chore: clean up code and comments

Signed-off-by: Christoph Auer <cau@zurich.ibm.com>

docling/datamodel/pipeline_options.py

@@ -333,6 +333,8 @@ class PaginatedPipelineOptions(PipelineOptions):
 
 class VlmPipelineOptions(PaginatedPipelineOptions):
     artifacts_path: Optional[Union[Path, str]] = None
+
+    generate_page_images: bool = True
     force_backend_text: bool = (
         False  # (To be used with vlms, or other generative models)
     )

docling/models/hf_vlm_model.py

@@ -116,7 +116,7 @@ class HuggingFaceVlmModel(BasePageModel):
             if not page._backend.is_valid():
                 yield page
             else:
-                with TimeRecorder(conv_res, "smolvlm"):
+                with TimeRecorder(conv_res, "vlm"):
                     assert page.size is not None
 
                     hi_res_image = page.get_image(scale=2.0)  # 144dpi

docling/pipeline/vlm_pipeline.py

@@ -51,7 +51,7 @@ class VlmPipeline(PaginatedPipeline):
         self.keep_backend = True
 
         warnings.warn(
-            "This API is currently experimental and may change in upcoming versions without notice.",
+            "The VlmPipeline is currently experimental and may change in upcoming versions without notice.",
             category=UserWarning,
             stacklevel=2,
         )
@@ -70,18 +70,18 @@ class VlmPipeline(PaginatedPipeline):
                 "When defined, it must point to a folder containing all models required by the pipeline."
             )
 
-        # force_backend_text = False - use text that is coming from SmolDocling
-        # force_backend_text = True - get text from backend using bounding boxes predicted by SmolDoclingss
-        self.force_backend_text = pipeline_options.force_backend_text
-
-        self.keep_images = (
-            self.pipeline_options.generate_page_images
-            or self.pipeline_options.generate_picture_images
+        # force_backend_text = False - use text that is coming from VLM response
+        # force_backend_text = True - get text from backend using bounding boxes predicted by SmolDocling doctags
+        self.force_backend_text = (
+            pipeline_options.force_backend_text
+            and pipeline_options.vlm_options.response_format == ResponseFormat.DOCTAGS
         )
 
+        self.keep_images = self.pipeline_options.generate_page_images
+
         self.build_pipe = [
             HuggingFaceVlmModel(
-                enabled=True,
+                enabled=True,  # must be always enabled for this pipeline to make sense.
                 artifacts_path=artifacts_path,
                 accelerator_options=pipeline_options.accelerator_options,
                 vlm_options=self.pipeline_options.vlm_options,
@@ -397,6 +397,7 @@ class VlmPipeline(PaginatedPipeline):
             if page.predictions.vlm_response:
                 predicted_text = page.predictions.vlm_response.text
             image = page.image
+
             page_no = pg_idx + 1
             bounding_boxes = []
 
@@ -448,12 +449,13 @@ class VlmPipeline(PaginatedPipeline):
                     text_caption_content = extract_inner_text(full_chunk)
                     if image:
                         if bbox:
-                            width, height = image.size
+                            im_width, im_height = image.size
+
                             crop_box = (
-                                int(bbox.l * width),
-                                int(bbox.t * height),
-                                int(bbox.r * width),
-                                int(bbox.b * height),
+                                int(bbox.l * im_width),
+                                int(bbox.t * im_height),
+                                int(bbox.r * im_width),
+                                int(bbox.b * im_height),
                             )
                             cropped_image = image.crop(crop_box)
                             pic = doc.add_picture(
@@ -461,7 +463,9 @@ class VlmPipeline(PaginatedPipeline):
                                 image=ImageRef.from_pil(image=cropped_image, dpi=72),
                                 prov=(
                                     ProvenanceItem(
-                                        bbox=bbox, charspan=(0, 0), page_no=page_no
+                                        bbox=bbox.resize_by_scale(pg_width, pg_height),
+                                        charspan=(0, 0),
+                                        page_no=page_no,
                                     )
                                 ),
                             )
@@ -501,7 +505,7 @@ class VlmPipeline(PaginatedPipeline):
                         text=text_content,
                         prov=(
                             ProvenanceItem(
-                                bbox=bbox,
+                                bbox=bbox.resize_by_scale(pg_width, pg_height),
                                 charspan=(0, len(text_content)),
                                 page_no=page_no,
                             )

docs/examples/minimal_vlm_pipeline.py

@@ -11,32 +11,34 @@ from docling.datamodel.pipeline_options import (
     granite_vision_vlm_conversion_options,
     smoldocling_vlm_conversion_options,
 )
+from docling.datamodel.settings import settings
 from docling.document_converter import DocumentConverter, PdfFormatOption
 from docling.pipeline.vlm_pipeline import VlmPipeline
 
 sources = [
-    # "https://arxiv.org/pdf/2408.09869",
     "tests/data/2305.03393v1-pg9-img.png",
-    # "tests/data/2305.03393v1-pg9.pdf",
 ]
 
-pipeline_options = VlmPipelineOptions()  # artifacts_path="~/local_model_artifacts/"
-pipeline_options.generate_page_images = True
+settings.debug.profile_pipeline_timings = True
+## Use experimental VlmPipeline
+pipeline_options = VlmPipelineOptions()
 # If force_backend_text = True, text from backend will be used instead of generated text
 pipeline_options.force_backend_text = False
 
-## Enable flash_attention_2 with CUDA:
+## On GPU systems, enable flash_attention_2 with CUDA:
 # pipeline_options.accelerator_options.device = AcceleratorDevice.CUDA
 # pipeline_options.accelerator_options.cuda_use_flash_attention2 = True
 
+## Pick a VLM model. We choose SmolDocling-256M by default
 pipeline_options.vlm_options = smoldocling_vlm_conversion_options
 
-## Choose alternative VLM models:
+## Alternative VLM models:
 # pipeline_options.vlm_options = granite_vision_vlm_conversion_options
 
 from docling_core.types.doc import DocItemLabel, ImageRefMode
 from docling_core.types.doc.document import DEFAULT_EXPORT_LABELS
 
+## Set up pipeline for PDF or image inputs
 converter = DocumentConverter(
     format_options={
         InputFormat.PDF: PdfFormatOption(
@@ -68,6 +70,12 @@ for source in sources:
     print("")
     print(res.document.export_to_markdown())
 
+    print("------------------------------------------------")
+    print("Timings:")
+    print("------------------------------------------------")
+    print("")
+    print(res.timings)
+
     for page in res.pages:
         print("")
         print("Predicted page in DOCTAGS:")
@@ -82,9 +90,6 @@ for source in sources:
     with (out_path / f"{res.input.file.stem}.json").open("w") as fp:
         fp.write(json.dumps(res.document.export_to_dict()))
 
-    with (out_path / f"{res.input.file.stem}.yaml").open("w") as fp:
-        fp.write(yaml.safe_dump(res.document.export_to_dict()))
-
     pg_num = res.document.num_pages()
 
     print("")

pyproject.toml

@@ -69,9 +69,6 @@ accelerate = [
 pillow = ">=10.0.0,<12.0.0"
 tqdm = "^4.65.0"
 
-# transformers = "^4.47.1"
-# accelerate = "^1.2.1"
-
 [tool.poetry.group.dev.dependencies]
 black = {extras = ["jupyter"], version = "^24.4.2"}
 pytest = "^7.2.2"