Remove epoch from trainer.logged_metrics (Lightning-AI#9904)

CHANGELOG.md

@@ -451,6 +451,9 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Removed a redundant warning with `ModelCheckpoint(monitor=None)` callback ([#9875](https://github.com/PyTorchLightning/pytorch-lightning/pull/9875))
 
 
+- Remove `epoch` from `trainer.logged_metrics` ([#9904](https://github.com/PyTorchLightning/pytorch-lightning/pull/9904))
+
+
 ### Fixed
 
 

docs/source/common/trainer.rst

@@ -1706,7 +1706,7 @@ The metrics sent to the logger (visualizer).
 .. code-block:: python
 
     def training_step(self, batch, batch_idx):
-        self.log("a_val", 2, log=True)
+        self.log("a_val", 2, logger=True)
 
 
     logged_metrics = trainer.logged_metrics

pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py

@@ -103,6 +103,9 @@ def log_metrics(self, metrics: _OUT_DICT, step: Optional[int] = None) -> None:
 
         if step is None:
             step = scalar_metrics.pop("step", None)
+
+        self._logged_metrics.update(scalar_metrics)
+
         if step is None:
             # added metrics for convenience
             scalar_metrics.setdefault("epoch", self.trainer.current_epoch)
@@ -112,8 +115,6 @@ def log_metrics(self, metrics: _OUT_DICT, step: Optional[int] = None) -> None:
         self.trainer.logger.agg_and_log_metrics(scalar_metrics, step=step)
         self.trainer.logger.save()
 
-        self._logged_metrics.update(scalar_metrics)
-
     """
     Evaluation metric updates
     """

tests/accelerators/test_multi_nodes_gpu.py

@@ -109,7 +109,7 @@ def backward(self, loss, optimizer, optimizer_idx):
     trainer.fit(model)
 
     # make sure all the metrics are available for callbacks
-    assert set(trainer.logged_metrics) == {"a2", "a_step", "a_epoch", "b_step", "b_epoch", "epoch"}
+    assert set(trainer.logged_metrics) == {"a2", "a_step", "a_epoch", "b_step", "b_epoch"}
 
     # we don't want to enable val metrics during steps because it is not something that users should do
     # on purpose DO NOT allow b_step... it's silly to monitor val step metrics

tests/loops/batch/test_truncated_bptt.py

@@ -169,4 +169,4 @@ def training_step(self, *args, **kwargs):
         enable_checkpointing=False,
     )
     trainer.fit(model)
-    assert set(trainer.logged_metrics) == {"loss_step", "loss_epoch", "epoch"}
+    assert set(trainer.logged_metrics) == {"loss_step", "loss_epoch"}

tests/trainer/logging_/test_eval_loop_logging.py

@@ -55,7 +55,7 @@ def validation_step(self, batch, batch_idx):
     )
     trainer.fit(model)
 
-    assert set(trainer.logged_metrics) == {"a2", "a_step", "a_epoch", "b_step", "b_epoch", "epoch"}
+    assert set(trainer.logged_metrics) == {"a2", "a_step", "a_epoch", "b_step", "b_epoch"}
 
     # we don't want to enable val metrics during steps because it is not something that users should do
     # on purpose DO NOT allow b_step... it's silly to monitor val step metrics
@@ -94,7 +94,7 @@ def validation_epoch_end(self, outputs):
     trainer.fit(model)
 
     # make sure all the metrics are available for loggers
-    assert set(trainer.logged_metrics) == {"epoch", "a", "b_step", "b_epoch", "c", "d_step", "d_epoch", "g"}
+    assert set(trainer.logged_metrics) == {"a", "b_step", "b_epoch", "c", "d_step", "d_epoch", "g"}
 
     assert not trainer.progress_bar_metrics
 
@@ -123,14 +123,14 @@ def validation_epoch_end(self, outputs):
 
     # assert the loggers received the expected number
     logged_metrics = set(trainer.logged_metrics)
-    assert logged_metrics == {"c", "d/e/f", "epoch"}
+    assert logged_metrics == {"c", "d/e/f"}
 
     pbar_metrics = set(trainer.progress_bar_metrics)
     assert pbar_metrics == {"c"}
 
