fix error when logging to progress bar with reserved name

CHANGELOG.md

@@ -23,6 +23,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Fixed an error when logging a progress bar metric with a reserved name ([#5620](https://github.com/PyTorchLightning/pytorch-lightning/pull/5620))
+
 
 
 

pytorch_lightning/trainer/properties.py

@@ -183,9 +183,23 @@ def progress_bar_callback(self):
     @property
     def progress_bar_dict(self) -> dict:
         """ Read-only for progress bar metrics. """
-        ref_model = self.model if not self.data_parallel else self.model.module
+        ref_model = self.get_model()
         ref_model = cast(LightningModule, ref_model)
-        return dict(**ref_model.get_progress_bar_dict(), **self.logger_connector.progress_bar_metrics)
+
+        standard_metrics = ref_model.get_progress_bar_dict()
+        logged_metrics = self.progress_bar_metrics
+        duplicates = list(standard_metrics.keys() & logged_metrics.keys())
+        if duplicates:
+            rank_zero_warn(
+                f"The progress bar already tracks a metric with the name(s) '{', '.join(duplicates)}' and"
+                f" `self.log('{duplicates[0]}', ..., prog_bar=True)` will overwrite this value. "
+                f" If this is undesired, change the name or override `get_progress_bar_dict()`"
+                f" in `LightingModule`.",
+                UserWarning
+            )
+        all_metrics = dict(**standard_metrics)
+        all_metrics.update(**logged_metrics)
+        return all_metrics
 
     @property
     def disable_validation(self) -> bool:

tests/trainer/logging/test_logger_connector.py

@@ -425,3 +425,21 @@ def test_dataloader(self):
     )
     trainer.fit(model)
     trainer.test(model, ckpt_path=None)
+
+
+def test_logging_to_progress_bar_with_reserved_key(tmpdir):
+    """ Test that logging a metric with a reserved name to the progress bar raises a warning. """
+    class TestModel(BoringModel):
+
+        def training_step(self, *args, **kwargs):
+            output = super().training_step(*args, **kwargs)
+            self.log("loss", output["loss"], prog_bar=True)
+            return output
+
+    model = TestModel()
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        max_steps=2,
+    )
+    with pytest.warns(UserWarning, match="The progress bar already tracks a metric with the .* 'loss'"):
+        trainer.fit(model)