fix: Add unit tests (#51)

* add the pytests Signed-off-by: Peter Staar <taa@zurich.ibm.com> * renamed the test folder and added the toplevel test Signed-off-by: Peter Staar <taa@zurich.ibm.com> * updated the toplevel function test Signed-off-by: Peter Staar <taa@zurich.ibm.com> * need to start running all tests successfully Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added the reference converted documents Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added first test for json and md output Signed-off-by: Peter Staar <taa@zurich.ibm.com> * ran pre-commit Signed-off-by: Peter Staar <taa@zurich.ibm.com> * replaced deprecated json function with model_dump_json Signed-off-by: Peter Staar <taa@zurich.ibm.com> * replaced deprecated json function with model_dump_json Signed-off-by: Peter Staar <taa@zurich.ibm.com> * reformatted code Signed-off-by: Peter Staar <taa@zurich.ibm.com> * Fix backend tests Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * commented out the drawing Signed-off-by: Peter Staar <taa@zurich.ibm.com> * ci: avoid duplicate runs Signed-off-by: Michele Dolfi <97102151+dolfim-ibm@users.noreply.github.com> * commented out json verification for now Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added verification of input cells Signed-off-by: Peter Staar <taa@zurich.ibm.com> * reformat code Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added test to verify the cells in the pages Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added test to verify the cells in the pages (2) Signed-off-by: Peter Staar <taa@zurich.ibm.com> * added test to verify the cells in the pages (3) Signed-off-by: Peter Staar <taa@zurich.ibm.com> * run all examples in CI Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * make sure examples return failures Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * raise a failure if examples fail Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * fix examples Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * run examples after tests Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * Add tests and update top_level_tests using only datamodels Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * Remove unnecessary code Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * Validate conversion status on e2e test Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * package verify utils and add more tests Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * reduce docs in example, since they are already in the tests Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * skip batch_convert Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * pin docling-parse 1.1.2 Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> * updated the error messages Signed-off-by: Peter Staar <taa@zurich.ibm.com> * commented out the json verification for now Signed-off-by: Peter Staar <taa@zurich.ibm.com> * bumped GLM version Signed-off-by: Peter Staar <taa@zurich.ibm.com> * Fix lockfile Signed-off-by: Christoph Auer <cau@zurich.ibm.com> * Pin new docling-parse v1.1.3 Signed-off-by: Christoph Auer <cau@zurich.ibm.com> --------- Signed-off-by: Peter Staar <taa@zurich.ibm.com> Signed-off-by: Christoph Auer <cau@zurich.ibm.com> Signed-off-by: Michele Dolfi <97102151+dolfim-ibm@users.noreply.github.com> Signed-off-by: Michele Dolfi <dol@zurich.ibm.com> Co-authored-by: Christoph Auer <cau@zurich.ibm.com> Co-authored-by: Michele Dolfi <97102151+dolfim-ibm@users.noreply.github.com> Co-authored-by: Michele Dolfi <dol@zurich.ibm.com>
2024-08-30 14:08:20 +02:00 · 2024-08-30 14:08:20 +02:00 · 48f4d1ba52
commit 48f4d1ba52
parent 256f4d504e
43 changed files with 4999 additions and 353 deletions
--- a/.github/workflows/checks.yml
+++ b/.github/workflows/checks.yml
@ -14,3 +14,22 @@ jobs:
          python-version: ${{ matrix.python-version }}
      - name: Run styling check
        run: poetry run pre-commit run --all-files
+      - name: Install with poetry
+        run: poetry install --all-extras
+      - name: Testing
+        run: |
+          poetry run pytest -v tests
+      - name: Run examples
+        run: |
+          for file in examples/*.py; do
+            # Skip batch_convert.py
+            if [[ "$(basename "$file")" == "batch_convert.py" ]]; then
+                echo "Skipping $file"
+                continue
+            fi
+
+            echo "Running example $file"
+            poetry run python "$file" || exit 1
+          done
+      - name: Build with poetry
+        run: poetry build
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@ -2,7 +2,7 @@ name: "Run CI"

 on:
  pull_request:
-    types: [opened, reopened, synchronize, ready_for_review]
+    types: [opened, reopened]
  push:
    branches:
      - "**"
@ -25,4 +25,4 @@ jobs:
  #     - uses: ./.github/actions/setup-poetry
  #     - name: Build docs
  #       run: poetry run mkdocs build --verbose --clean
-  
+  
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@ -4,7 +4,7 @@ repos:
    hooks:
      - id: system
        name: Black
-        entry: poetry run black docling examples
+        entry: poetry run black docling examples tests
        pass_filenames: false
        language: system
        files: '\.py$'
@ -12,7 +12,7 @@ repos:
    hooks:
      - id: system
        name: isort
-        entry: poetry run isort docling examples
+        entry: poetry run isort docling examples tests
        pass_filenames: false
        language: system
        files: '\.py$'
--- a/docling/datamodel/base_models.py
+++ b/docling/datamodel/base_models.py
@ -238,9 +238,9 @@ class EquationPrediction(BaseModel):

 class PagePredictions(BaseModel):
    layout: LayoutPrediction = None
-    tablestructure: TableStructurePrediction = None
-    figures_classification: FigureClassificationPrediction = None
-    equations_prediction: EquationPrediction = None
+    tablestructure: Optional[TableStructurePrediction] = None
+    figures_classification: Optional[FigureClassificationPrediction] = None
+    equations_prediction: Optional[EquationPrediction] = None


 PageElement = Union[TextElement, TableElement, FigureElement]
--- a/docling/models/ds_glm_model.py
+++ b/docling/models/ds_glm_model.py
@ -16,8 +16,12 @@ from docling.datamodel.document import ConversionResult
 class GlmModel:
    def __init__(self, config):
        self.config = config
+        self.model_names = self.config.get(
+            "model_names", ""
+        )  # "language;term;reference"
        load_pretrained_nlp_models()
-        model = init_nlp_model(model_names="language;term;reference")
+        # model = init_nlp_model(model_names="language;term;reference")
+        model = init_nlp_model(model_names=self.model_names)
        self.model = model

    def __call__(self, conv_res: ConversionResult) -> DsDocument:
--- a/docling/models/table_structure_model.py
+++ b/docling/models/table_structure_model.py
@ -44,7 +44,16 @@ class TableStructureModel:

            for tc in table_element.table_cells:
                x0, y0, x1, y1 = tc.bbox.as_tuple()
-                draw.rectangle([(x0, y0), (x1, y1)], outline="blue")
+                if tc.column_header:
+                    width = 3
+                else:
+                    width = 1
+                draw.rectangle([(x0, y0), (x1, y1)], outline="blue", width=width)
+                draw.text(
+                    (x0 + 3, y0 + 3),
+                    text=f"{tc.start_row_offset_idx}, {tc.start_col_offset_idx}",
+                    fill="black",
+                )

        image.show()

--- a/examples/batch_convert.py
+++ b/examples/batch_convert.py
@ -49,17 +49,18 @@ def export_documents(
        f"of which {failure_count} failed "
        f"and {partial_success_count} were partially converted."
    )
+    return success_count, partial_success_count, failure_count


 def main():
    logging.basicConfig(level=logging.INFO)

    input_doc_paths = [
-        Path("./test/data/2206.01062.pdf"),
-        Path("./test/data/2203.01017v2.pdf"),
-        Path("./test/data/2305.03393v1.pdf"),
-        Path("./test/data/redp5110.pdf"),
-        Path("./test/data/redp5695.pdf"),
+        Path("./tests/data/2206.01062.pdf"),
+        Path("./tests/data/2203.01017v2.pdf"),
+        Path("./tests/data/2305.03393v1.pdf"),
+        Path("./tests/data/redp5110.pdf"),
+        Path("./tests/data/redp5695.pdf"),
    ]

    # buf = BytesIO(Path("./test/data/2206.01062.pdf").open("rb").read())
@ -73,12 +74,19 @@ def main():
    start_time = time.time()

    conv_results = doc_converter.convert(input)
-    export_documents(conv_results, output_dir=Path("./scratch"))
+    success_count, partial_success_count, failure_count = export_documents(
+        conv_results, output_dir=Path("./scratch")
+    )

    end_time = time.time() - start_time

    _log.info(f"All documents were converted in {end_time:.2f} seconds.")

+    if failure_count > 0:
+        raise RuntimeError(
+            f"The example failed converting {failure_count} on {len(input_doc_paths)}."
+        )
+

 if __name__ == "__main__":
    main()
--- a/examples/custom_convert.py
+++ b/examples/custom_convert.py
@ -42,14 +42,14 @@ def export_documents(
        f"Processed {success_count + failure_count} docs, of which {failure_count} failed"
    )

+    return success_count, failure_count
+

 def main():
    logging.basicConfig(level=logging.INFO)

    input_doc_paths = [
-        Path("./test/data/2206.01062.pdf"),
-        Path("./test/data/2203.01017v2.pdf"),
-        Path("./test/data/2305.03393v1.pdf"),
+        Path("./tests/data/2206.01062.pdf"),
    ]

    ###########################################################################
@ -114,12 +114,19 @@ def main():
    start_time = time.time()

    conv_results = doc_converter.convert(input)
-    export_documents(conv_results, output_dir=Path("./scratch"))
+    success_count, failure_count = export_documents(
+        conv_results, output_dir=Path("./scratch")
+    )

    end_time = time.time() - start_time

    _log.info(f"All documents were converted in {end_time:.2f} seconds.")

+    if failure_count > 0:
+        raise RuntimeError(
+            f"The example failed converting {failure_count} on {len(input_doc_paths)}."
+        )
+

 if __name__ == "__main__":
    main()
--- a/examples/export_figures.py
+++ b/examples/export_figures.py
@ -22,7 +22,7 @@ def main():
    logging.basicConfig(level=logging.INFO)

    input_doc_paths = [
-        Path("./test/data/2206.01062.pdf"),
+        Path("./tests/data/2206.01062.pdf"),
    ]
    output_dir = Path("./scratch")

@ -41,10 +41,13 @@ def main():

    conv_results = doc_converter.convert(input_files)

+    success_count = 0
+    failure_count = 0
    output_dir.mkdir(parents=True, exist_ok=True)
    for conv_res in conv_results:
        if conv_res.status != ConversionStatus.SUCCESS:
            _log.info(f"Document {conv_res.input.file} failed to convert.")
+            failure_count += 1
            continue

        doc_filename = conv_res.input.file.stem
@ -66,10 +69,17 @@ def main():
            with element_image_filename.open("wb") as fp:
                image.save(fp, "PNG")

+        success_count += 1
+
    end_time = time.time() - start_time

    _log.info(f"All documents were converted in {end_time:.2f} seconds.")

+    if failure_count > 0:
+        raise RuntimeError(
+            f"The example failed converting {failure_count} on {len(input_doc_paths)}."
+        )
+

 if __name__ == "__main__":
    main()
--- a/poetry.lock
+++ b/poetry.lock
@ -11,17 +11,6 @@ files = [
    {file = "annotated_types-0.7.0.tar.gz", hash = "sha256:aff07c09a53a08bc8cfccb9c85b05f1aa9a2a6f23728d790723543408344ce89"},
 ]

-[[package]]
-name = "apted"
-version = "1.0.3"
-description = "APTED algorithm for the Tree Edit Distance"
-optional = false
-python-versions = "*"
-files = [
-    {file = "apted-1.0.3-py3-none-any.whl", hash = "sha256:74193369d023649d335269e67c4df07f922959e5ac2597de1b79af4e694150e8"},
-    {file = "apted-1.0.3.tar.gz", hash = "sha256:befa5181e2d4457fa88e54995a82604ee048bb2fbc781ea97d8e1856b4715ce9"},
-]
-
 [[package]]
 name = "astroid"
 version = "2.15.8"
@ -37,8 +26,8 @@ files = [
 lazy-object-proxy = ">=1.4.0"
 typing-extensions = {version = ">=4.0.0", markers = "python_version < \"3.11\""}
 wrapt = [
-    {version = ">=1.11,<2", markers = "python_version < \"3.11\""},
    {version = ">=1.14,<2", markers = "python_version >= \"3.11\""},
+    {version = ">=1.11,<2", markers = "python_version < \"3.11\""},
 ]

 [[package]]
@ -78,17 +67,6 @@ docs = ["cogapp", "furo", "myst-parser", "sphinx", "sphinx-notfound-page", "sphi
 tests = ["cloudpickle", "hypothesis", "mypy (>=1.11.1)", "pympler", "pytest (>=4.3.0)", "pytest-mypy-plugins", "pytest-xdist[psutil]"]
 tests-mypy = ["mypy (>=1.11.1)", "pytest-mypy-plugins"]

-[[package]]
-name = "bashlex"
-version = "0.18"
-description = "Python parser for bash"
-optional = false
-python-versions = ">=2.7, !=3.0, !=3.1, !=3.2, !=3.3, !=3.4"
-files = [
-    {file = "bashlex-0.18-py2.py3-none-any.whl", hash = "sha256:91d73a23a3e51711919c1c899083890cdecffc91d8c088942725ac13e9dcfffa"},
-    {file = "bashlex-0.18.tar.gz", hash = "sha256:5bb03a01c6d5676338c36fd1028009c8ad07e7d61d8a1ce3f513b7fff52796ee"},
-]
-
 [[package]]
 name = "black"
 version = "24.8.0"
@ -137,17 +115,6 @@ d = ["aiohttp (>=3.7.4)", "aiohttp (>=3.7.4,!=3.9.0)"]
 jupyter = ["ipython (>=7.8.0)", "tokenize-rt (>=3.2.0)"]
 uvloop = ["uvloop (>=0.15.2)"]

-[[package]]
-name = "bracex"
-version = "2.5"
-description = "Bash style brace expander."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "bracex-2.5-py3-none-any.whl", hash = "sha256:d2fcf4b606a82ac325471affe1706dd9bbaa3536c91ef86a31f6b766f3dad1d0"},
-    {file = "bracex-2.5.tar.gz", hash = "sha256:0725da5045e8d37ea9592ab3614d8b561e22c3c5fde3964699be672e072ab611"},
-]
-
 [[package]]
 name = "build"
 version = "1.2.1"
@ -196,13 +163,13 @@ redis = ["redis (>=2.10.5)"]

 [[package]]
 name = "certifi"
-version = "2024.7.4"
+version = "2024.8.30"
 description = "Python package for providing Mozilla's CA Bundle."
 optional = false
 python-versions = ">=3.6"
 files = [
-    {file = "certifi-2024.7.4-py3-none-any.whl", hash = "sha256:c198e21b1289c2ab85ee4e67bb4b4ef3ead0892059901a8d5b622f24a1101e90"},
-    {file = "certifi-2024.7.4.tar.gz", hash = "sha256:5a1e7645bc0ec61a09e26c36f6106dd4cf40c6db3a1fb6352b0244e7fb057c7b"},
+    {file = "certifi-2024.8.30-py3-none-any.whl", hash = "sha256:922820b53db7a7257ffbda3f597266d435245903d80737e34f8a45ff3e3230d8"},
+    {file = "certifi-2024.8.30.tar.gz", hash = "sha256:bec941d2aa8195e248a60b31ff9f0558284cf01a52591ceda73ea9afffd69fd9"},
 ]

 [[package]]
@ -394,34 +361,6 @@ files = [
    {file = "charset_normalizer-3.3.2-py3-none-any.whl", hash = "sha256:3e4d1f6587322d2788836a99c69062fbb091331ec940e02d12d179c1d53e25fc"},
 ]

-[[package]]
-name = "cibuildwheel"
-version = "2.20.0"
-description = "Build Python wheels on CI with minimal configuration."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "cibuildwheel-2.20.0-py3-none-any.whl", hash = "sha256:d90719cc386af540b52f3cd8c733972c1fe222bbb2a941e5f5cd87215a0c82a3"},
-    {file = "cibuildwheel-2.20.0.tar.gz", hash = "sha256:5c3fd67e4417fe37021b595bedcaf0c87e5800ecf9d6096229967858a20cc6c8"},
-]
-
-[package.dependencies]
-bashlex = "!=0.13"
-bracex = "*"
-certifi = "*"
-filelock = "*"
-packaging = ">=20.9"
-platformdirs = "*"
-tomli = {version = "*", markers = "python_version < \"3.11\""}
-typing-extensions = {version = ">=4.1.0", markers = "python_version < \"3.11\""}
-
-[package.extras]
-bin = ["click", "packaging (>=21.0)", "pip-tools", "pygithub", "pyyaml", "requests", "rich (>=9.6)"]
-dev = ["build", "click", "jinja2", "packaging (>=21.0)", "pip-tools", "pygithub", "pytest (>=6)", "pytest-timeout", "pytest-xdist", "pyyaml", "requests", "rich (>=9.6)", "setuptools", "tomli-w", "validate-pyproject"]
-docs = ["jinja2 (>=3.1.2)", "mkdocs (==1.3.1)", "mkdocs-include-markdown-plugin (==2.8.0)", "mkdocs-macros-plugin", "pymdown-extensions"]
-test = ["build", "jinja2", "pytest (>=6)", "pytest-timeout", "pytest-xdist", "setuptools", "tomli-w", "validate-pyproject"]
-uv = ["uv"]
-
 [[package]]
 name = "cleo"
 version = "2.1.0"
@ -495,66 +434,87 @@ cron = ["capturer (>=2.4)"]

 [[package]]
 name = "contourpy"
-version = "1.2.1"
+version = "1.3.0"
 description = "Python library for calculating contours of 2D quadrilateral grids"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "contourpy-1.2.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:bd7c23df857d488f418439686d3b10ae2fbf9bc256cd045b37a8c16575ea1040"},
-    {file = "contourpy-1.2.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:5b9eb0ca724a241683c9685a484da9d35c872fd42756574a7cfbf58af26677fd"},
-    {file = "contourpy-1.2.1-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4c75507d0a55378240f781599c30e7776674dbaf883a46d1c90f37e563453480"},
-    {file = "contourpy-1.2.1-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:11959f0ce4a6f7b76ec578576a0b61a28bdc0696194b6347ba3f1c53827178b9"},
-    {file = "contourpy-1.2.1-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:eb3315a8a236ee19b6df481fc5f997436e8ade24a9f03dfdc6bd490fea20c6da"},
-    {file = "contourpy-1.2.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:39f3ecaf76cd98e802f094e0d4fbc6dc9c45a8d0c4d185f0f6c2234e14e5f75b"},
-    {file = "contourpy-1.2.1-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:94b34f32646ca0414237168d68a9157cb3889f06b096612afdd296003fdd32fd"},
-    {file = "contourpy-1.2.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:457499c79fa84593f22454bbd27670227874cd2ff5d6c84e60575c8b50a69619"},
-    {file = "contourpy-1.2.1-cp310-cp310-win32.whl", hash = "sha256:ac58bdee53cbeba2ecad824fa8159493f0bf3b8ea4e93feb06c9a465d6c87da8"},
-    {file = "contourpy-1.2.1-cp310-cp310-win_amd64.whl", hash = "sha256:9cffe0f850e89d7c0012a1fb8730f75edd4320a0a731ed0c183904fe6ecfc3a9"},
-    {file = "contourpy-1.2.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:6022cecf8f44e36af10bd9118ca71f371078b4c168b6e0fab43d4a889985dbb5"},
-    {file = "contourpy-1.2.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:ef5adb9a3b1d0c645ff694f9bca7702ec2c70f4d734f9922ea34de02294fdf72"},
-    {file = "contourpy-1.2.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:6150ffa5c767bc6332df27157d95442c379b7dce3a38dff89c0f39b63275696f"},
-    {file = "contourpy-1.2.1-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4c863140fafc615c14a4bf4efd0f4425c02230eb8ef02784c9a156461e62c965"},
-    {file = "contourpy-1.2.1-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:00e5388f71c1a0610e6fe56b5c44ab7ba14165cdd6d695429c5cd94021e390b2"},
-    {file = "contourpy-1.2.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d4492d82b3bc7fbb7e3610747b159869468079fe149ec5c4d771fa1f614a14df"},
-    {file = "contourpy-1.2.1-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:49e70d111fee47284d9dd867c9bb9a7058a3c617274900780c43e38d90fe1205"},
-    {file = "contourpy-1.2.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:b59c0ffceff8d4d3996a45f2bb6f4c207f94684a96bf3d9728dbb77428dd8cb8"},
-    {file = "contourpy-1.2.1-cp311-cp311-win32.whl", hash = "sha256:7b4182299f251060996af5249c286bae9361fa8c6a9cda5efc29fe8bfd6062ec"},
-    {file = "contourpy-1.2.1-cp311-cp311-win_amd64.whl", hash = "sha256:2855c8b0b55958265e8b5888d6a615ba02883b225f2227461aa9127c578a4922"},
-    {file = "contourpy-1.2.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:62828cada4a2b850dbef89c81f5a33741898b305db244904de418cc957ff05dc"},
-    {file = "contourpy-1.2.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:309be79c0a354afff9ff7da4aaed7c3257e77edf6c1b448a779329431ee79d7e"},
-    {file = "contourpy-1.2.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:2e785e0f2ef0d567099b9ff92cbfb958d71c2d5b9259981cd9bee81bd194c9a4"},
-    {file = "contourpy-1.2.1-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:1cac0a8f71a041aa587410424ad46dfa6a11f6149ceb219ce7dd48f6b02b87a7"},
-    {file = "contourpy-1.2.1-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:af3f4485884750dddd9c25cb7e3915d83c2db92488b38ccb77dd594eac84c4a0"},
-    {file = "contourpy-1.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ce6889abac9a42afd07a562c2d6d4b2b7134f83f18571d859b25624a331c90b"},
-    {file = "contourpy-1.2.1-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:a1eea9aecf761c661d096d39ed9026574de8adb2ae1c5bd7b33558af884fb2ce"},
-    {file = "contourpy-1.2.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:187fa1d4c6acc06adb0fae5544c59898ad781409e61a926ac7e84b8f276dcef4"},
-    {file = "contourpy-1.2.1-cp312-cp312-win32.whl", hash = "sha256:c2528d60e398c7c4c799d56f907664673a807635b857df18f7ae64d3e6ce2d9f"},
-    {file = "contourpy-1.2.1-cp312-cp312-win_amd64.whl", hash = "sha256:1a07fc092a4088ee952ddae19a2b2a85757b923217b7eed584fdf25f53a6e7ce"},
-    {file = "contourpy-1.2.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:bb6834cbd983b19f06908b45bfc2dad6ac9479ae04abe923a275b5f48f1a186b"},
-    {file = "contourpy-1.2.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:1d59e739ab0e3520e62a26c60707cc3ab0365d2f8fecea74bfe4de72dc56388f"},
-    {file = "contourpy-1.2.1-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:bd3db01f59fdcbce5b22afad19e390260d6d0222f35a1023d9adc5690a889364"},
-    {file = "contourpy-1.2.1-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:a12a813949e5066148712a0626895c26b2578874e4cc63160bb007e6df3436fe"},
-    {file = "contourpy-1.2.1-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:fe0ccca550bb8e5abc22f530ec0466136379c01321fd94f30a22231e8a48d985"},
-    {file = "contourpy-1.2.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e1d59258c3c67c865435d8fbeb35f8c59b8bef3d6f46c1f29f6123556af28445"},
-    {file = "contourpy-1.2.1-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:f32c38afb74bd98ce26de7cc74a67b40afb7b05aae7b42924ea990d51e4dac02"},
-    {file = "contourpy-1.2.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:d31a63bc6e6d87f77d71e1abbd7387ab817a66733734883d1fc0021ed9bfa083"},
-    {file = "contourpy-1.2.1-cp39-cp39-win32.whl", hash = "sha256:ddcb8581510311e13421b1f544403c16e901c4e8f09083c881fab2be80ee31ba"},
-    {file = "contourpy-1.2.1-cp39-cp39-win_amd64.whl", hash = "sha256:10a37ae557aabf2509c79715cd20b62e4c7c28b8cd62dd7d99e5ed3ce28c3fd9"},
-    {file = "contourpy-1.2.1-pp39-pypy39_pp73-macosx_10_9_x86_64.whl", hash = "sha256:a31f94983fecbac95e58388210427d68cd30fe8a36927980fab9c20062645609"},
-    {file = "contourpy-1.2.1-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ef2b055471c0eb466033760a521efb9d8a32b99ab907fc8358481a1dd29e3bd3"},
-    {file = "contourpy-1.2.1-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:b33d2bc4f69caedcd0a275329eb2198f560b325605810895627be5d4b876bf7f"},
-    {file = "contourpy-1.2.1.tar.gz", hash = "sha256:4d8908b3bee1c889e547867ca4cdc54e5ab6be6d3e078556814a22457f49423c"},
+    {file = "contourpy-1.3.0-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:880ea32e5c774634f9fcd46504bf9f080a41ad855f4fef54f5380f5133d343c7"},
+    {file = "contourpy-1.3.0-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:76c905ef940a4474a6289c71d53122a4f77766eef23c03cd57016ce19d0f7b42"},
+    {file = "contourpy-1.3.0-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:92f8557cbb07415a4d6fa191f20fd9d2d9eb9c0b61d1b2f52a8926e43c6e9af7"},
+    {file = "contourpy-1.3.0-cp310-cp310-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:36f965570cff02b874773c49bfe85562b47030805d7d8360748f3eca570f4cab"},
+    {file = "contourpy-1.3.0-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:cacd81e2d4b6f89c9f8a5b69b86490152ff39afc58a95af002a398273e5ce589"},
+    {file = "contourpy-1.3.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:69375194457ad0fad3a839b9e29aa0b0ed53bb54db1bfb6c3ae43d111c31ce41"},
+    {file = "contourpy-1.3.0-cp310-cp310-musllinux_1_2_aarch64.whl", hash = "sha256:7a52040312b1a858b5e31ef28c2e865376a386c60c0e248370bbea2d3f3b760d"},
+    {file = "contourpy-1.3.0-cp310-cp310-musllinux_1_2_x86_64.whl", hash = "sha256:3faeb2998e4fcb256542e8a926d08da08977f7f5e62cf733f3c211c2a5586223"},
+    {file = "contourpy-1.3.0-cp310-cp310-win32.whl", hash = "sha256:36e0cff201bcb17a0a8ecc7f454fe078437fa6bda730e695a92f2d9932bd507f"},
+    {file = "contourpy-1.3.0-cp310-cp310-win_amd64.whl", hash = "sha256:87ddffef1dbe5e669b5c2440b643d3fdd8622a348fe1983fad7a0f0ccb1cd67b"},
+    {file = "contourpy-1.3.0-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:0fa4c02abe6c446ba70d96ece336e621efa4aecae43eaa9b030ae5fb92b309ad"},
+    {file = "contourpy-1.3.0-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:834e0cfe17ba12f79963861e0f908556b2cedd52e1f75e6578801febcc6a9f49"},
+    {file = "contourpy-1.3.0-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:dbc4c3217eee163fa3984fd1567632b48d6dfd29216da3ded3d7b844a8014a66"},
+    {file = "contourpy-1.3.0-cp311-cp311-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4865cd1d419e0c7a7bf6de1777b185eebdc51470800a9f42b9e9decf17762081"},
+    {file = "contourpy-1.3.0-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:303c252947ab4b14c08afeb52375b26781ccd6a5ccd81abcdfc1fafd14cf93c1"},
+    {file = "contourpy-1.3.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:637f674226be46f6ba372fd29d9523dd977a291f66ab2a74fbeb5530bb3f445d"},
+    {file = "contourpy-1.3.0-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:76a896b2f195b57db25d6b44e7e03f221d32fe318d03ede41f8b4d9ba1bff53c"},
+    {file = "contourpy-1.3.0-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:e1fd23e9d01591bab45546c089ae89d926917a66dceb3abcf01f6105d927e2cb"},
+    {file = "contourpy-1.3.0-cp311-cp311-win32.whl", hash = "sha256:d402880b84df3bec6eab53cd0cf802cae6a2ef9537e70cf75e91618a3801c20c"},
+    {file = "contourpy-1.3.0-cp311-cp311-win_amd64.whl", hash = "sha256:6cb6cc968059db9c62cb35fbf70248f40994dfcd7aa10444bbf8b3faeb7c2d67"},
+    {file = "contourpy-1.3.0-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:570ef7cf892f0afbe5b2ee410c507ce12e15a5fa91017a0009f79f7d93a1268f"},
+    {file = "contourpy-1.3.0-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:da84c537cb8b97d153e9fb208c221c45605f73147bd4cadd23bdae915042aad6"},
+    {file = "contourpy-1.3.0-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0be4d8425bfa755e0fd76ee1e019636ccc7c29f77a7c86b4328a9eb6a26d0639"},
+    {file = "contourpy-1.3.0-cp312-cp312-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:9c0da700bf58f6e0b65312d0a5e695179a71d0163957fa381bb3c1f72972537c"},
+    {file = "contourpy-1.3.0-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:eb8b141bb00fa977d9122636b16aa67d37fd40a3d8b52dd837e536d64b9a4d06"},
+    {file = "contourpy-1.3.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3634b5385c6716c258d0419c46d05c8aa7dc8cb70326c9a4fb66b69ad2b52e09"},
+    {file = "contourpy-1.3.0-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:0dce35502151b6bd35027ac39ba6e5a44be13a68f55735c3612c568cac3805fd"},
+    {file = "contourpy-1.3.0-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:aea348f053c645100612b333adc5983d87be69acdc6d77d3169c090d3b01dc35"},
+    {file = "contourpy-1.3.0-cp312-cp312-win32.whl", hash = "sha256:90f73a5116ad1ba7174341ef3ea5c3150ddf20b024b98fb0c3b29034752c8aeb"},
+    {file = "contourpy-1.3.0-cp312-cp312-win_amd64.whl", hash = "sha256:b11b39aea6be6764f84360fce6c82211a9db32a7c7de8fa6dd5397cf1d079c3b"},
+    {file = "contourpy-1.3.0-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:3e1c7fa44aaae40a2247e2e8e0627f4bea3dd257014764aa644f319a5f8600e3"},
+    {file = "contourpy-1.3.0-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:364174c2a76057feef647c802652f00953b575723062560498dc7930fc9b1cb7"},
+    {file = "contourpy-1.3.0-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:32b238b3b3b649e09ce9aaf51f0c261d38644bdfa35cbaf7b263457850957a84"},
+    {file = "contourpy-1.3.0-cp313-cp313-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:d51fca85f9f7ad0b65b4b9fe800406d0d77017d7270d31ec3fb1cc07358fdea0"},
+    {file = "contourpy-1.3.0-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:732896af21716b29ab3e988d4ce14bc5133733b85956316fb0c56355f398099b"},
+    {file = "contourpy-1.3.0-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d73f659398a0904e125280836ae6f88ba9b178b2fed6884f3b1f95b989d2c8da"},
+    {file = "contourpy-1.3.0-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:c6c7c2408b7048082932cf4e641fa3b8ca848259212f51c8c59c45aa7ac18f14"},
+    {file = "contourpy-1.3.0-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:f317576606de89da6b7e0861cf6061f6146ead3528acabff9236458a6ba467f8"},
+    {file = "contourpy-1.3.0-cp313-cp313-win32.whl", hash = "sha256:31cd3a85dbdf1fc002280c65caa7e2b5f65e4a973fcdf70dd2fdcb9868069294"},
+    {file = "contourpy-1.3.0-cp313-cp313-win_amd64.whl", hash = "sha256:4553c421929ec95fb07b3aaca0fae668b2eb5a5203d1217ca7c34c063c53d087"},
+    {file = "contourpy-1.3.0-cp313-cp313t-macosx_10_13_x86_64.whl", hash = "sha256:345af746d7766821d05d72cb8f3845dfd08dd137101a2cb9b24de277d716def8"},
+    {file = "contourpy-1.3.0-cp313-cp313t-macosx_11_0_arm64.whl", hash = "sha256:3bb3808858a9dc68f6f03d319acd5f1b8a337e6cdda197f02f4b8ff67ad2057b"},
+    {file = "contourpy-1.3.0-cp313-cp313t-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:420d39daa61aab1221567b42eecb01112908b2cab7f1b4106a52caaec8d36973"},
+    {file = "contourpy-1.3.0-cp313-cp313t-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:4d63ee447261e963af02642ffcb864e5a2ee4cbfd78080657a9880b8b1868e18"},
+    {file = "contourpy-1.3.0-cp313-cp313t-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:167d6c890815e1dac9536dca00828b445d5d0df4d6a8c6adb4a7ec3166812fa8"},
+    {file = "contourpy-1.3.0-cp313-cp313t-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:710a26b3dc80c0e4febf04555de66f5fd17e9cf7170a7b08000601a10570bda6"},
+    {file = "contourpy-1.3.0-cp313-cp313t-musllinux_1_2_aarch64.whl", hash = "sha256:75ee7cb1a14c617f34a51d11fa7524173e56551646828353c4af859c56b766e2"},
+    {file = "contourpy-1.3.0-cp313-cp313t-musllinux_1_2_x86_64.whl", hash = "sha256:33c92cdae89ec5135d036e7218e69b0bb2851206077251f04a6c4e0e21f03927"},
+    {file = "contourpy-1.3.0-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:a11077e395f67ffc2c44ec2418cfebed032cd6da3022a94fc227b6faf8e2acb8"},
+    {file = "contourpy-1.3.0-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:e8134301d7e204c88ed7ab50028ba06c683000040ede1d617298611f9dc6240c"},
+    {file = "contourpy-1.3.0-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:e12968fdfd5bb45ffdf6192a590bd8ddd3ba9e58360b29683c6bb71a7b41edca"},
+    {file = "contourpy-1.3.0-cp39-cp39-manylinux_2_17_ppc64le.manylinux2014_ppc64le.whl", hash = "sha256:fd2a0fc506eccaaa7595b7e1418951f213cf8255be2600f1ea1b61e46a60c55f"},
+    {file = "contourpy-1.3.0-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:4cfb5c62ce023dfc410d6059c936dcf96442ba40814aefbfa575425a3a7f19dc"},
+    {file = "contourpy-1.3.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:68a32389b06b82c2fdd68276148d7b9275b5f5cf13e5417e4252f6d1a34f72a2"},
+    {file = "contourpy-1.3.0-cp39-cp39-musllinux_1_2_aarch64.whl", hash = "sha256:94e848a6b83da10898cbf1311a815f770acc9b6a3f2d646f330d57eb4e87592e"},
+    {file = "contourpy-1.3.0-cp39-cp39-musllinux_1_2_x86_64.whl", hash = "sha256:d78ab28a03c854a873787a0a42254a0ccb3cb133c672f645c9f9c8f3ae9d0800"},
+    {file = "contourpy-1.3.0-cp39-cp39-win32.whl", hash = "sha256:81cb5ed4952aae6014bc9d0421dec7c5835c9c8c31cdf51910b708f548cf58e5"},
+    {file = "contourpy-1.3.0-cp39-cp39-win_amd64.whl", hash = "sha256:14e262f67bd7e6eb6880bc564dcda30b15e351a594657e55b7eec94b6ef72843"},
+    {file = "contourpy-1.3.0-pp310-pypy310_pp73-macosx_10_15_x86_64.whl", hash = "sha256:fe41b41505a5a33aeaed2a613dccaeaa74e0e3ead6dd6fd3a118fb471644fd6c"},
+    {file = "contourpy-1.3.0-pp310-pypy310_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:eca7e17a65f72a5133bdbec9ecf22401c62bcf4821361ef7811faee695799779"},
+    {file = "contourpy-1.3.0-pp310-pypy310_pp73-win_amd64.whl", hash = "sha256:1ec4dc6bf570f5b22ed0d7efba0dfa9c5b9e0431aeea7581aa217542d9e809a4"},
+    {file = "contourpy-1.3.0-pp39-pypy39_pp73-macosx_10_15_x86_64.whl", hash = "sha256:00ccd0dbaad6d804ab259820fa7cb0b8036bda0686ef844d24125d8287178ce0"},
+    {file = "contourpy-1.3.0-pp39-pypy39_pp73-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8ca947601224119117f7c19c9cdf6b3ab54c5726ef1d906aa4a69dfb6dd58102"},
+    {file = "contourpy-1.3.0-pp39-pypy39_pp73-win_amd64.whl", hash = "sha256:c6ec93afeb848a0845a18989da3beca3eec2c0f852322efe21af1931147d12cb"},
+    {file = "contourpy-1.3.0.tar.gz", hash = "sha256:7ffa0db17717a8ffb127efd0c95a4362d996b892c2904db72428d5b52e1938a4"},
 ]

 [package.dependencies]
-numpy = ">=1.20"
+numpy = ">=1.23"

 [package.extras]
 bokeh = ["bokeh", "selenium"]
 docs = ["furo", "sphinx (>=7.2)", "sphinx-copybutton"]
-mypy = ["contourpy[bokeh,docs]", "docutils-stubs", "mypy (==1.8.0)", "types-Pillow"]
+mypy = ["contourpy[bokeh,docs]", "docutils-stubs", "mypy (==1.11.1)", "types-Pillow"]
 test = ["Pillow", "contourpy[test-no-images]", "matplotlib"]
-test-no-images = ["pytest", "pytest-cov", "pytest-xdist", "wurlitzer"]
+test-no-images = ["pytest", "pytest-cov", "pytest-rerunfailures", "pytest-xdist", "wurlitzer"]

