Also update progress_bar in training_epoch_end (#1724)

* update prog. bar metrics on train epoch end

* changelog

* wip test

* more thorough testing

* comments

* update docs

* move test

Co-authored-by: Jirka <jirka.borovec@seznam.cz>

CHANGELOG.md

@@ -14,6 +14,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Added type hints in `Trainer.fit()` and `Trainer.test()` to reflect that also a list of dataloaders can be passed in ([#1723](https://github.com/PyTorchLightning/pytorch-lightning/pull/1723)).
 
+- The progress bar metrics now also get updated in `training_epoch_end` ([#1724](https://github.com/PyTorchLightning/pytorch-lightning/pull/1724)).
+
 ### Changed
 
 - Reduction when `batch_size < num_gpus` ([#1609](https://github.com/PyTorchLightning/pytorch-lightning/pull/1609))

pytorch_lightning/core/lightning.py

@@ -257,6 +257,7 @@ class LightningModule(ABC, GradInformation, ModelIO, ModelHooks):
             May contain the following optional keys:
 
             - log (metrics to be added to the logger; only tensors)
+            - progress_bar (dict for progress bar display)
             - any metric used in a callback (e.g. early stopping).
 
         Note:
@@ -280,7 +281,8 @@ class LightningModule(ABC, GradInformation, ModelIO, ModelHooks):
 
                     # log training accuracy at the end of an epoch
                     results = {
-                        'log': {'train_acc': train_acc_mean.item()}
+                        'log': {'train_acc': train_acc_mean.item()},
+                        'progress_bar': {'train_acc': train_acc_mean},
                     }
                     return results
 
@@ -303,6 +305,7 @@ class LightningModule(ABC, GradInformation, ModelIO, ModelHooks):
                     # log training accuracy at the end of an epoch
                     results = {
                         'log': {'train_acc': train_acc_mean.item(), 'step': self.current_epoch}
+                        'progress_bar': {'train_acc': train_acc_mean},
                     }
                     return results
         """

pytorch_lightning/trainer/training_loop.py

@@ -491,6 +491,7 @@ class TrainerTrainLoopMixin(ABC):
             callback_epoch_metrics = _processed_outputs[3]
             self.log_metrics(log_epoch_metrics, {})
             self.callback_metrics.update(callback_epoch_metrics)
+            self.add_progress_bar_metrics(_processed_outputs[1])
 
         # when no val loop is present or fast-dev-run still need to call checkpoints
         if not self.is_overridden('validation_step') and not (self.fast_dev_run or should_check_val):

tests/models/test_module_hooks.py

@@ -0,0 +1,46 @@
+import torch
+
+from pytorch_lightning import Trainer
+from tests.base import EvalModelTemplate
+
+import tests.base.utils as tutils
+
+
+def test_training_epoch_end_metrics_collection(tmpdir):
+    """ Test that progress bar metrics also get collected at the end of an epoch. """
+    num_epochs = 3
+    class CurrentModel(EvalModelTemplate):
+
+        def training_step(self, *args, **kwargs):
+            output = super().training_step(*args, **kwargs)
+            output['progress_bar'].update({'step_metric': torch.tensor(-1)})
+            output['progress_bar'].update({'shared_metric': 100})
+            return output
+
+        def training_epoch_end(self, outputs):
+            epoch = self.current_epoch
+            # both scalar tensors and Python numbers are accepted
+            return {
+                'progress_bar': {
+                    f'epoch_metric_{epoch}': torch.tensor(epoch),  # add a new metric key every epoch
+                    'shared_metric': 111,
+                }
+            }
+
+    model = CurrentModel(tutils.get_default_hparams())
+    trainer = Trainer(
+        max_epochs=num_epochs,
+        default_root_dir=tmpdir,
+        overfit_pct=0.1,
+    )
+    result = trainer.fit(model)
+    assert result == 1
+    metrics = trainer.progress_bar_dict
+
+    # metrics added in training step should be unchanged by epoch end method
+    assert metrics['step_metric'] == -1
+    # a metric shared in both methods gets overwritten by epoch_end
+    assert metrics['shared_metric'] == 111
+    # metrics are kept after each epoch
+    for i in range(num_epochs):
+        assert metrics[f'epoch_metric_{i}'] == i