Fix load labeled data

results/results.md

@@ -0,0 +1,62 @@
+### Letter corpus
+
+F1 score 99.25%
+Accuracy score 99.52%
+finished in 54 seconds
+
+File Type          Mistakes
+-----------------  ----------------
+one_column_test    162/213 (23.94%)
+multi_column_test  56/81 (30.86%)
+
+Average Accuracy: 218/294 (25.85%)
+Total Time: 0.53
+
+
+
+### Unicode 2 first characters + 2 last characters
+
+F1 score 99.19%
+Accuracy score 99.48%
+finished in 66 seconds
+
+Process finished with exit code 0
+
+File Type          Mistakes
+-----------------  ----------------
+one_column_test    162/213 (23.94%)
+multi_column_test  55/81 (32.1%)
+
+Average Accuracy: 217/294 (26.19%)
+Total Time: 0.81
+
+
+
+### No Unicode features
+
+F1 score 99.17%
+Accuracy score 99.46%
+finished in 55 seconds
+
+File Type          Mistakes
+-----------------  ----------------
+one_column_test    162/213 (23.94%)
+multi_column_test  55/81 (32.1%)
+
+Average Accuracy: 217/294 (26.19%)
+Total Time: 0.87
+
+
+### Only first two letters for unicode
+
+F1 score 99.21%
+Accuracy score 99.49%
+finished in 62 seconds
+
+File Type          Mistakes
+-----------------  ----------------
+one_column_test    162/213 (23.94%)
+multi_column_test  55/81 (32.1%)
+
+Average Accuracy: 217/294 (26.19%)
+Total Time: 0.8

src/benchmark.py

@@ -10,7 +10,7 @@
 from paragraph_extraction_trainer.load_labeled_data import load_labeled_data
 from paragraph_extraction_trainer.model_configuration import MODEL_CONFIGURATION
 
-BENCHMARK_MODEL_PATH = Path(join(ROOT_PATH, "model", "paragraph_extraction_benchmark.model"))
+BENCHMARK_MODEL_PATH = Path(join(ROOT_PATH, "model", "benchmark.model"))
 
 
 def loop_pdf_paragraph_tokens(pdf_paragraph_tokens_list: list[PdfParagraphTokens]):
@@ -25,7 +25,7 @@ def loop_pdf_paragraph_tokens(pdf_paragraph_tokens_list: list[PdfParagraphTokens
 
 def train_for_benchmark():
     pdf_paragraph_tokens_list = load_labeled_data(PDF_LABELED_DATA_ROOT_PATH, filter_in="train")
-
+    print('length of pdf paragraphs for training', len(pdf_paragraph_tokens_list))
     pdf_features_list = [pdf_paragraph_tokens.pdf_features for pdf_paragraph_tokens in pdf_paragraph_tokens_list]
     trainer = ParagraphExtractorTrainer(pdfs_features=pdf_features_list, model_configuration=MODEL_CONFIGURATION)
 

src/benchmark_table.txt

@@ -0,0 +1,7 @@
+File Type          Mistakes
+-----------------  ----------------
+multi_column_test  55/81 (32.1%)
+one_column_test    162/213 (23.94%)
+
+Average Accuracy: 217/294 (26.19%)
+Total Time: 0.8

src/paragraph_extraction_trainer/ParagraphExtractorTrainer.py

@@ -5,6 +5,7 @@
 from pdf_token_type_labels.TokenType import TokenType
 from pdf_tokens_type_trainer.TokenFeatures import TokenFeatures
 from pdf_tokens_type_trainer.TokenTypeTrainer import TokenTypeTrainer
+from pdf_tokens_type_trainer.config import CHARACTER_TYPE
 
 
 class ParagraphExtractorTrainer(TokenTypeTrainer):

src/paragraph_extraction_trainer/PdfParagraphTokens.py

@@ -1,4 +1,6 @@
 from os.path import join
+from tkinter import Label
+
 from pdf_features.PdfToken import PdfToken
 from pdf_token_type_labels.PdfLabels import PdfLabels
 
@@ -14,17 +16,13 @@ def __init__(self, pdf_features: PdfFeatures, paragraphs: list[Paragraph]):
         self.paragraphs = paragraphs
 
     @staticmethod
-    def get_page_labels(paragraphs_extractions_labels: PdfLabels, page_number: int):
-        page_labels = []
+    def get_page_number_labels(paragraphs_extractions_labels: PdfLabels):
+        page_number_labels = {}
 
         for page in paragraphs_extractions_labels.pages:
-            if page.number != page_number:
-                continue
-
-            for label_index, label in enumerate(sorted(page.labels, key=lambda _label: _label.area())):
-                page_labels.append((label_index, label, page.number))
+            page_number_labels[page.number] = list(sorted(page.labels, key=lambda _label: (_label.area(), _label.top)))
 
-        return page_labels
+        return page_number_labels
 
     @staticmethod
     def from_labeled_data(pdf_labeled_data_root_path, dataset, pdf_name):
@@ -36,29 +34,32 @@ def from_labeled_data(pdf_labeled_data_root_path, dataset, pdf_name):
 
     @staticmethod
     def set_paragraphs(pdf_features: PdfFeatures, paragraphs_extractions_labels: PdfLabels):
-        tokens_by_labels: dict[int, Paragraph] = dict()
+        tokens_by_labels: dict[tuple[int, int], Paragraph] = dict()
+
+        page_number_labels = PdfParagraphTokens.get_page_number_labels(paragraphs_extractions_labels)
 
         for token_index, (page, token) in enumerate(pdf_features.loop_tokens()):
-            page_labels = PdfParagraphTokens.get_page_labels(paragraphs_extractions_labels, page.page_number)
+            page_labels = page_number_labels[page.page_number]
             intersection, best_label = PdfParagraphTokens.get_intersected_label(page_labels, token)
 
             if intersection:
-                tokens_by_labels.setdefault(best_label, Paragraph([])).add_token(token)
+                label_index = page_labels.index(best_label)
+                tokens_by_labels.setdefault((page.page_number, label_index), Paragraph([])).add_token(token)
             else:
-                tokens_by_labels[-token_index - 1] = Paragraph(tokens=[token])
+                tokens_by_labels[(page.page_number, -token_index - 1)] = Paragraph(tokens=[token])
 
         return PdfParagraphTokens(pdf_features, list(tokens_by_labels.values()))
 
     @staticmethod
-    def get_intersected_label(page_labels, token):
+    def get_intersected_label(page_labels: list[Label], token: PdfToken):
         max_intersection = 0
-        best_label = -1
-        for label_index, label, label_page_number in page_labels:
+        best_label = None
+        for label in page_labels:
             intersection = token.get_label_intersection_percentage(label)
 
             if intersection > max_intersection:
                 max_intersection = intersection
-                best_label = label_index
+                best_label = label
 
             if max_intersection > 99:
                 break