Refactor geometry feature loaders and enhance distributed client status reporting

- Updated GTGeometryLoader to streamline geometry vector computation and caching.
- Introduced UNetGeometryLoader for UNet-derived geometry vectors.
- Added sample collection method in PapilaBundle for better data handling.
- Refined GeometrySegEncoder and ImageEncoder to utilize new sample collection.
- Enhanced distributed client with heartbeat mechanism for improved job tracking.
- Added new configuration files for UNet-derived geometry integration.

v4/classes/towers/geometry_tower.py

@@ -126,16 +126,14 @@ class GeometrySegEncoder(TowerBase):
     # ── EPC early_pass ───────────────────────────────────────────────────────
 
     def early_pass(self, context) -> None:
-        data   = context.require("data")
-        split  = context.require("split")
+        data  = context.require("data")
+        split = context.require("split")
 
-        train_samples = self._collect_samples(split.train, data)
-        all_samples   = self._collect_samples(split.train, data)
-        all_samples  += self._collect_samples(split.val,   data)
+        train_samples = data.collect_samples(split.train)
+        all_samples   = train_samples + data.collect_samples(split.val)
         if split.test is not None:
-            all_samples += self._collect_samples(split.test, data)
+            all_samples += data.collect_samples(split.test)
 
-        # Reset per-fold state if loader supports it (UNet only).
         if hasattr(self._loader, "reset_cache"):
             self._loader.reset_cache()
         if hasattr(self._loader, "reset_weights"):
@@ -172,18 +170,6 @@ class GeometrySegEncoder(TowerBase):
 
     # ── Internals ────────────────────────────────────────────────────────────
 
-    def _collect_samples(self, df, data) -> list:
-        """Build (pid, eye, image_path) tuples from a split DataFrame."""
-        if df is None or len(df) == 0:
-            return []
-        pc = data.patient_col
-        out = []
-        for _, row in df.iterrows():
-            pid = int(row[pc])
-            eye = str(row.get("eyeID", "OD"))
-            out.append((pid, eye, data.image.get_image_path(pid, eye)))
-        return out
-
     @staticmethod
     def _augment_array(arr: np.ndarray) -> np.ndarray:
         """Random flip + 90° rotation on a (C, H, W) seg-map array."""

v4/classes/towers/image_tower.py

@@ -151,11 +151,12 @@ class ImageEncoder(TowerBase):
 
     def early_pass(self, context) -> None:
         """Per-fold setup: warm tensor cache (if enabled), publish geometry vectors."""
-        data = context.require("data")
+        data  = context.require("data")
+        split = context.require("split")
 
         if self._cache_transformed:
             self._tensor_cache.clear()
-            n = self._warm_tensor_cache(data, context.require("split"))
+            n = self._warm_tensor_cache(data, split)
             print(
                 f"[ImageEncoder] warmed transformed-tensor cache for {n} entries "
                 f"({self._name})",
@@ -164,7 +165,20 @@ class ImageEncoder(TowerBase):
 
         if self._geom_loader is None:
             return
-        self._geom_loader.precompute(data.df, patient_col=data.patient_col)
+
+        train_samples = data.collect_samples(split.train)
+        all_samples   = train_samples + data.collect_samples(split.val)
+        if split.test is not None:
+            all_samples += data.collect_samples(split.test)
+
+        if hasattr(self._geom_loader, "reset_cache"):
+            self._geom_loader.reset_cache()
+        if hasattr(self._geom_loader, "reset_weights"):
+            self._geom_loader.reset_weights()
+        if hasattr(self._geom_loader, "finetune"):
+            self._geom_loader.finetune(train_samples)
+
+        self._geom_loader.precompute(all_samples)
         vecs = self._geom_loader.all_vectors()
         context.put(self.EPC_GEOMETRY_KEY, vecs)
         print(