Fix for wrong type assignation at ingest-test (#172)

This PR restores some thresholds for getting the ingest test correct at
`unstructured`, however, the thresholds keep being distinct for detectron2_mask_rcnn.

---------

Co-authored-by: Alan Bertl <[email protected]>

Author: benjats07

.github/workflows/ci.yml

@@ -131,6 +131,7 @@ jobs:
         sudo add-apt-repository -y ppa:alex-p/tesseract-ocr5
         sudo apt-get install -y tesseract-ocr
         sudo apt-get install -y tesseract-ocr-kor
+        sudo apt-get install -y diffstat
         tesseract --version
         make install-ingest-s3
         make install-ingest-azure

CHANGELOG.md

@@ -1,3 +1,7 @@
+## 0.5.18
+
+* Fix for incorrect type assignation at ingest test
+
 ## 0.5.17
 
 * Use `OMP_THREAD_LIMIT` to improve tesseract performance

unstructured_inference/__version__.py

@@ -1 +1 @@
-__version__ = "0.5.17"  # pragma: no cover
+__version__ = "0.5.18"  # pragma: no cover

unstructured_inference/inference/layout.py

@@ -28,6 +28,9 @@
 from unstructured_inference.inference.ordering import order_layout
 from unstructured_inference.logger import logger
 from unstructured_inference.models.base import get_model
+from unstructured_inference.models.detectron2onnx import (
+    UnstructuredDetectronONNXModel,
+)
 from unstructured_inference.models.unstructuredmodel import (
     UnstructuredElementExtractionModel,
     UnstructuredObjectDetectionModel,
@@ -262,15 +265,33 @@ def get_elements_with_detection_model(
             raise ValueError("Invalid OCR mode")
 
         if self.layout is not None:
+            threshold_kwargs = {}
+            # NOTE(Benjamin): With this the thresholds are only changed for detextron2_mask_rcnn
+            # In other case the default values for the functions are used
+            if (
+                isinstance(self.detection_model, UnstructuredDetectronONNXModel)
+                and "R_50" not in self.detection_model.model_path
+            ):
+                threshold_kwargs = {"same_region_threshold": 0.5, "subregion_threshold": 0.5}
             inferred_layout = merge_inferred_layout_with_extracted_layout(
                 inferred_layout=inferred_layout,
                 extracted_layout=self.layout,
                 ocr_layout=ocr_layout,
+                **threshold_kwargs,
             )
         elif ocr_layout is not None:
+            threshold_kwargs = {}
+            # NOTE(Benjamin): With this the thresholds are only changed for detextron2_mask_rcnn
+            # In other case the default values for the functions are used
+            if (
+                isinstance(self.detection_model, UnstructuredDetectronONNXModel)
+                and "R_50" not in self.detection_model.model_path
+            ):
+                threshold_kwargs = {"subregion_threshold": 0.3}
             inferred_layout = merge_inferred_layout_with_ocr_layout(
                 inferred_layout=inferred_layout,
                 ocr_layout=ocr_layout,
+                **threshold_kwargs,
             )
 
         elements = self.get_elements_from_layout(cast(List[TextRegion], inferred_layout))