 [[package]]
 name = "crashtest"
@ -663,36 +623,36 @@ files = [

 [[package]]
 name = "deepsearch-glm"
-version = "0.19.0"
+version = "0.19.1"
 description = "Graph Language Models"
 optional = false
 python-versions = "<4.0,>=3.8"
 files = [
-    {file = "deepsearch_glm-0.19.0-cp310-cp310-macosx_13_6_arm64.whl", hash = "sha256:d420c7eb4e27b64cdc33c0beba159147fc4be14e141133f0f6ef080465b2529c"},
-    {file = "deepsearch_glm-0.19.0-cp310-cp310-macosx_13_6_x86_64.whl", hash = "sha256:8af4583ea6d914e87d6db96cae1d73272af6fe85193e67406f0c700064e794c2"},
-    {file = "deepsearch_glm-0.19.0-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:772e6e245b4d77d9df84af07693f9c19bc2f3dc6de4cb44deaf5fdd4a6c8e68d"},
-    {file = "deepsearch_glm-0.19.0-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:6a0c29f8cf8a1ee392c68985f8952a01b43dd8f2c5a1476b890f2c90d7ecbc96"},
-    {file = "deepsearch_glm-0.19.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:bb4f34bb5e45df2790eb6bbaf5caa625393d903da502b086de65df9ce4e3fff2"},
-    {file = "deepsearch_glm-0.19.0-cp311-cp311-macosx_13_6_arm64.whl", hash = "sha256:320195914e96b8197e53665594c4480b86f3fc4cacd5e6782befb2bb94494a40"},
-    {file = "deepsearch_glm-0.19.0-cp311-cp311-macosx_13_6_x86_64.whl", hash = "sha256:7221851c304ef364a13eeffa940a7c15592e9d5b0050b97904221a65be33f3ab"},
-    {file = "deepsearch_glm-0.19.0-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:b9b9e7102cf4355be8458569c7a44133b54446ae623923772db6942ce0fb2e87"},
-    {file = "deepsearch_glm-0.19.0-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:fa8d8d718149cfecd724a0eca246a3bd57588dffb757f204b629a35623d8f946"},
-    {file = "deepsearch_glm-0.19.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2ae251cf69b43d945fbf2cd41a89ba12312cd319a1c28d41c99d35cc476376b5"},
-    {file = "deepsearch_glm-0.19.0-cp312-cp312-macosx_13_6_arm64.whl", hash = "sha256:ad9a1fbf76e2561bc37e238ee9dd320b4b9cd49e61c55613e3977eedadee52bc"},
-    {file = "deepsearch_glm-0.19.0-cp312-cp312-macosx_13_6_x86_64.whl", hash = "sha256:2030aec8ce751927fe20ca1788e125e9b0c37f994c30062e59c4d7b7a87cbb64"},
-    {file = "deepsearch_glm-0.19.0-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:e57611b5d6cc2da91901e4b39fab6c9131dffe8766f43c20093bff75a0039100"},
-    {file = "deepsearch_glm-0.19.0-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:efbcf157cd6bd2dd6138312cef5df378598fd67e6c3f6f0b63ed3342c1de7f49"},
-    {file = "deepsearch_glm-0.19.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9febce49f365fcc5ada1ded720d387c51328ee470d4fcc89044b0684e074e699"},
-    {file = "deepsearch_glm-0.19.0-cp38-cp38-macosx_13_6_arm64.whl", hash = "sha256:04a29ba5e942f32659ae1a65cfe5e90e93d50e058d53b4763fe13df93f30492f"},
-    {file = "deepsearch_glm-0.19.0-cp38-cp38-macosx_13_6_x86_64.whl", hash = "sha256:a0024f42d6711f574dcab52ef2914a55f31b4fd804d3ad20ca7f211498e8a19b"},
-    {file = "deepsearch_glm-0.19.0-cp38-cp38-macosx_14_0_arm64.whl", hash = "sha256:513e5f1de14f0b12c916a52118083094a9ced439e4800d3442b2dd04f3cdbead"},
-    {file = "deepsearch_glm-0.19.0-cp38-cp38-macosx_14_0_x86_64.whl", hash = "sha256:6d3dd07a549b8cd4408308b0b6b8ca65397ce7e8c819d050d8b2deb03cd1977e"},
-    {file = "deepsearch_glm-0.19.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:7be341a85ce1ff164963a6d58b50955599dc33b34040975c972a798ae0f6f12c"},
-    {file = "deepsearch_glm-0.19.0-cp39-cp39-macosx_13_6_arm64.whl", hash = "sha256:b6c3c0d1295666a8a68f76262c020ffdc6de64cdd95671bf24c0592fa1317533"},
-    {file = "deepsearch_glm-0.19.0-cp39-cp39-macosx_13_6_x86_64.whl", hash = "sha256:cf290fe3824bd0de01b7c1d681aa14c89c5e60c6735fa471e04a985e55aead44"},
-    {file = "deepsearch_glm-0.19.0-cp39-cp39-macosx_14_0_arm64.whl", hash = "sha256:92943c495646660aef99ba64a7e3b77ffeca4866e96044f8be5e14dfa7ee660e"},
-    {file = "deepsearch_glm-0.19.0-cp39-cp39-macosx_14_0_x86_64.whl", hash = "sha256:a7c89d6fae4ed9dc960f9ee9734e91d321222080bf439e1d89e8c67270afc282"},
-    {file = "deepsearch_glm-0.19.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fb6398a34f5afac6282c4a5b7ea5a89f27fcf4c0adac43af27ecbac9e2731ce3"},
+    {file = "deepsearch_glm-0.19.1-cp310-cp310-macosx_13_6_arm64.whl", hash = "sha256:340dcf42e16e5d1ed7d16a4707d1ec20f5af864ffd24c5baedce92d98205f334"},
+    {file = "deepsearch_glm-0.19.1-cp310-cp310-macosx_13_6_x86_64.whl", hash = "sha256:f448a08c80c8cadda1e042bbcf63c38cc070c17093fd57a1a1b94cf44a1753cf"},
+    {file = "deepsearch_glm-0.19.1-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:9c1e22d5e21c285fc217343673788b969220645a42f7bd4e43d97d3d60f6e63d"},
+    {file = "deepsearch_glm-0.19.1-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:0c0654c71a19f0413717201f8b6c815387ffb7c3351a48db89829082e01b784a"},
+    {file = "deepsearch_glm-0.19.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cb1b74440228e621c83c4a19032c4cb71eb0a6037a7087f368679355b09d9d40"},
+    {file = "deepsearch_glm-0.19.1-cp311-cp311-macosx_13_6_arm64.whl", hash = "sha256:01bfd641a8dab9621fe9ef4fb66d40306279093942050fbf097f4a17985a7316"},
+    {file = "deepsearch_glm-0.19.1-cp311-cp311-macosx_13_6_x86_64.whl", hash = "sha256:0cf94ddc34a59595be62d4cab10076e5679531159e4a51d783d2265ed961551e"},
+    {file = "deepsearch_glm-0.19.1-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:e6f6bb4b3d4ba10e4cac4fc5b810021555374e97d55ee4af0cc9b5996e29174f"},
+    {file = "deepsearch_glm-0.19.1-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:cc9559218bc14e961a83bc5dfaeab01c1eea3f155ac78db3f1446cde0d6e48de"},
+    {file = "deepsearch_glm-0.19.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:fa0db3185c1ca6ec88638d112502348367229c04274560f1ea9cd1b68bb02400"},
+    {file = "deepsearch_glm-0.19.1-cp312-cp312-macosx_13_6_arm64.whl", hash = "sha256:c22da061176311edf44fce3ce19b8709ab85be41a3550f591ec9e30757a5dce0"},
+    {file = "deepsearch_glm-0.19.1-cp312-cp312-macosx_13_6_x86_64.whl", hash = "sha256:e28327cf0642f0c541ebda3533e890b1fcf8d2435e9bbb6e34214426744238a4"},
+    {file = "deepsearch_glm-0.19.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:a660b85466acec2fb0e2682e840687f87dda79ad0d4c8cc0b3dbe3f689759a13"},
+    {file = "deepsearch_glm-0.19.1-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:d9a351e18677fb0f04c399b661ca5c1227c61d970ec193a8a557cd29c4b382a8"},
+    {file = "deepsearch_glm-0.19.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:f5b9cc480b3f689b637ebcb03720c75ddea3da3973cdad17cf098be9c38db575"},
+    {file = "deepsearch_glm-0.19.1-cp38-cp38-macosx_13_6_arm64.whl", hash = "sha256:5adeb3eacd41fbf3d4c5f6f62159c41a101d8baafda5466878f221d0dfade64a"},
+    {file = "deepsearch_glm-0.19.1-cp38-cp38-macosx_13_6_x86_64.whl", hash = "sha256:039f93b37a84813e397ae5861c84acdd7d32863ebbc2426d379598eaa3a5cbfb"},
+    {file = "deepsearch_glm-0.19.1-cp38-cp38-macosx_14_0_arm64.whl", hash = "sha256:7bcf4b0e96bd5a7770750e4ae7f58c7f3032fee1606fef18d9a2e209c996aedb"},
+    {file = "deepsearch_glm-0.19.1-cp38-cp38-macosx_14_0_x86_64.whl", hash = "sha256:727208c494469c49ae240a40bb383f3f421eacee4d684d56ae63fd12d73fbca9"},
+    {file = "deepsearch_glm-0.19.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:8dcacbebb079674e028bf0f5d4fc058574d49c54855d6c13ec0eedafbab4a8e3"},
+    {file = "deepsearch_glm-0.19.1-cp39-cp39-macosx_13_6_arm64.whl", hash = "sha256:5d038657bd384ae6676b99cf40c64377992c3e512e6bda6cb0e024027828ab2f"},
+    {file = "deepsearch_glm-0.19.1-cp39-cp39-macosx_13_6_x86_64.whl", hash = "sha256:a8571ea049c5533b71f7e7f911190f6d9b5ac43bfd938393f2a7ea5d02894c41"},
+    {file = "deepsearch_glm-0.19.1-cp39-cp39-macosx_14_0_arm64.whl", hash = "sha256:8e87a68f44187c28c265c051d35b5312c918330fa2c809955bde43137267e81b"},
+    {file = "deepsearch_glm-0.19.1-cp39-cp39-macosx_14_0_x86_64.whl", hash = "sha256:02496ebbce192a0745b46bdd992d1bc41f27345a272eaedd689c3b27aaa12f63"},
+    {file = "deepsearch_glm-0.19.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9ee4ae30f4c571da6169d6dc5340758d6720c01bf7ca909f3c37c07ecf9c2076"},
 ]

 [package.dependencies]
@ -754,16 +714,6 @@ files = [
 graph = ["objgraph (>=1.7.2)"]
 profile = ["gprof2dot (>=2022.7.29)"]

-[[package]]
-name = "distance"
-version = "0.1.3"
-description = "Utilities for comparing sequences"
-optional = false
-python-versions = "*"
-files = [
-    {file = "Distance-0.1.3.tar.gz", hash = "sha256:60807584f5b6003f5c521aa73f39f51f631de3be5cccc5a1d67166fcbf0d4551"},
-]
-
 [[package]]
 name = "distlib"
 version = "0.3.8"
@ -777,13 +727,13 @@ files = [

 [[package]]
 name = "docling-core"
-version = "1.1.2"
+version = "1.1.3"
 description = "A python library to define and validate data types in Docling."
 optional = false
 python-versions = "<4.0,>=3.9"
 files = [
-    {file = "docling_core-1.1.2-py3-none-any.whl", hash = "sha256:bdff5643e3e37a24204449eee99505db0f1cf620b8e1ce4cf4b71850bf49496b"},
-    {file = "docling_core-1.1.2.tar.gz", hash = "sha256:969cde6795631a5f5f8cbb5e7ca0e4032864c1abc8fff762415a09a9b1f7146c"},
+    {file = "docling_core-1.1.3-py3-none-any.whl", hash = "sha256:7d475e1b787a831c08d6260459747f645efd6044b1a82e2ebbe49f3bcb5e1ff8"},
+    {file = "docling_core-1.1.3.tar.gz", hash = "sha256:32304d5a8ebc77f9e98acde53c42d3dfdd2ee25adb7eedbf07e70e003ff7378f"},
 ]

 [package.dependencies]
@ -797,18 +747,16 @@ tabulate = ">=0.9.0,<0.10.0"

 [[package]]
 name = "docling-ibm-models"
-version = "1.1.3"
+version = "1.1.5"
 description = "This package contains the AI models used by the Docling PDF conversion package"
 optional = false
 python-versions = "<4.0,>=3.10"
 files = [
-    {file = "docling_ibm_models-1.1.3-py3-none-any.whl", hash = "sha256:d7dfbacf5f6c63a150cb2270d67d0d6e892763af99d2a7fd434b0b240da67df3"},
-    {file = "docling_ibm_models-1.1.3.tar.gz", hash = "sha256:dd89476d152c74c1b0a9c445fe31ebca4390d9b23c568bbd175b92a0ee112e77"},
+    {file = "docling_ibm_models-1.1.5-py3-none-any.whl", hash = "sha256:657f2e5798444f8792248e221906acea2b984d1793c1261a32d50b72d25df88a"},
+    {file = "docling_ibm_models-1.1.5.tar.gz", hash = "sha256:2a2fcf557d7c33878d3dc6ae90597b65728bbc240bd410faa178d1b8edc5a070"},
 ]

 [package.dependencies]
-apted = ">=1.0.3,<2.0.0"
-Distance = ">=0.1.3,<0.2.0"
 jsonlines = ">=3.1.0,<4.0.0"
 lxml = ">=4.9.1,<5.0.0"
 mean_average_precision = ">=2021.4.26.0,<2022.0.0.0"
@ -816,41 +764,46 @@ numpy = ">=1.24.4,<2.0.0"
 onnxruntime = ">=1.16.2,<2.0.0"
 opencv-python-headless = ">=4.9.0.80,<5.0.0.0"
 Pillow = ">=10.0.0,<11.0.0"
-torch = ">=2.2.2,<3.0.0"
-torchvision = ">=0.17.2"
+torch = [
+    {version = ">=2.2.2,<3.0.0", markers = "sys_platform != \"darwin\" or platform_machine != \"x86_64\""},
+    {version = ">=2.2.2,<2.3.0", markers = "sys_platform == \"darwin\" and platform_machine == \"x86_64\""},
+]
+torchvision = [
+    {version = ">=0,<1", markers = "sys_platform != \"darwin\" or platform_machine != \"x86_64\""},
+    {version = ">=0.17.2,<0.18.0", markers = "sys_platform == \"darwin\" and platform_machine == \"x86_64\""},
+]
 tqdm = ">=4.64.0,<5.0.0"

 [[package]]
 name = "docling-parse"
-version = "1.1.1"
+version = "1.1.3"
 description = "Simple package to extract text with coordinates from programmatic PDFs"
 optional = false
 python-versions = "<4.0,>=3.9"
 files = [
-    {file = "docling_parse-1.1.1-cp310-cp310-macosx_13_6_arm64.whl", hash = "sha256:a692eb79f173cec449eb66f618a1bc3dd66d13c8948d9a975cfba533b4ac5ff5"},
-    {file = "docling_parse-1.1.1-cp310-cp310-macosx_13_6_x86_64.whl", hash = "sha256:a369c91b04852ff21fca27834f2f7db8fa024fd037f6089dd46943e3ca2d2a61"},
-    {file = "docling_parse-1.1.1-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:b57b64ea2f33cc51f26f520cb69246c3a9bd06ac8b199f3decf02f8cd875446a"},
-    {file = "docling_parse-1.1.1-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:a07ffcd3341f9609dcbb942e3e60fa7eab8fb3cb15507efae73a939a31ca8ed9"},
-    {file = "docling_parse-1.1.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3fbf402666b429a290d0a1054f713aa8ebc390b29682c471acf98e0da996164f"},
-    {file = "docling_parse-1.1.1-cp311-cp311-macosx_13_6_arm64.whl", hash = "sha256:82d5719df763bca8d13acc7c5dc006fc05140f50b80ab063307e846c9272fc5c"},
-    {file = "docling_parse-1.1.1-cp311-cp311-macosx_13_6_x86_64.whl", hash = "sha256:537cdec2abb6e24124da5cfbbf67e3a56c3d61f32bffd0f8f0323107addbb343"},
-    {file = "docling_parse-1.1.1-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:4e0f7965b5389f3c657841d1e04680899a9caf431c13e020b8c4c1bac637bc6c"},
-    {file = "docling_parse-1.1.1-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:e37a36aa1f66d44d4a47d6412a19f1ffd5f44d6d7f18b7638e3e6125d83b453a"},
-    {file = "docling_parse-1.1.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ba139bfafce7dd281d0d0551415e915bbba4ed64f0827b752f99a0e717a13cd1"},
-    {file = "docling_parse-1.1.1-cp312-cp312-macosx_13_6_arm64.whl", hash = "sha256:0d62ffc592017826d1bff6dad0c97d05129c118b0b37d724c643fed2f5c77798"},
-    {file = "docling_parse-1.1.1-cp312-cp312-macosx_13_6_x86_64.whl", hash = "sha256:d2be36904005ccf5c4d44370ecd449f4e2d4df73c98c7dc88165b11028a8b6d8"},
-    {file = "docling_parse-1.1.1-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:f8caf7d08ac96929eb59009ad397c4143ef21024829a91a19d07571f0d70d2bf"},
-    {file = "docling_parse-1.1.1-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:a96286beabe65df64bc01285ecc893fae1513f6dda39898484da0fa7fb019123"},
-    {file = "docling_parse-1.1.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:74fcbccbed154a3e3e76471273cd62daf99f736c965d05a7fa5b9f4b1b446c5f"},
-    {file = "docling_parse-1.1.1-cp39-cp39-macosx_13_6_arm64.whl", hash = "sha256:133af429a329dad2c309ef3ed7538474c89c3a81e36adc720eeb62de7fff5a07"},
-    {file = "docling_parse-1.1.1-cp39-cp39-macosx_13_6_x86_64.whl", hash = "sha256:181e7537e6118706697ffa120670b10d312ace2ae35d308d10264b4e722758a2"},
-    {file = "docling_parse-1.1.1-cp39-cp39-macosx_14_0_arm64.whl", hash = "sha256:a26745edf9d8651b4a625ebf667422292420ce31d7ba1c26bd78c8b4ea15cb53"},
-    {file = "docling_parse-1.1.1-cp39-cp39-macosx_14_0_x86_64.whl", hash = "sha256:be93d954a29d38daa9c0485ef5c0b383c1f64d4dd4a6cdf22cd9d5fd782ccc9e"},
-    {file = "docling_parse-1.1.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:64ef45fc42e1c6a4a1e03c394e25ab7ed13191ba5b4994922efee02c79c51c19"},
+    {file = "docling_parse-1.1.3-cp310-cp310-macosx_13_0_arm64.whl", hash = "sha256:443e633085a0f9c7b397f64a83b8c3d7f75e43457cb91d561661286b10c6bf11"},
+    {file = "docling_parse-1.1.3-cp310-cp310-macosx_13_0_x86_64.whl", hash = "sha256:a221b3ac4c473c21c3fc75022ca83fb41e2064cef13f7a513b099617aba2141e"},
+    {file = "docling_parse-1.1.3-cp310-cp310-macosx_14_0_arm64.whl", hash = "sha256:aa8d60d7297dccc6cd494d8643a4a9c17dfccd28725745e0416f3f2176154ff0"},
+    {file = "docling_parse-1.1.3-cp310-cp310-macosx_14_0_x86_64.whl", hash = "sha256:b653f596692476e4f37446e1f87baa910c8d9c076e9666908a11623a19f493e1"},
+    {file = "docling_parse-1.1.3-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:2c34d833e1e3d812c07bd4b3451c911d1534426056b013e422660f27bd7f5d6b"},
+    {file = "docling_parse-1.1.3-cp311-cp311-macosx_13_0_arm64.whl", hash = "sha256:b42e2d46c19f4dd8fef2b9099f6f3a1edbb1a3a6c3b06d05323ddfd8ae9edd49"},
+    {file = "docling_parse-1.1.3-cp311-cp311-macosx_13_0_x86_64.whl", hash = "sha256:b45f5ba91c98d6ead15e63e5a0c114f48c1ed382b25be8c5f7cb8b0c319509ca"},
+    {file = "docling_parse-1.1.3-cp311-cp311-macosx_14_0_arm64.whl", hash = "sha256:6be36e3fc342428734f79d3cc7ef1328972d90df7e190273b7c6eba5daf03eea"},
+    {file = "docling_parse-1.1.3-cp311-cp311-macosx_14_0_x86_64.whl", hash = "sha256:da8d4c9222ab78592197bdf80877d90aa3f7735becd15dd62368a3bd76127b05"},
+    {file = "docling_parse-1.1.3-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:d882945ba55821952b90445158abb3897f3b0ec557a3ba309daa171988614fe8"},
+    {file = "docling_parse-1.1.3-cp312-cp312-macosx_13_0_arm64.whl", hash = "sha256:1628ab9b553018ba4c4d9a95bc38b7781455d7fa4111a83ae2863b0424ed95ef"},
+    {file = "docling_parse-1.1.3-cp312-cp312-macosx_13_0_x86_64.whl", hash = "sha256:fddbae1dcc66f6d3cdb19657349cebee41dbfe7b5b4ed65aed7292f8494338a9"},
+    {file = "docling_parse-1.1.3-cp312-cp312-macosx_14_0_arm64.whl", hash = "sha256:c470f9f30247db21dd75af9c45f1ff43eda760a000c7bd6b6eb5783b4291a94a"},
+    {file = "docling_parse-1.1.3-cp312-cp312-macosx_14_0_x86_64.whl", hash = "sha256:58a3342a1eaac1d7a53a700eaea58118f3f5c8bdf1658f2aedc0841bb0fbb9aa"},
+    {file = "docling_parse-1.1.3-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:5c23124fcce85618b0c2e19c8c776ee70ccc25d4bde0df9afc3ff93f68ee8133"},
+    {file = "docling_parse-1.1.3-cp39-cp39-macosx_13_0_arm64.whl", hash = "sha256:3fb872afc001ba08b121e71526035a076e1149afe00a03c513c902483dd70fbc"},
+    {file = "docling_parse-1.1.3-cp39-cp39-macosx_13_0_x86_64.whl", hash = "sha256:16ffe03379de8adbb8d86a1304220aeca268c775afe1b8453ff7623abc90eb19"},
+    {file = "docling_parse-1.1.3-cp39-cp39-macosx_14_0_arm64.whl", hash = "sha256:4343713e5d1c31669983716947014d889bbd19cce8e277b8600f3844c2d2d721"},
+    {file = "docling_parse-1.1.3-cp39-cp39-macosx_14_0_x86_64.whl", hash = "sha256:5322ec9d8389cd563518b9e462060f7c27dba147947c0639d73d7440d4c264bb"},
+    {file = "docling_parse-1.1.3-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9d3f7c068a59d66bde2bdd6312c602f2b3a71eb33801f5cc7151ead5cda3e7a7"},
 ]

 [package.dependencies]
-cibuildwheel = ">=2.20.0,<3.0.0"
 tabulate = ">=0.9.0,<1.0.0"

 [[package]]
@ -1323,13 +1276,13 @@ license = ["ukkonen"]

 [[package]]
 name = "idna"
-version = "3.7"
+version = "3.8"
 description = "Internationalized Domain Names in Applications (IDNA)"
 optional = false
-python-versions = ">=3.5"
+python-versions = ">=3.6"
 files = [
-    {file = "idna-3.7-py3-none-any.whl", hash = "sha256:82fee1fc78add43492d3a1898bfa6d8a904cc97d8427f683ed8e798d07761aa0"},
-    {file = "idna-3.7.tar.gz", hash = "sha256:028ff3aadf0609c1fd278d8ea3089299412a7a8b9bd005dd08b9f8285bcb5cfc"},
+    {file = "idna-3.8-py3-none-any.whl", hash = "sha256:050b4e5baadcd44d760cedbd2b8e639f2ff89bbc7a5730fcc662954303377aac"},
+    {file = "idna-3.8.tar.gz", hash = "sha256:d838c2c0ed6fced7693d5e8ab8e734d5f8fda53a039c0164afb0b82e771e3603"},
 ]

 [[package]]
@ -1419,13 +1372,13 @@ files = [

 [[package]]
 name = "ipython"
-version = "8.26.0"
+version = "8.27.0"
 description = "IPython: Productive Interactive Computing"
 optional = false
 python-versions = ">=3.10"
 files = [
-    {file = "ipython-8.26.0-py3-none-any.whl", hash = "sha256:e6b347c27bdf9c32ee9d31ae85defc525755a1869f14057e900675b9e8d6e6ff"},
-    {file = "ipython-8.26.0.tar.gz", hash = "sha256:1cec0fbba8404af13facebe83d04436a7434c7400e59f47acf467c64abd0956c"},
+    {file = "ipython-8.27.0-py3-none-any.whl", hash = "sha256:f68b3cb8bde357a5d7adc9598d57e22a45dfbea19eb6b98286fa3b288c9cd55c"},
+    {file = "ipython-8.27.0.tar.gz", hash = "sha256:0b99a2dc9f15fd68692e898e5568725c6d49c527d36a9fb5960ffbdeaa82ff7e"},
 ]

 [package.dependencies]
@ -2282,38 +2235,38 @@ files = [

 [[package]]
 name = "mypy"
-version = "1.11.1"
+version = "1.11.2"
 description = "Optional static typing for Python"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "mypy-1.11.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:a32fc80b63de4b5b3e65f4be82b4cfa362a46702672aa6a0f443b4689af7008c"},
-    {file = "mypy-1.11.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c1952f5ea8a5a959b05ed5f16452fddadbaae48b5d39235ab4c3fc444d5fd411"},
-    {file = "mypy-1.11.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:e1e30dc3bfa4e157e53c1d17a0dad20f89dc433393e7702b813c10e200843b03"},
-    {file = "mypy-1.11.1-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:2c63350af88f43a66d3dfeeeb8d77af34a4f07d760b9eb3a8697f0386c7590b4"},
-    {file = "mypy-1.11.1-cp310-cp310-win_amd64.whl", hash = "sha256:a831671bad47186603872a3abc19634f3011d7f83b083762c942442d51c58d58"},
-    {file = "mypy-1.11.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:7b6343d338390bb946d449677726edf60102a1c96079b4f002dedff375953fc5"},
-    {file = "mypy-1.11.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e4fe9f4e5e521b458d8feb52547f4bade7ef8c93238dfb5bbc790d9ff2d770ca"},
-    {file = "mypy-1.11.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:886c9dbecc87b9516eff294541bf7f3655722bf22bb898ee06985cd7269898de"},
-    {file = "mypy-1.11.1-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:fca4a60e1dd9fd0193ae0067eaeeb962f2d79e0d9f0f66223a0682f26ffcc809"},
-    {file = "mypy-1.11.1-cp311-cp311-win_amd64.whl", hash = "sha256:0bd53faf56de9643336aeea1c925012837432b5faf1701ccca7fde70166ccf72"},
-    {file = "mypy-1.11.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:f39918a50f74dc5969807dcfaecafa804fa7f90c9d60506835036cc1bc891dc8"},
-    {file = "mypy-1.11.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:0bc71d1fb27a428139dd78621953effe0d208aed9857cb08d002280b0422003a"},
-    {file = "mypy-1.11.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:b868d3bcff720dd7217c383474008ddabaf048fad8d78ed948bb4b624870a417"},
-    {file = "mypy-1.11.1-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:a707ec1527ffcdd1c784d0924bf5cb15cd7f22683b919668a04d2b9c34549d2e"},
-    {file = "mypy-1.11.1-cp312-cp312-win_amd64.whl", hash = "sha256:64f4a90e3ea07f590c5bcf9029035cf0efeae5ba8be511a8caada1a4893f5525"},
-    {file = "mypy-1.11.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:749fd3213916f1751fff995fccf20c6195cae941dc968f3aaadf9bb4e430e5a2"},
-    {file = "mypy-1.11.1-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:b639dce63a0b19085213ec5fdd8cffd1d81988f47a2dec7100e93564f3e8fb3b"},
-    {file = "mypy-1.11.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:4c956b49c5d865394d62941b109728c5c596a415e9c5b2be663dd26a1ff07bc0"},
-    {file = "mypy-1.11.1-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:45df906e8b6804ef4b666af29a87ad9f5921aad091c79cc38e12198e220beabd"},
-    {file = "mypy-1.11.1-cp38-cp38-win_amd64.whl", hash = "sha256:d44be7551689d9d47b7abc27c71257adfdb53f03880841a5db15ddb22dc63edb"},
-    {file = "mypy-1.11.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:2684d3f693073ab89d76da8e3921883019ea8a3ec20fa5d8ecca6a2db4c54bbe"},
-    {file = "mypy-1.11.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:79c07eb282cb457473add5052b63925e5cc97dfab9812ee65a7c7ab5e3cb551c"},
-    {file = "mypy-1.11.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:11965c2f571ded6239977b14deebd3f4c3abd9a92398712d6da3a772974fad69"},
-    {file = "mypy-1.11.1-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:a2b43895a0f8154df6519706d9bca8280cda52d3d9d1514b2d9c3e26792a0b74"},
-    {file = "mypy-1.11.1-cp39-cp39-win_amd64.whl", hash = "sha256:1a81cf05975fd61aec5ae16501a091cfb9f605dc3e3c878c0da32f250b74760b"},
-    {file = "mypy-1.11.1-py3-none-any.whl", hash = "sha256:0624bdb940255d2dd24e829d99a13cfeb72e4e9031f9492148f410ed30bcab54"},
-    {file = "mypy-1.11.1.tar.gz", hash = "sha256:f404a0b069709f18bbdb702eb3dcfe51910602995de00bd39cea3050b5772d08"},
+    {file = "mypy-1.11.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:d42a6dd818ffce7be66cce644f1dff482f1d97c53ca70908dff0b9ddc120b77a"},
+    {file = "mypy-1.11.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:801780c56d1cdb896eacd5619a83e427ce436d86a3bdf9112527f24a66618fef"},
+    {file = "mypy-1.11.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:41ea707d036a5307ac674ea172875f40c9d55c5394f888b168033177fce47383"},
+    {file = "mypy-1.11.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:6e658bd2d20565ea86da7d91331b0eed6d2eee22dc031579e6297f3e12c758c8"},
+    {file = "mypy-1.11.2-cp310-cp310-win_amd64.whl", hash = "sha256:478db5f5036817fe45adb7332d927daa62417159d49783041338921dcf646fc7"},
+    {file = "mypy-1.11.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:75746e06d5fa1e91bfd5432448d00d34593b52e7e91a187d981d08d1f33d4385"},
+    {file = "mypy-1.11.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:a976775ab2256aadc6add633d44f100a2517d2388906ec4f13231fafbb0eccca"},
+    {file = "mypy-1.11.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:cd953f221ac1379050a8a646585a29574488974f79d8082cedef62744f0a0104"},
+    {file = "mypy-1.11.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:57555a7715c0a34421013144a33d280e73c08df70f3a18a552938587ce9274f4"},
+    {file = "mypy-1.11.2-cp311-cp311-win_amd64.whl", hash = "sha256:36383a4fcbad95f2657642a07ba22ff797de26277158f1cc7bd234821468b1b6"},
+    {file = "mypy-1.11.2-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:e8960dbbbf36906c5c0b7f4fbf2f0c7ffb20f4898e6a879fcf56a41a08b0d318"},
+    {file = "mypy-1.11.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:06d26c277962f3fb50e13044674aa10553981ae514288cb7d0a738f495550b36"},
+    {file = "mypy-1.11.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:6e7184632d89d677973a14d00ae4d03214c8bc301ceefcdaf5c474866814c987"},
+    {file = "mypy-1.11.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:3a66169b92452f72117e2da3a576087025449018afc2d8e9bfe5ffab865709ca"},
+    {file = "mypy-1.11.2-cp312-cp312-win_amd64.whl", hash = "sha256:969ea3ef09617aff826885a22ece0ddef69d95852cdad2f60c8bb06bf1f71f70"},
+    {file = "mypy-1.11.2-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:37c7fa6121c1cdfcaac97ce3d3b5588e847aa79b580c1e922bb5d5d2902df19b"},
+    {file = "mypy-1.11.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:4a8a53bc3ffbd161b5b2a4fff2f0f1e23a33b0168f1c0778ec70e1a3d66deb86"},
+    {file = "mypy-1.11.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:2ff93107f01968ed834f4256bc1fc4475e2fecf6c661260066a985b52741ddce"},
+    {file = "mypy-1.11.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:edb91dded4df17eae4537668b23f0ff6baf3707683734b6a818d5b9d0c0c31a1"},
+    {file = "mypy-1.11.2-cp38-cp38-win_amd64.whl", hash = "sha256:ee23de8530d99b6db0573c4ef4bd8f39a2a6f9b60655bf7a1357e585a3486f2b"},
+    {file = "mypy-1.11.2-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:801ca29f43d5acce85f8e999b1e431fb479cb02d0e11deb7d2abb56bdaf24fd6"},
+    {file = "mypy-1.11.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:af8d155170fcf87a2afb55b35dc1a0ac21df4431e7d96717621962e4b9192e70"},
+    {file = "mypy-1.11.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.manylinux_2_28_x86_64.whl", hash = "sha256:f7821776e5c4286b6a13138cc935e2e9b6fde05e081bdebf5cdb2bb97c9df81d"},
+    {file = "mypy-1.11.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:539c570477a96a4e6fb718b8d5c3e0c0eba1f485df13f86d2970c91f0673148d"},
+    {file = "mypy-1.11.2-cp39-cp39-win_amd64.whl", hash = "sha256:3f14cd3d386ac4d05c5a39a51b84387403dadbd936e17cb35882134d4f8f0d24"},
+    {file = "mypy-1.11.2-py3-none-any.whl", hash = "sha256:b499bc07dbdcd3de92b0a8b29fdf592c111276f6a12fe29c30f6c417dd546d12"},
+    {file = "mypy-1.11.2.tar.gz", hash = "sha256:7f9993ad3e0ffdc95c2a14b66dee63729f021968bff8ad911867579c65d13a79"},
 ]

 [package.dependencies]
@ -2534,6 +2487,19 @@ files = [
    {file = "nvidia_cuda_runtime_cu12-12.1.105-py3-none-win_amd64.whl", hash = "sha256:dfb46ef84d73fababab44cf03e3b83f80700d27ca300e537f85f636fac474344"},
 ]

+[[package]]
+name = "nvidia-cudnn-cu12"
+version = "8.9.2.26"
+description = "cuDNN runtime libraries"
+optional = false
+python-versions = ">=3"
+files = [
+    {file = "nvidia_cudnn_cu12-8.9.2.26-py3-none-manylinux1_x86_64.whl", hash = "sha256:5ccb288774fdfb07a7e7025ffec286971c06d8d7b4fb162525334616d7629ff9"},
+]
+
+[package.dependencies]
+nvidia-cublas-cu12 = "*"
+
 [[package]]
 name = "nvidia-cudnn-cu12"
 version = "9.1.0.70"
@ -2600,6 +2566,16 @@ files = [
 [package.dependencies]
 nvidia-nvjitlink-cu12 = "*"

+[[package]]
+name = "nvidia-nccl-cu12"
+version = "2.19.3"
+description = "NVIDIA Collective Communication Library (NCCL) Runtime"
+optional = false
+python-versions = ">=3"
+files = [
+    {file = "nvidia_nccl_cu12-2.19.3-py3-none-manylinux1_x86_64.whl", hash = "sha256:a9734707a2c96443331c1e48c717024aa6678a0e2a4cb66b2c364d18cee6b48d"},
+]
+
 [[package]]
 name = "nvidia-nccl-cu12"
 version = "2.20.5"
@ -2613,14 +2589,14 @@ files = [

 [[package]]
 name = "nvidia-nvjitlink-cu12"
-version = "12.6.20"
+version = "12.6.68"
 description = "Nvidia JIT LTO Library"
 optional = false
 python-versions = ">=3"
 files = [
-    {file = "nvidia_nvjitlink_cu12-12.6.20-py3-none-manylinux2014_aarch64.whl", hash = "sha256:84fb38465a5bc7c70cbc320cfd0963eb302ee25a5e939e9f512bbba55b6072fb"},
-    {file = "nvidia_nvjitlink_cu12-12.6.20-py3-none-manylinux2014_x86_64.whl", hash = "sha256:562ab97ea2c23164823b2a89cb328d01d45cb99634b8c65fe7cd60d14562bd79"},
-    {file = "nvidia_nvjitlink_cu12-12.6.20-py3-none-win_amd64.whl", hash = "sha256:ed3c43a17f37b0c922a919203d2d36cbef24d41cc3e6b625182f8b58203644f6"},
+    {file = "nvidia_nvjitlink_cu12-12.6.68-py3-none-manylinux2014_aarch64.whl", hash = "sha256:b3fd0779845f68b92063ab1393abab1ed0a23412fc520df79a8190d098b5cd6b"},
+    {file = "nvidia_nvjitlink_cu12-12.6.68-py3-none-manylinux2014_x86_64.whl", hash = "sha256:125a6c2a44e96386dda634e13d944e60b07a0402d391a070e8fb4104b34ea1ab"},
+    {file = "nvidia_nvjitlink_cu12-12.6.68-py3-none-win_amd64.whl", hash = "sha256:a55744c98d70317c5e23db14866a8cc2b733f7324509e941fc96276f9f37801d"},
 ]

