docling-ibm-models 1.1.5__tar.gz → 1.1.7__tar.gz

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: docling-ibm-models
-Version: 1.1.5
+Version: 1.1.7
 Summary: This package contains the AI models used by the Docling PDF conversion package
 License: MIT
 Keywords: docling,convert,document,pdf,layout model,segmentation,table structure,table former

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/docling_ibm_models/tableformer/data_management/matching_post_processor.py

@@ -4,6 +4,7 @@
 #
 import json
 import logging
+import math
 import statistics
 
 import docling_ibm_models.tableformer.settings as s
@@ -403,45 +404,63 @@ class MatchingPostProcessor:
                 # Push horizontally
                 if x1_min < x2_min:
                     # Move box1 to the left and box2 to the right
-                    box1["bbox"][2] -= overlap_x
-                    box2["bbox"][0] += overlap_x
+                    box1["bbox"][2] -= math.ceil(overlap_x / 2) + 2
+                    box2["bbox"][0] += math.floor(overlap_x / 2)
                 else:
                     # Move box2 to the left and box1 to the right
-                    box2["bbox"][2] -= overlap_x
-                    box1["bbox"][0] += overlap_x
+                    box2["bbox"][2] -= math.ceil(overlap_x / 2) + 2
+                    box1["bbox"][0] += math.floor(overlap_x / 2)
             else:
                 # Push vertically
                 if y1_min < y2_min:
                     # Move box1 up and box2 down
-                    box1["bbox"][3] -= overlap_y
-                    box2["bbox"][1] += overlap_y
+                    box1["bbox"][3] -= math.ceil(overlap_y / 2) + 2
+                    box2["bbox"][1] += math.floor(overlap_y / 2)
                 else:
                     # Move box2 up and box1 down
-                    box2["bbox"][3] -= overlap_y
-                    box1["bbox"][1] += overlap_y
+                    box2["bbox"][3] -= math.ceil(overlap_y / 2) + 2
+                    box1["bbox"][1] += math.floor(overlap_y / 2)
+
+            # Will flip coordinates in proper order, if previous operations reversed it
+            box1["bbox"] = [
+                min(box1["bbox"][0], box1["bbox"][2]),
+                min(box1["bbox"][1], box1["bbox"][3]),
+                max(box1["bbox"][0], box1["bbox"][2]),
+                max(box1["bbox"][1], box1["bbox"][3]),
+            ]
+            box2["bbox"] = [
+                min(box2["bbox"][0], box2["bbox"][2]),
+                min(box2["bbox"][1], box2["bbox"][3]),
+                max(box2["bbox"][0], box2["bbox"][2]),
+                max(box2["bbox"][1], box2["bbox"][3]),
+            ]
 
             return box1, box2
 
         def do_boxes_overlap(box1, box2):
-            # print("{} - {}".format(box1["bbox"], box2["bbox"]))
-            # Extract coordinates from the bounding boxes
-            x1_min, y1_min, x1_max, y1_max = box1["bbox"]
-            x2_min, y2_min, x2_max, y2_max = box2["bbox"]
-            # Check if one box is to the left of the other
-            if x1_max < x2_min or x2_max < x1_min:
+            B1 = box1["bbox"]
+            B2 = box2["bbox"]
+            if (
+                (B1[0] >= B2[2])
+                or (B1[2] <= B2[0])
+                or (B1[3] <= B2[1])
+                or (B1[1] >= B2[3])
+            ):
                 return False
-            # Check if one box is above the other
-            if y1_max < y2_min or y2_max < y1_min:
-                return False
-            return True
+            else:
+                return True
 
         def find_overlapping_pairs_indexes(bboxes):
             overlapping_indexes = []
             # Compare each box with every other box (combinations)
             for i in range(len(bboxes)):
                 for j in range(i + 1, len(bboxes)):
-                    if do_boxes_overlap(bboxes[i], bboxes[j]):
-                        bboxes[i], bboxes[j] = correct_overlap(bboxes[i], bboxes[j])
+                    if i != j:
+                        if bboxes[i] != bboxes[j]:
+                            if do_boxes_overlap(bboxes[i], bboxes[j]):
+                                bboxes[i], bboxes[j] = correct_overlap(
+                                    bboxes[i], bboxes[j]
+                                )
 
             return overlapping_indexes, bboxes
 
@@ -1144,7 +1163,7 @@ class MatchingPostProcessor:
                 new_pdf_cells.append(pdf_cells[i])
         return new_pdf_cells
 
