Update BubbleML_2.py

BubbleML_2.py

@@ -11,6 +11,7 @@ from datasets import (
     GeneratorBasedBuilder,
     DatasetInfo,
     Features,
+    Array4D,
     Sequence,
     Value,
     Split,
@@ -24,13 +25,16 @@ Pre-defined train/test splits across two directories.
 """
 
 class BubbleMLConfig(BuilderConfig):
-    """BuilderConfig for BubbleML_2."""
+    """BuilderConfig for BubbleML_2 with fixed T, H, W."""
     def __init__(
         self,
         *,
         name: str,
         description: str,
         data_files: dict,
+        timesteps: int,
+        height: int,
+        width: int,
         data_dir: str = "",
         version: Version = Version("1.0.0"),
         **kwargs,
@@ -38,6 +42,9 @@ class BubbleMLConfig(BuilderConfig):
         super().__init__(name=name, version=version, description=description, **kwargs)
         self.data_files = data_files
         self.data_dir = data_dir
+        self.timesteps = timesteps
+        self.height = height
+        self.width = width
 
 class BubbleMLDataset(GeneratorBasedBuilder):
     """BubbleML_2: combined single-bubble dataset."""
@@ -47,7 +54,10 @@ class BubbleMLDataset(GeneratorBasedBuilder):
         BubbleMLConfig(
             name="single-bubble",
             description="Single-bubble (FC72 & R515B) train/test split",
-            data_dir="",  # use repo root locally; if empty, remote download is used
+            data_dir="",  # local root; overridden by --data_dir
+            timesteps=5,
+            height=288,
+            width=192,
             data_files={
                 "train": [
                     # FC72 train
@@ -84,11 +94,15 @@ class BubbleMLDataset(GeneratorBasedBuilder):
     ]
 
     def _info(self) -> DatasetInfo:
+        cfg = self.config
+        # C = number of fields
+        C = 4
+        T, H, W = cfg.timesteps, cfg.height, cfg.width
         features = Features({
-            "input": Sequence(Sequence(Sequence(Sequence(Value("float32"))))),
-            "output": Sequence(Sequence(Sequence(Sequence(Value("float32"))))),
+            "input":  Array4D(dtype="float32", shape=(T, C, H, W)),
+            "output": Array4D(dtype="float32", shape=(T, C, H, W)),
             "fluid_params": Sequence(Value("float32")),
-            "filename": Value("string"),
+            "filename":    Value("string"),
         })
         return DatasetInfo(
             description=_DESCRIPTION,
@@ -102,50 +116,54 @@ class BubbleMLDataset(GeneratorBasedBuilder):
         cfg = self.config
 
         if cfg.data_dir:
-            # Local testing: use the provided data_dir
+            # Local: read files from disk
             base_dir = cfg.data_dir
             def resolve_local(split):
-                paths = []
+                pairs = []
                 for rel in cfg.data_files[split]:
-                    full = os.path.join(base_dir, rel)
-                    if not os.path.isfile(full):
-                        raise FileNotFoundError(f"Expected {full}, but it was not found.")
-                    paths.append((full, full.replace(".hdf5", ".json")))
-                return paths
-            train_pairs = resolve_local("train")
-            test_pairs  = resolve_local("test")
+                    h5p = os.path.join(base_dir, rel)
+                    jp  = h5p.replace(".hdf5", ".json")
+                    if not os.path.isfile(h5p):
+                        raise FileNotFoundError(f"{h5p} not found.")
+                    if not os.path.isfile(jp):
+                        raise FileNotFoundError(f"{jp} not found.")
+                    pairs.append((h5p, jp))
+                return pairs
+            train = resolve_local("train")
+            test  = resolve_local("test")
         else:
-            # Remote download: fetch each .hdf5 and .json from the Hub
+            # Remote: download from Hub
             base_url = "https://huggingface.co/datasets/hpcforge/BubbleML_2/resolve/main/"
-            # build map of rel -> url
             url_map = {}
-            for split in ["train", "test"]:
+            for split in ["train","test"]:
                 for rel in cfg.data_files[split]:
                     url_map[rel] = base_url + rel
-                    meta = rel.replace(".hdf5", ".json")
-                    url_map[meta] = base_url + meta
-            # download all
-            downloaded = dl_manager.download(url_map)
-            # resolve into pairs
-            train_pairs = []
-            test_pairs  = []
-            for split in ["train", "test"]:
+                    url_map[rel.replace(".hdf5", ".json")] = base_url + rel.replace(".hdf5", ".json")
+            dl = dl_manager.download(url_map)
+            train, test = [], []
+            for split in ["train","test"]:
                 for rel in cfg.data_files[split]:
-                    h5_path   = downloaded[rel]
-                    json_path = downloaded[rel.replace(".hdf5", ".json")]
-                    train_pairs.append((h5_path, json_path)) if split == "train" else test_pairs.append((h5_path, json_path))
+                    h5p = dl[rel]
+                    jp  = dl[rel.replace(".hdf5", ".json")]
+                    (train if split=="train" else test).append((h5p,jp))
 