 [[package]]
@ -2694,9 +2670,9 @@ files = [

 [package.dependencies]
 numpy = [
+    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
    {version = ">=1.21.4", markers = "python_version >= \"3.10\" and platform_system == \"Darwin\" and python_version < \"3.11\""},
    {version = ">=1.21.2", markers = "platform_system != \"Darwin\" and python_version >= \"3.10\" and python_version < \"3.11\""},
-    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
    {version = ">=1.23.5", markers = "python_version >= \"3.11\" and python_version < \"3.12\""},
 ]

@ -2751,8 +2727,8 @@ files = [

 [package.dependencies]
 numpy = [
-    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
    {version = ">=1.26.0", markers = "python_version >= \"3.12\""},
+    {version = ">=1.22.4", markers = "python_version < \"3.11\""},
    {version = ">=1.23.2", markers = "python_version == \"3.11\""},
 ]
 python-dateutil = ">=2.8.2"
@ -3061,22 +3037,22 @@ wcwidth = "*"

 [[package]]
 name = "protobuf"
-version = "5.27.3"
+version = "5.28.0"
 description = ""
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "protobuf-5.27.3-cp310-abi3-win32.whl", hash = "sha256:dcb307cd4ef8fec0cf52cb9105a03d06fbb5275ce6d84a6ae33bc6cf84e0a07b"},
-    {file = "protobuf-5.27.3-cp310-abi3-win_amd64.whl", hash = "sha256:16ddf3f8c6c41e1e803da7abea17b1793a97ef079a912e42351eabb19b2cffe7"},
-    {file = "protobuf-5.27.3-cp38-abi3-macosx_10_9_universal2.whl", hash = "sha256:68248c60d53f6168f565a8c76dc58ba4fa2ade31c2d1ebdae6d80f969cdc2d4f"},
-    {file = "protobuf-5.27.3-cp38-abi3-manylinux2014_aarch64.whl", hash = "sha256:b8a994fb3d1c11156e7d1e427186662b64694a62b55936b2b9348f0a7c6625ce"},
-    {file = "protobuf-5.27.3-cp38-abi3-manylinux2014_x86_64.whl", hash = "sha256:a55c48f2a2092d8e213bd143474df33a6ae751b781dd1d1f4d953c128a415b25"},
-    {file = "protobuf-5.27.3-cp38-cp38-win32.whl", hash = "sha256:043853dcb55cc262bf2e116215ad43fa0859caab79bb0b2d31b708f128ece035"},
-    {file = "protobuf-5.27.3-cp38-cp38-win_amd64.whl", hash = "sha256:c2a105c24f08b1e53d6c7ffe69cb09d0031512f0b72f812dd4005b8112dbe91e"},
-    {file = "protobuf-5.27.3-cp39-cp39-win32.whl", hash = "sha256:c84eee2c71ed83704f1afbf1a85c3171eab0fd1ade3b399b3fad0884cbcca8bf"},
-    {file = "protobuf-5.27.3-cp39-cp39-win_amd64.whl", hash = "sha256:af7c0b7cfbbb649ad26132e53faa348580f844d9ca46fd3ec7ca48a1ea5db8a1"},
-    {file = "protobuf-5.27.3-py3-none-any.whl", hash = "sha256:8572c6533e544ebf6899c360e91d6bcbbee2549251643d32c52cf8a5de295ba5"},
-    {file = "protobuf-5.27.3.tar.gz", hash = "sha256:82460903e640f2b7e34ee81a947fdaad89de796d324bcbc38ff5430bcdead82c"},
+    {file = "protobuf-5.28.0-cp310-abi3-win32.whl", hash = "sha256:66c3edeedb774a3508ae70d87b3a19786445fe9a068dd3585e0cefa8a77b83d0"},
+    {file = "protobuf-5.28.0-cp310-abi3-win_amd64.whl", hash = "sha256:6d7cc9e60f976cf3e873acb9a40fed04afb5d224608ed5c1a105db4a3f09c5b6"},
+    {file = "protobuf-5.28.0-cp38-abi3-macosx_10_9_universal2.whl", hash = "sha256:532627e8fdd825cf8767a2d2b94d77e874d5ddb0adefb04b237f7cc296748681"},
+    {file = "protobuf-5.28.0-cp38-abi3-manylinux2014_aarch64.whl", hash = "sha256:018db9056b9d75eb93d12a9d35120f97a84d9a919bcab11ed56ad2d399d6e8dd"},
+    {file = "protobuf-5.28.0-cp38-abi3-manylinux2014_x86_64.whl", hash = "sha256:6206afcb2d90181ae8722798dcb56dc76675ab67458ac24c0dd7d75d632ac9bd"},
+    {file = "protobuf-5.28.0-cp38-cp38-win32.whl", hash = "sha256:eef7a8a2f4318e2cb2dee8666d26e58eaf437c14788f3a2911d0c3da40405ae8"},
+    {file = "protobuf-5.28.0-cp38-cp38-win_amd64.whl", hash = "sha256:d001a73c8bc2bf5b5c1360d59dd7573744e163b3607fa92788b7f3d5fefbd9a5"},
+    {file = "protobuf-5.28.0-cp39-cp39-win32.whl", hash = "sha256:dde9fcaa24e7a9654f4baf2a55250b13a5ea701493d904c54069776b99a8216b"},
+    {file = "protobuf-5.28.0-cp39-cp39-win_amd64.whl", hash = "sha256:853db610214e77ee817ecf0514e0d1d052dff7f63a0c157aa6eabae98db8a8de"},
+    {file = "protobuf-5.28.0-py3-none-any.whl", hash = "sha256:510ed78cd0980f6d3218099e874714cdf0d8a95582e7b059b06cabad855ed0a0"},
+    {file = "protobuf-5.28.0.tar.gz", hash = "sha256:dde74af0fa774fa98892209992295adbfb91da3fa98c8f67a88afe8f5a349add"},
 ]

 [[package]]
@ -3208,8 +3184,8 @@ files = [
 annotated-types = ">=0.4.0"
 pydantic-core = "2.20.1"
 typing-extensions = [
-    {version = ">=4.6.1", markers = "python_version < \"3.13\""},
    {version = ">=4.12.2", markers = "python_version >= \"3.13\""},
+    {version = ">=4.6.1", markers = "python_version < \"3.13\""},
 ]

 [package.extras]
@ -3376,8 +3352,8 @@ files = [
 astroid = ">=2.15.8,<=2.17.0-dev0"
 colorama = {version = ">=0.4.5", markers = "sys_platform == \"win32\""}
 dill = [
-    {version = ">=0.2", markers = "python_version < \"3.11\""},
    {version = ">=0.3.6", markers = "python_version >= \"3.11\""},
+    {version = ">=0.2", markers = "python_version < \"3.11\""},
 ]
 isort = ">=4.2.5,<6"
 mccabe = ">=0.6,<0.8"
@ -3391,13 +3367,13 @@ testutils = ["gitpython (>3)"]

 [[package]]
 name = "pyparsing"
-version = "3.1.2"
+version = "3.1.4"
 description = "pyparsing module - Classes and methods to define and execute parsing grammars"
 optional = false
 python-versions = ">=3.6.8"
 files = [
-    {file = "pyparsing-3.1.2-py3-none-any.whl", hash = "sha256:f9db75911801ed778fe61bb643079ff86601aca99fcae6345aa67292038fb742"},
-    {file = "pyparsing-3.1.2.tar.gz", hash = "sha256:a1bac0ce561155ecc3ed78ca94d3c9378656ad4c94c1270de543f621420f94ad"},
+    {file = "pyparsing-3.1.4-py3-none-any.whl", hash = "sha256:a6a7ee4235a3f944aa1fa2249307708f893fe5717dc603503c6c7969c070fb7c"},
+    {file = "pyparsing-3.1.4.tar.gz", hash = "sha256:f86ec8d1a83f11977c9a6ea7598e8c27fc5cddfa5b07ea2241edbbde1d7bc032"},
 ]

 [package.extras]
@ -4017,13 +3993,13 @@ idna2008 = ["idna"]

 [[package]]
 name = "rich"
-version = "13.7.1"
+version = "13.8.0"
 description = "Render rich text, tables, progress bars, syntax highlighting, markdown and more to the terminal"
 optional = false
 python-versions = ">=3.7.0"
 files = [
-    {file = "rich-13.7.1-py3-none-any.whl", hash = "sha256:4edbae314f59eb482f54e9e30bf00d33350aaa94f4bfcd4e9e3110e64d0d7222"},
-    {file = "rich-13.7.1.tar.gz", hash = "sha256:9be308cb1fe2f1f57d67ce99e95af38a1e2bc71ad9813b0e247cf7ffbcc3a432"},
+    {file = "rich-13.8.0-py3-none-any.whl", hash = "sha256:2e85306a063b9492dffc86278197a60cbece75bcb766022f3436f567cae11bdc"},
+    {file = "rich-13.8.0.tar.gz", hash = "sha256:a5ac1f1cd448ade0d59cc3356f7db7a7ccda2c8cbae9c7a90c28ff463d3e91f4"},
 ]

 [package.dependencies]
@ -4290,19 +4266,23 @@ files = [

 [[package]]
 name = "setuptools"
-version = "73.0.1"
+version = "74.0.0"
 description = "Easily download, build, install, upgrade, and uninstall Python packages"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "setuptools-73.0.1-py3-none-any.whl", hash = "sha256:b208925fcb9f7af924ed2dc04708ea89791e24bde0d3020b27df0e116088b34e"},
-    {file = "setuptools-73.0.1.tar.gz", hash = "sha256:d59a3e788ab7e012ab2c4baed1b376da6366883ee20d7a5fc426816e3d7b1193"},
+    {file = "setuptools-74.0.0-py3-none-any.whl", hash = "sha256:0274581a0037b638b9fc1c6883cc71c0210865aaa76073f7882376b641b84e8f"},
+    {file = "setuptools-74.0.0.tar.gz", hash = "sha256:a85e96b8be2b906f3e3e789adec6a9323abf79758ecfa3065bd740d81158b11e"},
 ]

 [package.extras]
+check = ["pytest-checkdocs (>=2.4)", "pytest-ruff (>=0.2.1)", "ruff (>=0.5.2)"]
 core = ["importlib-metadata (>=6)", "importlib-resources (>=5.10.2)", "jaraco.text (>=3.7)", "more-itertools (>=8.8)", "packaging (>=24)", "platformdirs (>=2.6.2)", "tomli (>=2.0.1)", "wheel (>=0.43.0)"]
+cover = ["pytest-cov"]
 doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "pygments-github-lexers (==0.0.5)", "pyproject-hooks (!=1.1)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-favicon", "sphinx-inline-tabs", "sphinx-lint", "sphinx-notfound-page (>=1,<2)", "sphinx-reredirects", "sphinxcontrib-towncrier", "towncrier (<24.7)"]
-test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "importlib-metadata", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "jaraco.test", "mypy (==1.11.*)", "packaging (>=23.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-home (>=0.5)", "pytest-mypy", "pytest-perf", "pytest-ruff (<0.4)", "pytest-ruff (>=0.2.1)", "pytest-ruff (>=0.3.2)", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel (>=0.44.0)"]
+enabler = ["pytest-enabler (>=2.2)"]
+test = ["build[virtualenv] (>=1.0.3)", "filelock (>=3.4.0)", "ini2toml[lite] (>=0.14)", "jaraco.develop (>=7.21)", "jaraco.envs (>=2.2)", "jaraco.path (>=3.2.0)", "jaraco.test", "packaging (>=23.2)", "pip (>=19.1)", "pyproject-hooks (!=1.1)", "pytest (>=6,!=8.1.*)", "pytest-home (>=0.5)", "pytest-perf", "pytest-subprocess", "pytest-timeout", "pytest-xdist (>=3)", "tomli-w (>=1.0.0)", "virtualenv (>=13.0.0)", "wheel (>=0.44.0)"]
+type = ["importlib-metadata (>=7.0.2)", "jaraco.develop (>=7.21)", "mypy (==1.11.*)", "pytest-mypy"]

 [[package]]
 name = "shapely"
@ -4577,13 +4557,13 @@ files = [

 [[package]]
 name = "tifffile"
-version = "2024.8.10"
+version = "2024.8.28"
 description = "Read and write TIFF files"
 optional = false
 python-versions = ">=3.9"
 files = [
-    {file = "tifffile-2024.8.10-py3-none-any.whl", hash = "sha256:1c224564fa92e7e9f9a0ed65880b2ece97c3f0d10029ffbebfa5e62b3f6b343d"},
-    {file = "tifffile-2024.8.10.tar.gz", hash = "sha256:fdc12124f1478a07b1524641dc6b50cf6bde0483011a63fd2a773094090c3dcf"},
+    {file = "tifffile-2024.8.28-py3-none-any.whl", hash = "sha256:7b618100c9043f2b7f196287ae5e9d139a6f4a0fa603d3698fd5b8f679ca2c66"},
+    {file = "tifffile-2024.8.28.tar.gz", hash = "sha256:a84056c02865d20090235816cddabc99a34493e870effcd119499548583f3204"},
 ]

 [package.dependencies]
@ -4641,6 +4621,64 @@ files = [
    {file = "tomlkit-0.13.2.tar.gz", hash = "sha256:fff5fe59a87295b278abd31bec92c15d9bc4a06885ab12bcea52c71119392e79"},
 ]

+[[package]]
+name = "torch"
+version = "2.2.2"
+description = "Tensors and Dynamic neural networks in Python with strong GPU acceleration"
+optional = false
+python-versions = ">=3.8.0"
+files = [
+    {file = "torch-2.2.2-cp310-cp310-manylinux1_x86_64.whl", hash = "sha256:bc889d311a855dd2dfd164daf8cc903a6b7273a747189cebafdd89106e4ad585"},
+    {file = "torch-2.2.2-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:15dffa4cc3261fa73d02f0ed25f5fa49ecc9e12bf1ae0a4c1e7a88bbfaad9030"},
+    {file = "torch-2.2.2-cp310-cp310-win_amd64.whl", hash = "sha256:11e8fe261233aeabd67696d6b993eeb0896faa175c6b41b9a6c9f0334bdad1c5"},
+    {file = "torch-2.2.2-cp310-none-macosx_10_9_x86_64.whl", hash = "sha256:b2e2200b245bd9f263a0d41b6a2dab69c4aca635a01b30cca78064b0ef5b109e"},
+    {file = "torch-2.2.2-cp310-none-macosx_11_0_arm64.whl", hash = "sha256:877b3e6593b5e00b35bbe111b7057464e76a7dd186a287280d941b564b0563c2"},
+    {file = "torch-2.2.2-cp311-cp311-manylinux1_x86_64.whl", hash = "sha256:ad4c03b786e074f46606f4151c0a1e3740268bcf29fbd2fdf6666d66341c1dcb"},
+    {file = "torch-2.2.2-cp311-cp311-manylinux2014_aarch64.whl", hash = "sha256:32827fa1fbe5da8851686256b4cd94cc7b11be962862c2293811c94eea9457bf"},
+    {file = "torch-2.2.2-cp311-cp311-win_amd64.whl", hash = "sha256:f9ef0a648310435511e76905f9b89612e45ef2c8b023bee294f5e6f7e73a3e7c"},
+    {file = "torch-2.2.2-cp311-none-macosx_10_9_x86_64.whl", hash = "sha256:95b9b44f3bcebd8b6cd8d37ec802048c872d9c567ba52c894bba90863a439059"},
+    {file = "torch-2.2.2-cp311-none-macosx_11_0_arm64.whl", hash = "sha256:49aa4126ede714c5aeef7ae92969b4b0bbe67f19665106463c39f22e0a1860d1"},
+    {file = "torch-2.2.2-cp312-cp312-manylinux1_x86_64.whl", hash = "sha256:cf12cdb66c9c940227ad647bc9cf5dba7e8640772ae10dfe7569a0c1e2a28aca"},
+    {file = "torch-2.2.2-cp312-cp312-manylinux2014_aarch64.whl", hash = "sha256:89ddac2a8c1fb6569b90890955de0c34e1724f87431cacff4c1979b5f769203c"},
+    {file = "torch-2.2.2-cp312-cp312-win_amd64.whl", hash = "sha256:451331406b760f4b1ab298ddd536486ab3cfb1312614cfe0532133535be60bea"},
+    {file = "torch-2.2.2-cp312-none-macosx_10_9_x86_64.whl", hash = "sha256:eb4d6e9d3663e26cd27dc3ad266b34445a16b54908e74725adb241aa56987533"},
+    {file = "torch-2.2.2-cp312-none-macosx_11_0_arm64.whl", hash = "sha256:bf9558da7d2bf7463390b3b2a61a6a3dbb0b45b161ee1dd5ec640bf579d479fc"},
+    {file = "torch-2.2.2-cp38-cp38-manylinux1_x86_64.whl", hash = "sha256:cd2bf7697c9e95fb5d97cc1d525486d8cf11a084c6af1345c2c2c22a6b0029d0"},
+    {file = "torch-2.2.2-cp38-cp38-manylinux2014_aarch64.whl", hash = "sha256:b421448d194496e1114d87a8b8d6506bce949544e513742b097e2ab8f7efef32"},
+    {file = "torch-2.2.2-cp38-cp38-win_amd64.whl", hash = "sha256:3dbcd563a9b792161640c0cffe17e3270d85e8f4243b1f1ed19cca43d28d235b"},
+    {file = "torch-2.2.2-cp38-none-macosx_10_9_x86_64.whl", hash = "sha256:31f4310210e7dda49f1fb52b0ec9e59382cfcb938693f6d5378f25b43d7c1d29"},
+    {file = "torch-2.2.2-cp38-none-macosx_11_0_arm64.whl", hash = "sha256:c795feb7e8ce2e0ef63f75f8e1ab52e7fd5e1a4d7d0c31367ade1e3de35c9e95"},
+    {file = "torch-2.2.2-cp39-cp39-manylinux1_x86_64.whl", hash = "sha256:a6e5770d68158d07456bfcb5318b173886f579fdfbf747543901ce718ea94782"},
+    {file = "torch-2.2.2-cp39-cp39-manylinux2014_aarch64.whl", hash = "sha256:67dcd726edff108e2cd6c51ff0e416fd260c869904de95750e80051358680d24"},
+    {file = "torch-2.2.2-cp39-cp39-win_amd64.whl", hash = "sha256:539d5ef6c4ce15bd3bd47a7b4a6e7c10d49d4d21c0baaa87c7d2ef8698632dfb"},
+    {file = "torch-2.2.2-cp39-none-macosx_10_9_x86_64.whl", hash = "sha256:dff696de90d6f6d1e8200e9892861fd4677306d0ef604cb18f2134186f719f82"},
+    {file = "torch-2.2.2-cp39-none-macosx_11_0_arm64.whl", hash = "sha256:3a4dd910663fd7a124c056c878a52c2b0be4a5a424188058fe97109d4436ee42"},
+]
+
+[package.dependencies]
+filelock = "*"
+fsspec = "*"
+jinja2 = "*"
+networkx = "*"
+nvidia-cublas-cu12 = {version = "12.1.3.1", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-cupti-cu12 = {version = "12.1.105", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-nvrtc-cu12 = {version = "12.1.105", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cuda-runtime-cu12 = {version = "12.1.105", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cudnn-cu12 = {version = "8.9.2.26", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cufft-cu12 = {version = "11.0.2.54", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-curand-cu12 = {version = "10.3.2.106", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cusolver-cu12 = {version = "11.4.5.107", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-cusparse-cu12 = {version = "12.1.0.106", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-nccl-cu12 = {version = "2.19.3", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+nvidia-nvtx-cu12 = {version = "12.1.105", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\""}
+sympy = "*"
+triton = {version = "2.2.0", markers = "platform_system == \"Linux\" and platform_machine == \"x86_64\" and python_version < \"3.12\""}
+typing-extensions = ">=4.8.0"
+
+[package.extras]
+opt-einsum = ["opt-einsum (>=3.3)"]
+optree = ["optree (>=0.9.1)"]
+
 [[package]]
 name = "torch"
 version = "2.4.0"
@ -4694,6 +4732,48 @@ typing-extensions = ">=4.8.0"
 opt-einsum = ["opt-einsum (>=3.3)"]
 optree = ["optree (>=0.11.0)"]

+[[package]]
+name = "torchvision"
+version = "0.17.2"
+description = "image and video datasets and models for torch deep learning"
+optional = false
+python-versions = ">=3.8"
+files = [
+    {file = "torchvision-0.17.2-cp310-cp310-macosx_10_13_x86_64.whl", hash = "sha256:1f2910fe3c21ad6875b2720d46fad835b2e4b336e9553d31ca364d24c90b1d4f"},
+    {file = "torchvision-0.17.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:ecc1c503fa8a54fbab777e06a7c228032b8ab78efebf35b28bc8f22f544f51f1"},
+    {file = "torchvision-0.17.2-cp310-cp310-manylinux1_x86_64.whl", hash = "sha256:f400145fc108833e7c2fc28486a04989ca742146d7a2a2cc48878ebbb40cdbbd"},
+    {file = "torchvision-0.17.2-cp310-cp310-manylinux2014_aarch64.whl", hash = "sha256:e9e4bed404af33dfc92eecc2b513d21ddc4c242a7fd8708b3b09d3a26aa6f444"},
+    {file = "torchvision-0.17.2-cp310-cp310-win_amd64.whl", hash = "sha256:ba2e62f233eab3d42b648c122a3a29c47cc108ca314dfd5cbb59cd3a143fd623"},
+    {file = "torchvision-0.17.2-cp311-cp311-macosx_10_13_x86_64.whl", hash = "sha256:9b83e55ee7d0a1704f52b9c0ac87388e7a6d1d98a6bde7b0b35f9ab54d7bda54"},
+    {file = "torchvision-0.17.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:e031004a1bc432c980a7bd642f6c189a3efc316e423fc30b5569837166a4e28d"},
+    {file = "torchvision-0.17.2-cp311-cp311-manylinux1_x86_64.whl", hash = "sha256:3bbc24b7713e8f22766992562547d8b4b10001208d372fe599255af84bfd1a69"},
+    {file = "torchvision-0.17.2-cp311-cp311-manylinux2014_aarch64.whl", hash = "sha256:833fd2e4216ced924c8aca0525733fe727f9a1af66dfad7c5be7257e97c39678"},
+    {file = "torchvision-0.17.2-cp311-cp311-win_amd64.whl", hash = "sha256:6835897df852fad1015e6a106c167c83848114cbcc7d86112384a973404e4431"},
+    {file = "torchvision-0.17.2-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:14fd1d4a033c325bdba2d03a69c3450cab6d3a625f85cc375781d9237ca5d04d"},
+    {file = "torchvision-0.17.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:9c3acbebbe379af112b62b535820174277b1f3eed30df264a4e458d58ee4e5b2"},
+    {file = "torchvision-0.17.2-cp312-cp312-manylinux1_x86_64.whl", hash = "sha256:77d680adf6ce367166a186d2c7fda3a73807ab9a03b2c31a03fa8812c8c5335b"},
+    {file = "torchvision-0.17.2-cp312-cp312-manylinux2014_aarch64.whl", hash = "sha256:f1c9ab3152cfb27f83aca072cac93a3a4c4e4ab0261cf0f2d516b9868a4e96f3"},
+    {file = "torchvision-0.17.2-cp312-cp312-win_amd64.whl", hash = "sha256:3f784381419f3ed3f2ec2aa42fb4aeec5bf4135e298d1631e41c926e6f1a0dff"},
+    {file = "torchvision-0.17.2-cp38-cp38-macosx_10_13_x86_64.whl", hash = "sha256:b83aac8d78f48981146d582168d75b6c947cfb0a7693f76e219f1926f6e595a3"},
+    {file = "torchvision-0.17.2-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:1ece40557e122d79975860a005aa7e2a9e2e6c350a03e78a00ec1450083312fd"},
+    {file = "torchvision-0.17.2-cp38-cp38-manylinux1_x86_64.whl", hash = "sha256:32dbeba3987e20f2dc1bce8d1504139fff582898346dfe8ad98d649f97ca78fa"},
+    {file = "torchvision-0.17.2-cp38-cp38-manylinux2014_aarch64.whl", hash = "sha256:35ba5c1600c3203549d2316422a659bd20c0cfda1b6085eec94fb9f35f55ca43"},
+    {file = "torchvision-0.17.2-cp38-cp38-win_amd64.whl", hash = "sha256:2f69570f50b1d195e51bc03feffb7b7728207bc36efcfb1f0813712b2379d881"},
+    {file = "torchvision-0.17.2-cp39-cp39-macosx_10_13_x86_64.whl", hash = "sha256:4868bbfa55758c8107e69a0e7dd5e77b89056035cd38b767ad5b98cdb71c0f0d"},
+    {file = "torchvision-0.17.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:efd6d0dd0668e15d01a2cffadc74068433b32cbcf5692e0c4aa15fc5cb250ce7"},
+    {file = "torchvision-0.17.2-cp39-cp39-manylinux1_x86_64.whl", hash = "sha256:7dc85b397f6c6d9ef12716ce0d6e11ac2b803f5cccff6fe3966db248e7774478"},
+    {file = "torchvision-0.17.2-cp39-cp39-manylinux2014_aarch64.whl", hash = "sha256:d506854c5acd69b20a8b6641f01fe841685a21c5406b56813184f1c9fc94279e"},
+    {file = "torchvision-0.17.2-cp39-cp39-win_amd64.whl", hash = "sha256:067095e87a020a7a251ac1d38483aa591c5ccb81e815527c54db88a982fc9267"},
+]
+
+[package.dependencies]
+numpy = "*"
+pillow = ">=5.3.0,<8.3.dev0 || >=8.4.dev0"
+torch = "2.2.2"
+
+[package.extras]
+scipy = ["scipy"]
+
 [[package]]
 name = "torchvision"
 version = "0.19.0"
@ -4772,6 +4852,29 @@ files = [
 docs = ["myst-parser", "pydata-sphinx-theme", "sphinx"]
 test = ["argcomplete (>=3.0.3)", "mypy (>=1.7.0)", "pre-commit", "pytest (>=7.0,<8.2)", "pytest-mock", "pytest-mypy-testing"]

+[[package]]
+name = "triton"
+version = "2.2.0"
+description = "A language and compiler for custom Deep Learning operations"
+optional = false
+python-versions = "*"
+files = [
+    {file = "triton-2.2.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a2294514340cfe4e8f4f9e5c66c702744c4a117d25e618bd08469d0bfed1e2e5"},
+    {file = "triton-2.2.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:da58a152bddb62cafa9a857dd2bc1f886dbf9f9c90a2b5da82157cd2b34392b0"},
+    {file = "triton-2.2.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0af58716e721460a61886668b205963dc4d1e4ac20508cc3f623aef0d70283d5"},
+    {file = "triton-2.2.0-cp37-cp37m-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e8fe46d3ab94a8103e291bd44c741cc294b91d1d81c1a2888254cbf7ff846dab"},
+    {file = "triton-2.2.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b8ce26093e539d727e7cf6f6f0d932b1ab0574dc02567e684377630d86723ace"},
+    {file = "triton-2.2.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:227cc6f357c5efcb357f3867ac2a8e7ecea2298cd4606a8ba1e931d1d5a947df"},
+]
+
+[package.dependencies]
+filelock = "*"
+
+[package.extras]
+build = ["cmake (>=3.20)", "lit"]
+tests = ["autopep8", "flake8", "isort", "numpy", "pytest", "scipy (>=1.7.1)", "torch"]
+tutorials = ["matplotlib", "pandas", "tabulate", "torch"]
+
 [[package]]
 name = "triton"
 version = "3.0.0"
@ -4784,11 +4887,6 @@ files = [
    {file = "triton-3.0.0-1-cp312-cp312-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:34e509deb77f1c067d8640725ef00c5cbfcb2052a1a3cb6a6d343841f92624eb"},
    {file = "triton-3.0.0-1-cp38-cp38-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:bcbf3b1c48af6a28011a5c40a5b3b9b5330530c3827716b5fbf6d7adcc1e53e9"},
    {file = "triton-3.0.0-1-cp39-cp39-manylinux2014_x86_64.manylinux_2_17_x86_64.whl", hash = "sha256:6e5727202f7078c56f91ff13ad0c1abab14a0e7f2c87e91b12b6f64f3e8ae609"},
-    {file = "triton-3.0.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:39b052da883351fdf6be3d93cedae6db3b8e3988d3b09ed221bccecfa9612230"},
-    {file = "triton-3.0.0-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:cd34f19a8582af96e6291d4afce25dac08cb2a5d218c599163761e8e0827208e"},
-    {file = "triton-3.0.0-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:0d5e10de8c011adeb7c878c6ce0dd6073b14367749e34467f1cff2bde1b78253"},
-    {file = "triton-3.0.0-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:e8903767951bf86ec960b4fe4e21bc970055afc65e9d57e916d79ae3c93665e3"},
-    {file = "triton-3.0.0-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:41004fb1ae9a53fcb3e970745feb87f0e3c94c6ce1ba86e95fa3b8537894bef7"},
 ]

 [package.dependencies]
@ -4835,13 +4933,13 @@ urllib3 = ">=1.26.0"

 [[package]]
 name = "typer"
-version = "0.12.4"
+version = "0.12.5"
 description = "Typer, build great CLIs. Easy to code. Based on Python type hints."
 optional = false
 python-versions = ">=3.7"
 files = [
-    {file = "typer-0.12.4-py3-none-any.whl", hash = "sha256:819aa03699f438397e876aa12b0d63766864ecba1b579092cc9fe35d886e34b6"},
-    {file = "typer-0.12.4.tar.gz", hash = "sha256:c9c1613ed6a166162705b3347b8d10b661ccc5d95692654d0fb628118f2c34e6"},
+    {file = "typer-0.12.5-py3-none-any.whl", hash = "sha256:62fe4e471711b147e3365034133904df3e235698399bc4de2b36c8579298d52b"},
+    {file = "typer-0.12.5.tar.gz", hash = "sha256:f592f089bedcc8ec1b974125d64851029c3b1af145f04aca64d69410f0c9b722"},
 ]

 [package.dependencies]
@ -4914,13 +5012,13 @@ files = [

 [[package]]
 name = "urllib3"
-version = "1.26.19"
+version = "1.26.20"
 description = "HTTP library with thread-safe connection pooling, file post, and more."
 optional = false
 python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,>=2.7"
 files = [
-    {file = "urllib3-1.26.19-py2.py3-none-any.whl", hash = "sha256:37a0344459b199fce0e80b0d3569837ec6b6937435c5244e7fd73fa6006830f3"},
-    {file = "urllib3-1.26.19.tar.gz", hash = "sha256:3e3d753a8618b86d7de333b4223005f68720bcd6a7d2bcb9fbd2229ec7c1e429"},
+    {file = "urllib3-1.26.20-py2.py3-none-any.whl", hash = "sha256:0ed14ccfbf1c30a9072c7ca157e4319b70d65f623e91e7b32fadb2853431016e"},
+    {file = "urllib3-1.26.20.tar.gz", hash = "sha256:40c2dc0c681e47eb8f90e7e27bf6ff7df2e677421fd46756da1161c39ca70d32"},
 ]

 [package.extras]
@ -5127,20 +5225,24 @@ test = ["pytest"]

 [[package]]
 name = "zipp"
-version = "3.20.0"
+version = "3.20.1"
 description = "Backport of pathlib-compatible object wrapper for zip files"
 optional = false
 python-versions = ">=3.8"
 files = [
-    {file = "zipp-3.20.0-py3-none-any.whl", hash = "sha256:58da6168be89f0be59beb194da1250516fdaa062ccebd30127ac65d30045e10d"},
-    {file = "zipp-3.20.0.tar.gz", hash = "sha256:0145e43d89664cfe1a2e533adc75adafed82fe2da404b4bbb6b026c0157bdb31"},
+    {file = "zipp-3.20.1-py3-none-any.whl", hash = "sha256:9960cd8967c8f85a56f920d5d507274e74f9ff813a0ab8889a5b5be2daf44064"},
+    {file = "zipp-3.20.1.tar.gz", hash = "sha256:c22b14cc4763c5a5b04134207736c107db42e9d3ef2d9779d465f5f1bcba572b"},
 ]

 [package.extras]
+check = ["pytest-checkdocs (>=2.4)", "pytest-ruff (>=0.2.1)"]
+cover = ["pytest-cov"]
 doc = ["furo", "jaraco.packaging (>=9.3)", "jaraco.tidelift (>=1.4)", "rst.linker (>=1.9)", "sphinx (>=3.5)", "sphinx-lint"]
-test = ["big-O", "importlib-resources", "jaraco.functools", "jaraco.itertools", "jaraco.test", "more-itertools", "pytest (>=6,!=8.1.*)", "pytest-checkdocs (>=2.4)", "pytest-cov", "pytest-enabler (>=2.2)", "pytest-ignore-flaky", "pytest-mypy", "pytest-ruff (>=0.2.1)"]
+enabler = ["pytest-enabler (>=2.2)"]
+test = ["big-O", "importlib-resources", "jaraco.functools", "jaraco.itertools", "jaraco.test", "more-itertools", "pytest (>=6,!=8.1.*)", "pytest-ignore-flaky"]
+type = ["pytest-mypy"]