-    def process(self, matching_details):
+    def process(self, matching_details, correct_overlapping_cells=False):
         r"""
         Do post processing, see details in the comments below
 
@@ -1348,9 +1367,10 @@ class MatchingPostProcessor:
         table_cells_wo = po2
         max_cell_id = po3
 
-        # As the last step - correct cell bboxes in a way that they don't overlap:
-        if len(table_cells_wo) <= 300:  # For performance reasons
-            table_cells_wo = self._find_overlapping(table_cells_wo)
+        if correct_overlapping_cells:
+            # As the last step - correct cell bboxes in a way that they don't overlap:
+            if len(table_cells_wo) <= 300:  # For performance reasons
+                table_cells_wo = self._find_overlapping(table_cells_wo)
 
         self._log().debug("*** final_matches_wo")
         self._log().debug(final_matches_wo)

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/docling_ibm_models/tableformer/data_management/tf_cell_matcher.py

@@ -127,13 +127,14 @@ class CellMatcher:
             Dictionary with all details about the mathings between the table and pdf cells
         """
         pdf_cells = copy.deepcopy(iocr_page["tokens"])
-        for word in pdf_cells:
-            word["bbox"] = [
-                word["bbox"]["l"],
-                word["bbox"]["t"],
-                word["bbox"]["r"],
-                word["bbox"]["b"],
-            ]
+        if len(pdf_cells) > 0:
+            for word in pdf_cells:
+                word["bbox"] = [
+                    word["bbox"]["l"],
+                    word["bbox"]["t"],
+                    word["bbox"]["r"],
+                    word["bbox"]["b"],
+                ]
         table_bboxes = prediction["bboxes"]
         table_classes = prediction["classes"]
         # BBOXES transformed...
@@ -145,9 +146,13 @@ class CellMatcher:
         table_cells = self._build_table_cells(
             html_seq, otsl_seq, table_bboxes_page, table_classes
         )
-        matches, matches_counter = self._intersection_over_pdf_match(
-            table_cells, pdf_cells
-        )
+
+        matches = {}
+        matches_counter = 0
+        if len(pdf_cells) > 0:
+            matches, matches_counter = self._intersection_over_pdf_match(
+                table_cells, pdf_cells
+            )
 
         self._log().debug("matches_counter: {}".format(matches_counter))
 
@@ -188,13 +193,14 @@ class CellMatcher:
             Dictionary with all details about the mathings between the table and pdf cells
         """
         pdf_cells = copy.deepcopy(iocr_page["tokens"])
-        for word in pdf_cells:
-            word["bbox"] = [
-                word["bbox"]["l"],
-                word["bbox"]["t"],
-                word["bbox"]["r"],
-                word["bbox"]["b"],
-            ]
+        if len(pdf_cells) > 0:
+            for word in pdf_cells:
+                word["bbox"] = [
+                    word["bbox"]["l"],
+                    word["bbox"]["t"],
+                    word["bbox"]["r"],
+                    word["bbox"]["b"],
+                ]
 
         table_bboxes = prediction["bboxes"]
         table_classes = prediction["classes"]

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/docling_ibm_models/tableformer/data_management/tf_predictor.py

@@ -523,8 +523,9 @@ class TFPredictor:
         # return the resized image
         return resized, sf
 
-    def multi_table_predict(self, iocr_page, table_bboxes, do_matching=True):
-        # def multi_table_predict(self, iocr_page, page_image, table_bboxes):
+    def multi_table_predict(
+        self, iocr_page, table_bboxes, do_matching=True, correct_overlapping_cells=False
+    ):
         multi_tf_output = []
         page_image = iocr_page["image"]
 
@@ -546,7 +547,12 @@ class TFPredictor:
             # Predict
             if do_matching:
                 tf_responses, predict_details = self.predict(
-                    iocr_page, table_bbox, table_image, scale_factor, None
+                    iocr_page,
+                    table_bbox,
+                    table_image,
+                    scale_factor,
+                    None,
+                    correct_overlapping_cells,
                 )
             else:
                 tf_responses, predict_details = self.predict_dummy(
@@ -696,7 +702,12 @@ class TFPredictor:
             prediction["bboxes"] = corrected_bboxes
 
