Upload markushgrapher-datasets.py

markushgrapher-datasets.py

@@ -0,0 +1,57 @@
+import os
+from datasets import DatasetInfo, GeneratorBasedBuilder, SplitGenerator, Split, BuilderConfig, Features, Value, Array2D, Array3D, Sequence, Image, DatasetDict
+import datasets
+
+class MarkushGrapherConfig(BuilderConfig):
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+
+class MarkushGrapherDataset(GeneratorBasedBuilder):
+    BUILDER_CONFIGS = [
+        MarkushGrapherConfig(name="markushgrapher-synthetic-training", description="Synthetic training set"),
+        MarkushGrapherConfig(name="m2s", description="Multi-modal real-world benchmark set"),        
+        MarkushGrapherConfig(name="markushgrapher-synthetic", description="Synthetic benchmark set"),
+        MarkushGrapherConfig(name="uspto-markush", description="Image-only real-world benchmark set"),
+    ]
+
+    def _info(self):
+        return DatasetInfo(
+            description="MarkushGrapher-Datasets",
+            features=Features({
+                "id": Value("int64"),
+                "image_name": Value("string"),
+                "page_image": Image(mode=None, decode=True, id=None),
+                "description": Value("string"),
+                "annotation": Value("string"),
+                "mol": Value("string"),
+                "cxsmiles_dataset": Value("string"),
+                "cxsmiles": Value("string"),
+                "cxsmiles_opt": Value("string"),
+                "keypoints": Value(dtype='null', id=None),
+                "cells": [{'bbox': Sequence(feature=Value(dtype='float64', id=None), length=-1, id=None), 'text': Value(dtype='string', id=None)}],
+            }),
+            supervised_keys=None,
+        )
+
+    def _split_generators(self, dl_manager):
+        data_dir = dl_manager.download_and_extract(self.config.name)
+
+        splits = []
+        for split in ["train", "test"]:
+            split_path = os.path.join(data_dir, split)
+            if os.path.exists(split_path):
+                splits.append(SplitGenerator(
+                    name=Split.TRAIN if split == "train" else Split.TEST,
+                    gen_kwargs={"data_dir": split_path},
+                ))
+
+        return splits
+
+    def _generate_examples(self, data_dir):
+        # Automatically loads .arrow files
+        arrow_files = [os.path.join(data_dir, f) for f in os.listdir(data_dir) if f.endswith(".arrow")]
+        for file in arrow_files:
+            dataset = datasets.Dataset.from_file(file)
+            for i, row in enumerate(dataset):
+                yield f"{file}_{i}", row