Upload FinDoc-Robust.py with huggingface_hub

FinDoc-Robust.py

@@ -0,0 +1,48 @@
+import os
+import json
+import datasets
+
+class FinDocRobust(datasets.GeneratorBasedBuilder):
+    def _info(self):
+        return datasets.DatasetInfo(
+            features=datasets.Features({
+                "file_name": datasets.Image(),
+                "document_type": datasets.Value("string"),
+                "document_id": datasets.Value("int64"),
+                "clean_pdf": datasets.Value("string"),
+                "clean_xlsx": datasets.Value("string"),
+                "clean_bbox_px": datasets.Value("string"),
+                "clean_bbox_pdf_pt": datasets.Value("string"),
+                **{f"dirty_{i}_image": datasets.Image() for i in range(1, 6)},
+                **{f"dirty_{i}_bbox": datasets.Value("string") for i in range(1, 6)}
+            })
+        )
+
+    def _split_generators(self, dl_manager):
+        # Скачиваем индексный файл и весь репозиторий
+        meta_path = dl_manager.download_and_extract("dataset_index.csv")
+        archive_path = dl_manager.download_and_extract(".")
+        return [
+            datasets.SplitGenerator(
+                name=datasets.Split.TRAIN,
+                gen_kwargs={"meta_path": meta_path, "base_path": archive_path}
+            )
+        ]
+
+    def _generate_examples(self, meta_path, base_path):
+        import pandas as pd
+        df = pd.read_csv(meta_path)
+        for idx, row in df.iterrows():
+            res = {
+                "file_name": os.path.join(base_path, row["file_name"]),
+                "document_type": row["document_type"],
+                "document_id": int(row["document_id"]),
+                "clean_pdf": row["clean_pdf"],
+                "clean_xlsx": row["clean_xlsx"],
+                "clean_bbox_px": row["clean_bbox_px"],
+                "clean_bbox_pdf_pt": row["clean_bbox_pdf_pt"]
+            }
+            for i in range(1, 6):
+                res[f"dirty_{i}_image"] = os.path.join(base_path, row[f"dirty_{i}_image"])
+                res[f"dirty_{i}_bbox"] = row[f"dirty_{i}_bbox"]
+            yield idx, res