 [metadata]
 lock-version = "2.0"
 python-versions = "^3.10"
-content-hash = "35a2c8652173107818d14f38266af07b689fa80a5f0e8f3e06f5708797511cf6"
+content-hash = "9f876a0fe3d1f350eb9279f6e8ec8c481d2ce196d1bdd9834f4b89c881658074"
--- a/pyproject.toml
+++ b/pyproject.toml
@ -25,14 +25,14 @@ python = "^3.10"
 pydantic = "^2.0.0"
 docling-core = "^1.1.2"
 docling-ibm-models = "^1.1.3"
-deepsearch-glm = ">=0.19.0,<1"
+deepsearch-glm = "^0.19.1"
 filetype = "^1.2.0"
 pypdfium2 = "^4.30.0"
 pydantic-settings = "^2.3.0"
 huggingface_hub = ">=0.23,<1"
 requests = "^2.32.3"
 easyocr = "^1.7"
-docling-parse = "^1.1.1"
+docling-parse = "^1.1.3"
 certifi = ">=2024.7.4"
 rtree = "^1.3.0"
 scipy = "^1.14.1"
--- a/test/test_backend_docling_parse.py
+++ b/test/test_backend_docling_parse.py
@ -1,33 +0,0 @@
-from pathlib import Path
-
-import pytest
-
-from docling.backend.docling_parse_backend import DoclingParseDocumentBackend, DoclingParsePageBackend
-from docling.datamodel.base_models import BoundingBox
-
-
-@pytest.fixture
-def test_doc_path():
-    return Path("./data/2206.01062.pdf")
-
-def test_get_text_from_rect(test_doc_path):
-    doc_backend = DoclingParseDocumentBackend(test_doc_path)
-    page_backend: DoclingParsePageBackend = doc_backend.load_page(0)
-
-    # Get the title text of the DocLayNet paper
-    textpiece = page_backend.get_text_in_rect(bbox=BoundingBox(l=102,t=77,r=511,b=124))
-    ref = "DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis"
-
-    assert textpiece.strip() == ref
-
-def test_crop_page_image(test_doc_path):
-    doc_backend = DoclingParseDocumentBackend(test_doc_path)
-    page_backend: DoclingParsePageBackend = doc_backend.load_page(0)
-
-    # Crop out "Figure 1" from the DocLayNet paper
-    im = page_backend.get_page_image(scale=2, cropbox=BoundingBox(l=317,t=246,r=574,b=527))
-    # im.show()
-
-def test_num_pages(test_doc_path):
-    doc_backend = DoclingParseDocumentBackend(test_doc_path)
-    doc_backend.page_count() == 9
--- a/test/test_backend_pdfium.py
+++ b/test/test_backend_pdfium.py
@ -1,33 +0,0 @@
-from pathlib import Path
-
-import pytest
-
-from docling.backend.pypdfium2_backend import PyPdfiumDocumentBackend, PyPdfiumPageBackend
-from docling.datamodel.base_models import BoundingBox
-
-
-@pytest.fixture
-def test_doc_path():
-    return Path("./data/2206.01062.pdf")
-
-def test_get_text_from_rect(test_doc_path):
-    doc_backend = PyPdfiumDocumentBackend(test_doc_path)
-    page_backend: PyPdfiumPageBackend = doc_backend.load_page(0)
-
-    # Get the title text of the DocLayNet paper
-    textpiece = page_backend.get_text_in_rect(bbox=BoundingBox(l=102,t=77,r=511,b=124))
-    ref = "DocLayNet: A Large Human-Annotated Dataset for\r\nDocument-Layout Analysis"
-
-    assert textpiece.strip() == ref
-
-def test_crop_page_image(test_doc_path):
-    doc_backend = PyPdfiumDocumentBackend(test_doc_path)
-    page_backend: PyPdfiumPageBackend = doc_backend.load_page(0)
-
-    # Crop out "Figure 1" from the DocLayNet paper
-    im = page_backend.get_page_image(scale=2, cropbox=BoundingBox(l=317,t=246,r=574,b=527))
-    # im.show()
-
-def test_num_pages(test_doc_path):
-    doc_backend = PyPdfiumDocumentBackend(test_doc_path)
-    doc_backend.page_count() == 9
--- a/tests/init.py
+++ b/tests/init.py
--- a/tests/data/2203.01017v2.json
+++ b/tests/data/2203.01017v2.json
--- a/tests/data/2203.01017v2.md
+++ b/tests/data/2203.01017v2.md
@ -0,0 +1,450 @@
+## TableFormer: Table Structure Understanding with Transformers.
+
+## Ahmed Nassar, Nikolaos Livathinos, Maksym Lysak, Peter Staar IBM Research
+
+{ ahn,nli,mly,taa } @zurich.ibm.com
+
+## Abstract
+
+## a. Picture of a table:
+
+## 1. Introduction
+
+The occurrence of tables in documents is ubiquitous. They often summarise quantitative or factual data, which is cumbersome to describe in verbose text but nevertheless extremely valuable. Unfortunately, this compact representation is often not easy to parse by machines. There are many implicit conventions used to obtain a compact table representation. For example, tables often have complex columnand row-headers in order to reduce duplicated cell content. Lines of different shapes and sizes are leveraged to separate content or indicate a tree structure. Additionally, tables can also have empty/missing table-entries or multi-row textual table-entries. Fig. 1 shows a table which presents all these issues.
+
+Tables organize valuable content in a concise and compact representation. This content is extremely valuable for systems such as search engines, Knowledge Graph's, etc, since they enhance their predictive capabilities. Unfortunately, tables come in a large variety of shapes and sizes. Furthermore, they can have complex column/row-header configurations, multiline rows, different variety of separation lines, missing entries, etc. As such, the correct identification of the table-structure from an image is a nontrivial task. In this paper, we present a new table-structure identification model. The latter improves the latest end-toend deep learning model (i.e. encoder-dual-decoder from PubTabNet) in two significant ways. First, we introduce a new object detection decoder for table-cells. In this way, we can obtain the content of the table-cells from programmatic PDF's directly from the PDF source and avoid the training of the custom OCR decoders. This architectural change leads to more accurate table-content extraction and allows us to tackle non-english tables. Second, we replace the LSTM decoders with transformer based decoders. This upgrade improves significantly the previous state-of-the-art tree-editing-distance-score (TEDS) from 91% to 98.5% on simple tables and from 88.7% to 95% on complex tables.
+
+|    | 3   | 1   |
+|----|-----|-----|
+|  2 |     |     |
+
+b. Red-annotation of bounding boxes, Blue-predictions by TableFormer
+
+c. Structure predicted by TableFormer:
+
+Figure 1: Picture of a table with subtle, complex features such as (1) multi-column headers, (2) cell with multi-row text and (3) cells with no content. Image from PubTabNet evaluation set, filename: 'PMC2944238 004 02'.
+
+| 0   |   1 | 1   |   2 1 |   2 1 |    |
+|-----|-----|-----|-------|-------|----|
+| 3   |   4 | 5 3 |     6 |     7 |    |
+| 8   |   9 | 10  |    11 |    12 | 2  |
+|     |  13 | 14  |    15 |    16 | 2  |
+|     |  17 | 18  |    19 |    20 | 2  |
+
+Recently, significant progress has been made with vision based approaches to extract tables in documents. For the sake of completeness, the issue of table extraction from documents is typically decomposed into two separate challenges, i.e. (1) finding the location of the table(s) on a document-page and (2) finding the structure of a given table in the document.
+
+The first problem is called table-location and has been previously addressed [30, 38, 19, 21, 23, 26, 8] with stateof-the-art object-detection networks (e.g. YOLO and later on Mask-RCNN [9]). For all practical purposes, it can be
+
+considered as a solved problem, given enough ground-truth data to train on.
+
+The second problem is called table-structure decomposition. The latter is a long standing problem in the community of document understanding [6, 4, 14]. Contrary to the table-location problem, there are no commonly used approaches that can easily be re-purposed to solve this problem. Lately, a set of new model-architectures has been proposed by the community to address table-structure decomposition [37, 36, 18, 20]. All these models have some weaknesses (see Sec. 2). The common denominator here is the reliance on textual features and/or the inability to provide the bounding box of each table-cell in the original image.
+
+In this paper, we want to address these weaknesses and present a robust table-structure decomposition algorithm. The design criteria for our model are the following. First, we want our algorithm to be language agnostic. In this way, we can obtain the structure of any table, irregardless of the language. Second, we want our algorithm to leverage as much data as possible from the original PDF document. For programmatic PDF documents, the text-cells can often be extracted much faster and with higher accuracy compared to OCR methods. Last but not least, we want to have a direct link between the table-cell and its bounding box in the image.
+
+To meet the design criteria listed above, we developed a new model called TableFormer and a synthetically generated table structure dataset called SynthTabNet $^{1}$. In particular, our contributions in this work can be summarised as follows:
+
+· We propose TableFormer , a transformer based model that predicts tables structure and bounding boxes for the table content simultaneously in an end-to-end approach.
+
+· Across all benchmark datasets TableFormer significantly outperforms existing state-of-the-art metrics, while being much more efficient in training and inference to existing works.
+
+· We present SynthTabNet a synthetically generated dataset, with various appearance styles and complexity.
+
+· An augmented dataset based on PubTabNet [37], FinTabNet [36], and TableBank [17] with generated ground-truth for reproducibility.
+
+The paper is structured as follows. In Sec. 2, we give a brief overview of the current state-of-the-art. In Sec. 3, we describe the datasets on which we train. In Sec. 4, we introduce the TableFormer model-architecture and describe
+
+its results & performance in Sec. 5. As a conclusion, we describe how this new model-architecture can be re-purposed for other tasks in the computer-vision community.
+
+## 2. Previous work and State of the Art
+
+Identifying the structure of a table has been an outstanding problem in the document-parsing community, that motivates many organised public challenges [6, 4, 14]. The difficulty of the problem can be attributed to a number of factors. First, there is a large variety in the shapes and sizes of tables. Such large variety requires a flexible method. This is especially true for complex column- and row headers, which can be extremely intricate and demanding. A second factor of complexity is the lack of data with regard to table-structure. Until the publication of PubTabNet [37], there were no large datasets (i.e. > 100 K tables) that provided structure information. This happens primarily due to the fact that tables are notoriously time-consuming to annotate by hand. However, this has definitely changed in recent years with the deliverance of PubTabNet [37], FinTabNet [36], TableBank [17] etc.
+
+Before the rising popularity of deep neural networks, the community relied heavily on heuristic and/or statistical methods to do table structure identification [3, 7, 11, 5, 13, 28]. Although such methods work well on constrained tables [12], a more data-driven approach can be applied due to the advent of convolutional neural networks (CNNs) and the availability of large datasets. To the best-of-our knowledge, there are currently two different types of network architecture that are being pursued for state-of-the-art tablestructure identification.
+
+Image-to-Text networks : In this type of network, one predicts a sequence of tokens starting from an encoded image. Such sequences of tokens can be HTML table tags [37, 17] or LaTeX symbols[10]. The choice of symbols is ultimately not very important, since one can be transformed into the other. There are however subtle variations in the Image-to-Text networks. The easiest network architectures are "image-encoder → text-decoder" (IETD), similar to network architectures that try to provide captions to images [32]. In these IETD networks, one expects as output the LaTeX/HTML string of the entire table, i.e. the symbols necessary for creating the table with the content of the table. Another approach is the "image-encoder → dual decoder" (IEDD) networks. In these type of networks, one has two consecutive decoders with different purposes. The first decoder is the tag-decoder , i.e. it only produces the HTML/LaTeX tags which construct an empty table. The second content-decoder uses the encoding of the image in combination with the output encoding of each cell-tag (from the tag-decoder ) to generate the textual content of each table cell. The network architecture of IEDD is certainly more elaborate, but it has the advantage that one can pre-train the
+
+tag-decoder which is constrained to the table-tags.
+
+In practice, both network architectures (IETD and IEDD) require an implicit, custom trained object-characterrecognition (OCR) to obtain the content of the table-cells. In the case of IETD, this OCR engine is implicit in the decoder similar to [24]. For the IEDD, the OCR is solely embedded in the content-decoder. This reliance on a custom, implicit OCR decoder is of course problematic. OCR is a well known and extremely tough problem, that often needs custom training for each individual language. However, the limited availability for non-english content in the current datasets, makes it impractical to apply the IETD and IEDD methods on tables with other languages. Additionally, OCR can be completely omitted if the tables originate from programmatic PDF documents with known positions of each cell. The latter was the inspiration for the work of this paper.
+
+Graph Neural networks : Graph Neural networks (GNN's) take a radically different approach to tablestructure extraction. Note that one table cell can constitute out of multiple text-cells. To obtain the table-structure, one creates an initial graph, where each of the text-cells becomes a node in the graph similar to [33, 34, 2]. Each node is then associated with en embedding vector coming from the encoded image, its coordinates and the encoded text. Furthermore, nodes that represent adjacent text-cells are linked. Graph Convolutional Networks (GCN's) based methods take the image as an input, but also the position of the text-cells and their content [18]. The purpose of a GCN is to transform the input graph into a new graph, which replaces the old links with new ones. The new links then represent the table-structure. With this approach, one can avoid the need to build custom OCR decoders. However, the quality of the reconstructed structure is not comparable to the current state-of-the-art [18].
+
+Hybrid Deep Learning-Rule-Based approach : A popular current model for table-structure identification is the use of a hybrid Deep Learning-Rule-Based approach similar to [27, 29]. In this approach, one first detects the position of the table-cells with object detection (e.g. YoloVx or MaskRCNN), then classifies the table into different types (from its images) and finally uses different rule-sets to obtain its table-structure. Currently, this approach achieves stateof-the-art results, but is not an end-to-end deep-learning method. As such, new rules need to be written if different types of tables are encountered.
+
+## 3. Datasets
+
+We rely on large-scale datasets such as PubTabNet [37], FinTabNet [36], and TableBank [17] datasets to train and evaluate our models. These datasets span over various appearance styles and content. We also introduce our own synthetically generated SynthTabNet dataset to fix an im-
+
+Figure 2: Distribution of the tables across different table dimensions in PubTabNet + FinTabNet datasets
+
+balance in the previous datasets.
+
+The PubTabNet dataset contains 509k tables delivered as annotated PNG images. The annotations consist of the table structure represented in HTML format, the tokenized text and its bounding boxes per table cell. Fig. 1 shows the appearance style of PubTabNet. Depending on its complexity, a table is characterized as "simple" when it does not contain row spans or column spans, otherwise it is "complex". The dataset is divided into Train and Val splits (roughly 98% and 2%). The Train split consists of 54% simple and 46% complex tables and the Val split of 51% and 49% respectively. The FinTabNet dataset contains 112k tables delivered as single-page PDF documents with mixed table structures and text content. Similarly to the PubTabNet, the annotations of FinTabNet include the table structure in HTML, the tokenized text and the bounding boxes on a table cell basis. The dataset is divided into Train, Test and Val splits (81%, 9.5%, 9.5%), and each one is almost equally divided into simple and complex tables (Train: 48% simple, 52% complex, Test: 48% simple, 52% complex, Test: 53% simple, 47% complex). Finally the TableBank dataset consists of 145k tables provided as JPEG images. The latter has annotations for the table structure, but only few with bounding boxes of the table cells. The entire dataset consists of simple tables and it is divided into 90% Train, 3% Test and 7% Val splits.
+
+Due to the heterogeneity across the dataset formats, it was necessary to combine all available data into one homogenized dataset before we could train our models for practical purposes. Given the size of PubTabNet, we adopted its annotation format and we extracted and converted all tables as PNG images with a resolution of 72 dpi. Additionally, we have filtered out tables with extreme sizes due to small
+
+amount of such tables, and kept only those ones ranging between 1*1 and 20*10 (rows/columns).
+
+The availability of the bounding boxes for all table cells is essential to train our models. In order to distinguish between empty and non-empty bounding boxes, we have introduced a binary class in the annotation. Unfortunately, the original datasets either omit the bounding boxes for whole tables (e.g. TableBank) or they narrow their scope only to non-empty cells. Therefore, it was imperative to introduce a data pre-processing procedure that generates the missing bounding boxes out of the annotation information. This procedure first parses the provided table structure and calculates the dimensions of the most fine-grained grid that covers the table structure. Notice that each table cell may occupy multiple grid squares due to row or column spans. In case of PubTabNet we had to compute missing bounding boxes for 48% of the simple and 69% of the complex tables. Regarding FinTabNet, 68% of the simple and 98% of the complex tables require the generation of bounding boxes.
+
+As it is illustrated in Fig. 2, the table distributions from all datasets are skewed towards simpler structures with fewer number of rows/columns. Additionally, there is very limited variance in the table styles, which in case of PubTabNet and FinTabNet means one styling format for the majority of the tables. Similar limitations appear also in the type of table content, which in some cases (e.g. FinTabNet) is restricted to a certain domain. Ultimately, the lack of diversity in the training dataset damages the ability of the models to generalize well on unseen data.
+
+Motivated by those observations we aimed at generating a synthetic table dataset named SynthTabNet . This approach offers control over: 1) the size of the dataset, 2) the table structure, 3) the table style and 4) the type of content. The complexity of the table structure is described by the size of the table header and the table body, as well as the percentage of the table cells covered by row spans and column spans. A set of carefully designed styling templates provides the basis to build a wide range of table appearances. Lastly, the table content is generated out of a curated collection of text corpora. By controlling the size and scope of the synthetic datasets we are able to train and evaluate our models in a variety of different conditions. For example, we can first generate a highly diverse dataset to train our models and then evaluate their performance on other synthetic datasets which are focused on a specific domain.
+
+In this regard, we have prepared four synthetic datasets, each one containing 150k examples. The corpora to generate the table text consists of the most frequent terms appearing in PubTabNet and FinTabNet together with randomly generated text. The first two synthetic datasets have been fine-tuned to mimic the appearance of the original datasets but encompass more complicated table structures. The third
+
+Table 1: Both "Combined-Tabnet" and "CombinedTabnet" are variations of the following: (*) The CombinedTabnet dataset is the processed combination of PubTabNet and Fintabnet. (**) The combined dataset is the processed combination of PubTabNet, Fintabnet and TableBank.
+
+|                    |   Tags |   Bbox | Size   | Format   |
+|--------------------|--------|--------|--------|----------|
+| PubTabNet          |      3 |      3 | 509k   | PNG      |
+| FinTabNet          |      3 |      3 | 112k   | PDF      |
+| TableBank          |      3 |      7 | 145k   | JPEG     |
+| Combined-Tabnet(*) |      3 |      3 | 400k   | PNG      |
+| Combined(**)       |      3 |      3 | 500k   | PNG      |
+| SynthTabNet        |      3 |      3 | 600k   | PNG      |
+
+one adopts a colorful appearance with high contrast and the last one contains tables with sparse content. Lastly, we have combined all synthetic datasets into one big unified synthetic dataset of 600k examples.
+
+Tab. 1 summarizes the various attributes of the datasets.
+
+## 4. The TableFormer model
+
+Given the image of a table, TableFormer is able to predict: 1) a sequence of tokens that represent the structure of a table, and 2) a bounding box coupled to a subset of those tokens. The conversion of an image into a sequence of tokens is a well-known task [35, 16]. While attention is often used as an implicit method to associate each token of the sequence with a position in the original image, an explicit association between the individual table-cells and the image bounding boxes is also required.
+
+## 4.1. Model architecture.
+
+We now describe in detail the proposed method, which is composed of three main components, see Fig. 4. Our CNN Backbone Network encodes the input as a feature vector of predefined length. The input feature vector of the encoded image is passed to the Structure Decoder to produce a sequence of HTML tags that represent the structure of the table. With each prediction of an HTML standard data cell (' < td > ') the hidden state of that cell is passed to the Cell BBox Decoder. As for spanning cells, such as row or column span, the tag is broken down to ' < ', 'rowspan=' or 'colspan=', with the number of spanning cells (attribute), and ' > '. The hidden state attached to ' < ' is passed to the Cell BBox Decoder. A shared feed forward network (FFN) receives the hidden states from the Structure Decoder, to provide the final detection predictions of the bounding box coordinates and their classification.
+
+CNN Backbone Network. A ResNet-18 CNN is the backbone that receives the table image and encodes it as a vector of predefined length. The network has been modified by removing the linear and pooling layer, as we are not per-
+
+Figure 3: TableFormer takes in an image of the PDF and creates bounding box and HTML structure predictions that are synchronized. The bounding boxes grabs the content from the PDF and inserts it in the structure.
+
+Figure 4: Given an input image of a table, the Encoder produces fixed-length features that represent the input image. The features are then passed to both the Structure Decoder and Cell BBox Decoder . During training, the Structure Decoder receives 'tokenized tags' of the HTML code that represent the table structure. Afterwards, a transformer encoder and decoder architecture is employed to produce features that are received by a linear layer, and the Cell BBox Decoder. The linear layer is applied to the features to predict the tags. Simultaneously, the Cell BBox Decoder selects features referring to the data cells (' < td > ', ' < ') and passes them through an attention network, an MLP, and a linear layer to predict the bounding boxes.
+
+forming classification, and adding an adaptive pooling layer of size 28*28. ResNet by default downsamples the image resolution by 32 and then the encoded image is provided to both the Structure Decoder , and Cell BBox Decoder .
+
+Structure Decoder. The transformer architecture of this component is based on the work proposed in [31]. After extensive experimentation, the Structure Decoder is modeled as a transformer encoder with two encoder layers and a transformer decoder made from a stack of 4 decoder layers that comprise mainly of multi-head attention and feed forward layers. This configuration uses fewer layers and heads in comparison to networks applied to other problems (e.g. "Scene Understanding", "Image Captioning"), something which we relate to the simplicity of table images.
+
+The transformer encoder receives an encoded image from the CNN Backbone Network and refines it through a multi-head dot-product attention layer, followed by a Feed Forward Network. During training, the transformer decoder receives as input the output feature produced by the transformer encoder, and the tokenized input of the HTML ground-truth tags. Using a stack of multi-head attention layers, different aspects of the tag sequence could be inferred. This is achieved by each attention head on a layer operating in a different subspace, and then combining altogether their attention score.
+
+Cell BBox Decoder. Our architecture allows to simultaneously predict HTML tags and bounding boxes for each table cell without the need of a separate object detector end to end. This approach is inspired by DETR [1] which employs a Transformer Encoder, and Decoder that looks for a specific number of object queries (potential object detections). As our model utilizes a transformer architecture, the hidden state of the < td > ' and ' < ' HTML structure tags become the object query.
+
+The encoding generated by the CNN Backbone Network along with the features acquired for every data cell from the Transformer Decoder are then passed to the attention network. The attention network takes both inputs and learns to provide an attention weighted encoding. This weighted at-
+
+tention encoding is then multiplied to the encoded image to produce a feature for each table cell. Notice that this is different than the typical object detection problem where imbalances between the number of detections and the amount of objects may exist. In our case, we know up front that the produced detections always match with the table cells in number and correspondence.
+
+The output features for each table cell are then fed into the feed-forward network (FFN). The FFN consists of a Multi-Layer Perceptron (3 layers with ReLU activation function) that predicts the normalized coordinates for the bounding box of each table cell. Finally, the predicted bounding boxes are classified based on whether they are empty or not using a linear layer.
+
+Loss Functions. We formulate a multi-task loss Eq. 2 to train our network. The Cross-Entropy loss (denoted as l$_{s}$ ) is used to train the Structure Decoder which predicts the structure tokens. As for the Cell BBox Decoder it is trained with a combination of losses denoted as l$_{box}$ . l$_{box}$ consists of the generally used l$_{1}$ loss for object detection and the IoU loss ( l$_{iou}$ ) to be scale invariant as explained in [25]. In comparison to DETR, we do not use the Hungarian algorithm [15] to match the predicted bounding boxes with the ground-truth boxes, as we have already achieved a one-toone match through two steps: 1) Our token input sequence is naturally ordered, therefore the hidden states of the table data cells are also in order when they are provided as input to the Cell BBox Decoder , and 2) Our bounding boxes generation mechanism (see Sec. 3) ensures a one-to-one mapping between the cell content and its bounding box for all post-processed datasets.
+
+The loss used to train the TableFormer can be defined as following:
+
+where λ ∈ [0, 1], and λ$_{iou}$, λ$_{l}$$_{1}$ ∈$_{R}$ are hyper-parameters.
+
+## 5. Experimental Results
+
+## 5.1. Implementation Details
+
+TableFormer uses ResNet-18 as the CNN Backbone Network . The input images are resized to 448*448 pixels and the feature map has a dimension of 28*28. Additionally, we enforce the following input constraints:
+
+Although input constraints are used also by other methods, such as EDD, ours are less restrictive due to the improved
+
+runtime performance and lower memory footprint of TableFormer. This allows to utilize input samples with longer sequences and images with larger dimensions.
+
+The Transformer Encoder consists of two "Transformer Encoder Layers", with an input feature size of 512, feed forward network of 1024, and 4 attention heads. As for the Transformer Decoder it is composed of four "Transformer Decoder Layers" with similar input and output dimensions as the "Transformer Encoder Layers". Even though our model uses fewer layers and heads than the default implementation parameters, our extensive experimentation has proved this setup to be more suitable for table images. We attribute this finding to the inherent design of table images, which contain mostly lines and text, unlike the more elaborate content present in other scopes (e.g. the COCO dataset). Moreover, we have added ResNet blocks to the inputs of the Structure Decoder and Cell BBox Decoder. This prevents a decoder having a stronger influence over the learned weights which would damage the other prediction task (structure vs bounding boxes), but learn task specific weights instead. Lastly our dropout layers are set to 0.5.
+
+For training, TableFormer is trained with 3 Adam optimizers, each one for the CNN Backbone Network , Structure Decoder , and Cell BBox Decoder . Taking the PubTabNet as an example for our parameter set up, the initializing learning rate is 0.001 for 12 epochs with a batch size of 24, and λ set to 0.5. Afterwards, we reduce the learning rate to 0.0001, the batch size to 18 and train for 12 more epochs or convergence.
+
+TableFormer is implemented with PyTorch and Torchvision libraries [22]. To speed up the inference, the image undergoes a single forward pass through the CNN Backbone Network and transformer encoder. This eliminates the overhead of generating the same features for each decoding step. Similarly, we employ a 'caching' technique to preform faster autoregressive decoding. This is achieved by storing the features of decoded tokens so we can reuse them for each time step. Therefore, we only compute the attention for each new tag.
+
+## 5.2. Generalization
+
+TableFormer is evaluated on three major publicly available datasets of different nature to prove the generalization and effectiveness of our model. The datasets used for evaluation are the PubTabNet, FinTabNet and TableBank which stem from the scientific, financial and general domains respectively.
+
+We also share our baseline results on the challenging SynthTabNet dataset. Throughout our experiments, the same parameters stated in Sec. 5.1 are utilized.
+
+## 5.3. Datasets and Metrics
+
+The Tree-Edit-Distance-Based Similarity (TEDS) metric was introduced in [37]. It represents the prediction, and ground-truth as a tree structure of HTML tags. This similarity is calculated as:
+
+where T$_{a}$ and T$_{b}$ represent tables in tree structure HTML format. EditDist denotes the tree-edit distance, and | T | represents the number of nodes in T .
+
+## 5.4. Quantitative Analysis
+
+Structure. As shown in Tab. 2, TableFormer outperforms all SOTA methods across different datasets by a large margin for predicting the table structure from an image. All the more, our model outperforms pre-trained methods. During the evaluation we do not apply any table filtering. We also provide our baseline results on the SynthTabNet dataset. It has been observed that large tables (e.g. tables that occupy half of the page or more) yield poor predictions. We attribute this issue to the image resizing during the preprocessing step, that produces downsampled images with indistinguishable features. This problem can be addressed by treating such big tables with a separate model which accepts a large input image size.
+
+Table 2: Structure results on PubTabNet (PTN), FinTabNet (FTN), TableBank (TB) and SynthTabNet (STN).
+
+| Model       | Dataset   | Simple   | TEDS Complex   |   All |
+|-------------|-----------|----------|----------------|-------|
+| EDD         | PTN       | 91.1     | 88.7           | 89.9  |
+| GTE         | PTN       | -        | -              | 93.01 |
+| TableFormer | PTN       | 98.5     | 95.0           | 96.75 |
+| EDD         | FTN       | 88.4     | 92.08          | 90.6  |
+| GTE         | FTN       | -        | -              | 87.14 |
+| GTE (FT)    | FTN       | -        | -              | 91.02 |
+| TableFormer | FTN       | 97.5     | 96.0           | 96.8  |
+| EDD         | TB        | 86.0     | -              | 86    |
+| TableFormer | TB        | 89.6     | -              | 89.6  |
+| TableFormer | STN       | 96.9     | 95.7           | 96.7  |
+
+FT: Model was trained on PubTabNet then finetuned.
+
+Cell Detection. Like any object detector, our Cell BBox Detector provides bounding boxes that can be improved with post-processing during inference. We make use of the grid-like structure of tables to refine the predictions. A detailed explanation on the post-processing is available in the supplementary material. As shown in Tab. 3, we evaluate
+
+our Cell BBox Decoder accuracy for cells with a class label of 'content' only using the PASCAL VOC mAP metric for pre-processing and post-processing. Note that we do not have post-processing results for SynthTabNet as images are only provided. To compare the performance of our proposed approach, we've integrated TableFormer's Cell BBox Decoder into EDD architecture. As mentioned previously, the Structure Decoder provides the Cell BBox Decoder with the features needed to predict the bounding box predictions. Therefore, the accuracy of the Structure Decoder directly influences the accuracy of the Cell BBox Decoder . If the Structure Decoder predicts an extra column, this will result in an extra column of predicted bounding boxes.
+
+Table 3: Cell Bounding Box detection results on PubTabNet, and FinTabNet. PP: Post-processing.
+
+| Model       | Dataset     |   mAP | mAP (PP)   |
+|-------------|-------------|-------|------------|
+| EDD+BBox    | PubTabNet   |  79.2 | 82.7       |
+| TableFormer | PubTabNet   |  82.1 | 86.8       |
+| TableFormer | SynthTabNet |  87.7 | -          |
+
+Cell Content. In this section, we evaluate the entire pipeline of recovering a table with content. Here we put our approach to test by capitalizing on extracting content from the PDF cells rather than decoding from images. Tab. 4 shows the TEDs score of HTML code representing the structure of the table along with the content inserted in the data cell and compared with the ground-truth. Our method achieved a 5.3% increase over the state-of-the-art, and commercial solutions. We believe our scores would be higher if the HTML ground-truth matched the extracted PDF cell content. Unfortunately, there are small discrepancies such as spacings around words or special characters with various unicode representations.
+
+Table 4: Results of structure with content retrieved using cell detection on PubTabNet. In all cases the input is PDF documents with cropped tables.
+
+| Model       |   Simple |   TEDS Complex |   All |
+|-------------|----------|----------------|-------|
+| Tabula      |     78   |           57.8 |  67.9 |
+| Traprange   |     60.8 |           49.9 |  55.4 |
+| Camelot     |     80   |           66   |  73   |
+| Acrobat Pro |     68.9 |           61.8 |  65.3 |
+| EDD         |     91.2 |           85.4 |  88.3 |
+| TableFormer |     95.4 |           90.1 |  93.6 |
+
+a. Red - PDF cells, Green - predicted bounding boxes, Blue - post-processed predictions matched to PDF cells
+
+Japanese language (previously unseen by TableFormer):
+
+Example table from FinTabNet:
+
+b. Structure predicted by TableFormer, with superimposed matched PDF cell text:
+
+|                                                    |             | 論文ファイル   | 論文ファイル   | 参考文献   | 参考文献   |
+|----------------------------------------------------|-------------|----------------|----------------|------------|------------|
+| 出典                                               | ファイル 数 | 英語           | 日本語         | 英語       | 日本語     |
+| Association for Computational Linguistics(ACL2003) | 65          | 65             | 0              | 150        | 0          |
+| Computational Linguistics(COLING2002)              | 140         | 140            | 0              | 150        | 0          |
+| 電気情報通信学会 2003 年総合大会                   | 150         | 8              | 142            | 223        | 147        |
+| 情報処理学会第 65 回全国大会 (2003)                | 177         | 1              | 176            | 150        | 236        |
+| 第 17 回人工知能学会全国大会 (2003)                | 208         | 5              | 203            | 152        | 244        |
+| 自然言語処理研究会第 146 〜 155 回                 | 98          | 2              | 96             | 150        | 232        |
+| WWW から収集した論文                               | 107         | 73             | 34             | 147        | 96         |
+|                                                    | 945         | 294            | 651            | 1122       | 955        |
+
+Text is aligned to match original for ease of viewing
+
+|                          | Shares (in millions)   | Shares (in millions)   | Weighted Average Grant Date Fair Value   | Weighted Average Grant Date Fair Value   |
+|--------------------------|------------------------|------------------------|------------------------------------------|------------------------------------------|
+|                          | RS U s                 | PSUs                   | RSUs                                     | PSUs                                     |
+| Nonvested on Janua ry 1  | 1. 1                   | 0.3                    | 90.10 $                                  | $ 91.19                                  |
+| Granted                  | 0. 5                   | 0.1                    | 117.44                                   | 122.41                                   |
+| Vested                   | (0. 5 )                | (0.1)                  | 87.08                                    | 81.14                                    |
+| Canceled or forfeited    | (0. 1 )                | -                      | 102.01                                   | 92.18                                    |
+| Nonvested on December 31 | 1.0                    | 0.3                    | 104.85 $                                 | $ 104.51                                 |
+
+Figure 5: One of the benefits of TableFormer is that it is language agnostic, as an example, the left part of the illustration demonstrates TableFormer predictions on previously unseen language (Japanese). Additionally, we see that TableFormer is robust to variability in style and content, right side of the illustration shows the example of the TableFormer prediction from the FinTabNet dataset.
+
+Figure 6: An example of TableFormer predictions (bounding boxes and structure) from generated SynthTabNet table.
+
+## 5.5. Qualitative Analysis
+
+We showcase several visualizations for the different components of our network on various "complex" tables within datasets presented in this work in Fig. 5 and Fig. 6 As it is shown, our model is able to predict bounding boxes for all table cells, even for the empty ones. Additionally, our post-processing techniques can extract the cell content by matching the predicted bounding boxes to the PDF cells based on their overlap and spatial proximity. The left part of Fig. 5 demonstrates also the adaptability of our method to any language, as it can successfully extract Japanese text, although the training set contains only English content. We provide more visualizations including the intermediate steps in the supplementary material. Overall these illustrations justify the versatility of our method across a diverse range of table appearances and content type.
+
+## 6. Future Work & Conclusion
+
+In this paper, we presented TableFormer an end-to-end transformer based approach to predict table structures and bounding boxes of cells from an image. This approach enables us to recreate the table structure, and extract the cell content from PDF or OCR by using bounding boxes. Additionally, it provides the versatility required in real-world scenarios when dealing with various types of PDF documents, and languages. Furthermore, our method outperforms all state-of-the-arts with a wide margin. Finally, we introduce "SynthTabNet" a challenging synthetically generated dataset that reinforces missing characteristics from other datasets.
+
+## References
+
+[1] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-
+
+end object detection with transformers. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision - ECCV 2020 , pages 213-229, Cham, 2020. Springer International Publishing. 5
+
+[2] Zewen Chi, Heyan Huang, Heng-Da Xu, Houjin Yu, Wanxuan Yin, and Xian-Ling Mao. Complicated table structure recognition. arXiv preprint arXiv:1908.04729 , 2019. 3
+
+[3] Bertrand Couasnon and Aurelie Lemaitre. Recognition of Tables and Forms , pages 647-677. Springer London, London, 2014. 2
+
+[4] Herv'e D'ejean, Jean-Luc Meunier, Liangcai Gao, Yilun Huang, Yu Fang, Florian Kleber, and Eva-Maria Lang. ICDAR 2019 Competition on Table Detection and Recognition (cTDaR), Apr. 2019. http://sac.founderit.com/. 2
+
+[5] Basilios Gatos, Dimitrios Danatsas, Ioannis Pratikakis, and Stavros J Perantonis. Automatic table detection in document images. In International Conference on Pattern Recognition and Image Analysis , pages 609-618. Springer, 2005. 2
+
+[6] Max Gobel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In 2013 12th International Conference on Document Analysis and Recognition , pages 1449-1453, 2013. 2
+
+[7] EA Green and M Krishnamoorthy. Recognition of tables using table grammars. procs. In Symposium on Document Analysis and Recognition (SDAIR'95) , pages 261-277. 2
+
+[8] Khurram Azeem Hashmi, Alain Pagani, Marcus Liwicki, Didier Stricker, and Muhammad Zeshan Afzal. Castabdetectors: Cascade network for table detection in document images with recursive feature pyramid and switchable atrous convolution. Journal of Imaging , 7(10), 2021. 1
+
+[9] Kaiming He, Georgia Gkioxari, Piotr Dollar, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE International Conference on Computer Vision (ICCV) , Oct 2017. 1
+
+[10] Yelin He, X. Qi, Jiaquan Ye, Peng Gao, Yihao Chen, Bingcong Li, Xin Tang, and Rong Xiao. Pingan-vcgroup's solution for icdar 2021 competition on scientific table image recognition to latex. ArXiv , abs/2105.01846, 2021. 2
+
+[11] Jianying Hu, Ramanujan S Kashi, Daniel P Lopresti, and Gordon Wilfong. Medium-independent table detection. In Document Recognition and Retrieval VII , volume 3967, pages 291-302. International Society for Optics and Photonics, 1999. 2
+
+[12] Matthew Hurst. A constraint-based approach to table structure derivation. In Proceedings of the Seventh International Conference on Document Analysis and Recognition - Volume 2 , ICDAR '03, page 911, USA, 2003. IEEE Computer Society. 2
+
+[13] Thotreingam Kasar, Philippine Barlas, Sebastien Adam, Cl'ement Chatelain, and Thierry Paquet. Learning to detect tables in scanned document images using line information. In 2013 12th International Conference on Document Analysis and Recognition , pages 1185-1189. IEEE, 2013. 2
+
+[14] Pratik Kayal, Mrinal Anand, Harsh Desai, and Mayank Singh. Icdar 2021 competition on scientific table image recognition to latex, 2021. 2
+
+[15] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly , 2(1-2):83-97, 1955. 6
