Properly handle no validation split

silnlp/nmt/config.py

@@ -454,6 +454,12 @@ def stats_max_size(self) -> int:
     def has_parent(self) -> bool:
         return "parent" in self.data
 
+    @property
+    def has_val_split(self) -> bool:
+        return any(
+            pair.is_val and (pair.size if pair.val_size is None else pair.val_size) > 0 for pair in self.corpus_pairs
+        )
+
     def set_seed(self) -> None:
         seed = self.data["seed"]
         set_seed(seed)

silnlp/nmt/hugging_face_config.py

@@ -239,6 +239,12 @@ def __init__(self, exp_dir: Path, config: dict) -> None:
 
         super().__init__(exp_dir, config)
 
+        # disable evaluation if there is no validation split
+        if not self.has_val_split:
+            config["eval"]["evaluation_strategy"] = "no"
+            config["eval"]["load_best_model_at_end"] = False
+            config["eval"]["early_stopping"] = None
+
     @property
     def model_dir(self) -> Path:
         return Path(self.train["output_dir"])

silnlp/nmt/test.py

@@ -487,7 +487,6 @@ def test(
 
     tokenizer = config.create_tokenizer()
     model = config.create_model()
-    _, best_step = model.get_checkpoint_path(CheckpointType.BEST)
     results: Dict[int, List[PairScore]] = {}
     step: int
     if checkpoint is not None:
@@ -523,7 +522,9 @@ def test(
         except ValueError:
             LOGGER.warn("No average checkpoint available.")
 
-    if best:
+    best_step = 0
+    if best and config.has_val_split:
+        _, best_step = model.get_checkpoint_path(CheckpointType.BEST)
         step = best_step
         if step not in results:
             results[step] = test_checkpoint(