ref: (results 1/n) enable tracking original metric when step and epoc…

…h are both true (#3685)

* enable tracking original metric when step and epoch are both true

pytorch_lightning/core/datamodule.py

@@ -152,6 +152,8 @@ def __init__(
         self._has_setup_fit = False
         self._has_setup_test = False
 
+        self.trainer = None
+
     @property
     def train_transforms(self):
         """

pytorch_lightning/core/step_result.py

@@ -161,21 +161,22 @@ def log(
                 tbptt_pad_token=tbptt_pad_token,
             )
             self.__setitem__(epoch_name, value)
-        else:
-            self.__set_meta(
-                name,
-                value,
-                prog_bar,
-                logger,
-                on_step,
-                on_epoch,
-                reduce_fx,
-                tbptt_reduce_fx=tbptt_reduce_fx,
-                tbptt_pad_token=tbptt_pad_token,
-            )
 
-            # set the value
-            self.__setitem__(name, value)
+        # always log the original metric
+        self.__set_meta(
+            name,
+            value,
+            prog_bar,
+            logger,
+            on_step,
+            on_epoch,
+            reduce_fx,
+            tbptt_reduce_fx=tbptt_reduce_fx,
+            tbptt_pad_token=tbptt_pad_token,
+        )
+
+        # set the value
+        self.__setitem__(name, value)
 
     def __set_meta(
         self,
@@ -378,12 +379,17 @@ def reduce_on_epoch_end(cls, outputs):
     def reduce_across_time(cls, time_outputs):
         # auto-reduce across time for tbptt
         meta = time_outputs[0]['meta']
+
+        # in 1.0 the results have 'extra'. Once we deprecate 0.10.0 we may not need this
+        if 'extra' in time_outputs[0]:
+            [x.pop('extra', None) for x in time_outputs]
+
         result = cls()
         result = recursive_gather(time_outputs, result)
         recursive_stack(result)
 
         for k, value in result.items():
-            if k == 'meta':
+            if k in ['meta', 'extra']:
                 continue
 
             # pick the reduce fx

pytorch_lightning/utilities/model_utils.py

@@ -1,8 +1,9 @@
 from pytorch_lightning.core.lightning import LightningModule
 from pytorch_lightning.core.datamodule import LightningDataModule
+from typing import Union
 
 
-def is_overridden(method_name: str, model: LightningModule) -> bool:
+def is_overridden(method_name: str, model: Union[LightningModule, LightningDataModule]) -> bool:
     # if you pass DataModule instead of None or a LightningModule, we use LightningDataModule as super
     # TODO - refector this function to accept model_name, instance, parent so it makes more sense
     super_object = LightningModule if not isinstance(model, LightningDataModule) else LightningDataModule

tests/base/deterministic_model.py

@@ -188,6 +188,12 @@ def training_epoch_end_return_for_log_epoch_and_step(self, result):
             # only saw 4 batches
             assert isinstance(result, TrainResult)
 
+        result.step_epoch_log_acc2 = result.step_step_epoch_log_acc2.prod()
+        result.step_epoch_pbar_acc3 = result.step_step_epoch_pbar_acc3.prod()
+        result.step_epoch_log_and_pbar_acc1 = result.step_epoch_log_and_pbar_acc1.prod()
+        result.minimize = result.minimize.mean()
+        result.checkpoint_on = result.checkpoint_on.mean()
+
         result.step_step_epoch_log_and_pbar_acc1 = result.step_step_epoch_log_and_pbar_acc1.prod()
         result.epoch_step_epoch_log_and_pbar_acc1 = result.epoch_step_epoch_log_and_pbar_acc1.prod()
         result.step_step_epoch_log_acc2 = result.step_step_epoch_log_acc2.prod()

tests/base/model_train_steps.py

@@ -137,6 +137,7 @@ def eval_step_full_loop_result_obj_dp(self, batch, batch_idx, optimizer_idx=None
         result.log(f'{eval_name}_step_metric', loss_val + 1, on_step=True)
 
         setattr(self, f'{eval_name}_step_called', True)
+
         return result
 
     def eval_step_end_full_loop_result_obj_dp(self, result):
@@ -150,10 +151,14 @@ def eval_step_end_full_loop_result_obj_dp(self, result):
         reduced = getattr(result, f'epoch_{eval_name}_step_metric').mean()
         setattr(result, f'epoch_{eval_name}_step_metric', reduced)
 
