annotation(
-            PageType.LANGUAGE, PageType.LANGUAGE, 1, predict_result.text, predict_result.score
+            PageType.LANGUAGE, PageType.LANGUAGE, 1, predict_result.class_name, predict_result.score
         )
 
     def clone(self) -> PipelineComponent:

deepdoctection/pipe/lm.py

@@ -20,6 +20,7 @@ Module for token classification pipeline
 """
 from __future__ import annotations
 
+import inspect
 from copy import copy
 from typing import TYPE_CHECKING, Any, Callable, Literal, Optional, Sequence, Union
 
@@ -32,6 +33,7 @@ from .registry import pipeline_component_registry
 
 if TYPE_CHECKING:
     from ..extern.hflayoutlm import LayoutSequenceModels, LayoutTokenModels
+    from ..extern.hflm import LmSequenceModels, LmTokenModels
 
 
 @pipeline_component_registry.register("LMTokenClassifierService")
@@ -70,7 +72,7 @@ class LMTokenClassifierService(PipelineComponent):
     def __init__(
         self,
         tokenizer: Any,
-        language_model: LayoutTokenModels,
+        language_model: Union[LayoutTokenModels, LmTokenModels],
         padding: Literal["max_length", "do_not_pad", "longest"] = "max_length",
         truncation: bool = True,
         return_overflowing_tokens: bool = False,
@@ -124,7 +126,7 @@ class LMTokenClassifierService(PipelineComponent):
                                            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 the `BioTag.outside` token.
-            segment_positions: Using bounding boxes of segment instead of words improves model accuracy
+            segment_positions: Using bounding boxes of segments instead of words improves model accuracy
                                significantly for models that have been trained on segments rather than words.
                                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
@@ -271,6 +273,8 @@ class LMTokenClassifierService(PipelineComponent):
                 f"You want to use {type(self.tokenizer)} but you should use {tokenizer_class_name} "
                 f"in this framework"
             )
+        func_params = inspect.signature(self.mapping_to_lm_input_func).parameters
+        self.required_kwargs = {k: v for k, v in self.required_kwargs.items() if k in func_params}
 
     @staticmethod
     def image_to_features_func(mapping_str: str) -> Callable[..., Callable[[Image], Optional[Any]]]:
@@ -318,7 +322,7 @@ class LMSequenceClassifierService(PipelineComponent):
     def __init__(
         self,
         tokenizer: Any,
-        language_model: LayoutSequenceModels,
+        language_model: Union[LayoutSequenceModels, LmSequenceModels],
         padding: Literal["max_length", "do_not_pad", "longest"] = "max_length",
         truncation: bool = True,
         return_overflowing_tokens: bool = False,

deepdoctection/utils/file_utils.py

@@ -18,6 +18,7 @@ from types import ModuleType
 from typing import Any, Union, no_type_check
 
 import importlib_metadata
+import numpy as np
 from packaging import version
 
 from .error import DependencyError
@@ -249,6 +250,39 @@ def get_distance_requirement() -> Requirement:
     return "distance", distance_available(), _DISTANCE_ERR_MSG
 
 
+_NUMPY_V1_ERR_MSG = "numpy v1 must be installed."
+
+
+def numpy_v1_available() -> bool:
+    """
+    Check if the installed NumPy version is version 1.
+
+    This helper function determines whether the currently installed version
+    of NumPy is version 1 by inspecting its major version number.
+
+    Returns:
+        True if the installed NumPy version is 1, otherwise False
+    """
+    major_version = np.__version__.split(".", maxsplit=1)[0]
+    print(f"major version: {major_version}")
+    if major_version in (1, "1"):
+        return True
+    return False
+
+
+def get_numpy_v1_requirement() -> Requirement:
+    """
+    Retrieves the requirement details for numpy version 1.
+
+    Returns:
+        A tuple containing three elements:
+            - The requirement name for numpy version 1.
+            - A Boolean value indicating whether numpy version 1 is available.
+            - An error message in case numpy version 1 is not available.
+    """
+    return "numpy v1", numpy_v1_available(), _NUMPY_V1_ERR_MSG
+
+
 # Transformers
 _TRANSFORMERS_AVAILABLE = importlib.util.find_spec("transformers") is not None
 _TRANSFORMERS_ERR_MSG = f"transformers must be installed. {_GENERIC_ERR_MSG}"

deepdoctection/utils/settings.py

@@ -108,6 +108,7 @@ class DocumentType(ObjectTypes):
     GOVERNMENT_TENDERS = "government_tenders"
     MANUALS = "manuals"
     PATENTS = "patents"
+    BANK_STATEMENT = "bank_statement"
 
 
 @object_types_registry.register("LayoutType")
@@ -296,6 +297,7 @@ class Languages(ObjectTypes):
     BOSNIAN = "bos"
     NORWEGIAN_NOVOSIBIRSK = "nno"
     URDU = "urd"
+    SWAHILI = "swa"
     NOT_DEFINED = "nn"
 
 

deepdoctection/utils/types.py

@@ -70,7 +70,6 @@ AnnotationDict: TypeAlias = dict[str, Any]
 ImageDict: TypeAlias = dict[str, Any]
 
 # We use these types for output types of the Page object
-Text_: TypeAlias = dict[str, Any]
 HTML: TypeAlias = str
 csv: TypeAlias = list[list[str]]
 Chunks: TypeAlias = list[tuple[str, str, int, str, str, str, str]]