Skip to content
Draft
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions CHANGELOG.md
Original file line number Diff line number Diff line change
@@ -1,5 +1,9 @@
## 0.27.19-dev0

### Enhancements

- **Render the PDF once in `hi_res`.** Layout inference, OCR and image extraction each rendered every page of the PDF again. The pages are now rendered a single time and shared, and `save_elements` keeps one page image open instead of re-decoding the page for every element. Takes effect with `unstructured-inference>=1.6.14`; older versions keep rendering per stage.

### Maintenance

- **Start CI tests without waiting for setup and lint.** Lint installs only its own dependency group, extras jobs install only the system packages their tests use, and a new push cancels the pull request's previous CI run.
Expand Down
30 changes: 30 additions & 0 deletions test_unstructured/partition/pdf_image/test_ocr.py
Original file line number Diff line number Diff line change
Expand Up @@ -673,3 +673,33 @@ def test_pass_down_agents(mock_ocr_get_instance, mocker, mock_page):
"language": "eng",
"ocr_agent_module": OCR_AGENT_TESSERACT,
}


def test_process_file_with_ocr_uses_given_image_paths_without_rendering(mocker, tmp_path):
from unstructured.partition.pdf_image.ocr import PILImage

def fail_render(*args, **kwargs):
raise AssertionError("pdf must not be rendered when image_paths are given")

mocker.patch.object(ocr, "convert_pdf_to_image", side_effect=fail_render)
seen = []

def record(page_layout, image, **kwargs):
seen.append((page_layout, image.size))
return page_layout

mocker.patch.object(ocr, "supplement_page_layout_with_ocr", side_effect=record)
page_paths = []
for i, size in enumerate(((50, 50), (60, 70))):
page_path = tmp_path / f"page-{i}.png"
PILImage.new("RGB", size).save(page_path)
page_paths.append(str(page_path))
doc = MagicMock(DocumentLayout)
doc.pages = [MagicMock(), MagicMock()]

result = ocr.process_file_with_ocr(
"foo.pdf", doc, [[], []], is_image=False, image_paths=page_paths
)

assert seen == [(doc.pages[0], (50, 50)), (doc.pages[1], (60, 70))]
assert result.pages == doc.pages
143 changes: 141 additions & 2 deletions test_unstructured/partition/pdf_image/test_pdf.py
Original file line number Diff line number Diff line change
Expand Up @@ -44,7 +44,7 @@
)
from unstructured.errors import PageCountExceededError, UnprocessableEntityError
from unstructured.partition import pdf, strategies
from unstructured.partition.pdf_image import ocr, pdfminer_processing
from unstructured.partition.pdf_image import ocr, pdf_image_utils, pdfminer_processing
from unstructured.partition.pdf_image.pdfminer_processing import get_uris_from_annots
from unstructured.partition.utils import config as partition_config
from unstructured.partition.utils.constants import (
Expand Down Expand Up @@ -177,6 +177,7 @@ def pages(self):
],
)
def test_partition_pdf_local(monkeypatch, filename, file):
monkeypatch.setattr(pdf_image_utils, "convert_pdf_to_image", lambda *a, **k: [])
monkeypatch.setattr(
layout,
"process_data_with_model",
Expand Down Expand Up @@ -213,7 +214,7 @@ def test_partition_pdf_local(monkeypatch, filename, file):


def test_partition_pdf_local_raises_with_no_filename():
with pytest.raises((FileNotFoundError, PDFPageCountError, TypeError)):
with pytest.raises((FileNotFoundError, PDFPageCountError, TypeError, ValueError)):
pdf._partition_pdf_or_image_local(filename="", file=None, is_image=False)


Expand Down Expand Up @@ -252,6 +253,7 @@ def test_partition_pdf_local_threads_rotation_corrections_into_pdfminer(
rotated_layout = _layout_with_rotation_corrections(
[{"pdf_rotation_correction": 90}, {"pdf_rotation_correction": 0}]
)
monkeypatch.setattr(pdf_image_utils, "convert_pdf_to_image", lambda *a, **k: [])
monkeypatch.setattr(layout, model_target, lambda *a, **k: rotated_layout)

captured = {}
Expand All @@ -276,6 +278,143 @@ def _capture_pdfminer(*args, **kwargs):
assert captured["rotation_corrections"] == [90, 0]


@pytest.mark.parametrize(
("file_arg", "model_target", "pdfminer_target", "ocr_target"),
[
(None, "process_file_with_model", "process_file_with_pdfminer", "process_file_with_ocr"),
(
b"0000",
"process_data_with_model",
"process_data_with_pdfminer",
"process_data_with_ocr",
),
],
)
def test_partition_pdf_local_renders_pages_once_and_shares_them(
monkeypatch, file_arg, model_target, pdfminer_target, ocr_target
):
"""The PDF is rendered a single time and the page images reach layout inference, OCR
and image extraction instead of each of them rendering the PDF again."""
monkeypatch.setattr(pdf, "_inference_reuses_page_images", lambda: True)
rendered = ["page-1.png", "page-2.png"]
render_calls = []
received = {}
saved_with = []

def _render(*args, **kwargs):
render_calls.append(kwargs)
return rendered

def _capture(name):
def _inner(*args, **kwargs):
received[name] = kwargs.get("image_paths")
return MockDocumentLayout()