     # make sure all the metrics are available for callbacks
     callback_metrics = set(trainer.callback_metrics)
-    assert callback_metrics == (logged_metrics | pbar_metrics) - {"epoch"}
+    assert callback_metrics == (logged_metrics | pbar_metrics)
 
 
 def test_eval_float_logging(tmpdir):
@@ -153,7 +153,7 @@ def validation_step(self, batch, batch_idx):
     )
     trainer.fit(model)
 
-    assert set(trainer.logged_metrics) == {"a", "epoch"}
+    assert set(trainer.logged_metrics) == {"a"}
 
 
 def test_eval_logging_auto_reduce(tmpdir):

tests/trainer/logging_/test_train_loop_logging.py

@@ -84,12 +84,12 @@ def training_step(self, batch, batch_idx):
     trainer.fit(model)
 
     logged_metrics = set(trainer.logged_metrics)
-    assert logged_metrics == {"epoch", "default", "l_e", "l_s", "l_se_step", "l_se_epoch"}
+    assert logged_metrics == {"default", "l_e", "l_s", "l_se_step", "l_se_epoch"}
 
     pbar_metrics = set(trainer.progress_bar_metrics)
     assert pbar_metrics == {"p_e", "p_s", "p_se_step", "p_se_epoch"}
 
-    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"p_se", "l_se"}) - {"epoch"}
+    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"p_se", "l_se"})
 
 
 def test__training_step__epoch_end__log(tmpdir):
@@ -121,12 +121,12 @@ def training_epoch_end(self, outputs):
     trainer.fit(model)
 
     logged_metrics = set(trainer.logged_metrics)
-    assert logged_metrics == {"epoch", "a_step", "a_epoch", "b", "b1", "a1", "a2"}
+    assert logged_metrics == {"a_step", "a_epoch", "b", "b1", "a1", "a2"}
 
     pbar_metrics = set(trainer.progress_bar_metrics)
     assert pbar_metrics == {"b"}
 
-    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"a"}) - {"epoch"}
+    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"a"})
 
 
 @pytest.mark.parametrize(["batches", "log_interval", "max_epochs"], [(1, 1, 1), (64, 32, 2)])
@@ -162,12 +162,12 @@ def training_epoch_end(self, outputs):
 
     # make sure all the metrics are available for callbacks
     logged_metrics = set(trainer.logged_metrics)
-    assert logged_metrics == {"a_step", "a_epoch", "b_step", "b_epoch", "c", "d/e/f", "epoch"}
+    assert logged_metrics == {"a_step", "a_epoch", "b_step", "b_epoch", "c", "d/e/f"}
 
     pbar_metrics = set(trainer.progress_bar_metrics)
     assert pbar_metrics == {"c", "b_epoch", "b_step"}
 
-    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"a", "b"}) - {"epoch"}
+    assert set(trainer.callback_metrics) == (logged_metrics | pbar_metrics | {"a", "b"})
 
 
 @pytest.mark.parametrize(
@@ -237,7 +237,7 @@ def val_dataloader(self):
     )
     trainer.fit(model)
 
-    assert set(trainer.logged_metrics) == {"a_step", "a_epoch", "n_step", "n_epoch", "epoch"}
+    assert set(trainer.logged_metrics) == {"a_step", "a_epoch", "n_step", "n_epoch"}
 
 
 def test_log_works_in_train_callback(tmpdir):

tests/trainer/optimization/test_manual_optimization.py

@@ -193,7 +193,7 @@ def training_epoch_end(self, outputs) -> None:
     with mock.patch.object(Accelerator, "backward", wraps=trainer.accelerator.backward) as bwd_mock:
         trainer.fit(model)
     assert bwd_mock.call_count == limit_train_batches * 3
-    assert set(trainer.logged_metrics) == {"a_step", "a_epoch", "epoch"}
+    assert set(trainer.logged_metrics) == {"a_step", "a_epoch"}
 
 
 @RunIf(min_gpus=1)
@@ -1055,7 +1055,7 @@ def configure_optimizers(self):
 
     trainer.fit(model)
 
-    assert set(trainer.logged_metrics) == {"epoch", "loss_d", "loss_g"}
+    assert set(trainer.logged_metrics) == {"loss_d", "loss_g"}
     assert set(trainer.progress_bar_metrics) == {"loss_d", "loss_g"}