Auto-sync: 2026-06-28 01:00:39

dovla_cil/training/trainer.py

@@ -120,6 +120,11 @@ class DoVLATrainer:
         self.train_group_ids, self.val_group_ids = _split_group_ids(
             self.dataset.group_ids, val_fraction=config.val_fraction, seed=config.seed
         )
+        self.policy_target_map_coverage = _policy_target_map_coverage(
+            self.policy_target_record_ids,
+            train_group_ids=self.train_group_ids,
+            val_group_ids=self.val_group_ids,
+        )
         self.device = self._resolve_device(config.device)
         self.model_config = DoVLAConfig(
             obs_dim=config.obs_dim,
@@ -163,6 +168,14 @@ class DoVLATrainer:
 
     def train(self) -> dict[str, Any]:
         write_json(self._resolved_config(), self.output_dir / "resolved_config.json")
+        if self.policy_target_record_ids:
+            coverage = self.policy_target_map_coverage
+            print(
+                "policy_target_map_coverage "
+                "train={train_covered}/{train_groups} ({train_coverage:.1%}) "
+                "val={val_covered}/{val_groups} ({val_coverage:.1%}) "
+                "targets={num_targets}".format(**coverage)
+            )
         if torch is None:
             return self._train_without_torch()
 
@@ -556,6 +569,7 @@ class DoVLATrainer:
             payload["backbone_feature_cache"] = str(payload["backbone_feature_cache"])
         if payload["policy_target_map"] is not None:
             payload["policy_target_map"] = str(payload["policy_target_map"])
+        payload["policy_target_map_coverage"] = self.policy_target_map_coverage
         return payload
 
     def _initialize_frozen_backbone_cache(self) -> None:
@@ -729,6 +743,27 @@ def _load_policy_target_map(path: str | Path | None) -> dict[str, str]:
     return output
 
 
+def _policy_target_map_coverage(
+    target_record_ids: dict[str, str],
+    *,
+    train_group_ids: list[str],
+    val_group_ids: list[str],
+) -> dict[str, float | int]:
+    train_covered = sum(1 for group_id in train_group_ids if group_id in target_record_ids)
+    val_covered = sum(1 for group_id in val_group_ids if group_id in target_record_ids)
+    train_groups = len(train_group_ids)
+    val_groups = len(val_group_ids)
+    return {
+        "num_targets": len(target_record_ids),
+        "train_groups": train_groups,
+        "train_covered": train_covered,
+        "train_coverage": train_covered / train_groups if train_groups else 0.0,
+        "val_groups": val_groups,
+        "val_covered": val_covered,
+        "val_coverage": val_covered / val_groups if val_groups else 0.0,
+    }
+
+
 def _record_score(record: CILRecord) -> tuple[float, float, str]:
     rank_score = -float(record.rank_within_group) if record.rank_within_group is not None else 0.0
     return record.reward.score, rank_score, record.record_id