Properly handle empty metric_names passed to Trainer._filter_metrics

composer/core/evaluator.py

@@ -86,11 +86,10 @@ def __init__(
         self.label = label
         self.dataloader = ensure_data_spec(dataloader)
 
-        self.metric_names = []
         if metric_names is not None:
             if not isinstance(metric_names, list):
                 raise ValueError(f'``metric_names`` should be a list of strings, not a {type(metric_names)}')
-            self.metric_names = metric_names
+        self.metric_names = metric_names
 
         self.subset_num_batches = subset_num_batches
         self._eval_interval = None

composer/trainer/trainer.py

@@ -106,14 +106,13 @@ def _get_default_scheduler_frequency(schedulers: Optional[Union[Scheduler, Seque
 def _filter_metrics(metrics: Dict[str, Metric], metric_names: Optional[List[str]]) -> Dict[str, Metric]:
     """Filter the metrics based on the given metric_names as regex strings (e.g. 'Accuracy', 'f1' for 'BinaryF1Score', 'Top-.' for 'Top-1 Accuracy' and 'Top-2 Accuracy', etc). If no metric_names are provided, all metrics will be returned."""
     metrics = deepcopy(metrics)
-    if not metric_names:
+    if metric_names is None:
         return metrics
-    else:
-        filtered_metrics = {}
-        for name, metric in metrics.items():
-            if any(re.match(f'.*{metric_name}.*', name, re.IGNORECASE) for metric_name in metric_names):
-                filtered_metrics[name] = metric
-        return filtered_metrics
+    filtered_metrics = {}
+    for name, metric in metrics.items():
+        if any(re.match(f'.*{metric_name}.*', name, re.IGNORECASE) for metric_name in metric_names):
+            filtered_metrics[name] = metric
+    return filtered_metrics
 
 
 def _validate_precision(precision: Precision, device: Device):