+
+[16] Girish Kulkarni, Visruth Premraj, Vicente Ordonez, Sagnik Dhar, Siming Li, Yejin Choi, Alexander C. Berg, and Tamara L. Berg. Babytalk: Understanding and generating simple image descriptions. IEEE Transactions on Pattern Analysis and Machine Intelligence , 35(12):2891-2903, 2013. 4
+
+[17] Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou, and Zhoujun Li. Tablebank: A benchmark dataset for table detection and recognition, 2019. 2, 3
+
+[18] Yiren Li, Zheng Huang, Junchi Yan, Yi Zhou, Fan Ye, and Xianhui Liu. Gfte: Graph-based financial table extraction. In Alberto Del Bimbo, Rita Cucchiara, Stan Sclaroff, Giovanni Maria Farinella, Tao Mei, Marco Bertini, Hugo Jair Escalante, and Roberto Vezzani, editors, Pattern Recognition. ICPR International Workshops and Challenges , pages 644-658, Cham, 2021. Springer International Publishing. 2, 3
+
+[19] Nikolaos Livathinos, Cesar Berrospi, Maksym Lysak, Viktor Kuropiatnyk, Ahmed Nassar, Andre Carvalho, Michele Dolfi, Christoph Auer, Kasper Dinkla, and Peter Staar. Robust pdf document conversion using recurrent neural networks. Proceedings of the AAAI Conference on Artificial Intelligence , 35(17):15137-15145, May 2021. 1
+
+[20] Rujiao Long, Wen Wang, Nan Xue, Feiyu Gao, Zhibo Yang, Yongpan Wang, and Gui-Song Xia. Parsing table structures in the wild. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 944-952, 2021. 2
+
+[21] Shubham Singh Paliwal, D Vishwanath, Rohit Rahul, Monika Sharma, and Lovekesh Vig. Tablenet: Deep learning model for end-to-end table detection and tabular data extraction from scanned document images. In 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages 128-133. IEEE, 2019. 1
+
+[22] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alch'e-Buc, E. Fox, and R. Garnett, editors, Advances in Neural Information Processing Systems 32 , pages 8024-8035. Curran Associates, Inc., 2019. 6
+
+[23] Devashish Prasad, Ayan Gadpal, Kshitij Kapadni, Manish Visave, and Kavita Sultanpure. Cascadetabnet: An approach for end to end table detection and structure recognition from image-based documents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops , pages 572-573, 2020. 1
+
+[24] Shah Rukh Qasim, Hassan Mahmood, and Faisal Shafait. Rethinking table recognition using graph neural networks. In 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages 142-147. IEEE, 2019. 3
+
+[25] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In Proceedings of the IEEE/CVF Conference on
+
+Computer Vision and Pattern Recognition , pages 658-666, 2019. 6
+
+[26] Sebastian Schreiber, Stefan Agne, Ivo Wolf, Andreas Dengel, and Sheraz Ahmed. Deepdesrt: Deep learning for detection and structure recognition of tables in document images. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , volume 01, pages 11621167, 2017. 1
+
+[27] Sebastian Schreiber, Stefan Agne, Ivo Wolf, Andreas Dengel, and Sheraz Ahmed. Deepdesrt: Deep learning for detection and structure recognition of tables in document images. In 2017 14th IAPR international conference on document analysis and recognition (ICDAR) , volume 1, pages 1162-1167. IEEE, 2017. 3
+
+[28] Faisal Shafait and Ray Smith. Table detection in heterogeneous documents. In Proceedings of the 9th IAPR International Workshop on Document Analysis Systems , pages 6572, 2010. 2
+
+[29] Shoaib Ahmed Siddiqui, Imran Ali Fateh, Syed Tahseen Raza Rizvi, Andreas Dengel, and Sheraz Ahmed. Deeptabstr: Deep learning based table structure recognition. In 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages 1403-1409. IEEE, 2019. 3
+
+[30] Peter W J Staar, Michele Dolfi, Christoph Auer, and Costas Bekas. Corpus conversion service: A machine learning platform to ingest documents at scale. In Proceedings of the 24th ACM SIGKDD , KDD '18, pages 774-782, New York, NY, USA, 2018. ACM. 1
+
+[31] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Ł ukasz Kaiser, and Illia Polosukhin. Attention is all you need. In I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Information Processing Systems 30 , pages 5998-6008. Curran Associates, Inc., 2017. 5
+
+[32] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , June 2015. 2
+
+[33] Wenyuan Xue, Qingyong Li, and Dacheng Tao. Res2tim: reconstruct syntactic structures from table images. In 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages 749-755. IEEE, 2019. 3
+
+[34] Wenyuan Xue, Baosheng Yu, Wen Wang, Dacheng Tao, and Qingyong Li. Tgrnet: A table graph reconstruction network for table structure recognition. arXiv preprint arXiv:2106.10598 , 2021. 3
+
+[35] Quanzeng You, Hailin Jin, Zhaowen Wang, Chen Fang, and Jiebo Luo. Image captioning with semantic attention. In Proceedings of the IEEE conference on computer vision and pattern recognition , pages 4651-4659, 2016. 4
+
+[36] Xinyi Zheng, Doug Burdick, Lucian Popa, Peter Zhong, and Nancy Xin Ru Wang. Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. Winter Conference for Applications in Computer Vision (WACV) , 2021. 2, 3
+
+[37] Xu Zhong, Elaheh ShafieiBavani, and Antonio Jimeno Yepes. Image-based table recognition: Data, model,
+
+and evaluation. In Andrea Vedaldi, Horst Bischof, Thomas Brox, and Jan-Michael Frahm, editors, Computer Vision ECCV 2020 , pages 564-580, Cham, 2020. Springer International Publishing. 2, 3, 7
+
+[38] Xu Zhong, Jianbin Tang, and Antonio Jimeno Yepes. Publaynet: Largest dataset ever for document layout analysis. In 2019 International Conference on Document Analysis and Recognition (ICDAR) , pages 1015-1022, 2019. 1
+
+## TableFormer: Table Structure Understanding with Transformers
+
+Supplementary Material
+
+## 1. Details on the datasets
+
+## 1.1. Data preparation
+
+As a first step of our data preparation process, we have calculated statistics over the datasets across the following dimensions: (1) table size measured in the number of rows and columns, (2) complexity of the table, (3) strictness of the provided HTML structure and (4) completeness (i.e. no omitted bounding boxes). A table is considered to be simple if it does not contain row spans or column spans. Additionally, a table has a strict HTML structure if every row has the same number of columns after taking into account any row or column spans. Therefore a strict HTML structure looks always rectangular. However, HTML is a lenient encoding format, i.e. tables with rows of different sizes might still be regarded as correct due to implicit display rules. These implicit rules leave room for ambiguity, which we want to avoid. As such, we prefer to have "strict" tables, i.e. tables where every row has exactly the same length.
+
+We have developed a technique that tries to derive a missing bounding box out of its neighbors. As a first step, we use the annotation data to generate the most fine-grained grid that covers the table structure. In case of strict HTML tables, all grid squares are associated with some table cell and in the presence of table spans a cell extends across multiple grid squares. When enough bounding boxes are known for a rectangular table, it is possible to compute the geometrical border lines between the grid rows and columns. Eventually this information is used to generate the missing bounding boxes. Additionally, the existence of unused grid squares indicates that the table rows have unequal number of columns and the overall structure is non-strict. The generation of missing bounding boxes for non-strict HTML tables is ambiguous and therefore quite challenging. Thus, we have decided to simply discard those tables. In case of PubTabNet we have computed missing bounding boxes for 48% of the simple and 69% of the complex tables. Regarding FinTabNet, 68% of the simple and 98% of the complex tables require the generation of bounding boxes.
+
+Figure 7 illustrates the distribution of the tables across different dimensions per dataset.
+
+## 1.2. Synthetic datasets
+
+Aiming to train and evaluate our models in a broader spectrum of table data we have synthesized four types of datasets. Each one contains tables with different appear-
+
+ances in regard to their size, structure, style and content. Every synthetic dataset contains 150k examples, summing up to 600k synthetic examples. All datasets are divided into Train, Test and Val splits (80%, 10%, 10%).
+
+The process of generating a synthetic dataset can be decomposed into the following steps:
+
+1. Prepare styling and content templates: The styling templates have been manually designed and organized into groups of scope specific appearances (e.g. financial data, marketing data, etc.) Additionally, we have prepared curated collections of content templates by extracting the most frequently used terms out of non-synthetic datasets (e.g. PubTabNet, FinTabNet, etc.).
+
+2. Generate table structures: The structure of each synthetic dataset assumes a horizontal table header which potentially spans over multiple rows and a table body that may contain a combination of row spans and column spans. However, spans are not allowed to cross the header - body boundary. The table structure is described by the parameters: Total number of table rows and columns, number of header rows, type of spans (header only spans, row only spans, column only spans, both row and column spans), maximum span size and the ratio of the table area covered by spans.
+
+3. Generate content: Based on the dataset theme , a set of suitable content templates is chosen first. Then, this content can be combined with purely random text to produce the synthetic content.
+
+4. Apply styling templates: Depending on the domain of the synthetic dataset, a set of styling templates is first manually selected. Then, a style is randomly selected to format the appearance of the synthesized table.
+
+5. Render the complete tables: The synthetic table is finally rendered by a web browser engine to generate the bounding boxes for each table cell. A batching technique is utilized to optimize the runtime overhead of the rendering process.
+
+## 2. Prediction post-processing for PDF documents
+
+Although TableFormer can predict the table structure and the bounding boxes for tables recognized inside PDF documents, this is not enough when a full reconstruction of the original table is required. This happens mainly due the following reasons:
+
+Figure 7: Distribution of the tables across different dimensions per dataset. Simple vs complex tables per dataset and split, strict vs non strict html structures per dataset and table complexity, missing bboxes per dataset and table complexity.
+
+· TableFormer output does not include the table cell content.
+
+· There are occasional inaccuracies in the predictions of the bounding boxes.
+
+However, it is possible to mitigate those limitations by combining the TableFormer predictions with the information already present inside a programmatic PDF document. More specifically, PDF documents can be seen as a sequence of PDF cells where each cell is described by its content and bounding box. If we are able to associate the PDF cells with the predicted table cells, we can directly link the PDF cell content to the table cell structure and use the PDF bounding boxes to correct misalignments in the predicted table cell bounding boxes.
+
+Here is a step-by-step description of the prediction postprocessing:
+
+1. Get the minimal grid dimensions - number of rows and columns for the predicted table structure. This represents the most granular grid for the underlying table structure.
+
+2. Generate pair-wise matches between the bounding boxes of the PDF cells and the predicted cells. The Intersection Over Union (IOU) metric is used to evaluate the quality of the matches.
+
+3. Use a carefully selected IOU threshold to designate the matches as "good" ones and "bad" ones.
+
+3.a. If all IOU scores in a column are below the threshold, discard all predictions (structure and bounding boxes) for that column.
+
+4. Find the best-fitting content alignment for the predicted cells with good IOU per each column. The alignment of the column can be identified by the following formula:
+
+where c is one of { left, centroid, right } and x$_{c}$ is the xcoordinate for the corresponding point.
+
+5. Use the alignment computed in step 4, to compute the median x -coordinate for all table columns and the me-
+
+dian cell size for all table cells. The usage of median during the computations, helps to eliminate outliers caused by occasional column spans which are usually wider than the normal.
+
+6. Snap all cells with bad IOU to their corresponding median x -coordinates and cell sizes.
+
+7. Generate a new set of pair-wise matches between the corrected bounding boxes and PDF cells. This time use a modified version of the IOU metric, where the area of the intersection between the predicted and PDF cells is divided by the PDF cell area. In case there are multiple matches for the same PDF cell, the prediction with the higher score is preferred. This covers the cases where the PDF cells are smaller than the area of predicted or corrected prediction cells.
+
+8. In some rare occasions, we have noticed that TableFormer can confuse a single column as two. When the postprocessing steps are applied, this results with two predicted columns pointing to the same PDF column. In such case we must de-duplicate the columns according to highest total column intersection score.
+
+9. Pick up the remaining orphan cells. There could be cases, when after applying all the previous post-processing steps, some PDF cells could still remain without any match to predicted cells. However, it is still possible to deduce the correct matching for an orphan PDF cell by mapping its bounding box on the geometry of the grid. This mapping decides if the content of the orphan cell will be appended to an already matched table cell, or a new table cell should be created to match with the orphan.
+
+9a. Compute the top and bottom boundary of the horizontal band for each grid row (min/max y coordinates per row).
+
+9b. Intersect the orphan's bounding box with the row bands, and map the cell to the closest grid row.
+
+9c. Compute the left and right boundary of the vertical band for each grid column (min/max x coordinates per column).
+
+9d. Intersect the orphan's bounding box with the column bands, and map the cell to the closest grid column.
+
+9e. If the table cell under the identified row and column is not empty, extend its content with the content of the or-
+
+## phan cell.
+
+9f. Otherwise create a new structural cell and match it wit the orphan cell.
+
+Aditional images with examples of TableFormer predictions and post-processing can be found below.
+
+Figure 8: Example of a table with multi-line header.
+
+Figure 9: Example of a table with big empty distance between cells.
+
+Figure 10: Example of a complex table with empty cells.
+
+Figure 11: Simple table with different style and empty cells.
+
+Figure 12: Simple table predictions and post processing.
+
+Figure 13: Table predictions example on colorful table.
+
+Figure 14: Example with multi-line text.
+
+Figure 16: Example of how post-processing helps to restore mis-aligned bounding boxes prediction artifact.
+
+Figure 15: Example with triangular table.
+
+Figure 17: Example of long table. End-to-end example from initial PDF cells to prediction of bounding boxes, post processing and prediction of structure.
--- a/tests/data/2203.01017v2.pages.json
+++ b/tests/data/2203.01017v2.pages.json
--- a/tests/data/2203.01017v2.pdf
+++ b/tests/data/2203.01017v2.pdf
--- a/tests/data/2206.01062.json
+++ b/tests/data/2206.01062.json
--- a/tests/data/2206.01062.md
+++ b/tests/data/2206.01062.md
@ -0,0 +1,321 @@
+## DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis
+
+Birgit Pfitzmann IBM Research Rueschlikon, Switzerland bpf@zurich.ibm.com
+
+Christoph Auer IBM Research Rueschlikon, Switzerland cau@zurich.ibm.com
+
+Michele Dolfi IBM Research Rueschlikon, Switzerland dol@zurich.ibm.com
+
+Ahmed S. Nassar IBM Research Rueschlikon, Switzerland ahn@zurich.ibm.com
+
+Peter Staar IBM Research Rueschlikon, Switzerland taa@zurich.ibm.com
+
+## ABSTRACT
+
+Accurate document layout analysis is a key requirement for highquality PDF document conversion. With the recent availability of public, large ground-truth datasets such as PubLayNet and DocBank, deep-learning models have proven to be very effective at layout detection and segmentation. While these datasets are of adequate size to train such models, they severely lack in layout variability since they are sourced from scientific article repositories such as PubMed and arXiv only. Consequently, the accuracy of the layout segmentation drops significantly when these models are applied on more challenging and diverse layouts. In this paper, we present DocLayNet , a new, publicly available, document-layout annotation dataset in COCO format. It contains 80863 manually annotated pages from diverse data sources to represent a wide variability in layouts. For each PDF page, the layout annotations provide labelled bounding-boxes with a choice of 11 distinct classes. DocLayNet also provides a subset of double- and triple-annotated pages to determine the inter-annotator agreement. In multiple experiments, we provide baseline accuracy scores (in mAP) for a set of popular object detection models. We also demonstrate that these models fall approximately 10% behind the inter-annotator agreement. Furthermore, we provide evidence that DocLayNet is of sufficient size. Lastly, we compare models trained on PubLayNet, DocBank and DocLayNet, showing that layout predictions of the DocLayNettrained models are more robust and thus the preferred choice for general-purpose document-layout analysis.
+
+## CCS CONCEPTS
+
+· Information systems → Document structure ; · Applied computing → Document analysis ; · Computing methodologies → Machine learning ; Computer vision ; Object detection ;
+
+Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the owner/author(s).
+
+KDD '22, August 14-18, 2022, Washington, DC, USA © 2022 Copyright held by the owner/author(s). ACM ISBN 978-1-4503-9385-0/22/08. https://doi.org/10.1145/3534678.3539043
+
+Figure 1: Four examples of complex page layouts across different document categories
+
+## KEYWORDS
+
+PDF document conversion, layout segmentation, object-detection, data set, Machine Learning
+
+## ACM Reference Format:
+
+Birgit Pfitzmann, Christoph Auer, Michele Dolfi, Ahmed S. Nassar, and Peter Staar. 2022. DocLayNet: A Large Human-Annotated Dataset for DocumentLayout Analysis. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD '22), August 14-18, 2022, Washington, DC, USA. ACM, New York, NY, USA, 9 pages. https://doi.org/10.1145/ 3534678.3539043
+
+## 1 INTRODUCTION
+
+Despite the substantial improvements achieved with machine-learning (ML) approaches and deep neural networks in recent years, document conversion remains a challenging problem, as demonstrated by the numerous public competitions held on this topic [1-4]. The challenge originates from the huge variability in PDF documents regarding layout, language and formats (scanned, programmatic or a combination of both). Engineering a single ML model that can be applied on all types of documents and provides high-quality layout segmentation remains to this day extremely challenging [5]. To highlight the variability in document layouts, we show a few example documents from the DocLayNet dataset in Figure 1.
+
+A key problem in the process of document conversion is to understand the structure of a single document page, i.e. which segments of text should be grouped together in a unit. To train models for this task, there are currently two large datasets available to the community, PubLayNet [6] and DocBank [7]. They were introduced in 2019 and 2020 respectively and significantly accelerated the implementation of layout detection and segmentation models due to their sizes of 300K and 500K ground-truth pages. These sizes were achieved by leveraging an automation approach. The benefit of automated ground-truth generation is obvious: one can generate large ground-truth datasets at virtually no cost. However, the automation introduces a constraint on the variability in the dataset, because corresponding structured source data must be available. PubLayNet and DocBank were both generated from scientific document repositories (PubMed and arXiv), which provide XML or L A T E X sources. Those scientific documents present a limited variability in their layouts, because they are typeset in uniform templates provided by the publishers. Obviously, documents such as technical manuals, annual company reports, legal text, government tenders, etc. have very different and partially unique layouts. As a consequence, the layout predictions obtained from models trained on PubLayNet or DocBank is very reasonable when applied on scientific documents. However, for more artistic or free-style layouts, we see sub-par prediction quality from these models, which we demonstrate in Section 5.
+
+In this paper, we present the DocLayNet dataset. It provides pageby-page layout annotation ground-truth using bounding-boxes for 11 distinct class labels on 80863 unique document pages, of which a fraction carry double- or triple-annotations. DocLayNet is similar in spirit to PubLayNet and DocBank and will likewise be made available to the public 1 in order to stimulate the document-layout analysis community. It distinguishes itself in the following aspects:
+
+(1) Human Annotation : In contrast to PubLayNet and DocBank, we relied on human annotation instead of automation approaches to generate the data set.
+
+(2) Large Layout Variability : We include diverse and complex layouts from a large variety of public sources.
+
+(3) Detailed Label Set : We define 11 class labels to distinguish layout features in high detail. PubLayNet provides 5 labels; DocBank provides 13, although not a superset of ours.
+
+(4) Redundant Annotations : A fraction of the pages in the DocLayNet data set carry more than one human annotation.
+
+This enables experimentation with annotation uncertainty and quality control analysis.
+
+(5) Pre-defined Train-, Test- & Validation-set : Like DocBank, we provide fixed train-, test- & validation-sets to ensure proportional representation of the class-labels. Further, we prevent leakage of unique layouts across sets, which has a large effect on model accuracy scores.
+
+All aspects outlined above are detailed in Section 3. In Section 4, we will elaborate on how we designed and executed this large-scale human annotation campaign. We will also share key insights and lessons learned that might prove helpful for other parties planning to set up annotation campaigns.
+
+In Section 5, we will present baseline accuracy numbers for a variety of object detection methods (Faster R-CNN, Mask R-CNN and YOLOv5) trained on DocLayNet. We further show how the model performance is impacted by varying the DocLayNet dataset size, reducing the label set and modifying the train/test-split. Last but not least, we compare the performance of models trained on PubLayNet, DocBank and DocLayNet and demonstrate that a model trained on DocLayNet provides overall more robust layout recovery.
+
+## 2 RELATED WORK
+
+While early approaches in document-layout analysis used rulebased algorithms and heuristics [8], the problem is lately addressed with deep learning methods. The most common approach is to leverage object detection models [9-15]. In the last decade, the accuracy and speed of these models has increased dramatically. Furthermore, most state-of-the-art object detection methods can be trained and applied with very little work, thanks to a standardisation effort of the ground-truth data format [16] and common deep-learning frameworks [17]. Reference data sets such as PubLayNet [6] and DocBank provide their data in the commonly accepted COCO format [16].
+
+Lately, new types of ML models for document-layout analysis have emerged in the community [18-21]. These models do not approach the problem of layout analysis purely based on an image representation of the page, as computer vision methods do. Instead, they combine the text tokens and image representation of a page in order to obtain a segmentation. While the reported accuracies appear to be promising, a broadly accepted data format which links geometric and textual features has yet to establish.
+
+## 3 THE DOCLAYNET DATASET
+
+DocLayNet contains 80863 PDF pages. Among these, 7059 carry two instances of human annotations, and 1591 carry three. This amounts to 91104 total annotation instances. The annotations provide layout information in the shape of labeled, rectangular boundingboxes. We define 11 distinct labels for layout features, namely Caption , Footnote , Formula , List-item , Page-footer , Page-header , Picture , Section-header , Table , Text , and Title . Our reasoning for picking this particular label set is detailed in Section 4.
+
+In addition to open intellectual property constraints for the source documents, we required that the documents in DocLayNet adhere to a few conditions. Firstly, we kept scanned documents
+
+Figure 2: Distribution of DocLayNet pages across document categories.
+
+to a minimum, since they introduce difficulties in annotation (see Section 4). As a second condition, we focussed on medium to large documents ( > 10 pages) with technical content, dense in complex tables, figures, plots and captions. Such documents carry a lot of information value, but are often hard to analyse with high accuracy due to their challenging layouts. Counterexamples of documents not included in the dataset are receipts, invoices, hand-written documents or photographs showing "text in the wild".
+
+The pages in DocLayNet can be grouped into six distinct categories, namely Financial Reports , Manuals , Scientific Articles , Laws & Regulations , Patents and Government Tenders . Each document category was sourced from various repositories. For example, Financial Reports contain both free-style format annual reports 2 which expose company-specific, artistic layouts as well as the more formal SEC filings. The two largest categories ( Financial Reports and Manuals ) contain a large amount of free-style layouts in order to obtain maximum variability. In the other four categories, we boosted the variability by mixing documents from independent providers, such as different government websites or publishers. In Figure 2, we show the document categories contained in DocLayNet with their respective sizes.
+
+We did not control the document selection with regard to language. The vast majority of documents contained in DocLayNet (close to 95%) are published in English language. However, DocLayNet also contains a number of documents in other languages such as German (2.5%), French (1.0%) and Japanese (1.0%). While the document language has negligible impact on the performance of computer vision methods such as object detection and segmentation models, it might prove challenging for layout analysis methods which exploit textual features.
+
+To ensure that future benchmarks in the document-layout analysis community can be easily compared, we have split up DocLayNet into pre-defined train-, test- and validation-sets. In this way, we can avoid spurious variations in the evaluation scores due to random splitting in train-, test- and validation-sets. We also ensured that less frequent labels are represented in train and test sets in equal proportions.
+
+Table 1 shows the overall frequency and distribution of the labels among the different sets. Importantly, we ensure that subsets are only split on full-document boundaries. This avoids that pages of the same document are spread over train, test and validation set, which can give an undesired evaluation advantage to models and lead to overestimation of their prediction accuracy. We will show the impact of this decision in Section 5.
+
+In order to accommodate the different types of models currently in use by the community, we provide DocLayNet in an augmented COCO format [16]. This entails the standard COCO ground-truth file (in JSON format) with the associated page images (in PNG format, 1025 × 1025 pixels). Furthermore, custom fields have been added to each COCO record to specify document category, original document filename and page number. In addition, we also provide the original PDF pages, as well as sidecar files containing parsed PDF text and text-cell coordinates (in JSON). All additional files are linked to the primary page images by their matching filenames.
+
+Despite being cost-intense and far less scalable than automation, human annotation has several benefits over automated groundtruth generation. The first and most obvious reason to leverage human annotations is the freedom to annotate any type of document without requiring a programmatic source. For most PDF documents, the original source document is not available. The latter is not a hard constraint with human annotation, but it is for automated methods. A second reason to use human annotations is that the latter usually provide a more natural interpretation of the page layout. The human-interpreted layout can significantly deviate from the programmatic layout used in typesetting. For example, "invisible" tables might be used solely for aligning text paragraphs on columns. Such typesetting tricks might be interpreted by automated methods incorrectly as an actual table, while the human annotation will interpret it correctly as Text or other styles. The same applies to multi-line text elements, when authors decided to space them as "invisible" list elements without bullet symbols. A third reason to gather ground-truth through human annotation is to estimate a "natural" upper bound on the segmentation accuracy. As we will show in Section 4, certain documents featuring complex layouts can have different but equally acceptable layout interpretations. This natural upper bound for segmentation accuracy can be found by annotating the same pages multiple times by different people and evaluating the inter-annotator agreement. Such a baseline consistency evaluation is very useful to define expectations for a good target accuracy in trained deep neural network models and avoid overfitting (see Table 1). On the flip side, achieving high annotation consistency proved to be a key challenge in human annotation, as we outline in Section 4.
+
+## 4 ANNOTATION CAMPAIGN
+
+The annotation campaign was carried out in four phases. In phase one, we identified and prepared the data sources for annotation. In phase two, we determined the class labels and how annotations should be done on the documents in order to obtain maximum consistency. The latter was guided by a detailed requirement analysis and exhaustive experiments. In phase three, we trained the annotation staff and performed exams for quality assurance. In phase four,
+
+Table 1: DocLayNet dataset overview. Along with the frequency of each class label, we present the relative occurrence (as % of row "Total") in the train, test and validation sets. The inter-annotator agreement is computed as the mAP@0.5-0.95 metric between pairwise annotations from the triple-annotated pages, from which we obtain accuracy ranges.
+
+|                |         | % of Total   | % of Total   | % of Total   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   | triple inter-annotator mAP @ 0.5-0.95 (%)   |
+|----------------|---------|--------------|--------------|--------------|---------------------------------------------|---------------------------------------------|---------------------------------------------|---------------------------------------------|---------------------------------------------|---------------------------------------------|---------------------------------------------|
+| class label    | Count   | Train        | Test         | Val          | All                                         | Fin                                         | Man                                         | Sci                                         | Law                                         | Pat                                         | Ten                                         |
+| Caption        | 22524   | 2.04         | 1.77         | 2.32         | 84-89                                       | 40-61                                       | 86-92                                       | 94-99                                       | 95-99                                       | 69-78                                       | n/a                                         |
+| Footnote       | 6318    | 0.60         | 0.31         | 0.58         | 83-91                                       | n/a                                         | 100                                         | 62-88                                       | 85-94                                       | n/a                                         | 82-97                                       |
+| Formula        | 25027   | 2.25         | 1.90         | 2.96         | 83-85                                       | n/a                                         | n/a                                         | 84-87                                       | 86-96                                       | n/a                                         | n/a                                         |
+| List-item      | 185660  | 17.19        | 13.34        | 15.82        | 87-88                                       | 74-83                                       | 90-92                                       | 97-97                                       | 81-85                                       | 75-88                                       | 93-95                                       |
+| Page-footer    | 70878   | 6.51         | 5.58         | 6.00         | 93-94                                       | 88-90                                       | 95-96                                       | 100                                         | 92-97                                       | 100                                         | 96-98                                       |
+| Page-header    | 58022   | 5.10         | 6.70         | 5.06         | 85-89                                       | 66-76                                       | 90-94                                       | 98-100                                      | 91-92                                       | 97-99                                       | 81-86                                       |
+| Picture        | 45976   | 4.21         | 2.78         | 5.31         | 69-71                                       | 56-59                                       | 82-86                                       | 69-82                                       | 80-95                                       | 66-71                                       | 59-76                                       |
+| Section-header | 142884  | 12.60        | 15.77        | 12.85        | 83-84                                       | 76-81                                       | 90-92                                       | 94-95                                       | 87-94                                       | 69-73                                       | 78-86                                       |
+| Table          | 34733   | 3.20         | 2.27         | 3.60         | 77-81                                       | 75-80                                       | 83-86                                       | 98-99                                       | 58-80                                       | 79-84                                       | 70-85                                       |
+| Text           | 510377  | 45.82        | 49.28        | 45.00        | 84-86                                       | 81-86                                       | 88-93                                       | 89-93                                       | 87-92                                       | 71-79                                       | 87-95                                       |
+| Title          | 5071    | 0.47         | 0.30         | 0.50         | 60-72                                       | 24-63                                       | 50-63                                       | 94-100                                      | 82-96                                       | 68-79                                       | 24-56                                       |
+| Total          | 1107470 | 941123       | 99816        | 66531        | 82-83                                       | 71-74                                       | 79-81                                       | 89-94                                       | 86-91                                       | 71-76                                       | 68-85                                       |
+
+Figure 3: Corpus Conversion Service annotation user interface. The PDF page is shown in the background, with overlaid text-cells (in darker shades). The annotation boxes can be drawn by dragging a rectangle over each segment with the respective label from the palette on the right.
+
+we distributed the annotation workload and performed continuous quality controls. Phase one and two required a small team of experts only. For phases three and four, a group of 40 dedicated annotators were assembled and supervised.
+
+Phase 1: Data selection and preparation. Our inclusion criteria for documents were described in Section 3. A large effort went into ensuring that all documents are free to use. The data sources
+
+include publication repositories such as arXiv$^{3}$, government offices, company websites as well as data directory services for financial reports and patents. Scanned documents were excluded wherever possible because they can be rotated or skewed. This would not allow us to perform annotation with rectangular bounding-boxes and therefore complicate the annotation process.
+
+Preparation work included uploading and parsing the sourced PDF documents in the Corpus Conversion Service (CCS) [22], a cloud-native platform which provides a visual annotation interface and allows for dataset inspection and analysis. The annotation interface of CCS is shown in Figure 3. The desired balance of pages between the different document categories was achieved by selective subsampling of pages with certain desired properties. For example, we made sure to include the title page of each document and bias the remaining page selection to those with figures or tables. The latter was achieved by leveraging pre-trained object detection models from PubLayNet, which helped us estimate how many figures and tables a given page contains.
+
+Phase 2: Label selection and guideline. We reviewed the collected documents and identified the most common structural features they exhibit. This was achieved by identifying recurrent layout elements and lead us to the definition of 11 distinct class labels. These 11 class labels are Caption , Footnote , Formula , List-item , Pagefooter , Page-header , Picture , Section-header , Table , Text , and Title . Critical factors that were considered for the choice of these class labels were (1) the overall occurrence of the label, (2) the specificity of the label, (3) recognisability on a single page (i.e. no need for context from previous or next page) and (4) overall coverage of the page. Specificity ensures that the choice of label is not ambiguous, while coverage ensures that all meaningful items on a page can be annotated. We refrained from class labels that are very specific to a document category, such as Abstract in the Scientific Articles category. We also avoided class labels that are tightly linked to the semantics of the text. Labels such as Author and Affiliation , as seen in DocBank, are often only distinguishable by discriminating on
+
+the textual content of an element, which goes beyond visual layout recognition, in particular outside the Scientific Articles category.
+
+At first sight, the task of visual document-layout interpretation appears intuitive enough to obtain plausible annotations in most cases. However, during early trial-runs in the core team, we observed many cases in which annotators use different annotation styles, especially for documents with challenging layouts. For example, if a figure is presented with subfigures, one annotator might draw a single figure bounding-box, while another might annotate each subfigure separately. The same applies for lists, where one might annotate all list items in one block or each list item separately. In essence, we observed that challenging layouts would be annotated in different but plausible ways. To illustrate this, we show in Figure 4 multiple examples of plausible but inconsistent annotations on the same pages.
+
+Obviously, this inconsistency in annotations is not desirable for datasets which are intended to be used for model training. To minimise these inconsistencies, we created a detailed annotation guideline. While perfect consistency across 40 annotation staff members is clearly not possible to achieve, we saw a huge improvement in annotation consistency after the introduction of our annotation guideline. A few selected, non-trivial highlights of the guideline are:
+
+(1) Every list-item is an individual object instance with class label List-item . This definition is different from PubLayNet and DocBank, where all list-items are grouped together into one List object.
+
+(2) A List-item is a paragraph with hanging indentation. Singleline elements can qualify as List-item if the neighbour elements expose hanging indentation. Bullet or enumeration symbols are not a requirement.
+
+(3) For every Caption , there must be exactly one corresponding Picture or Table .
+
+(4) Connected sub-pictures are grouped together in one Picture object.
+
+(5) Formula numbers are included in a Formula object.
+
+(6) Emphasised text (e.g. in italic or bold) at the beginning of a paragraph is not considered a Section-header , unless it appears exclusively on its own line.
+
+The complete annotation guideline is over 100 pages long and a detailed description is obviously out of scope for this paper. Nevertheless, it will be made publicly available alongside with DocLayNet for future reference.
+
+Phase 3: Training. After a first trial with a small group of people, we realised that providing the annotation guideline and a set of random practice pages did not yield the desired quality level for layout annotation. Therefore we prepared a subset of pages with two different complexity levels, each with a practice and an exam part. 974 pages were reference-annotated by one proficient core team member. Annotation staff were then given the task to annotate the same subsets (blinded from the reference). By comparing the annotations of each staff member with the reference annotations, we could quantify how closely their annotations matched the reference. Only after passing two exam levels with high annotation quality, staff were admitted into the production phase. Practice iterations
+
+Figure 4: Examples of plausible annotation alternatives for the same page. Criteria in our annotation guideline can resolve cases A to C, while the case D remains ambiguous.
+
+were carried out over a timeframe of 12 weeks, after which 8 of the 40 initially allocated annotators did not pass the bar.
+