return _inner

monkeypatch.setattr(pdf_image_utils, "convert_pdf_to_image", _render)
monkeypatch.setattr(layout, model_target, _capture("model"))
monkeypatch.setattr(pdfminer_processing, pdfminer_target, lambda *a, **k: ([], []))
monkeypatch.setattr(
pdfminer_processing,
"merge_inferred_with_extracted_layout",
lambda **k: MockDocumentLayout(),
)
monkeypatch.setattr(ocr, ocr_target, _capture("ocr"))
monkeypatch.setattr(
pdf_image_utils,
"save_elements",
lambda **kwargs: saved_with.append(kwargs["image_paths"]),
)

pdf._partition_pdf_or_image_local(
filename=example_doc_path("pdf/layout-parser-paper-fast.pdf") if file_arg is None else "",
file=file_arg,
extract_image_block_types=["Image", "Table"],
)

assert len(render_calls) == 1
assert received == {"model": rendered, "ocr": rendered}
assert saved_with == [rendered, rendered]


def test_partition_pdf_local_renders_streams_without_readinto(monkeypatch):
"""pdfium can only open streams that implement `readinto`; other file-likes are rendered
from their bytes."""
monkeypatch.setattr(pdf, "_inference_reuses_page_images", lambda: True)

class ReadSeekOnly:
def __init__(self, data):
self._buffer = io.BytesIO(data)

def read(self, *args):
return self._buffer.read(*args)

def seek(self, *args):
return self._buffer.seek(*args)

def tell(self):
return self._buffer.tell()

rendered_inputs = []

def _render(filename, file=None, *args, **kwargs):
rendered_inputs.append(file)
return []

monkeypatch.setattr(pdf_image_utils, "convert_pdf_to_image", _render)
monkeypatch.setattr(layout, "process_data_with_model", lambda *a, **k: MockDocumentLayout())
monkeypatch.setattr(pdfminer_processing, "process_data_with_pdfminer", lambda *a, **k: ([], []))
monkeypatch.setattr(
pdfminer_processing,
"merge_inferred_with_extracted_layout",
lambda **k: MockDocumentLayout(),
)
monkeypatch.setattr(ocr, "process_data_with_ocr", lambda *a, **k: MockDocumentLayout())
with open(example_doc_path("pdf/layout-parser-paper-fast.pdf"), "rb") as f:
stream = ReadSeekOnly(f.read())

pdf._partition_pdf_or_image_local(filename="", file=stream)

assert isinstance(rendered_inputs[0], bytes)
assert stream.tell() == 0


def test_partition_pdf_local_does_not_prerender_when_inference_cannot_reuse_pages(monkeypatch):
"""Without inference support the pages would be rendered twice, so the pipeline keeps
rendering per stage."""
monkeypatch.setattr(pdf, "_inference_reuses_page_images", lambda: False)
monkeypatch.setattr(
pdf_image_utils,
"convert_pdf_to_image",
lambda *a, **k: pytest.fail("pdf must not be pre-rendered"),
)
received = {}

def _capture(name):
def _inner(*args, **kwargs):
received[name] = kwargs
return MockDocumentLayout()

return _inner

monkeypatch.setattr(layout, "process_file_with_model", _capture("model"))
monkeypatch.setattr(pdfminer_processing, "process_file_with_pdfminer", lambda *a, **k: ([], []))
monkeypatch.setattr(
pdfminer_processing,
"merge_inferred_with_extracted_layout",
lambda **k: MockDocumentLayout(),
)
monkeypatch.setattr(ocr, "process_file_with_ocr", _capture("ocr"))

pdf._partition_pdf_or_image_local(
filename=example_doc_path("pdf/layout-parser-paper-fast.pdf"),
)

assert received["model"].get("image_paths") is None
assert received["ocr"].get("image_paths") is None


@pytest.mark.parametrize("file_mode", ["filename", "rb", "spool"])
@pytest.mark.parametrize(
("strategy", "starting_page_number", "expected_page_numbers", "origin"),
Expand Down
38 changes: 38 additions & 0 deletions test_unstructured/partition/pdf_image/test_pdf_image_utils.py
Original file line number Diff line number Diff line change
Expand Up @@ -249,6 +249,44 @@ def test_save_elements(
assert not el.metadata.image_mime_type


def test_save_elements_uses_given_image_paths_without_rendering(monkeypatch, tmp_path):
def fail_render(*args, **kwargs):
raise AssertionError("pdf must not be rendered when image_paths are given")

monkeypatch.setattr(pdf_image_utils, "convert_pdf_to_image", fail_render)
page_paths = []
for i, color in enumerate(("red", "blue")):
page_path = tmp_path / f"page-{i}.png"
PILImg.new("RGB", (400, 400), color).save(page_path)
page_paths.append(str(page_path))
elements = [
Image(
text=f"Image {page}",
coordinates=((10, 10), (10, 100), (100, 100), (100, 10)),
coordinate_system=PixelSpace(width=400, height=400),
metadata=ElementMetadata(page_number=page),
)
for page in (1, 1, 2)
]

pdf_image_utils.save_elements(
elements=elements,
starting_page_number=1,
element_category_to_save=ElementType.IMAGE,
pdf_image_dpi=200,
filename="unused.pdf",
output_dir_path=str(tmp_path),
extract_image_block_to_payload=True,
image_paths=page_paths,
)

def dominant_channel(el):
crop = PILImg.open(io.BytesIO(base64.b64decode(el.metadata.image_base64))).convert("RGB")
return max(range(3), key=lambda channel: crop.getpixel((10, 10))[channel])

assert [dominant_channel(el) for el in elements] == [0, 0, 2]


def test_save_elements_with_inverted_point_ordering(monkeypatch):
"""Regression: points whose ordering puts points[0] below points[2] (as happens when
coordinates come from / were converted from a y-up CARTESIAN system) must not raise
Expand Down
Loading
Loading