Create dataset.py

dataset.py

@@ -0,0 +1,60 @@
+import os
+from datasets import GeneratorBasedBuilder, DatasetInfo, Features, Value, Image, SplitGenerator, Split
+
+
+class FFHQProxyReference(GeneratorBasedBuilder):
+    VERSION = "1.0.0"
+
+    def _info(self):
+        return DatasetInfo(
+            description="FFHQ Proxy Reference dataset for high-quality facial images.",
+            features=Features(
+                {
+                    "image": Image(),
+                    "resolution": Value("int32"),
+                    "source": Value("string"),
+                }
+            ),
+            supervised_keys=None,
+            homepage="https://github.com/NVlabs/ffhq-dataset",
+            license="creativeml-openrail-m",
+        )
+
+    def _split_generators(self, dl_manager):
+        """
+        User MUST pass data_dir when calling load_dataset.
+        """
+        if not self.config.data_dir:
+            raise ValueError(
+                "You must provide a local path to FFHQ images using `data_dir`."
+            )
+
+        return [
+            SplitGenerator(
+                name=Split.TRAIN,
+                gen_kwargs={"image_dir": self.config.data_dir},
+            )
+        ]
+
+    def _generate_examples(self, image_dir):
+        """
+        Expects a flat or nested directory of FFHQ images.
+        """
+        idx = 0
+        for root, _, files in os.walk(image_dir):
+            for file in files:
+                if file.lower().endswith((".png", ".jpg", ".jpeg")):
+                    path = os.path.join(root, file)
+
+                    yield idx, {
+                        "image": path,
+                        "resolution": self._infer_resolution(file),
+                        "source": "FFHQ",
+                    }
+                    idx += 1
+
+    def _infer_resolution(self, filename):
+        """
+        Optional heuristic – FFHQ commonly uses 1024x1024.
+        """
+        return 1024