+Phase 4: Production annotation. The previously selected 80K pages were annotated with the defined 11 class labels by 32 annotators. This production phase took around three months to complete. All annotations were created online through CCS, which visualises the programmatic PDF text-cells as an overlay on the page. The page annotation are obtained by drawing rectangular bounding-boxes, as shown in Figure 3. With regard to the annotation practices, we implemented a few constraints and capabilities on the tooling level. First, we only allow non-overlapping, vertically oriented, rectangular boxes. For the large majority of documents, this constraint was sufficient and it speeds up the annotation considerably in comparison with arbitrary segmentation shapes. Second, annotator staff were not able to see each other's annotations. This was enforced by design to avoid any bias in the annotation, which could skew the numbers of the inter-annotator agreement (see Table 1). We wanted
+
+Table 2: Prediction performance (mAP@0.5-0.95) of object detection networks on DocLayNet test set. The MRCNN (Mask R-CNN) and FRCNN (Faster R-CNN) models with ResNet-50 or ResNet-101 backbone were trained based on the network architectures from the detectron2 model zoo (Mask R-CNN R50, R101-FPN 3x, Faster R-CNN R101-FPN 3x), with default configurations. The YOLO implementation utilized was YOLOv5x6 [13]. All models were initialised using pre-trained weights from the COCO 2017 dataset.
+
+|                | human   | MRCNN   | MRCNN   | FRCNN   | YOLO   |
+|----------------|---------|---------|---------|---------|--------|
+|                | human   | R50     | R101    | R101    | v5x6   |
+| Caption        | 84-89   | 68.4    | 71.5    | 70.1    | 77.7   |
+| Footnote       | 83-91   | 70.9    | 71.8    | 73.7    | 77.2   |
+| Formula        | 83-85   | 60.1    | 63.4    | 63.5    | 66.2   |
+| List-item      | 87-88   | 81.2    | 80.8    | 81.0    | 86.2   |
+| Page-footer    | 93-94   | 61.6    | 59.3    | 58.9    | 61.1   |
+| Page-header    | 85-89   | 71.9    | 70.0    | 72.0    | 67.9   |
+| Picture        | 69-71   | 71.7    | 72.7    | 72.0    | 77.1   |
+| Section-header | 83-84   | 67.6    | 69.3    | 68.4    | 74.6   |
+| Table          | 77-81   | 82.2    | 82.9    | 82.2    | 86.3   |
+| Text           | 84-86   | 84.6    | 85.8    | 85.4    | 88.1   |
+| Title          | 60-72   | 76.7    | 80.4    | 79.9    | 82.7   |
+| All            | 82-83   | 72.4    | 73.5    | 73.4    | 76.8   |
+
+to avoid this at any cost in order to have clear, unbiased baseline numbers for human document-layout annotation. Third, we introduced the feature of snapping boxes around text segments to obtain a pixel-accurate annotation and again reduce time and effort. The CCS annotation tool automatically shrinks every user-drawn box to the minimum bounding-box around the enclosed text-cells for all purely text-based segments, which excludes only Table and Picture . For the latter, we instructed annotation staff to minimise inclusion of surrounding whitespace while including all graphical lines. A downside of snapping boxes to enclosed text cells is that some wrongly parsed PDF pages cannot be annotated correctly and need to be skipped. Fourth, we established a way to flag pages as rejected for cases where no valid annotation according to the label guidelines could be achieved. Example cases for this would be PDF pages that render incorrectly or contain layouts that are impossible to capture with non-overlapping rectangles. Such rejected pages are not contained in the final dataset. With all these measures in place, experienced annotation staff managed to annotate a single page in a typical timeframe of 20s to 60s, depending on its complexity.
+
+## 5 EXPERIMENTS
+
+The primary goal of DocLayNet is to obtain high-quality ML models capable of accurate document-layout analysis on a wide variety of challenging layouts. As discussed in Section 2, object detection models are currently the easiest to use, due to the standardisation of ground-truth data in COCO format [16] and the availability of general frameworks such as detectron2 [17]. Furthermore, baseline numbers in PubLayNet and DocBank were obtained using standard object detection models such as Mask R-CNN and Faster R-CNN. As such, we will relate to these object detection methods in this
+
+Figure 5: Prediction performance (mAP@0.5-0.95) of a Mask R-CNN network with ResNet50 backbone trained on increasing fractions of the DocLayNet dataset. The learning curve flattens around the 80% mark, indicating that increasing the size of the DocLayNet dataset with similar data will not yield significantly better predictions.
+
+paper and leave the detailed evaluation of more recent methods mentioned in Section 2 for future work.
+
+In this section, we will present several aspects related to the performance of object detection models on DocLayNet. Similarly as in PubLayNet, we will evaluate the quality of their predictions using mean average precision (mAP) with 10 overlaps that range from 0.5 to 0.95 in steps of 0.05 (mAP@0.5-0.95). These scores are computed by leveraging the evaluation code provided by the COCO API [16].
+
+## Baselines for Object Detection
+
+In Table 2, we present baseline experiments (given in mAP) on Mask R-CNN [12], Faster R-CNN [11], and YOLOv5 [13]. Both training and evaluation were performed on RGB images with dimensions of 1025 × 1025 pixels. For training, we only used one annotation in case of redundantly annotated pages. As one can observe, the variation in mAP between the models is rather low, but overall between 6 and 10% lower than the mAP computed from the pairwise human annotations on triple-annotated pages. This gives a good indication that the DocLayNet dataset poses a worthwhile challenge for the research community to close the gap between human recognition and ML approaches. It is interesting to see that Mask R-CNN and Faster R-CNN produce very comparable mAP scores, indicating that pixel-based image segmentation derived from bounding-boxes does not help to obtain better predictions. On the other hand, the more recent Yolov5x model does very well and even out-performs humans on selected labels such as Text , Table and Picture . This is not entirely surprising, as Text , Table and Picture are abundant and the most visually distinctive in a document.
+
+Table 3: Performance of a Mask R-CNN R50 network in mAP@0.5-0.95 scores trained on DocLayNet with different class label sets. The reduced label sets were obtained by either down-mapping or dropping labels.
+
+| Class-count    |   11 | 6       | 5       | 4       |
+|----------------|------|---------|---------|---------|
+| Caption        |   68 | Text    | Text    | Text    |
+| Footnote       |   71 | Text    | Text    | Text    |
+| Formula        |   60 | Text    | Text    | Text    |
+| List-item      |   81 | Text    | 82      | Text    |
+| Page-footer    |   62 | 62      | -       | -       |
+| Page-header    |   72 | 68      | -       | -       |
+| Picture        |   72 | 72      | 72      | 72      |
+| Section-header |   68 | 67      | 69      | 68      |
+| Table          |   82 | 83      | 82      | 82      |
+| Text           |   85 | 84      | 84      | 84      |
+| Title          |   77 | Sec.-h. | Sec.-h. | Sec.-h. |
+| Overall        |   72 | 73      | 78      | 77      |
+
+## Learning Curve
+
+One of the fundamental questions related to any dataset is if it is "large enough". To answer this question for DocLayNet, we performed a data ablation study in which we evaluated a Mask R-CNN model trained on increasing fractions of the DocLayNet dataset. As can be seen in Figure 5, the mAP score rises sharply in the beginning and eventually levels out. To estimate the error-bar on the metrics, we ran the training five times on the entire data-set. This resulted in a 1% error-bar, depicted by the shaded area in Figure 5. In the inset of Figure 5, we show the exact same data-points, but with a logarithmic scale on the x-axis. As is expected, the mAP score increases linearly as a function of the data-size in the inset. The curve ultimately flattens out between the 80% and 100% mark, with the 80% mark falling within the error-bars of the 100% mark. This provides a good indication that the model would not improve significantly by yet increasing the data size. Rather, it would probably benefit more from improved data consistency (as discussed in Section 3), data augmentation methods [23], or the addition of more document categories and styles.
+
+## Impact of Class Labels
+
+The choice and number of labels can have a significant effect on the overall model performance. Since PubLayNet, DocBank and DocLayNet all have different label sets, it is of particular interest to understand and quantify this influence of the label set on the model performance. We investigate this by either down-mapping labels into more common ones (e.g. Caption → Text ) or excluding them from the annotations entirely. Furthermore, it must be stressed that all mappings and exclusions were performed on the data before model training. In Table 3, we present the mAP scores for a Mask R-CNN R50 network on different label sets. Where a label is down-mapped, we show its corresponding label, otherwise it was excluded. We present three different label sets, with 6, 5 and 4 different labels respectively. The set of 5 labels contains the same labels as PubLayNet. However, due to the different definition of
+
+Table 4: Performance of a Mask R-CNN R50 network with document-wise and page-wise split for different label sets. Naive page-wise split will result in GLYPH<tildelow> 10% point improvement.
+
+| Class-count    | 11   | 11   | 5   | 5    |
+|----------------|------|------|-----|------|
+| Split          | Doc  | Page | Doc | Page |
+| Caption        | 68   | 83   |     |      |
+| Footnote       | 71   | 84   |     |      |
+| Formula        | 60   | 66   |     |      |
+| List-item      | 81   | 88   | 82  | 88   |
+| Page-footer    | 62   | 89   |     |      |
+| Page-header    | 72   | 90   |     |      |
+| Picture        | 72   | 82   | 72  | 82   |
+| Section-header | 68   | 83   | 69  | 83   |
+| Table          | 82   | 89   | 82  | 90   |
+| Text           | 85   | 91   | 84  | 90   |
+| Title          | 77   | 81   |     |      |
+| All            | 72   | 84   | 78  | 87   |
+
+lists in PubLayNet (grouped list-items) versus DocLayNet (separate list-items), the label set of size 4 is the closest to PubLayNet, in the assumption that the List is down-mapped to Text in PubLayNet. The results in Table 3 show that the prediction accuracy on the remaining class labels does not change significantly when other classes are merged into them. The overall macro-average improves by around 5%, in particular when Page-footer and Page-header are excluded.
+
+## Impact of Document Split in Train and Test Set
+
+Many documents in DocLayNet have a unique styling. In order to avoid overfitting on a particular style, we have split the train-, test- and validation-sets of DocLayNet on document boundaries, i.e. every document contributes pages to only one set. To the best of our knowledge, this was not considered in PubLayNet or DocBank. To quantify how this affects model performance, we trained and evaluated a Mask R-CNN R50 model on a modified dataset version. Here, the train-, test- and validation-sets were obtained by a randomised draw over the individual pages. As can be seen in Table 4, the difference in model performance is surprisingly large: pagewise splitting gains ˜ 10% in mAP over the document-wise splitting. Thus, random page-wise splitting of DocLayNet can easily lead to accidental overestimation of model performance and should be avoided.
+
+## Dataset Comparison
+
+Throughout this paper, we claim that DocLayNet's wider variety of document layouts leads to more robust layout detection models. In Table 5, we provide evidence for that. We trained models on each of the available datasets (PubLayNet, DocBank and DocLayNet) and evaluated them on the test sets of the other datasets. Due to the different label sets and annotation styles, a direct comparison is not possible. Hence, we focussed on the common labels among the datasets. Between PubLayNet and DocLayNet, these are Picture ,
+
+Table 5: Prediction Performance (mAP@0.5-0.95) of a Mask R-CNN R50 network across the PubLayNet, DocBank & DocLayNet data-sets. By evaluating on common label classes of each dataset, we observe that the DocLayNet-trained model has much less pronounced variations in performance across all datasets.
+
+|            | Testing on   | Testing on   | Testing on   |
+|------------|--------------|--------------|--------------|
+| labels     | PLN          | DB           | DLN          |
+| Figure     | 96           | 43           | 23           |
+| Sec-header | 87           | -            | 32           |
+| Table      | 95           | 24           | 49           |
+| Text       | 96           | -            | 42           |
+| total      | 93           | 34           | 30           |
+| Figure     | 77           | 71           | 31           |
+| Table      | 19           | 65           | 22           |
+| total      | 48           | 68           | 27           |
+| Figure     | 67           | 51           | 72           |
+| Sec-header | 53           | -            | 68           |
+| Table      | 87           | 43           | 82           |
+| Text       | 77           | -            | 84           |
+| total      | 59           | 47           | 78           |
+
+Section-header , Table and Text . Before training, we either mapped or excluded DocLayNet's other labels as specified in table 3, and also PubLayNet's List to Text . Note that the different clustering of lists (by list-element vs. whole list objects) naturally decreases the mAP score for Text .
+
+For comparison of DocBank with DocLayNet, we trained only on Picture and Table clusters of each dataset. We had to exclude Text because successive paragraphs are often grouped together into a single object in DocBank. This paragraph grouping is incompatible with the individual paragraphs of DocLayNet. As can be seen in Table 5, DocLayNet trained models yield better performance compared to the previous datasets. It is noteworthy that the models trained on PubLayNet and DocBank perform very well on their own test set, but have a much lower performance on the foreign datasets. While this also applies to DocLayNet, the difference is far less pronounced. Thus we conclude that DocLayNet trained models are overall more robust and will produce better results for challenging, unseen layouts.
+
+## Example Predictions
+
+To conclude this section, we illustrate the quality of layout predictions one can expect from DocLayNet-trained models by providing a selection of examples without any further post-processing applied. Figure 6 shows selected layout predictions on pages from the test-set of DocLayNet. Results look decent in general across document categories, however one can also observe mistakes such as overlapping clusters of different classes, or entirely missing boxes due to low confidence.
+
+## 6 CONCLUSION
+
+In this paper, we presented the DocLayNet dataset. It provides the document conversion and layout analysis research community a new and challenging dataset to improve and fine-tune novel ML methods on. In contrast to many other datasets, DocLayNet was created by human annotation in order to obtain reliable layout ground-truth on a wide variety of publication- and typesettingstyles. Including a large proportion of documents outside the scientific publishing domain adds significant value in this respect.
+
+From the dataset, we have derived on the one hand reference metrics for human performance on document-layout annotation (through double and triple annotations) and on the other hand evaluated the baseline performance of commonly used object detection methods. We also illustrated the impact of various dataset-related aspects on model performance through data-ablation experiments, both from a size and class-label perspective. Last but not least, we compared the accuracy of models trained on other public datasets and showed that DocLayNet trained models are more robust.
+
+To date, there is still a significant gap between human and ML accuracy on the layout interpretation task, and we hope that this work will inspire the research community to close that gap.
+
+## REFERENCES
+
+[1] Max Göbel, Tamir Hassan, Ermelinda Oro, and Giorgio Orsi. Icdar 2013 table competition. In 2013 12th International Conference on Document Analysis and Recognition , pages 1449-1453, 2013.
+
+[2] Christian Clausner, Apostolos Antonacopoulos, and Stefan Pletschacher. Icdar2017 competition on recognition of documents with complex layouts rdcl2017. In 2017 14th IAPR International Conference on Document Analysis and Recognition (ICDAR) , volume 01, pages 1404-1410, 2017.
+
+[3] Hervé Déjean, Jean-Luc Meunier, Liangcai Gao, Yilun Huang, Yu Fang, Florian Kleber, and Eva-Maria Lang. ICDAR 2019 Competition on Table Detection and Recognition (cTDaR), April 2019. http://sac.founderit.com/.
+
+[4] Antonio Jimeno Yepes, Peter Zhong, and Douglas Burdick. Competition on scientific literature parsing. In Proceedings of the International Conference on Document Analysis and Recognition , ICDAR, pages 605-617. LNCS 12824, SpringerVerlag, sep 2021.
+
+[5] Logan Markewich, Hao Zhang, Yubin Xing, Navid Lambert-Shirzad, Jiang Zhexin, Roy Lee, Zhi Li, and Seok-Bum Ko. Segmentation for document layout analysis: not dead yet. International Journal on Document Analysis and Recognition (IJDAR) , pages 1-11, 01 2022.
+
+[6] Xu Zhong, Jianbin Tang, and Antonio Jimeno-Yepes. Publaynet: Largest dataset ever for document layout analysis. In Proceedings of the International Conference on Document Analysis and Recognition , ICDAR, pages 1015-1022, sep 2019.
+
+[7] Minghao Li, Yiheng Xu, Lei Cui, Shaohan Huang, Furu Wei, Zhoujun Li, and Ming Zhou. Docbank: A benchmark dataset for document layout analysis. In Proceedings of the 28th International Conference on Computational Linguistics , COLING, pages 949-960. International Committee on Computational Linguistics, dec 2020.
+
+[8] Riaz Ahmad, Muhammad Tanvir Afzal, and M. Qadir. Information extraction from pdf sources based on rule-based system using integrated formats. In SemWebEval@ESWC , 2016.
+
+[9] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In IEEE Conference on Computer Vision and Pattern Recognition , CVPR, pages 580-587. IEEE Computer Society, jun 2014.
+
+[10] Ross B. Girshick. Fast R-CNN. In 2015 IEEE International Conference on Computer Vision , ICCV, pages 1440-1448. IEEE Computer Society, dec 2015.
+
+[11] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence , 39(6):1137-1149, 2017.
+
+[12] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross B. Girshick. Mask R-CNN. In IEEE International Conference on Computer Vision , ICCV, pages 2980-2988. IEEE Computer Society, Oct 2017.
+
+[13] Glenn Jocher, Alex Stoken, Ayush Chaurasia, Jirka Borovec, NanoCode012, TaoXie, Yonghye Kwon, Kalen Michael, Liu Changyu, Jiacong Fang, Abhiram V, Laughing, tkianai, yxNONG, Piotr Skalski, Adam Hogan, Jebastin Nadar, imyhxy, Lorenzo Mammana, Alex Wang, Cristi Fati, Diego Montes, Jan Hajek, Laurentiu
+
+Figure 6: Example layout predictions on selected pages from the DocLayNet test-set. (A, D) exhibit favourable results on coloured backgrounds. (B, C) show accurate list-item and paragraph differentiation despite densely-spaced lines. (E) demonstrates good table and figure distinction. (F) shows predictions on a Chinese patent with multiple overlaps, label confusion and missing boxes.
+
+Diaconu, Mai Thanh Minh, Marc, albinxavi, fatih, oleg, and wanghao yang. ultralytics/yolov5: v6.0 - yolov5n nano models, roboflow integration, tensorflow export, opencv dnn support, October 2021.
+
+[14] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. CoRR , abs/2005.12872, 2020.
+
+[15] Mingxing Tan, Ruoming Pang, and Quoc V. Le. Efficientdet: Scalable and efficient object detection. CoRR , abs/1911.09070, 2019.
+
+[16] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, Lubomir D. Bourdev, Ross B. Girshick, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context, 2014.
+
+[17] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detectron2, 2019.
+
+[18] Nikolaos Livathinos, Cesar Berrospi, Maksym Lysak, Viktor Kuropiatnyk, Ahmed Nassar, Andre Carvalho, Michele Dolfi, Christoph Auer, Kasper Dinkla, and Peter W. J. Staar. Robust pdf document conversion using recurrent neural networks. In Proceedings of the 35th Conference on Artificial Intelligence , AAAI, pages 1513715145, feb 2021.
+
+[19] Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and Ming Zhou. Layoutlm: Pre-training of text and layout for document image understanding. In Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , KDD, pages 1192-1200, New York, USA, 2020. Association for Computing Machinery.
+
+[20] Shoubin Li, Xuyan Ma, Shuaiqun Pan, Jun Hu, Lin Shi, and Qing Wang. Vtlayout: Fusion of visual and text features for document layout analysis, 2021.
+
+[21] Peng Zhang, Can Li, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Fei Wu. Vsr: A unified framework for document layout analysis combining vision, semantics and relations, 2021.
+
+[22] Peter W J Staar, Michele Dolfi, Christoph Auer, and Costas Bekas. Corpus conversion service: A machine learning platform to ingest documents at scale. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining , KDD, pages 774-782. ACM, 2018.
+
+[23] Connor Shorten and Taghi M. Khoshgoftaar. A survey on image data augmentation for deep learning. Journal of Big Data , 6(1):60, 2019.
--- a/tests/data/2206.01062.pages.json
+++ b/tests/data/2206.01062.pages.json
--- a/tests/data/2206.01062.pdf
+++ b/tests/data/2206.01062.pdf
--- a/tests/data/2305.03393v1-pg9.json
+++ b/tests/data/2305.03393v1-pg9.json
--- a/tests/data/2305.03393v1-pg9.md
+++ b/tests/data/2305.03393v1-pg9.md
@ -0,0 +1,22 @@
+order to compute the TED score. Inference timing results for all experiments were obtained from the same machine on a single core with AMD EPYC 7763 CPU @2.45 GHz.
+
+## 5.1 Hyper Parameter Optimization
+
+We have chosen the PubTabNet data set to perform HPO, since it includes a highly diverse set of tables. Also we report TED scores separately for simple and complex tables (tables with cell spans). Results are presented in Table. 1. It is evident that with OTSL, our model achieves the same TED score and slightly better mAP scores in comparison to HTML. However OTSL yields a 2x speed up in the inference runtime over HTML.
+
+Table 1. HPO performed in OTSL and HTML representation on the same transformer-based TableFormer [9] architecture, trained only on PubTabNet [22]. Effects of reducing the # of layers in encoder and decoder stages of the model show that smaller models trained on OTSL perform better, especially in recognizing complex table structures, and maintain a much higher mAP score than the HTML counterpart.
+
+| #          | #          | Language   | TEDs        | TEDs              | TEDs        | mAP         | Inference   |
+|------------|------------|------------|-------------|-------------------|-------------|-------------|-------------|
+| enc-layers | dec-layers | Language   | simple      | complex           | all         | (0.75)      | time (secs) |
+| 6          | 6          | OTSL HTML  | 0.965 0.969 | 0.934 0.927       | 0.955 0.955 | 0.88 0.857  | 2.73 5.39   |
+| 4          | 4          | OTSL HTML  | 0.938 0.952 | 0.904             | 0.927       | 0.853       | 1.97        |
+|            |            | OTSL HTML  | 0.923       | 0.909 0.897 0.901 | 0.938 0.915 | 0.843       | 3.77        |
+| 2          | 4          |            | 0.945       |                   | 0.931       | 0.859 0.834 | 1.91 3.81   |
+| 4          | 2          | OTSL HTML  | 0.952 0.944 | 0.92 0.903        | 0.942 0.931 | 0.857 0.824 | 1.22 2      |
+
+## 5.2 Quantitative Results
+
+We picked the model parameter configuration that produced the best prediction quality (enc=6, dec=6, heads=8) with PubTabNet alone, then independently trained and evaluated it on three publicly available data sets: PubTabNet (395k samples), FinTabNet (113k samples) and PubTables-1M (about 1M samples). Performance results are presented in Table. 2. It is clearly evident that the model trained on OTSL outperforms HTML across the board, keeping high TEDs and mAP scores even on difficult financial tables (FinTabNet) that contain sparse and large tables.
+
+Additionally, the results show that OTSL has an advantage over HTML when applied on a bigger data set like PubTables-1M and achieves significantly improved scores. Finally, OTSL achieves faster inference due to fewer decoding steps which is a result of the reduced sequence representation.
--- a/tests/data/2305.03393v1-pg9.pages.json
+++ b/tests/data/2305.03393v1-pg9.pages.json
--- a/tests/data/2305.03393v1-pg9.pdf
+++ b/tests/data/2305.03393v1-pg9.pdf
--- a/tests/data/2305.03393v1.json
+++ b/tests/data/2305.03393v1.json
--- a/tests/data/2305.03393v1.md
+++ b/tests/data/2305.03393v1.md
@ -0,0 +1,218 @@
+## Optimized Table Tokenization for Table Structure Recognition
+
+Maksym Lysak [0000 - 0002 - 3723 - $^{6960]}$, Ahmed Nassar[0000 - 0002 - 9468 - $^{0822]}$, Nikolaos Livathinos [0000 - 0001 - 8513 - $^{3491]}$, Christoph Auer[0000 - 0001 - 5761 - $^{0422]}$, and Peter Staar [0000 - 0002 - 8088 - 0823]
+
+IBM Research {mly,ahn,nli,cau,taa}@zurich.ibm.com
+
+Abstract. Extracting tables from documents is a crucial task in any document conversion pipeline. Recently, transformer-based models have demonstrated that table-structure can be recognized with impressive accuracy using Image-to-Markup-Sequence (Im2Seq) approaches. Taking only the image of a table, such models predict a sequence of tokens (e.g. in HTML, LaTeX) which represent the structure of the table. Since the token representation of the table structure has a significant impact on the accuracy and run-time performance of any Im2Seq model, we investigate in this paper how table-structure representation can be optimised. We propose a new, optimised table-structure language (OTSL) with a minimized vocabulary and specific rules. The benefits of OTSL are that it reduces the number of tokens to 5 (HTML needs 28+) and shortens the sequence length to half of HTML on average. Consequently, model accuracy improves significantly, inference time is halved compared to HTML-based models, and the predicted table structures are always syntactically correct. This in turn eliminates most post-processing needs. Popular table structure data-sets will be published in OTSL format to the community.
+
+Keywords: Table Structure Recognition · Data Representation · Transformers · Optimization.
+
+## 1 Introduction
+
+Tables are ubiquitous in documents such as scientific papers, patents, reports, manuals, specification sheets or marketing material. They often encode highly valuable information and therefore need to be extracted with high accuracy. Unfortunately, tables appear in documents in various sizes, styling and structure, making it difficult to recover their correct structure with simple analytical methods. Therefore, accurate table extraction is achieved these days with machine-learning based methods.
+
+In modern document understanding systems [1,15], table extraction is typically a two-step process. Firstly, every table on a page is located with a bounding box, and secondly, their logical row and column structure is recognized. As of
+
+Fig. 1. Comparison between HTML and OTSL table structure representation: (A) table-example with complex row and column headers, including a 2D empty span, (B) minimal graphical representation of table structure using rectangular layout, (C) HTML representation, (D) OTSL representation. This example demonstrates many of the key-features of OTSL, namely its reduced vocabulary size (12 versus 5 in this case), its reduced sequence length (55 versus 30) and a enhanced internal structure (variable token sequence length per row in HTML versus a fixed length of rows in OTSL).
+
+today, table detection in documents is a well understood problem, and the latest state-of-the-art (SOTA) object detection methods provide an accuracy comparable to human observers [7,8,10,14,23]. On the other hand, the problem of table structure recognition (TSR) is a lot more challenging and remains a very active area of research, in which many novel machine learning algorithms are being explored [3,4,5,9,11,12,13,14,17,18,21,22].
+
+Recently emerging SOTA methods for table structure recognition employ transformer-based models, in which an image of the table is provided to the network in order to predict the structure of the table as a sequence of tokens. These image-to-sequence (Im2Seq) models are extremely powerful, since they allow for a purely data-driven solution. The tokens of the sequence typically belong to a markup language such as HTML, Latex or Markdown, which allow to describe table structure as rows, columns and spanning cells in various configurations. In Figure 1, we illustrate how HTML is used to represent the table-structure of a particular example table. Public table-structure data sets such as PubTabNet [22], and FinTabNet [21], which were created in a semi-automated way from paired PDF and HTML sources (e.g. PubMed Central), popularized primarily the use of HTML as ground-truth representation format for TSR.
+
+While the majority of research in TSR is currently focused on the development and application of novel neural model architectures, the table structure representation language (e.g. HTML in PubTabNet and FinTabNet) is usually adopted as is for the sequence tokenization in Im2Seq models. In this paper, we aim for the opposite and investigate the impact of the table structure representation language with an otherwise unmodified Im2Seq transformer-based architecture. Since the current state-of-the-art Im2Seq model is TableFormer [9], we select this model to perform our experiments.
+
+The main contribution of this paper is the introduction of a new optimised table structure language (OTSL), specifically designed to describe table-structure in an compact and structured way for Im2Seq models. OTSL has a number of key features, which make it very attractive to use in Im2Seq models. Specifically, compared to other languages such as HTML, OTSL has a minimized vocabulary which yields short sequence length, strong inherent structure (e.g. strict rectangular layout) and a strict syntax with rules that only look backwards. The latter allows for syntax validation during inference and ensures a syntactically correct table-structure. These OTSL features are illustrated in Figure 1, in comparison to HTML.
+
+The paper is structured as follows. In section 2, we give an overview of the latest developments in table-structure reconstruction. In section 3 we review the current HTML table encoding (popularised by PubTabNet and FinTabNet) and discuss its flaws. Subsequently, we introduce OTSL in section 4, which includes the language definition, syntax rules and error-correction procedures. In section 5, we apply OTSL on the TableFormer architecture, compare it to TableFormer models trained on HTML and ultimately demonstrate the advantages of using OTSL. Finally, in section 6 we conclude our work and outline next potential steps.
+
+## 2 Related Work
+
+Approaches to formalize the logical structure and layout of tables in electronic documents date back more than two decades [16]. In the recent past, a wide variety of computer vision methods have been explored to tackle the problem of table structure recognition, i.e. the correct identification of columns, rows and spanning cells in a given table. Broadly speaking, the current deeplearning based approaches fall into three categories: object detection (OD) methods, Graph-Neural-Network (GNN) methods and Image-to-Markup-Sequence (Im2Seq) methods. Object-detection based methods [11,12,13,14,21] rely on tablestructure annotation using (overlapping) bounding boxes for training, and produce bounding-box predictions to define table cells, rows, and columns on a table image. Graph Neural Network (GNN) based methods [3,6,17,18], as the name suggests, represent tables as graph structures. The graph nodes represent the content of each table cell, an embedding vector from the table image, or geometric coordinates of the table cell. The edges of the graph define the relationship between the nodes, e.g. if they belong to the same column, row, or table cell.
+
+Other work [20] aims at predicting a grid for each table and deciding which cells must be merged using an attention network. Im2Seq methods cast the problem as a sequence generation task [4,5,9,22], and therefore need an internal tablestructure representation language, which is often implemented with standard markup languages (e.g. HTML, LaTeX, Markdown). In theory, Im2Seq methods have a natural advantage over the OD and GNN methods by virtue of directly predicting the table-structure. As such, no post-processing or rules are needed in order to obtain the table-structure, which is necessary with OD and GNN approaches. In practice, this is not entirely true, because a predicted sequence of table-structure markup does not necessarily have to be syntactically correct. Hence, depending on the quality of the predicted sequence, some post-processing needs to be performed to ensure a syntactically valid (let alone correct) sequence.
+
+Within the Im2Seq method, we find several popular models, namely the encoder-dual-decoder model (EDD) [22], TableFormer [9], Tabsplitter[2] and Ye et. al. [19]. EDD uses two consecutive long short-term memory (LSTM) decoders to predict a table in HTML representation. The tag decoder predicts a sequence of HTML tags. For each decoded table cell ( <td> ), the attention is passed to the cell decoder to predict the content with an embedded OCR approach. The latter makes it susceptible to transcription errors in the cell content of the table. TableFormer address this reliance on OCR and uses two transformer decoders for HTML structure and cell bounding box prediction in an end-to-end architecture. The predicted cell bounding box is then used to extract text tokens from an originating (digital) PDF page, circumventing any need for OCR. TabSplitter [2] proposes a compact double-matrix representation of table rows and columns to do error detection and error correction of HTML structure sequences based on predictions from [19]. This compact double-matrix representation can not be used directly by the Img2seq model training, so the model uses HTML as an intermediate form. Chi et. al. [4] introduce a data set and a baseline method using bidirectional LSTMs to predict LaTeX code. Kayal [5] introduces Gated ResNet transformers to predict LaTeX code, and a separate OCR module to extract content.
+
+Im2Seq approaches have shown to be well-suited for the TSR task and allow a full end-to-end network design that can output the final table structure without pre- or post-processing logic. Furthermore, Im2Seq models have demonstrated to deliver state-of-the-art prediction accuracy [9]. This motivated the authors to investigate if the performance (both in accuracy and inference time) can be further improved by optimising the table structure representation language. We believe this is a necessary step before further improving neural network architectures for this task.
+
+## 3 Problem Statement
+
+All known Im2Seq based models for TSR fundamentally work in similar ways. Given an image of a table, the Im2Seq model predicts the structure of the table by generating a sequence of tokens. These tokens originate from a finite vocab-
+
+ulary and can be interpreted as a table structure. For example, with the HTML tokens <table> , </table> , <tr> , </tr> , <td> and </td> , one can construct simple table structures without any spanning cells. In reality though, one needs at least 28 HTML tokens to describe the most common complex tables observed in real-world documents [21,22], due to a variety of spanning cells definitions in the HTML token vocabulary.
+
+Fig. 2. Frequency of tokens in HTML and OTSL as they appear in PubTabNet.
+
+Obviously, HTML and other general-purpose markup languages were not designed for Im2Seq models. As such, they have some serious drawbacks. First, the token vocabulary needs to be artificially large in order to describe all plausible tabular structures. Since most Im2Seq models use an autoregressive approach, they generate the sequence token by token. Therefore, to reduce inference time, a shorter sequence length is critical. Every table-cell is represented by at least two tokens ( <td> and </td> ). Furthermore, when tokenizing the HTML structure, one needs to explicitly enumerate possible column-spans and row-spans as words. In practice, this ends up requiring 28 different HTML tokens (when including column- and row-spans up to 10 cells) just to describe every table in the PubTabNet dataset. Clearly, not every token is equally represented, as is depicted in Figure 2. This skewed distribution of tokens in combination with variable token row-length makes it challenging for models to learn the HTML structure.
+
+Additionally, it would be desirable if the representation would easily allow an early detection of invalid sequences on-the-go, before the prediction of the entire table structure is completed. HTML is not well-suited for this purpose as the verification of incomplete sequences is non-trivial or even impossible.
+
+In a valid HTML table, the token sequence must describe a 2D grid of table cells, serialised in row-major ordering, where each row and each column have the same length (while considering row- and column-spans). Furthermore, every opening tag in HTML needs to be matched by a closing tag in a correct hierarchical manner. Since the number of tokens for each table row and column can vary significantly, especially for large tables with many row- and column-spans, it is complex to verify the consistency of predicted structures during sequence
+
+generation. Implicitly, this also means that Im2Seq models need to learn these complex syntax rules, simply to deliver valid output.
+
+In practice, we observe two major issues with prediction quality when training Im2Seq models on HTML table structure generation from images. On the one hand, we find that on large tables, the visual attention of the model often starts to drift and is not accurately moving forward cell by cell anymore. This manifests itself in either in an increasing location drift for proposed table-cells in later rows on the same column or even complete loss of vertical alignment, as illustrated in Figure 5. Addressing this with post-processing is partially possible, but clearly undesired. On the other hand, we find many instances of predictions with structural inconsistencies or plain invalid HTML output, as shown in Figure 6, which are nearly impossible to properly correct. Both problems seriously impact the TSR model performance, since they reflect not only in the task of pure structure recognition but also in the equally crucial recognition or matching of table cell content.
+
+## 4 Optimised Table Structure Language
+
+To mitigate the issues with HTML in Im2Seq-based TSR models laid out before, we propose here our Optimised Table Structure Language (OTSL). OTSL is designed to express table structure with a minimized vocabulary and a simple set of rules, which are both significantly reduced compared to HTML. At the same time, OTSL enables easy error detection and correction during sequence generation. We further demonstrate how the compact structure representation and minimized sequence length improves prediction accuracy and inference time in the TableFormer architecture.
+
+## 4.1 Language Definition
+
+In Figure 3, we illustrate how the OTSL is defined. In essence, the OTSL defines only 5 tokens that directly describe a tabular structure based on an atomic 2D grid.
+
+The OTSL vocabulary is comprised of the following tokens:
+
+-"C" cell a new table cell that either has or does not have cell content
+
+-"L" cell left-looking cell , merging with the left neighbor cell to create a span
+
+-"U" cell up-looking cell , merging with the upper neighbor cell to create a span
+
+-"X" cell cross cell , to merge with both left and upper neighbor cells
+
+-"NL" new-line , switch to the next row.
+
+A notable attribute of OTSL is that it has the capability of achieving lossless conversion to HTML.
+
+Fig. 3. OTSL description of table structure: A - table example; B - graphical representation of table structure; C - mapping structure on a grid; D - OTSL structure encoding; E - explanation on cell encoding
+
+## 4.2 Language Syntax
+
+The OTSL representation follows these syntax rules:
+
+1. Left-looking cell rule : The left neighbour of an "L" cell must be either another "L" cell or a "C" cell.
+
+2. Up-looking cell rule : The upper neighbour of a "U" cell must be either another "U" cell or a "C" cell.
+
+## 3. Cross cell rule :
+
+: The left neighbour of an "X" cell must be either another "X" cell or a "U" cell, and the upper neighbour of an "X" cell must be either another "X" cell or an "L" cell.
+
+4. First row rule : Only "L" cells and "C" cells are allowed in the first row.
+
+5. First column rule : Only "U" cells and "C" cells are allowed in the first column.
+
+6. Rectangular rule : The table representation is always rectangular - all rows must have an equal number of tokens, terminated with "NL" token.
+
+The application of these rules gives OTSL a set of unique properties. First of all, the OTSL enforces a strictly rectangular structure representation, where every new-line token starts a new row. As a consequence, all rows and all columns have exactly the same number of tokens, irrespective of cell spans. Secondly, the OTSL representation is unambiguous: Every table structure is represented in one way. In this representation every table cell corresponds to a "C"-cell token, which in case of spans is always located in the top-left corner of the table cell definition. Third, OTSL syntax rules are only backward-looking. As a consequence, every predicted token can be validated straight during sequence generation by looking at the previously predicted sequence. As such, OTSL can guarantee that every predicted sequence is syntactically valid.
+
+These characteristics can be easily learned by sequence generator networks, as we demonstrate further below. We find strong indications that this pattern
+
+reduces significantly the column drift seen in the HTML based models (see Figure 5).
+
+## 4.3 Error-detection and -mitigation
+
+The design of OTSL allows to validate a table structure easily on an unfinished sequence. The detection of an invalid sequence token is a clear indication of a prediction mistake, however a valid sequence by itself does not guarantee prediction correctness. Different heuristics can be used to correct token errors in an invalid sequence and thus increase the chances for accurate predictions. Such heuristics can be applied either after the prediction of each token, or at the end on the entire predicted sequence. For example a simple heuristic which can correct the predicted OTSL sequence on-the-fly is to verify if the token with the highest prediction confidence invalidates the predicted sequence, and replace it by the token with the next highest confidence until OTSL rules are satisfied.
+
+## 5 Experiments
+
+To evaluate the impact of OTSL on prediction accuracy and inference times, we conducted a series of experiments based on the TableFormer model (Figure 4) with two objectives: Firstly we evaluate the prediction quality and performance of OTSL vs. HTML after performing Hyper Parameter Optimization (HPO) on the canonical PubTabNet data set. Secondly we pick the best hyper-parameters found in the first step and evaluate how OTSL impacts the performance of TableFormer after training on other publicly available data sets (FinTabNet, PubTables-1M [14]). The ground truth (GT) from all data sets has been converted into OTSL format for this purpose, and will be made publicly available.
