Implemented updating callback metric only if key not exists (less aff…

…ecting other users code).

Moved CHANGELOG.md entry to unreleased section.

CHANGELOG.md

@@ -16,6 +16,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
+- Fixed callback metric getting overwritten by progress bar or log metric ([#1800](https://github.com/PyTorchLightning/pytorch-lightning/pull/1800))
 
 ## [0.8.0] - 2020-06-DD
 
@@ -70,7 +71,6 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed an issue with `_auto_collect_arguments` collecting local variables that are not constructor arguments and not working for signatures that have the instance not named `self` ([#2048](https://github.com/PyTorchLightning/pytorch-lightning/pull/2048))
 - Fixed mistake in parameters' grad norm tracking ([#2012](https://github.com/PyTorchLightning/pytorch-lightning/pull/2012))
 - Fixed CPU and hanging GPU crash ([#2118](https://github.com/PyTorchLightning/pytorch-lightning/pull/2118))
-- Fixed callback metric getting overwritten by progress bar or log metric ([#1800](https://github.com/PyTorchLightning/pytorch-lightning/pull/1800))
 
 ## [0.7.6] - 2020-05-16
 

pytorch_lightning/trainer/logging.py

@@ -168,6 +168,19 @@ def process_output(self, output, train=False):
         # ---------------
         hiddens = output.get('hiddens')
 
+        # iterate over log_metric and progress_bar metric values
+        # and add it to the callback metric dict because every 
+        # metric value of logging or progressbar could be a candidate 
+        # for early stopping or similar
+        # 
+        # NOTE: through the dict looping sequence a priority is defined
+        # so first log metrics values will be added if not existing and
+        # then progress bar values if not existing in callback and log metric 
+        for metric_dict in [log_metrics, progress_bar_metrics]:
+            for key in metric_dict.keys():
+                if key not in callback_metrics.keys():
+                    callback_metrics[key] = metric_dict[key]
+
         # detach all metrics for callbacks to prevent memory leaks
         # no .item() because it will slow things down
         callback_metrics = recursive_detach(callback_metrics)