         # Match the cells
-        matching_details = {"table_cells": [], "matches": {}}
+        matching_details = {
+            "table_cells": [],
+            "matches": {},
+            "pdf_cells": [],
+            "prediction_bboxes_page": [],
+        }
 
         # Table bbox upscaling will scale predicted bboxes too within cell matcher
         scaled_table_bbox = [
@@ -728,7 +739,13 @@ class TFPredictor:
         return tf_output, matching_details
 
     def predict(
-        self, iocr_page, table_bbox, table_image, scale_factor, eval_res_preds=None
+        self,
+        iocr_page,
+        table_bbox,
+        table_image,
+        scale_factor,
+        eval_res_preds=None,
+        correct_overlapping_cells=False,
     ):
         r"""
         Predict the table out of an image in memory
@@ -739,6 +756,8 @@ class TFPredictor:
             Docling provided table data
         eval_res_preds : dict
             Ready predictions provided by the evaluation results
+        correct_overlapping_cells : boolean
+            Enables or disables last post-processing step, that fixes cell bboxes to remove overlap
 
         Returns
         -------
@@ -803,7 +822,12 @@ class TFPredictor:
             prediction["bboxes"] = corrected_bboxes
 
         # Match the cells
-        matching_details = {"table_cells": [], "matches": {}}
+        matching_details = {
+            "table_cells": [],
+            "matches": {},
+            "pdf_cells": [],
+            "prediction_bboxes_page": [],
+        }
 
         # Table bbox upscaling will scale predicted bboxes too within cell matcher
         scaled_table_bbox = [
@@ -819,10 +843,15 @@ class TFPredictor:
             )
         # Post-processing
         if len(prediction["bboxes"]) > 0:
-            if self.enable_post_process:
-                AggProfiler().begin("post_process", self._prof)
-                matching_details = self._post_processor.process(matching_details)
-                AggProfiler().end("post_process", self._prof)
+            if (
+                len(iocr_page["tokens"]) > 0
+            ):  # There are at least some pdf cells to match with
+                if self.enable_post_process:
+                    AggProfiler().begin("post_process", self._prof)
+                    matching_details = self._post_processor.process(
+                        matching_details, correct_overlapping_cells
+                    )
+                    AggProfiler().end("post_process", self._prof)
 
         # Generate the expected Docling responses
         AggProfiler().begin("generate_docling_response", self._prof)

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/docling_ibm_models/tableformer/models/table04_rs/bbox_decoder_rs.py

@@ -157,7 +157,12 @@ class BBoxDecoder(nn.Module):
             predictions_classes.append(self._class_embed(h))
         if len(predictions_bboxes) > 0:
             predictions_bboxes = torch.stack([x[0] for x in predictions_bboxes])
+        else:
+            predictions_bboxes = torch.empty(0)
+
         if len(predictions_classes) > 0:
             predictions_classes = torch.stack([x[0] for x in predictions_classes])
+        else:
+            predictions_classes = torch.empty(0)
 
         return predictions_classes, predictions_bboxes

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/docling_ibm_models/tableformer/otsl.py

@@ -123,6 +123,9 @@ def otsl_check_right(rs_split, x, y):
 
 
 def otsl_to_html(rs_list, logdebug):
+    if len(rs_list) == 0:
+        return []
+
     if rs_list[0] not in ["fcel", "ched", "rhed", "srow", "ecel"]:
         # Most likely already HTML...
         return rs_list

{docling_ibm_models-1.1.5 → docling_ibm_models-1.1.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "docling-ibm-models"
-version = "1.1.5"  # DO NOT EDIT, updated automatically
+version = "1.1.7"  # DO NOT EDIT, updated automatically
 description = "This package contains the AI models used by the Docling PDF conversion package"
 authors = ["Nikos Livathinos <nli@zurich.ibm.com>", "Maxim Lysak <mly@zurich.ibm.com>", "Ahmed Nassar <ahn@zurich.ibm.com>", "Christoph Auer <cau@zurich.ibm.com>", "Michele Dolfi <dol@zurich.ibm.com>", "Peter Staar <taa@zurich.ibm.com>"]
 license = "MIT"