+
+Fig. 4. Architecture sketch of the TableFormer model, which is a representative for the Im2Seq approach.
+
+We rely on standard metrics such as Tree Edit Distance score (TEDs) for table structure prediction, and Mean Average Precision (mAP) with 0.75 Intersection Over Union (IOU) threshold for the bounding-box predictions of table cells. The predicted OTSL structures were converted back to HTML format in
+
+order to compute the TED score. Inference timing results for all experiments were obtained from the same machine on a single core with AMD EPYC 7763 CPU @2.45 GHz.
+
+## 5.1 Hyper Parameter Optimization
+
+We have chosen the PubTabNet data set to perform HPO, since it includes a highly diverse set of tables. Also we report TED scores separately for simple and complex tables (tables with cell spans). Results are presented in Table. 1. It is evident that with OTSL, our model achieves the same TED score and slightly better mAP scores in comparison to HTML. However OTSL yields a 2x speed up in the inference runtime over HTML.
+
+Table 1. HPO performed in OTSL and HTML representation on the same transformer-based TableFormer [9] architecture, trained only on PubTabNet [22]. Effects of reducing the # of layers in encoder and decoder stages of the model show that smaller models trained on OTSL perform better, especially in recognizing complex table structures, and maintain a much higher mAP score than the HTML counterpart.
+
+| #          | #          | Language   | TEDs        | TEDs        | TEDs        | mAP         | Inference   |
+|------------|------------|------------|-------------|-------------|-------------|-------------|-------------|
+| enc-layers | dec-layers | Language   | simple      | complex     | all         | (0.75)      | time (secs) |
+| 6          | 6          | OTSL HTML  | 0.965 0.969 | 0.934 0.927 | 0.955 0.955 | 0.88 0.857  | 2.73 5.39   |
+| 4          | 4          | OTSL HTML  | 0.938 0.952 | 0.904 0.909 | 0.927       | 0.853       | 1.97        |
+| 2          | 4          | OTSL HTML  | 0.923       | 0.897 0.901 | 0.938 0.915 | 0.843       | 3.77        |
+|            |            |            | 0.945       |             | 0.931       | 0.859 0.834 | 1.91 3.81   |
+| 4          | 2          | OTSL HTML  | 0.952 0.944 | 0.92 0.903  | 0.942 0.931 | 0.857 0.824 | 1.22 2      |
+
+## 5.2 Quantitative Results
+
+We picked the model parameter configuration that produced the best prediction quality (enc=6, dec=6, heads=8) with PubTabNet alone, then independently trained and evaluated it on three publicly available data sets: PubTabNet (395k samples), FinTabNet (113k samples) and PubTables-1M (about 1M samples). Performance results are presented in Table. 2. It is clearly evident that the model trained on OTSL outperforms HTML across the board, keeping high TEDs and mAP scores even on difficult financial tables (FinTabNet) that contain sparse and large tables.
+
+Additionally, the results show that OTSL has an advantage over HTML when applied on a bigger data set like PubTables-1M and achieves significantly improved scores. Finally, OTSL achieves faster inference due to fewer decoding steps which is a result of the reduced sequence representation.
+
+Table 2. TSR and cell detection results compared between OTSL and HTML on the PubTabNet [22], FinTabNet [21] and PubTables-1M [14] data sets using TableFormer [9] (with enc=6, dec=6, heads=8).
+
+|              | Language   | TEDs   | TEDs    | TEDs   | mAP(0.75)   | Inference time (secs)   |
+|--------------|------------|--------|---------|--------|-------------|-------------------------|
+|              | Language   | simple | complex | all    | mAP(0.75)   | Inference time (secs)   |
+| PubTabNet    | OTSL       | 0.965  | 0.934   | 0.955  | 0.88        | 2.73                    |
+| PubTabNet    | HTML       | 0.969  | 0.927   | 0.955  | 0.857       | 5.39                    |
+| FinTabNet    | OTSL       | 0.955  | 0.961   | 0.959  | 0.862       | 1.85                    |
+| FinTabNet    | HTML       | 0.917  | 0.922   | 0.92   | 0.722       | 3.26                    |
+| PubTables-1M | OTSL       | 0.987  | 0.964   | 0.977  | 0.896       | 1.79                    |
+| PubTables-1M | HTML       | 0.983  | 0.944   | 0.966  | 0.889       | 3.26                    |
+
+## 5.3 Qualitative Results
+
+To illustrate the qualitative differences between OTSL and HTML, Figure 5 demonstrates less overlap and more accurate bounding boxes with OTSL. In Figure 6, OTSL proves to be more effective in handling tables with longer token sequences, resulting in even more precise structure prediction and bounding boxes.
+
+Fig. 5. The OTSL model produces more accurate bounding boxes with less overlap (E) than the HTML model (D), when predicting the structure of a sparse table (A), at twice the inference speed because of shorter sequence length (B),(C). "PMC2807444_006_00.png" PubTabNet. μ
+
+μ
+
+≥
+
+Fig. 6. Visualization of predicted structure and detected bounding boxes on a complex table with many rows. The OTSL model (B) captured repeating pattern of horizontally merged cells from the GT (A), unlike the HTML model (C). The HTML model also didn't complete the HTML sequence correctly and displayed a lot more of drift and overlap of bounding boxes. "PMC5406406_003_01.png" PubTabNet.
+
+## 6 Conclusion
+
+We demonstrated that representing tables in HTML for the task of table structure recognition with Im2Seq models is ill-suited and has serious limitations. Furthermore, we presented in this paper an Optimized Table Structure Language (OTSL) which, when compared to commonly used general purpose languages, has several key benefits.
+
+First and foremost, given the same network configuration, inference time for a table-structure prediction is about 2 times faster compared to the conventional HTML approach. This is primarily owed to the shorter sequence length of the OTSL representation. Additional performance benefits can be obtained with HPO (hyper parameter optimization). As we demonstrate in our experiments, models trained on OTSL can be significantly smaller, e.g. by reducing the number of encoder and decoder layers, while preserving comparatively good prediction quality. This can further improve inference performance, yielding 5-6 times faster inference speed in OTSL with prediction quality comparable to models trained on HTML (see Table 1).
+
+Secondly, OTSL has more inherent structure and a significantly restricted vocabulary size. This allows autoregressive models to perform better in the TED metric, but especially with regards to prediction accuracy of the table-cell bounding boxes (see Table 2). As shown in Figure 5, we observe that the OTSL drastically reduces the drift for table cell bounding boxes at high row count and in sparse tables. This leads to more accurate predictions and a significant reduction in post-processing complexity, which is an undesired necessity in HTML-based Im2Seq models. Significant novelty lies in OTSL syntactical rules, which are few, simple and always backwards looking. Each new token can be validated only by analyzing the sequence of previous tokens, without requiring the entire sequence to detect mistakes. This in return allows to perform structural error detection and correction on-the-fly during sequence generation.
+
+## References
+
+1. Auer, C., Dolfi, M., Carvalho, A., Ramis, C.B., Staar, P.W.J.: Delivering document conversion as a cloud service with high throughput and responsiveness. CoRR abs/2206.00785 (2022). https://doi.org/10.48550/arXiv.2206.00785 , https://doi.org/10.48550/arXiv.2206.00785
+
+2. Chen, B., Peng, D., Zhang, J., Ren, Y., Jin, L.: Complex table structure recognition in the wild using transformer and identity matrix-based augmentation. In: Porwal, U., Fornés, A., Shafait, F. (eds.) Frontiers in Handwriting Recognition. pp. 545561. Springer International Publishing, Cham (2022)
+
+3. Chi, Z., Huang, H., Xu, H.D., Yu, H., Yin, W., Mao, X.L.: Complicated table structure recognition. arXiv preprint arXiv:1908.04729 (2019)
+
+4. Deng, Y., Rosenberg, D., Mann, G.: Challenges in end-to-end neural scientific table recognition. In: 2019 International Conference on Document Analysis and Recognition (ICDAR). pp. 894-901. IEEE (2019)
+
+5. Kayal, P., Anand, M., Desai, H., Singh, M.: Tables to latex: structure and content extraction from scientific tables. International Journal on Document Analysis and Recognition (IJDAR) pp. 1-10 (2022)
+
+6. Lee, E., Kwon, J., Yang, H., Park, J., Lee, S., Koo, H.I., Cho, N.I.: Table structure recognition based on grid shape graph. In: 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC). pp. 18681873. IEEE (2022)
+
+7. Li, M., Cui, L., Huang, S., Wei, F., Zhou, M., Li, Z.: Tablebank: A benchmark dataset for table detection and recognition (2019)
+
+8. Livathinos, N., Berrospi, C., Lysak, M., Kuropiatnyk, V., Nassar, A., Carvalho, A., Dolfi, M., Auer, C., Dinkla, K., Staar, P.: Robust pdf document conversion using recurrent neural networks. Proceedings of the AAAI Conference on Artificial Intelligence 35 (17), 15137-15145 (May 2021), https://ojs.aaai.org/index.php/ AAAI/article/view/17777
+
+9. Nassar, A., Livathinos, N., Lysak, M., Staar, P.: Tableformer: Table structure understanding with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4614-4623 (June 2022)
+
+10. Pfitzmann, B., Auer, C., Dolfi, M., Nassar, A.S., Staar, P.W.J.: Doclaynet: A large human-annotated dataset for document-layout segmentation. In: Zhang, A., Rangwala, H. (eds.) KDD '22: The 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, Washington, DC, USA, August 14 - 18, 2022. pp. 3743-3751. ACM (2022). https://doi.org/10.1145/3534678.3539043 , https:// doi.org/10.1145/3534678.3539043
+
+11. Prasad, D., Gadpal, A., Kapadni, K., Visave, M., Sultanpure, K.: Cascadetabnet: An approach for end to end table detection and structure recognition from imagebased documents. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops. pp. 572-573 (2020)
+
+12. Schreiber, S., Agne, S., Wolf, I., Dengel, A., Ahmed, S.: Deepdesrt: Deep learning for detection and structure recognition of tables in document images. In: 2017 14th IAPR international conference on document analysis and recognition (ICDAR). vol. 1, pp. 1162-1167. IEEE (2017)
+
+13. Siddiqui, S.A., Fateh, I.A., Rizvi, S.T.R., Dengel, A., Ahmed, S.: Deeptabstr: Deep learning based table structure recognition. In: 2019 International Conference on Document Analysis and Recognition (ICDAR). pp. 1403-1409 (2019). https:// doi.org/10.1109/ICDAR.2019.00226
+
+14. Smock, B., Pesala, R., Abraham, R.: PubTables-1M: Towards comprehensive table extraction from unstructured documents. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 4634-4642 (June 2022)
+
+15. Staar, P.W.J., Dolfi, M., Auer, C., Bekas, C.: Corpus conversion service: A machine learning platform to ingest documents at scale. In: Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. pp. 774-782. KDD '18, Association for Computing Machinery, New York, NY, USA (2018). https://doi.org/10.1145/3219819.3219834 , https://doi.org/10. 1145/3219819.3219834
+
+16. Wang, X.: Tabular Abstraction, Editing, and Formatting. Ph.D. thesis, CAN (1996), aAINN09397
+
+17. Xue, W., Li, Q., Tao, D.: Res2tim: Reconstruct syntactic structures from table images. In: 2019 International Conference on Document Analysis and Recognition (ICDAR). pp. 749-755. IEEE (2019)
+
+18. Xue, W., Yu, B., Wang, W., Tao, D., Li, Q.: Tgrnet: A table graph reconstruction network for table structure recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 1295-1304 (2021)
+
+19. Ye, J., Qi, X., He, Y., Chen, Y., Gu, D., Gao, P., Xiao, R.: Pingan-vcgroup's solution for icdar 2021 competition on scientific literature parsing task b: Table recognition to html (2021). https://doi.org/10.48550/ARXIV.2105.01848 , https://arxiv.org/abs/2105.01848
+
+20. Zhang, Z., Zhang, J., Du, J., Wang, F.: Split, embed and merge: An accurate table structure recognizer. Pattern Recognition 126 , 108565 (2022)
+
+21. Zheng, X., Burdick, D., Popa, L., Zhong, X., Wang, N.X.R.: Global table extractor (gte): A framework for joint table identification and cell structure recognition using visual context. In: 2021 IEEE Winter Conference on Applications of Computer Vision (WACV). pp. 697-706 (2021). https://doi.org/10.1109/WACV48630.2021. 00074
+
+22. Zhong, X., ShafieiBavani, E., Jimeno Yepes, A.: Image-based table recognition: Data, model, and evaluation. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J.M. (eds.) Computer Vision - ECCV 2020. pp. 564-580. Springer International Publishing, Cham (2020)
+
+23. Zhong, X., Tang, J., Yepes, A.J.: Publaynet: largest dataset ever for document layout analysis. In: 2019 International Conference on Document Analysis and Recognition (ICDAR). pp. 1015-1022. IEEE (2019)
--- a/tests/data/2305.03393v1.pages.json
+++ b/tests/data/2305.03393v1.pages.json
--- a/tests/data/2305.03393v1.pdf
+++ b/tests/data/2305.03393v1.pdf
--- a/tests/data/redp5110.json
+++ b/tests/data/redp5110.json
--- a/tests/data/redp5110.md
+++ b/tests/data/redp5110.md
--- a/tests/data/redp5110.pages.json
+++ b/tests/data/redp5110.pages.json
--- a/tests/data/redp5110.pdf
+++ b/tests/data/redp5110.pdf
--- a/tests/data/redp5695.json
+++ b/tests/data/redp5695.json
--- a/tests/data/redp5695.md
+++ b/tests/data/redp5695.md
@ -0,0 +1,649 @@
+Front cover
+
+## IBM Cloud Pak for Data on IBM Z
+
+## Executive overview
+
+Most industries are susceptible to fraud, which poses a risk to both businesses and consumers. According to The National Health Care Anti-Fraud Association, health care fraud alone causes the nation around $68 billion annually.$^{1}$ This statistic does not include the numerous other industries where fraudulent activities occur daily. In addition, the growing amount of data that enterprises own makes it difficult for them to detect fraud. Businesses can benefit by using an analytical platform to fully integrate their data with artificial intelligence (AI) technology.
+
+With IBM Cloud Pakfi for Data on IBM Z, enterprises can modernize their data infrastructure, develop, and deploy machine learning (ML) and AI models, and instantiate highly efficient analytics deployment on IBM LinuxONE. Enterprises can create cutting-edge, intelligent, and interactive applications with embedded AI, colocate data with commercial applications, and use AI to make inferences.
+
+This IBM Redguide publication presents a high-level overview of IBM Z. It describes IBM Cloud Pak for Data (CP4D) on IBM Z and IBM LinuxONE, the different features that are supported on the platform, and how the associated features can help enterprise customers in building AI and ML models by using core transactional data, which results in decreased latency and increased throughput.
+
+This publication highlights real-time CP4D on IBM Z use cases. Real-time Clearing and Settlement Transactions, Trustworthy AI and its Role in Day-To-Day Monitoring, and the Prevention of Retail Crimes are use cases that are described in this publication. Using CP4D on IBM Z and LinuxONE, this publication shows how businesses can implement a highly efficient analytics deployment that minimizes latency, cost inefficiencies, and potential security exposures that are connected with data transportation.
+
+## IBM Z: An overview
+
+Ever wonder how many transactions a bank processes per day? What about the pace at which these transactions happen? According to an IBMfi report, 44 of 50 of the world's top banks use IBM Z mainframes for these daily transactions.$^{2}$ IBM Z is a platform that is designed for voluminous data, maximum security, real-time transaction analysis, and cost efficiency.
+
+The most recent platform for IBM Z is IBM z16™. The IBM z16 supports the following features:
+
+GLYPH<SM590000> On-chip AI acceleration
+
+GLYPH<SM590000> Quantum-safe crypto discovery
+
+GLYPH<SM590000> Simplified compliance
+
+GLYPH<SM590000> Flexible capacity
+
+GLYPH<SM590000> Modernization of applications
+
+GLYPH<SM590000> Sustainability
+
+With these features, enterprises can upgrade applications while preserving secure and resilient data.
+
+To learn more about these features, see the IBM z16 product page.
+
+Figure 1 on page 3 shows a picture of the IBM z16 mainframe.
+
+Figure 1 IBM z16
+
+## IBM z16 and IBM LinuxONE Emperor 4 features
+
+IBM Z are based on enterprise mainframe technology. Starting with transaction-based workloads and databases, IBM Z has undergone tremendous transformations in its system design for many generations to build servers that cater to Linux-based workloads and security with a cyberresilient system, and support quantum computing and modernization by using a hybrid cloud with a focus on data and AI.
+
+Figure 2 provides a snapshot of the IBM Z processor roadmap, which depicts the journey of transformation and improvement.
+
+Figure 2 IBM Z: Processor roadmap
+
+The IBM z16 and IBM LinuxONE Emperor 4 are the latest of the IBM Z, and they are developed with a 'built to build' focus to provide a powerful, cyberresilient, open, and secure platform for business with an extra focus on sustainability to help build sustainable data centers. Although the z16 server can host both IBM z/OSfi and Linux workloads, LinuxONE Emperor 4 is built to host Linux only workloads with a focus on consolidation and resiliency. Depending on the workload, consolidation from numerous x86 servers into a LinuxONE Emperor 4 can help reduce energy consumption by 75% and data center floor space by 50%, which helps to achieve the sustainability goals of the organization.
+
+Figure 3 on page 5 shows a summary of the system design of IBM LinuxONE Emperor 4 with the IBM Telum™ processor. The IBM Telum processor chip is designed to run enterprise applications efficiently where their data resides to embed AI with super low latency. The support for higher bandwidth and I/O rates is supported through FCP Express cards with an endpoint security solution. The memory subsystem supports up to 40 TB of memory.
+
+Figure 3 System design of IBM z16 LinuxONE Emperor 4
+
+The IBM z16 and IBM LinuxONE Emperor 4 servers are built with 7-nm technology at a 5.2 GHz speed. They consist of four dual-chip modules (DCMs) per central processor complex (CPC) drawer, each of which is built with two 8-core Telum processor chips that has "first in the industry" on-chip acceleration for mid-transaction, real-time AI inferencing, which supports many different use cases, including fraud detection.
+
+Each core has access to a huge private 32 MB L2 cache where up to 16 MB of the L2 cache of an inactive core can be used as virtual cache (L3 / L4) by neighboring active cores on the chip. This cache helps address translation and access checking by prefetching the same virtual cache into the L2 cache. The virtual cache also includes Neural Network Processing Assist instructions and direct memory access with protection, and per chip GZIP compression.
+
+Figure 4 provides more information about the features of AI Accelerator integration with the IBM Z processor cores.
+
+Figure 4 IBM z16 on-chip AI Accelerator integration with IBM Z processor cores
+
+The IBM z16 and IBM LinuxONE Emperor 4 server platforms are built with the hardware features that are shown in Figure 4 with addressing data and AI workloads in mind. Regardless of where the ML and deep learning (DL) frameworks are used to build and train data and AI models, the inferencing on existing enterprise application data can happen along currently running enterprise business applications. CP4D 4.6 supports Tensorflow and IBM Snap ML frameworks, which are optimized to use the on-chip AI Accelerator during inferencing. Support for various other frameworks is planned for future releases.
+
+Figure 5 on page 7 shows the seamless integration of AI into existing enterprises workloads on the IBM z16 while leveraging the underlying hardware capabilities.
+
+Figure 5 Seamless integration
+
+## What is Cloud Pak for Data on IBM Z
+
+IBM Cloud Pak for Data allows enterprises to simplify, unify, and automate the delivery of data and AI. It categorizes the activities within the journey to AI as four rungs of the AI Ladder: Collect, Organize, Analyze, and Infuse. For more information about each of the AI Ladder rungs, see Become Data Driven with IBM Z Infused Data Fabric , REDP-5680.
+
+CP4D on IBM Z provides enterprises with a resilient and secure private cloud platform. You can use it to create ML and AI models that may be included into modern intelligent applications. You also can use it to use and construct applications for mission-critical data. With CP4D on IBM Z, enterprises can lower data movement latency, cost inefficiencies, and potential security exposures. Enterprises can safely store and access their most important company data, and leverage their current infrastructure by using cutting-edge hybrid cloud applications. Enterprises can combine their current database applications without any rewrites, which results in reduced cost and complexity. Lastly, by using CP4D on IBM Z, enterprises can update their database infrastructure to benefit from easier management, a quicker time to value, and lower operating expenses.
+
+Figure 6 shows a solution overview of CP4D. The infrastructure alternatives are shown at the bottom, and they include IBM Z and LinuxONE. They all leverage Red Hat OpenShift. Common Foundational Services come next, which offer clarity throughout the data and AI lifecycle, that is, from user access management to monitoring and service provisioning. A high-level view of the services is shown in the middle section. The services have several different capabilities that span the AI hierarchy. The platform can be expanded, and it offers a seamless user experience for all distinct personas across the AI lifecycle, from data gathering through AI infusion.
+
+Figure 6 Solution overview of Cloud Pak for Data
+
+We highlight the four main pillars that make IBM Z the correct infrastructure for CP4D:
+
+GLYPH<SM590000> Performance and Scale
+
+GLYPH<SM590000> Embedded Accelerators
+
+GLYPH<SM590000> Reliability and Availability
+
+GLYPH<SM590000> Security and Governance.
+
+From a performance perspective, CP4D on IBM Z provides your data and AI with high transaction processing and a powerful infrastructure. From the embedded accelerators perspective, CP4D on IBM Z can investigate each transaction thanks to a cutting-edge DL inference technology even in the most demanding, sensitive, and latency-prone real-time workloads. From a reliability perspective, CP4D on IBM Z provides high availability and resiliency. Lastly from the security perspective, CP4D on IBM Z is suitable for protecting sensitive data and AI models for enterprises in highly regulated industries or those industries that are worried about security.
+
+## Cloud Pak for Data capabilities on IBM Z and IBM LinuxONE
+
+With CP4D on IBM Z and IBM LinuxONE, users can develop, train, and deploy AI and ML models. Users can accomplish this task by using the CP4D IBM Watsonfi Studio and IBM Watson Machine Learning (WLM) services. By using these two fundamental services, users can accomplish the following tasks:
+
+GLYPH<SM590000> Provision various containerized databases.
+
+GLYPH<SM590000> Explore, clean, shape, and alter data by using Data Refinery.
+
+GLYPH<SM590000> Use project-specific data that is uploaded, or connect to distant data.
+
+GLYPH<SM590000> Create Spark run times and applications.
+
+GLYPH<SM590000> Create, build, evaluate, and deploy analytics and ML models with trust and transparency.
+
+GLYPH<SM590000> Leverage the AI Integrated Accelerator for TensorFlow 2.7.2 and Snap ML 1.9.
+
+For more information about the specifics of these capabilities, see Capabilities on Linux on IBM Z and IBM LinuxONE.
+
+## Open-source ecosystem
+
+These days, innovation and product development are not limited to closed doors within an organization. In any industry sector, the solutions include a mix of proprietary code addressing the core business solution that is supported or integrated into other software components from open source. In some cases, enterprises business solutions also are built from open-source community offerings. Thus, open-source software becomes an important ingredient in modern-day solution building.
+
+IBM actively participates in various open-source communities as part of steering boards defining the roadmap of the community, and also in contributing code to make the community a better place for everyone to participate. Red Hat also actively participates in various open-source communities and makes extensive contributions. In open-source communities, although most open-source development happens on x86 / amd64 or the Intel architecture, the same open-source software is used by other architectures, such as IBM Power (ppc64le), IBM Z and IBM LInuxONE (s390x), ARM, and Sparc. So, the availability of an open-source ecosystem on any architecture is key and critical to business.
+
+On IBM Z and IBM LinuxONE (s390x) architecture, there is a huge open-source support ecosystem that ranges from operating systems such as Linux; application run times; cloud and container services; DevOps and automation; big data; observability; analytics; databases; and storage. The ecosystem on IBM Z and IBM LinuxONE is growing.
+
+IBM Z and IBM LinuxONE include much open-source software in their ecosystem. You can see the growing list of open-source software for IBM Z and LinuxONE at The Growing Ecosystem of Open-Source Software for IBM Z and LinuxONE.
+
+IBM Z and IBM LinuxONE are available to various communities to include support for s390x builds as part of their community's continuous integration and continuous delivery (CI/CD). Also, for open-source community developers, infrastructure resources are available on a no-charge basis through the IBM LinuxONE community cloud.
+
+CP4D includes a mix of open-source and proprietary data and AI runtime databases; open-source run times like Python; open-source data platforms like Anaconda; ML and DL frameworks like Pytorch and Tensorflow; and thousands of reusable Python packages. All of them are available and supported on s390x architecture to provide seamless parity with x86 architecture and a seamless experience for enterprise data scientists, architects, and data and AI solution developers on IBM Z and IBM LinuxONE platforms.
+
+Anaconda is one of the open-source data platforms that provide Python and R based data science ML frameworks; analytics and data visualization tools; and open-source data science tools and libraries like Conda, XGBoost, and SciKit-Learn. Anaconda runs natively on Linux on IBM Z and IBM LinuxONE, and on IBM z/OS Container Extensions (zcX) on z/OS. For more information, see Announcing Anaconda for Linux on IBM Z and LinuxONE.
+
+In addition to strong, open-source ecosystem support for application development on Linux and enterprise operating systems, a new generation of IBM Z and IBM LinuxONE servers (IBM z16™) also have strong platform support, and AI acceleration capabilities that can be leveraged by open-source software to perform better on the server infrastructure. For example, the recently released CP4D 4.6 has Tensorflow and IBM SnapML frameworks that leverage the AI accelerators when running on an IBM z16 server.
+
+So, to summarize, there is a huge, growing data and AI open source ecosystem that is supported and optimized on IBM Z and IBM LinuxONE servers.
+
+## Why AI on IBM Z
+
+Data and AI playing a major role in the modernization story to enable the digital transformation journey of every organization. Many organizations recognize the business value of infusing AI into their infrastructure. CP4D provides the cloud-native solution to put your data to work. With CP4D, all your data users can collaborate from a single, unified interface that supports many services that work together, including collecting data, organizing the data, analyzing the data, and infusing AI.
+
+Traditional ML models' power most of today's ML applications in business and among AI practitioners. CP4D supports traditional ML frameworks for training and inferencing, such as Scikit-learn, Snap ML, and XGBoost. Snap ML is a library that provides high-speed training and inferencing of ML models that leverage the AI accelerator while running on an IBM z16 (Linux on IBM Z). CP4D supports DL frameworks such as TensorFlow and PyTorch. TensorFlow is a DL framework that leverages the AI accelerator while running on an IBM z16 (Linux on IBM Z).
+
+Figure 7 on page 11 provides an overview of the components that are supported on CP4D on IBM Z. You can leverage Watson Studio for model building, training, and validation, and WML for deployment of the model. Eventually, applications can use the AI inference endpoint to score the model.
+
+Figure 7 Developing, training, and deploying an AI model on Cloud Pak for Data on IBM Z and IBM LinuxONE
+
+In summary, here are some of the reasons why you should choose AI on IBM Z:
+
+GLYPH<SM590000> World-class AI inference platform for enterprise workloads:
+
+-Embedded accelerators: A centralized on-chip AI accelerator that is shared by all cores.
+
+-Industry standard AI ecosystem: Many industry open-source data science frameworks are available on the platform.
+
+-Seamlessly integrate AI into existing enterprise workload stacks: Train anywhere, and then deploy on IBM Z.
+
+GLYPH<SM590000> Security: Encrypted memory, and improved trusted execution environments.
+
+GLYPH<SM590000> Sustainability: Reduce your energy consumption with real-time monitoring tools about the energy consumption of the system.
+
+## AI use cases
+
+With billions of transactions per day in many of today's industries, it is key to get real-time insights about what is happening in your data. AI on the IBM Z stack understands these situations, and it delivers in-transaction inference in real time and at scale.
+
+Core banking solutions running on IBM Z that are involved in processing inbound transactions need real-time fraud detection to prevent fraud. Other types of possible use cases might be credit risk analysis, anti-money laundering, loan approval, fraud detection in payments, and instant payments.
+
+For insurance companies, a pressing use case would be claims processing. For markets and trading, clearing and settlement use cases are paramount.
+
+For the health care industry, medical image processing (such as MRIs and x-rays), skin cancer detection, and patient monitoring activities such as infant motion analysis, is important.
+
+For the airline industry, processes such as air traffic management, flight management systems, and flight maintenance predictions are use cases that are ideal candidates for using AI on IBM Z.
+
+In the following sections, we describe the following use cases:
+
+GLYPH<SM590000> "Use case 1: Responsible AI augmented with risk and regulatory compliance" on page 12 AI model lifecycle governance, risk management, and regulatory compliance are key to the success of the enterprises. It is imperative to adopt a typical AI model lifecycle to protect new end-to-end risks.
+
+GLYPH<SM590000> "Use case 2: Credit default risk assessment" on page 22
+
+Core banking solutions running on IBM Z that are involved in processing inbound transactions need real-time fraud detection to prevent fraud. Other types of possible use cases might be credit risk analysis, anti-money laundering, loan approval, fraud detection in payments, and instant payments.
+
+GLYPH<SM590000> "Use case 3: Clearing and settlement" on page 25
+
+The use of AI can help to predict which trades or transactions have high risk exposures, and propose solutions for a more efficient settlement process.
+
+GLYPH<SM590000> "Use case 4: Remaining Useful Life of an aircraft engine" on page 27 We describe how AI can help to avoid unplanned aircraft downtime by determining the remaining time or cycles that an aircraft engine is likely to operate before failure.
+
+GLYPH<SM590000> "Use case 5: AI-powered video analytics on an infant's motions for health prediction" on page 30
+
+In this section, we describe how AI can predict an infant's health conditions by monitoring real-time body movements.
+
+## Use case 1: Responsible AI augmented with risk and regulatory compliance
+
+Advancement in AI is changing the world, and organizations must adopt AI to embrace new challenges daily. Many enterprises see tremendous value in adopting AI and ML technologies while establishing organization trust in the models, underlying data, and the process to be followed. An AI model lifecycle can be a daunting task.
+
+How mature is your AI governance? In this section, we provide a use case demonstrating the trustworthiness of AI and its importance in daily monitoring.
+
+## Industry challenges
+
+Here are the three main reasons why organizations struggle with the adoption of AI:
+
+GLYPH<SM590000> Scaling with growing regulations
+
+GLYPH<SM590000> Lack of confidence in operationalized AI (making responsible AI)
+
+GLYPH<SM590000> Challenges around managing the risk throughout the entire AI workflow
+
+## Scaling with growing regulations
+
+Laws and regulations in the data and AI space are accelerating, and many countries are proposing strict AI policies. Countries are monitoring adherence of these policies by the enterprises and imposing fines for any violations. Responding to these regulations are challenging global organizations where multiple regulations apply. For enterprises, it is important to adopt AI policies when there is change, and to validate explainable models to protect against discrimination.
+
+## Responsible AI
+
+Responsible AI protects against loss of data privacy, and reduced customer loyalty and trust. A data scientist cannot maximize accuracy and model performance above all other concerns. Practicing responsible AI is a best practice, and you must establish protection and validation to ensure that any models that are placed into production are fair and explainable.
+
+## Risks throughout the entire AI workflow
+
+Organizations need to mitigate risk of the following items:
+
+GLYPH<SM590000> Deciding not to use certain technologies or practices
+
+GLYPH<SM590000> Using personal information when needed and with a user's consent
+
+GLYPH<SM590000> Ensuring automated decisions are free from bias
+
+GLYPH<SM590000> Customer confidence by providing explanations for business decisions
+
+GLYPH<SM590000> Fraud to the organization and to customer's accounts
+
+GLYPH<SM590000> Delays in putting models into production
+
+In fact, in a recent survey, these concerns were echoed by real AI adopters when asked what aspects of trust are most important to them. Although explaining how AI decides is the primary concern, all of these concerns are important.
+
+The key point here is that risk exists throughout the entire AI lifecycle starting with the underlying data and the business justification behind the "why" of the project and continuing into production. Without a formalized process, there is no way to mitigate these risks to unlock the scale that is required to make automated decisions profitable. With these decisions, the business can operate proactively instead of reactively.
+
+For example, a business can start testing a model before production for fairness metrics. For this task, enterprises need an end-to-end workflow with approvals to mitigate these risks and increase the scale of AI investments, as shown in Figure 8, which presents a typical AI model lifecycle in an enterprise.
+
+Figure 8 Typical AI model lifecycle
+
+Due to regulations, more stakeholders adopt the typical AI model lifecycle to protect their brand from new end-to-end risks. To ensure various aspects of both regulatory compliance and security, the personas that must be involved include the chief financial officer (CFO), chief marketing officer (CMO), chief data officer (CDO), HR, and chief regulatory officer (CRO), along with the data engineers, data scientists, and business analysts, who build AI workflows.
+
+## IBM governance solution for IBM Z
+
+AI model lifecycle governance, risk management, and regulatory compliance are key to the success of enterprises.
+
+AI governance is a comprehensive framework that uses a set of automated processes, methodologies, and tools to manage an organization's use of AI. Consistent principles guiding the design, development, deployment, and monitoring of models are critical in driving responsible and trustworthy AI. AI governance includes processes that trace and record the origin of data, models (including associated metadata), and pipelines for audits. The details of entry should include the techniques that trained each model, the hyperparameters that were used, and the metrics from testing phases. These details provide increased transparency into the model's behavior throughout the lifecycle, the data that was influential in its development, and the possible risks.
+
+In a world where trust, transparency and explainable AI matters, every organization wants compliance along with the comfort of understanding how analytic insights and decisions are made. The following sections describe some of the principles and organizational requirements for AI governance.
+
+## Lifecycle governance
+
+Lifecycle governance helps you manage your business information throughout its lifecycle, that is, from creation to deletion. IBM AI governance addresses the problems that challenge records managements:
+
+GLYPH<SM590000> Monitor, catalog, and govern AI models from anywhere throughout the AI lifecycle.
+
+GLYPH<SM590000> Automate the capture of model metadata for report generation.
+
+GLYPH<SM590000> Drive transparent and explainable AI at scale.
+
+GLYPH<SM590000> Increase accuracy of predictions by identifying how AI is used and where it is lagging.
+
+## Risk management
+
+Risk management is used in IBM AI governance to identify, manage, monitor, and report on risk and compliance initiatives at scale:
+
+GLYPH<SM590000> Automate facts and workflow management to comply with business standards.
+
+GLYPH<SM590000> Use dynamic dashboards for clear and concise customizable results.
+
+GLYPH<SM590000> Enhanced collaboration across multiple regions and geographies.
+
+## Regulatory compliance
+
+Regulatory compliance is a set of rules that organizations must follow to protect sensitive information and ensure human safety. Any business that works with digital assets, consumer data, health regulations, employee safety, and private communications is subject to regulatory compliance.$^{3}$ The IBM AI governance solution for IBM Z includes the following tasks:
+
+GLYPH<SM590000> Help adhere to external AI regulations for audit and compliance.
+
+GLYPH<SM590000> Convert external AI regulations into policies for automatic enforcement.
+
+GLYPH<SM590000> Use dynamic dashboards for compliance status across policies and regulations.
+
+Enterprises can develop AI models and deploy them by using IBM Watson Studio or WML on CP4D on Red Hat OpenShift on a virtual machine that is based on IBM z/VM or Red Hat Enterprise Linux KVM on IBM Z. AI governance on IBM LinuxONE is supported in the following two ways:
+
+GLYPH<SM590000> Monitor the AI models with Watson OpenScale on CP4D on Red Hat OpenShift on a virtual machine on IBM Z.
+
+GLYPH<SM590000> Enterprises can develop AI models by creating and training models by using Watson Studio and development tools such as Jupyter Notebook or JupyterLab, and then deploying the model onto WML on CP4D on Red Hat OpenShift on a virtual machine on IBM Z. Then, these enterprises can achieve end-end AI governance by running AI Factsheets, IBM Watson OpenScale, and IBM Watson OpenPagesfi on CP4D on x86.
+
+Figure 9 on page 16 shows the end-to-end flow for a remote AI governance solution.
+
+Figure 9 Remote AI governance solution end-to-end flow
+
+To achieve end-to-end AI governance, complete the following steps:
+
+1. Create a model entry in IBM OpenPages by using CP4D on a x86 platform, as shown in Figure 10.
+
+Figure 10 Creating a model entry in IBM OpenPages
+
+2. Train a model by using Watson Studio and by using development tools such as Jupyter Notebook or JupyterLab on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, as shown in Figure 11.
+
+Figure 11 Training an AI model by using Watson Studio
+
+3. Deploy the model by using WML on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, as shown in Figure 12.
+
+Figure 12 Deploying an AI model by using WML on Cloud Pak for Data
+
+4. Track the external model lifecycle by browsing through the Catalogs/Platform assets catalog by using AI Factsheets and OpenPages while using CP4D on an x86 platform, as shown in Figure 13. The external model (deployed on CP4D on Red Hat OpenShift on a virtual machine on IBM Z) is saved as a platform asset catalog on the x86 platform.
+
+Figure 13 External model
+
+You can track the model through each stage of the model lifecycle, as shown in Figure 14, by using AI Factsheets and OpenPages.
+
+Figure 14 Tracking the model
+
+You can see that the model facts are tracked and synchronized to IBM OpenPages for risk management, as shown in Figure 15.
+
+Figure 15 Model facts that are tracked and synchronized to IBM OpenPages on an x86 platform
+
+5. Create an external model by using IBM OpenScale on the x86 platform, as shown in Figure 16.
+
+Figure 16 Creating an external model on an x86 platform
+
+IBM OpenScale provides a comprehensive dashboard that tracks fairness, quality monitoring, drift, and explainability of a model. Fairness determines whether your model produces biased outcomes. Quality determines how well your model predicts outcomes. Drift is the degradation of predictive performance over time. A sample is shown in Figure 17 on page 21.
+
+Figure 17 IBM OpenScale dashboard that is used to monitor the external model
+
+You developed and deployed the AI model by using Watson Studio, WML on CP4D on Red Hat OpenShift on a virtual machine on IBM Z, and end-to-end AI model governance by leveraging AI Factsheets, OpenScale, and OpenPages on CP4D on a x86 platform. Figure 18 shows end-to-end AI governance when using IBM OpenPages, AI Factsheets, and OpenScale.
+
+Figure 18 Final result: End-to-end AI governance when using IBM OpenPages, AI Factsheets, and OpenScale
+
+## Use case 2: Credit default risk assessment
+
+In today's world, many individuals or businesses seeking loans to meet their growing business needs often look to financial institutions. Financial institutions can offer loans to individuals or businesses and charge interest based on the current market situations.
+
+## Industry challenges
+
+Financial institutions must make an accurate decision about whether to sanction a loan or not, and judging the likelihood of default is the difference between a successful and unsuccessful loan portfolio. In a traditional scenario, an experienced banker can judge someone's likelihood of default, but that is not an efficient method for judgment as a business grows.
+
+## Predictions of credit default risk assessment
+
+In the modern world, growing business institutions can no longer rely on only experienced bankers to decide whether to sanction a loan knowing that there is a probability that the borrower might default on their loans. A better choice is to rely on technological advancements that can help with reasoning based on facts, such as leveraging credit risk modeling techniques to process the historical data of past borrowers to understand their credit behavior and make a more informed decision about whether to lend money, how much money, and decide on the tenure to close the loan.