         return [
-            datasets.SplitGenerator(name=Split.TRAIN, gen_kwargs={"file_pairs": train_pairs}),
-            datasets.SplitGenerator(name=Split.TEST,  gen_kwargs={"file_pairs": test_pairs}),
+            datasets.SplitGenerator(name=Split.TRAIN, gen_kwargs={"file_pairs": train}),
+            datasets.SplitGenerator(name=Split.TEST,  gen_kwargs={"file_pairs": test}),
         ]
 
     def _generate_examples(self, file_pairs):
-        """Yield examples from each (hdf5,json) pair."""
-        for idx, (h5_path, json_path) in enumerate(file_pairs):
+        """Yield sliding-window examples, yielding NumPy arrays directly."""
+        idx = 0
+        fields = ["dfun", "temperature", "velx", "vely"]
+        tw = self.config.timesteps
+        for h5_path, json_path in file_pairs:
             with h5py.File(h5_path, "r") as h5f:
-                inp = np.stack([h5f[k][:5] for k in ["dfun", "temperature", "velx", "vely"]])
-                out = np.stack([h5f[k][5:10] for k in ["dfun", "temperature", "velx", "vely"]])
+                arrays = {k: h5f[k][...] for k in fields}  # (T, H, W)
+            T = arrays[fields[0]].shape[0]
+            max_start = T - 2*tw + 1
+
+            # metadata
             with open(json_path, "r") as jf:
                 p = json.load(jf)
             fluid_params = [
@@ -154,14 +172,27 @@ class BubbleMLDataset(GeneratorBasedBuilder):
                 p["mugas"],
                 p["rhogas"],
                 p["thcogas"],
-                p["stefan"],
+                p.get("stefan", 0.0),
                 p["prandtl"],
                 p["heater"]["nucWaitTime"],
-                p["heater"]["wallTemp"],
+                p["heater"].get("wallTemp", 0.0),
             ]
-            yield idx, {
-                "input": inp.astype(np.float32).tolist(),
-                "output": out.astype(np.float32).tolist(),
-                "fluid_params": fluid_params,
-                "filename": os.path.basename(h5_path),
-            }
+
+            for start in range(max_start):
+                ei = start + tw
+                eo = ei + tw
+
+                inp_c_t_h_w = np.stack([arrays[k][start:ei] for k in fields], axis=0)
+                out_c_t_h_w = np.stack([arrays[k][ei:eo] for k in fields], axis=0)
+
+                # transpose (C,T,H,W) -> (T,C,H,W)
+                inp = inp_c_t_h_w.transpose(1,0,2,3).astype("float32")
+                out = out_c_t_h_w.transpose(1,0,2,3).astype("float32")
+
+                yield idx, {
+                    "input": inp,
+                    "output": out,
+                    "fluid_params": fluid_params,
+                    "filename": os.path.basename(h5_path),
+                }
+                idx += 1