+
+Financial institutions can leverage AI solutions by using ML techniques to predict the credit risk. Applying AI to credit risk modeling techniques can benefit institutions in decision-making, and thus can help better manage the exposure to credit risk.
+
+Figure 19 on page 23 shows a sample architecture about how to design and develop an AI model for credit risk assessment on IBM Z. An IBM WebSpherefi Application Server is used for handling in-bound transactions, and CP4D is used for AI model lifecycle management that includes building, training, and deploying the model.
+
+Figure 19 Architecture for credit risk prediction by using an ML AI model on IBM Z
+
+A data scientist can leverage Watson Studio to develop and train an AI model and WML to deploy and score the model. In this sample architecture, the WML Python run time leverages the ML framework, IBM Snap Machine Learning (Snap ML), for scoring, can leverage an integrated AI accelerator at the time of model import.
+
+Then, the banking loan approval team can send a loan applicant request to the IBM WebSphere Application Server, which can make a request to the AI inference endpoint. The AI inference engine scores the transaction and sends the result back to the loan approval team. Based on the results, the approval team can decide on whether to approve a loan or not, and also decide how much they can lend, timelines, and other factors.
+
+The transaction system that is shown in Figure 19 uses IBM WebSphere Liberty as an application server, but you also can use an IBM Open Libertyfi application server or any application server that can send RESTful API communications.
+
+Models are frequently developed and tested in many platforms and languages, such as Python, Scala, R, and Go. Models can leverage ML frameworks like scikit-learn, Snap ML, or XGBoost, or DL frameworks like TensorFlow or PyTorch. Training a model can be done on any platform if you have enough computing power for complex models, but moving that model into production requires careful testing to ensure that transactions are not delayed, especially if you plan to run the model within a transaction.
+
+We showed how IBM Z enable customers to use AI frameworks to detect credit risk. Now, we look at how you can leverage CP4D and TensorFlow on IBM Z to detect the credit risk.
+
+Figure 20 shows an architecture for predicting credit risk by using DL on IBM Z.
+
+Figure 20 Architecture for credit risk prediction by using DL on IBM Z
+
+Data scientists can start creating and training a DL AI model by using a Jupyter Notebook instance and Watson Studio. Then, they can deploy the model by using WML on CP4D running on IBM Z, which provides an endpoint. Other applications, including the IBM WebSphere server, can produce credit risk results by using the model's endpoint.
+
+In summary, here are some considerations for developing real-time AI models, such as credit risk assessment:
+
+GLYPH<SM590000> A preference for in-platform run times of the model, such as faster execution results.
+
+GLYPH<SM590000> Less overhead in the end-to-end flows might improve scoring time.
+
+GLYPH<SM590000> If you are using models that are not deployable, CP4D offers a custom Python run time to build your own stack if they are not available on the platform.
+
+GLYPH<SM590000> AI inferencing based on ML or DL models can increase the accuracy of better credit risk assessment.
+
+GLYPH<SM590000> Using IBM z16 and on-chip AI acceleration with the Telum chip that is embedded with regular Integrated Facility for Linux (IFLs) provides an execution speed for your transactions that cannot be achieved by other means.
+
+## Use case 3: Clearing and settlement
+
+Clearing and settlements involve banks or financial institutions sending and receiving wire transfers by using secure interbank payments networks that can clear or settle numerous transactions. When an individual or business entity initiates a wire transfer, clearing begins the fund delivery process. Banks can begin the settlement phase either immediately after clearing takes place or later, mostly at the end of the business day.
+
+## Industry challenge
+
+Banks and financial institutions must deal with high-risk transactions that can lead to loss. Moreover, these transactions can lead to regulatory violations and extra compliance costs.
+
+## Clearing and settlement solution
+
+Use AI to predict which trades or transactions have high risk exposures, and propose solutions for a more efficient settlement process. The expedited remediation of questionable transactions can prevent costly consequences, regulatory violations, and negative business impacts.
+
+In financial institutions, finding which financial transactions are legitimate and which transactions are fraudulent is of paramount importance. In this section, we go through a use case where we use AI to predict which trades or transactions have high risk exposures, and propose solutions for a more efficient settlement process. The expedited remediation of questionable transactions can prevent costly consequences, regulatory violations, and negative business impacts to financial institutions.
+
+The goal is to predict in real time whether the transaction being processed might be a fraudulent transaction or not. To achieve this goal, we build an ML model that can do this prediction for the financial institution. Because there would be many transactions being processed at any point by the financial institution, it is important to perform this prediction of fraudulent transactions in near-real time in a few milliseconds.
+
+One possible solution is to build and train a TensorFlow based DL model that learns from the historical data and predicts the fraudulent transactions. CP4D on IBM Z and IBM LinuxONE is a suitable product where this task can be achieved and the model deployed, and coming up with a serving endpoint.
+
+Figure 21 provides a high-level diagram of a clearing and settlement use case for financial transactions that uses CP4D on IBM Z and IBM LinuxONE.
+
+Figure 21 Clearing and settlement use case for financial transactions by using Cloud Pak for Data
+
+Here are the steps of the high-level process flow:
+
+1. Create a connection to a database (for example, an IBM Db2fi database) where the historical data will be used for ML model building.
+
+2. Read the data from the database and prepare the data for AI by using the Data Refinery tool in CP4D.
+
+3. A Jupyter Notebook or JupyterLab IDE that is provided by the Watson Studio component in CP4D helps us build and train the AI model. The trained model can be saved into a WML repository.
+
+4. Deploy the saved model into a deployment space for batch deployment.
+
+5. Create a batch deployment by using any of these interfaces:
+
+a. Watson Studio user interface from an Analytics deployment space.
+
+b. WML Python client.
+
+c. WML REST APIs.
+
+6. A hardware configuration can be chosen for the deployment.
+
+7. A batch deployment processes input data from a file, data connection, or connected data in a storage bucket, and writes the output to a selected destination.
+
+8. One way to run batch deployment to predict or score is to create and run a batch deployment job.
+
+9. Provide an input data type:
+
+a. Inline data for entering a JSON format payload.
+
+b. Select Data asset , click Select data source , and then specify your asset.
+
+10.The output data type can be a new output file or a connected data asset.
+
+11.A Kubernetes admin can change the maximum number of concurrent batch jobs that can be run.
+
+12.Get the deployment endpoint URL. For more information, see Getting the deployment endpoint URL.
+
+## Summary
+
+With this use case, we attempted to demonstrate how to predict, in real time, whether the transaction that is being processed might be a fraudulent transaction or not. By using the method, you have the following advantages:
+
+GLYPH<SM590000> No Impact to SLAs and the batch process window.
+
+GLYPH<SM590000> Proactively stop losses, and lower operational, regulatory, and compliance costs.
+
+GLYPH<SM590000> The solution is using a DL framework like TensorFlow for high-performing, low latency scoring.
+
+## Use case 4: Remaining Useful Life of an aircraft engine
+
+In this use case, we describe how an airline can deploy an AI model for inferencing by using IBMfi zSystems.
+
+Remaining Useful Life (RUL) is the remaining time or cycles that an aircraft engine is likely to operate without any failure. In this case, it is the equivalent of the number of flights remaining for the engine after the last flight. By estimating RUL, the operator can decide on the next maintenance schedule and avoid unplanned downtime.
+
+Figure 22 provides an overview of the inferencing architecture for the RUL of an aircraft engine when using IBM Z.
+
+Figure 22 Inferencing architecture on IBM Z
+
+Because we are looking into data-driven model development, the data set of our target is the run-to-failure data of the engine. We are looking into a supervised learning problem, and we use regression techniques to learn from the data. DL techniques such as Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRU) are our choice because we are looking into a time series data set. TensorFlow or PyTorch frameworks are leveraged to create models. AI governance monitors the data and model drift to maintain the model quality throughout the model's life.
+
+Open-source data from NASA was used to build the AI model, which then was deployed on CP4D. CP4D enables the data-scientist's journey from modeling to deployment in a seamless process. Data engineers leverage Db2 to host the data set, which includes the training, testing, and validation of a data set. Since data is hosted on Db2, you can expect low latency while retrieving the data and serve data security needs because Db2 is hosted on the IBM Z platform. Data is fetched by the data refinery to do the necessary pre-processing and data imputations. You can use the programming languages Golang or C++ for real-time predictions, depending on customer needs. For more information about this topic, see "Use case 3: Clearing and settlement" on page 25.
+
+Model building is done on Watson Studio, leveraging the high-performance computing hardware on IBM Z. You can train the model anywhere (on your own hardware or the cloud) and bring the model directly into CP4D, which provides data scientists with the flexibility of implementation choices.
+
+We used LSTM to build the AI model and used the training data. The model was continuously evaluated to model convergence. The final model is tested with the test data, which is never exposed at the time of training to make sure that the model works.
+
+This model is deployed on WML on CP4D and runs on IBM Z. If required, the trained model can be converted to the Open Neural Network Exchange (ONNX) format before deployment. Based on project requirements, IBM Z supports high-throughput, low latency inference requirements by leveraging an AI accelerator.
+
+For decision-making about an aircraft engine's life, it is important to be able to explain the model predictions from end to end. This explainability may be global or local. Global explainability enables decision-makers to evaluate the trained model in general from the subject matter expert (SME) point of view. Local explainability enables the operator to validate the reasons behind the present inference and relate it to the past data points, which are an indicative cause of the prediction.
+
+The AI governance components such as IBM OpenScale on CP4D support explainability and manages the drifts in data and concept. OpenPages and AI FactSheet together can alert the stakeholders about important events through a dashboard and allow course correction at any point.
+
+Client-side applications can invoke a REST apiserver that handles some preprocessing of an incoming request before initiating the inference pipeline. Efficiencies might be needed in real-time applications, and inference response time can be reduced by adopting low-level programming while components are communicating.
+
+Figure 23 on page 29 provides a more in-depth view of the architecture of an AI-based predictive maintenance application.
+
+Figure 23 In-depth architectural view
+
+In summary, consider the following points while developing an AI-based predictive maintenance application:
+
+GLYPH<SM590000> CP4D offers a Python run time to build a custom solution stack, but also supports different components like Watson Studio, WML, Db2, Data Refinery, OpenScale, AI Factsheets, and OpenPages.
+
+GLYPH<SM590000> The trustworthiness of the predicted output is important for critical use cases.
+
+GLYPH<SM590000> IBM Z provides high data security and low latency requirements at scale for the critical applications.
+
+GLYPH<SM590000> A data scientist can choose to train the model and deploy it on CP4D seamlessly with the latest tech stack that is available.
+
+GLYPH<SM590000> The AIOps and MLOps supported by CP4D to track AI model and data lifecycle throughout the application lifecycle.
+
+## Use case 5: AI-powered video analytics on an infant's motions for health prediction
+
+Each year, approximately 5 million newborns worldwide are suffering from a neuro-developmental disorder. Due to the lack of early diagnoses and intervention, many infants are disabled and abandoned, especially in countries with limited numbers of pediatricians with extensive experience in neuro-developmental disorders. This situation is a conundrum that plagues many families around the world.
+
+Infant motion analysis plays critical importance to understanding and comprehending healthy childhood development. In infants, monitoring their poses provides information about their health that can lead to a better prediction of early developmental risk assessment and diagnosis.
+
+Adults use different techniques and methods to express their feelings (like sick, happy, stressed, or hungry), but this case is usually different for infants who cannot express their feelings. Based on the baby movements, AI can predict their expression or health.
+
+In this use case, we examine how AI-powered video analytics can assist new parents and hospitals by addressing pose-based real-time body movements of the infants (such as arching back, head banging, kicking legs, rubbing eyes, stretching, and sucking fingers). During the initial months of a baby's life, spontaneous movements might indicate later developmental disorders, such as cerebral palsy, Rett syndrome, and autism spectrum disorders.
+
+## Industry challenges
+
+There are video surveillance systems that are installed for monitoring an infant's movement in many hospitals or homes so that any problem can be witnessed and potentially even stopped before they take place. These systems require much manual work to monitor the real-stream videos and intervene when a problem is detected.
+
+There is a certain amount of trust that you must place on the person who monitors a surveillance system to ensure that the job is being done effectively and efficiently, and that the surveillance system is being vigilantly watched. Because of the dependency on these manual efforts, you need something "smart" that monitors constantly the surveillance system and detect problems effectively.
+
+AI is shaping the controls of surveillance that can map and track occurrences with self-learning abilities, AI can improve on human operations and analyze video footage in real time to alert the hospitals or parents if any anomalies are identified.
+
+Video processing a stream of data from surveillance systems and then performing advance analytics and detecting anomalies quickly is a significant challenge in the industry.
+
+## Infant motion analytics in real time
+
+AI is the current "market trend evolution" in video analytics and advancing the decision-making capabilities of the human mind. DL-based computer vision AI techniques are being widely adopted by various industries to solve real-time problems. These techniques improve the detection and prediction accuracy without increasing the hardware cost exponentially. For users, AI greatly reduces the workload of the monitoring staff and provides benefits by detecting unusual incidents and solving many video forensic problems.
+
+S
+
+Figure 24 Architecture for AI-powered video analytics
+
+Live camera feeds or recorded videos of an infant's movement are the inputs for a pose detection model. This video streaming data was stored in IBM Cloudfi Object Storage for image processing. Video data must be transformed into frames so that the infant's body poses can be detected. These post-estimation components of the pipeline predict the location of all 17-person key points with 3 degrees of freedom each (x, y location and visibility) plus two virtual alignment key points. This approach also embraces a compute-intensive heat map prediction of infant body posture.
+
+When changes in body posture or movement happen, analytics can be performed, and a threshold can be set for the angle of the body and posture movements. An analysis can be performed on movement that is based on that threshold to help to predict an infant's health index in the output video stream by leveraging the IBM z16 on-chip AI acceleration, which provides an execution speed in real time on an edge device, which cannot be achieved by other means.
+
+We can leverage the following AI technology stack for this use case:
+
+GLYPH<SM590000> Convolutional neural network: Build an artificial neural network model on video streaming and images.
+
+GLYPH<SM590000> TensorFlow: A DL back-end framework that is based on TensorFlow.
+
+GLYPH<SM590000> Mediapipe: A library that helps with video streaming processing and prediction of human pose estimation.
+
+GLYPH<SM590000> OpenCV: A real-time computer vision library that helps perform image processing.
+
+CP4D was used to build and deploy the AI-powered video analytics on infant's motion for health prediction use case on IBM Z. IBM Z with AI accelerator enables faster inference for detecting face and body movements and performing angle analytics in real time.
+
+Figure 24 shows an architectural diagram about how to design and develop an AI model for real-time body pose detection on IBM Z. A deep convolutional neural network architecture was trained on the task of infant pose estimation on the custom data set by leveraging IBM Cloud Pak for Data.
+
+WML was used for deployment of the pose detection model and generated notifications to users with web and mobile applications, and it integrates with Fitbit for push notifications so that hospitals and parents can take preventive actions.
+
+## Additional resources
+
+GLYPH<SM590000> The Cloud Pak for Data 4.5 on IBM Z Overview Demo video provides an overview of some of the more important features of CP4D on IBM Z.
+
+GLYPH<SM590000> IBM Cloud Pak for Data Tutorials.
+
+GLYPH<SM590000> Here are some additional use cases that use the data science frameworks that are available as part of CP4D on IBM Z and IBM LinuxONE:
+
+-Payment Card Fraud Detection by using TensorFlow on CP4D on IBM Z and IBM LinuxONE is a payment card fraud detection use case.
+
+-Fashion-MNIST clothing classification with PyTorch on Cloud Pak for Data on IBM Z and IBM LinuxONE is a Fashion-MNIST clothing classification use case.
+
+-Payment Card Fraud Prevention by using Snap ML on IBM Cloud Pak for Data on Red Hat OpenShift on a virtual machine on IBM Z and IBM LinuxONE, which leverage the z16 integrated AI accelerator describes a use case that uses Snap Machine Learning in Cloud Pak for Data on IBM Z and IBM LinuxONE. It is a Snap ML use case.
+
+A companion video can be found at Credit Card Fraud Detection by using Snap ML on IBM Cloud Pak for Data on IBM Z and IBM LinuxONE.
+
+## Summary
+
+This IBM Redbooksfi publication presented an overview of how IBM Cloud Pak for Data on IBM Z can modernize your data infrastructure; develop and deploy ML and AI models; and instantiate highly efficient analytics deployment on IBM LinuxONE. This publication demonstrated these tasks by guiding the reader through five common use cases where CP4D on IBM Z and IBM LinuxONE uses the different features that are supported on the platform, and showing how the associated features can help an enterprise to build AI and ML models with core transactional data, which results in a highly efficient analytics deployment that minimizes latency, cost inefficiencies, and potential security exposures that are connected with data transportation.
+
+## Authors
+
+This publication was produced by a team of specialists from around the world working with the IBM Redbooks team:
+
+Jasmeet Bhatia is an AI on IBM Z Product Manager who supports CP4D on IBM Z. She has 2.5 years of combined experience as a data scientist and a product manager. Jasmeet lives in San Francisco, California and holds a Bachelor of Arts degree in Data Science. She is working on her Master of Science degree in Data Science. Her area of expertise includes AI, data science, and product management.
+
+Ravi Gummadi is a Technical Leader for CP4D on Linux on IBM Z and IBM LinuxONE in India. He has 18+ years of experience in the design and development of enterprise software for various platforms, including IBM Z and IBM LinuxONE. He holds a master's degree in computer science and engineering from the Indian Institute of Technology Madras (IIT Madras). His areas of expertise include compilers, virtualization, big data analytics, containers, data, and AI, with a special focus on open-source ecosystems.
+
+Chandra Shekhar Reddy Potula is a Lead AI on zSystems team Architect for Linux on IBM Z and LinuxONE in India. He has 18+ years of experience in the design and development of enterprise software and firmware for various platforms, including IBM Z and LinuxONE. He holds a degree in computer science of engineering from Jawaharlal Nehru Technological University (JNTU). His areas of expertise include networking, virtualization, containers, data, and AI, with a special focus on open-source ecosystems.
+
+Srirama Sharma is a Lead Technical Architect for IBM Cloud Pak, IBM Instanafi, IBM Turbonomicfi, and Red Hat Advanced Cluster Management for Kubernetes (RHACM) on IBM Z and LinuxONE. He has 18+ years of experience in UNIX and Linux application and device driver development. He designs ISV solutions on IBM Systems and IBM Blockchainfi. He also works on cloud-native adoption of enterprise solutions on IBM Z and LinuxONE. Srirama holds a Bachelor of Engineering degree in computer science from Visvesvaraya Technological University (VTU). He lives in Bangalore, Karnataka. His areas of expertise include UNIX and Linux systems programming, virtualization, performance benchmarking of Financial Services Sector (FSS) industry solutions, open-source ecosystems, server infrastructure, and cloud-native adoption and modernization.
+
+Thanks to the following people for their contributions to this project:
+
+Lydia Parziale, Project Manager IBM Redbooks, Poughkeepsie Center
+
+Shin Kelly Yang, AI on IBM Z Product Management IBM US
+
+Tom Ramey, Anna Shugol, Andrew Sica, Jonathan Sloan, Elpida Tzortzatos, Meeta Vouk, IBM
+
+## Now you can become a published author, too!
+
+Here's an opportunity to spotlight your skills, grow your career, and become a published author-all at the same time! Join an IBM Redbooks residency project and help write a book in your area of expertise, while honing your experience using leading-edge technologies. Your efforts will help to increase product acceptance and customer satisfaction, as you expand your network of technical contacts and relationships. Residencies run from two to six weeks in length, and you can participate either in person or as a remote resident working from your home base.
+
+Find out more about the residency program, browse the residency index, and apply online at:
+
+ibm.com /redbooks/residencies.html
+
+## Stay connected to IBM Redbooks
+
+GLYPH<SM590000> Find us on LinkedIn:
+
+http://www.linkedin.com/groups?home=&gid=2130806
+
+GLYPH<SM590000> Explore new Redbooks publications, residencies, and workshops with the IBM Redbooks weekly newsletter:
+
+https://www.redbooks.ibm.com/Redbooks.nsf/subscribe?OpenForm
+
+GLYPH<SM590000> Stay current on recent Redbooks publications with RSS Feeds:
+
+http://www.redbooks.ibm.com/rss.html
+
+## Notices
+
+This information was developed for products and services offered in the US. This material might be available from IBM in other languages. However, you may be required to own a copy of the product or product version in that language in order to access it.
+
+IBM may not offer the products, services, or features discussed in this document in other countries. Consult your local IBM representative for information on the products and services currently available in your area. Any reference to an IBM product, program, or service is not intended to state or imply that only that IBM product, program, or service may be used. Any functionally equivalent product, program, or service that does not infringe any IBM intellectual property right may be used instead. However, it is the user's responsibility to evaluate and verify the operation of any non-IBM product, program, or service.
+
+IBM may have patents or pending patent applications covering subject matter described in this document. The furnishing of this document does not grant you any license to these patents. You can send license inquiries, in writing, to:
+
+IBM Director of Licensing, IBM Corporation, North Castle Drive, MD-NC119, Armonk, NY 10504-1785, US
+
+INTERNATIONAL BUSINESS MACHINES CORPORATION PROVIDES THIS PUBLICATION "AS IS" WITHOUT WARRANTY OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF NON-INFRINGEMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Some jurisdictions do not allow disclaimer of express or implied warranties in certain transactions, therefore, this statement may not apply to you.
+
+This information could include technical inaccuracies or typographical errors. Changes are periodically made to the information herein; these changes will be incorporated in new editions of the publication. IBM may make improvements and/or changes in the product(s) and/or the program(s) described in this publication at any time without notice.
+
+Any references in this information to non-IBM websites are provided for convenience only and do not in any manner serve as an endorsement of those websites. The materials at those websites are not part of the materials for this IBM product and use of those websites is at your own risk.
+
+IBM may use or distribute any of the information you provide in any way it believes appropriate without incurring any obligation to you.
+
+The performance data and client examples cited are presented for illustrative purposes only. Actual performance results may vary depending on specific configurations and operating conditions.
+
+Information concerning non-IBM products was obtained from the suppliers of those products, their published announcements or other publicly available sources. IBM has not tested those products and cannot confirm the accuracy of performance, compatibility or any other claims related to non-IBM products. Questions on the capabilities of non-IBM products should be addressed to the suppliers of those products.
+
+Statements regarding IBM's future direction or intent are subject to change or withdrawal without notice, and represent goals and objectives only.
+
+This information contains examples of data and reports used in daily business operations. To illustrate them as completely as possible, the examples include the names of individuals, companies, brands, and products. All of these names are fictitious and any similarity to actual people or business enterprises is entirely coincidental.
+
+## COPYRIGHT LICENSE:
+
+This information contains sample application programs in source language, which illustrate programming techniques on various operating platforms. You may copy, modify, and distribute these sample programs in any form without payment to IBM, for the purposes of developing, using, marketing or distributing application programs conforming to the application programming interface for the operating platform for which the sample programs are written. These examples have not been thoroughly tested under all conditions. IBM, therefore, cannot guarantee or imply reliability, serviceability, or function of these programs. The sample programs are provided "AS IS", without warranty of any kind. IBM shall not be liable for any damages arising out of your use of the sample programs.
+
+## Trademarks
+
+IBM, the IBM logo, and ibm.com are trademarks or registered trademarks of International Business Machines Corporation, registered in many jurisdictions worldwide. Other product and service names might be trademarks of IBM or other companies. A current list of IBM trademarks is available on the web at "Copyright and trademark information" at http://www.ibm.com/legal/copytrade.shtml
+
+The following terms are trademarks or registered trademarks of International Business Machines Corporation, and might also be trademarks or registered trademarks in other countries.
+
+| Db2fi IBMfi          | IBM Watsonfi   | Redbooks (log o) fi Turbon   |
+|----------------------|----------------|------------------------------|
+|                      | IBM z16™       | omicfi                       |
+| IBM Blockchainfi     | Instanafi      | WebSpherefi                  |
+| IBM Cloudfi IBM Clou | Open Libertyfi | z/OSfi                       |
+| d Pakfi              | OpenPagesfi    | z16™                         |
+| IBM Telum™           | Redbooksfi     |                              |
+
+The following terms are trademarks of other companies:
+
+Intel, Intel logo, Intel Inside logo, and Intel Centrino logo are trademarks or registered trademarks of Intel Corporation or its subsidiaries in the United States and other countries.
+
+The registered trademark Linuxfi is used pursuant to a sublicense from the Linux Foundation, the exclusive licensee of Linus Torvalds, owner of the mark on a worldwide basis.
+
+Red Hat and OpenShift are trademarks or registered trademarks of Red Hat, Inc. or its subsidiaries in the United States and other countries.
+
+UNIX is a registered trademark of The Open Group in the United States and other countries.
+
+Other company, product, or service names may be trademarks or service marks of others.
+
+Back cover
+
+REDP-5695-00
+
+ISBN 0738461067
+
+Printed in U.S.A.
--- a/tests/data/redp5695.pages.json
+++ b/tests/data/redp5695.pages.json
--- a/tests/data/redp5695.pdf
+++ b/tests/data/redp5695.pdf
--- a/tests/test_backend_docling_parse.py
+++ b/tests/test_backend_docling_parse.py
@ -0,0 +1,64 @@
+from pathlib import Path
+
+import pytest
+
+from docling.backend.docling_parse_backend import (
+    DoclingParseDocumentBackend,
+    DoclingParsePageBackend,
+)
+from docling.datamodel.base_models import BoundingBox
+
+
+@pytest.fixture
+def test_doc_path():
+    return Path("./tests/data/2206.01062.pdf")
+
+
+def test_text_cell_counts():
+    pdf_doc = Path("./tests/data/redp5695.pdf")
+
+    doc_backend = DoclingParseDocumentBackend(pdf_doc, "123456xyz")
+
+    for page_index in range(0, doc_backend.page_count()):
+        last_cell_count = None
+        for i in range(10):
+            page_backend: DoclingParsePageBackend = doc_backend.load_page(0)
+            cells = list(page_backend.get_text_cells())
+
+            if last_cell_count is None:
+                last_cell_count = len(cells)
+
+            if len(cells) != last_cell_count:
+                assert (
+                    False
+                ), "Loading page multiple times yielded non-identical text cell counts"
+            last_cell_count = len(cells)
+
+
+def test_get_text_from_rect(test_doc_path):
+    doc_backend = DoclingParseDocumentBackend(test_doc_path, "123456xyz")
+    page_backend: DoclingParsePageBackend = doc_backend.load_page(0)
+
+    # Get the title text of the DocLayNet paper
+    textpiece = page_backend.get_text_in_rect(
+        bbox=BoundingBox(l=102, t=77, r=511, b=124)
+    )
+    ref = "DocLayNet: A Large Human-Annotated Dataset for Document-Layout Analysis"
+
+    assert textpiece.strip() == ref
+
+
+def test_crop_page_image(test_doc_path):
+    doc_backend = DoclingParseDocumentBackend(test_doc_path, "123456xyz")
+    page_backend: DoclingParsePageBackend = doc_backend.load_page(0)
+
+    # Crop out "Figure 1" from the DocLayNet paper
+    im = page_backend.get_page_image(
+        scale=2, cropbox=BoundingBox(l=317, t=246, r=574, b=527)
+    )
+    # im.show()
+
+
+def test_num_pages(test_doc_path):
+    doc_backend = DoclingParseDocumentBackend(test_doc_path, "123456xyz")
+    doc_backend.page_count() == 9
--- a/tests/test_backend_pdfium.py
+++ b/tests/test_backend_pdfium.py
@ -0,0 +1,64 @@
+from pathlib import Path
+
+import pytest
+
+from docling.backend.pypdfium2_backend import (
+    PyPdfiumDocumentBackend,
+    PyPdfiumPageBackend,
+)
+from docling.datamodel.base_models import BoundingBox
+
+
+@pytest.fixture
+def test_doc_path():
+    return Path("./tests/data/2206.01062.pdf")
+
+
+def test_text_cell_counts():
+    pdf_doc = Path("./tests/data/redp5695.pdf")
+
+    doc_backend = PyPdfiumDocumentBackend(pdf_doc, "123456xyz")
+
+    for page_index in range(0, doc_backend.page_count()):
+        last_cell_count = None
+        for i in range(10):
+            page_backend: PyPdfiumPageBackend = doc_backend.load_page(0)
+            cells = list(page_backend.get_text_cells())
+
+            if last_cell_count is None:
+                last_cell_count = len(cells)
+
+            if len(cells) != last_cell_count:
+                assert (
+                    False
+                ), "Loading page multiple times yielded non-identical text cell counts"
+            last_cell_count = len(cells)
+
+
+def test_get_text_from_rect(test_doc_path):
+    doc_backend = PyPdfiumDocumentBackend(test_doc_path, "123456xyz")
+    page_backend: PyPdfiumPageBackend = doc_backend.load_page(0)
+
+    # Get the title text of the DocLayNet paper
+    textpiece = page_backend.get_text_in_rect(
+        bbox=BoundingBox(l=102, t=77, r=511, b=124)
+    )
+    ref = "DocLayNet: A Large Human-Annotated Dataset for\r\nDocument-Layout Analysis"
+
+    assert textpiece.strip() == ref
+
+
+def test_crop_page_image(test_doc_path):
+    doc_backend = PyPdfiumDocumentBackend(test_doc_path, "123456xyz")
+    page_backend: PyPdfiumPageBackend = doc_backend.load_page(0)
+
+    # Crop out "Figure 1" from the DocLayNet paper
+    im = page_backend.get_page_image(
+        scale=2, cropbox=BoundingBox(l=317, t=246, r=574, b=527)
+    )
+    # im.show()
+
+
+def test_num_pages(test_doc_path):
+    doc_backend = PyPdfiumDocumentBackend(test_doc_path, "123456xyz")
+    doc_backend.page_count() == 9
--- a/tests/test_e2e_conversion.py
+++ b/tests/test_e2e_conversion.py
@ -0,0 +1,51 @@
+from pathlib import Path
+
+from docling.backend.docling_parse_backend import DoclingParseDocumentBackend
+from docling.backend.pypdfium2_backend import PyPdfiumDocumentBackend
+from docling.datamodel.base_models import PipelineOptions
+from docling.datamodel.document import ConversionResult
+from docling.document_converter import DocumentConverter
+
+from .verify_utils import verify_conversion_result
+
+GENERATE = False
+
+
+def get_pdf_paths():
+
+    # Define the directory you want to search
+    directory = Path("./tests/data")
+
+    # List all PDF files in the directory and its subdirectories
+    pdf_files = sorted(directory.rglob("*.pdf"))
+    return pdf_files
+
+
+def get_converter():
+
+    pipeline_options = PipelineOptions()
+    pipeline_options.do_ocr = False
+    pipeline_options.do_table_structure = True
+    pipeline_options.table_structure_options.do_cell_matching = True
+
+    converter = DocumentConverter(
+        pipeline_options=pipeline_options,
+        pdf_backend=DoclingParseDocumentBackend,
+    )
+
+    return converter
+
+
+def test_e2e_conversions():
+
+    pdf_paths = get_pdf_paths()
+    converter = get_converter()
+
+    for pdf_path in pdf_paths:
+        print(f"converting {pdf_path}")
+
+        doc_result: ConversionResult = converter.convert_single(pdf_path)
+
+        verify_conversion_result(
+            input_path=pdf_path, doc_result=doc_result, generate=GENERATE
+        )
--- a/tests/test_interfaces.py
+++ b/tests/test_interfaces.py
@ -0,0 +1,69 @@
+from io import BytesIO
+from pathlib import Path
+
+import pytest
+
+from docling.backend.docling_parse_backend import DoclingParseDocumentBackend
+from docling.backend.pypdfium2_backend import PyPdfiumDocumentBackend
+from docling.datamodel.base_models import DocumentStream, PipelineOptions
+from docling.datamodel.document import ConversionResult, DocumentConversionInput
+from docling.document_converter import DocumentConverter
+
+from .verify_utils import verify_conversion_result
+
+
+def get_pdf_path():
+
+    pdf_path = Path("./tests/data/2305.03393v1-pg9.pdf")
+    return pdf_path
+
+
+@pytest.fixture
+def converter():
+
+    pipeline_options = PipelineOptions()
+    pipeline_options.do_ocr = False
+    pipeline_options.do_table_structure = True
+    pipeline_options.table_structure_options.do_cell_matching = True
+
+    converter = DocumentConverter(
+        pipeline_options=pipeline_options,
+        pdf_backend=DoclingParseDocumentBackend,
+    )
+
+    return converter
+
+
+def test_convert_single(converter: DocumentConverter):
+
+    pdf_path = get_pdf_path()
+    print(f"converting {pdf_path}")
+
+    doc_result: ConversionResult = converter.convert_single(pdf_path)
+    verify_conversion_result(input_path=pdf_path, doc_result=doc_result)
+
+
+def test_batch_path(converter: DocumentConverter):
+
+    pdf_path = get_pdf_path()
+    print(f"converting {pdf_path}")
+
+    conv_input = DocumentConversionInput.from_paths([pdf_path])
+
+    results = converter.convert(conv_input)
+    for doc_result in results:
+        verify_conversion_result(input_path=pdf_path, doc_result=doc_result)
+
+
+def test_batch_bytes(converter: DocumentConverter):
+
+    pdf_path = get_pdf_path()
+    print(f"converting {pdf_path}")
+
+    buf = BytesIO(pdf_path.open("rb").read())
+    docs = [DocumentStream(filename=pdf_path.name, stream=buf)]
+    conv_input = DocumentConversionInput.from_streams(docs)
+
+    results = converter.convert(conv_input)
+    for doc_result in results:
+        verify_conversion_result(input_path=pdf_path, doc_result=doc_result)
--- a/tests/verify_utils.py
+++ b/tests/verify_utils.py
@ -0,0 +1,148 @@
+import json
+from pathlib import Path
+from typing import List
+
+from docling_core.types import BaseText
+from docling_core.types import Document as DsDocument
+from pydantic import TypeAdapter
+from pydantic.json import pydantic_encoder
+
+from docling.datamodel.base_models import ConversionStatus, Page
+from docling.datamodel.document import ConversionResult
+
+
+def verify_cells(doc_pred_pages: List[Page], doc_true_pages: List[Page]):
+
+    assert len(doc_pred_pages) == len(
+        doc_true_pages
+    ), "pred- and true-doc do not have the same number of pages"
+
+    for pid, page_true_item in enumerate(doc_true_pages):
+
+        num_true_cells = len(page_true_item.cells)
+        num_pred_cells = len(doc_pred_pages[pid].cells)
+
+        assert (
+            num_true_cells == num_pred_cells
+        ), f"num_true_cells!=num_pred_cells {num_true_cells}!={num_pred_cells}"
+
+        for cid, cell_true_item in enumerate(page_true_item.cells):
+
+            cell_pred_item = doc_pred_pages[pid].cells[cid]
+
+            true_text = cell_true_item.text
+            pred_text = cell_pred_item.text
+
+            assert true_text == pred_text, f"{true_text}!={pred_text}"
+
+            true_bbox = cell_true_item.bbox.as_tuple()
+            pred_bbox = cell_pred_item.bbox.as_tuple()
+            assert (
+                true_bbox == pred_bbox
+            ), f"bbox is not the same: {true_bbox} != {pred_bbox}"
+
+    return True
+
+
+def verify_maintext(doc_pred: DsDocument, doc_true: DsDocument):
+
+    assert len(doc_true.main_text) == len(
+        doc_pred.main_text
+    ), f"document has different length of main-text than expected. {len(doc_true.main_text)}!={len(doc_pred.main_text)}"
+
+    for l, true_item in enumerate(doc_true.main_text):
+        if isinstance(true_item, BaseText):
+            pred_item = doc_pred.main_text[l]
+
+            assert isinstance(
+                pred_item, BaseText
+            ), f"{pred_item} is not a BaseText element, but {true_item} is."
+            assert true_item.text == pred_item.text
+
+    return True
+
+
+def verify_tables(doc_pred: DsDocument, doc_true: DsDocument):
+    assert len(doc_true.tables) == len(
+        doc_pred.tables
+    ), "document has different count of tables than expected."
+
+    for l, true_item in enumerate(doc_true.tables):
+        pred_item = doc_pred.tables[l]
+
+        assert (
+            true_item.num_rows == pred_item.num_rows
+        ), "table does not have the same #-rows"
+        assert (
+            true_item.num_cols == pred_item.num_cols
+        ), "table does not have the same #-cols"
+
+        for i, row in enumerate(true_item.data):
+            for j, col in enumerate(true_item.data[i]):
+
+                assert (
+                    true_item.data[i][j].text == pred_item.data[i][j].text
+                ), "table-cell does not have the same text"
+
+    return True
+
+
+def verify_output(doc_pred: DsDocument, doc_true: DsDocument):
+
+    assert verify_maintext(doc_pred, doc_true), "verify_maintext(doc_pred, doc_true)"
+    assert verify_tables(doc_pred, doc_true), "verify_tables(doc_pred, doc_true)"
+
+    return True
+
+
+def verify_md(doc_pred_md, doc_true_md):
+    return doc_pred_md == doc_true_md
+
+
+def verify_conversion_result(
+    input_path: Path, doc_result: ConversionResult, generate=False
+):
+    PageList = TypeAdapter(List[Page])
+
+    assert (
+        doc_result.status == ConversionStatus.SUCCESS
+    ), f"Doc {input_path} did not convert successfully."
+
+    doc_pred_pages: List[Page] = doc_result.pages
+    doc_pred: DsDocument = doc_result.output
+    doc_pred_md = doc_result.render_as_markdown()
+
+    pages_path = input_path.with_suffix(".pages.json")
+    json_path = input_path.with_suffix(".json")
+    md_path = input_path.with_suffix(".md")
+
+    if generate:  # only used when re-generating truth
+        with open(pages_path, "w") as fw:
+            fw.write(json.dumps(doc_pred_pages, default=pydantic_encoder))
+
+        with open(json_path, "w") as fw:
+            fw.write(json.dumps(doc_pred, default=pydantic_encoder))
+
+        with open(md_path, "w") as fw:
+            fw.write(doc_pred_md)
+    else:  # default branch in test
+        with open(pages_path, "r") as fr:
+            doc_true_pages = PageList.validate_json(fr.read())
+
+        with open(json_path, "r") as fr:
+            doc_true = DsDocument.model_validate_json(fr.read())
+
+        with open(md_path, "r") as fr:
+            doc_true_md = fr.read()
+
+        assert verify_cells(
+            doc_pred_pages, doc_true_pages
+        ), f"Mismatch in PDF cell prediction for {input_path}"
+
+        # assert verify_output(
+        #     doc_pred, doc_true
+        # ), f"Mismatch in JSON prediction for {input_path}"
+
+        assert verify_md(
+            doc_pred_md, doc_true_md
+        ), f"Mismatch in Markdown prediction for {input_path}"