+        reduced = getattr(result, f'{eval_name}_step_metric').mean()
+        setattr(result, f'{eval_name}_step_metric', reduced)
+
         result.checkpoint_on = result.checkpoint_on.mean()
         result.early_stop_on = result.early_stop_on.mean()
         result.log(f'{eval_name}_step_end_metric', torch.tensor(1).type_as(result.checkpoint_on))
         setattr(self, f'{eval_name}_step_end_called', True)
+
         return result
 
     def eval_epoch_end_full_loop_result_obj_dp(self, result):
@@ -176,6 +181,9 @@ def eval_epoch_end_full_loop_result_obj_dp(self, result):
         reduced = getattr(result, f'{eval_name}_step_end_metric').mean()
         setattr(result, f'{eval_name}_step_end_metric', reduced)
 
+        reduced = getattr(result, f'{eval_name}_step_metric').mean()
+        setattr(result, f'{eval_name}_step_metric', reduced)
+
         return result
 
     def training_step__using_metrics(self, batch, batch_idx, optimizer_idx=None):

tests/trainer/test_trainer_steps_result_return.py

@@ -201,7 +201,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
     assert not model.training_step_end_called
     assert not model.training_epoch_end_called
 
-    assert len(trainer.logger_connector.callback_metrics) == 8
+    assert len(trainer.logger_connector.callback_metrics) == 11
 
     # make sure correct metrics are logged (one per batch step as requested)
     assert len(trainer.dev_debugger.logged_metrics) == (epochs * batches) + epochs
@@ -227,7 +227,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
             assert logged_metrics['step_step_epoch_log_and_pbar_acc1'] == expected_val_1
             assert logged_metrics['step_step_epoch_log_acc2'] == expected_val_2
             assert 'step_epoch_pbar_acc3' not in logged_metrics
-            assert len(logged_metrics) == 4
+            assert len(logged_metrics) == 6
 
         # make sure the metrics for the epoch end are actual means (the default reduce fx) or all the batches
         epoch_end_metrics = epoch_outputs[-1]
@@ -236,7 +236,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
         assert epoch_end_metrics['epoch_step_epoch_log_and_pbar_acc1'] == eval_1
         assert epoch_end_metrics['epoch_step_epoch_log_acc2'] == eval_2
         assert 'step_epoch_pbar_acc3' not in epoch_end_metrics
-        assert len(logged_metrics) == 4
+        assert len(logged_metrics) == 6
 
     # make sure we are using the correct metrics for callbacks
     assert trainer.logger_connector.callback_metrics['checkpoint_on'] == 171
@@ -268,7 +268,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
             assert logged_metrics['step_step_epoch_log_and_pbar_acc1'] == expected_val_1
             assert logged_metrics['step_step_epoch_pbar_acc3'] == expected_val_2
             assert 'step_epoch_log_acc2' not in logged_metrics
-            assert len(logged_metrics) == 3
+            assert len(logged_metrics) == 5
 
         # make sure the metrics for the epoch end are actual means (the default reduce fx) or all the batches
         epoch_end_metrics = epoch_outputs[-1]
@@ -277,7 +277,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
         assert epoch_end_metrics['epoch_step_epoch_log_and_pbar_acc1'] == eval_1
         assert epoch_end_metrics['epoch_step_epoch_pbar_acc3'] == eval_2
         assert 'step_epoch_log_acc2' not in epoch_end_metrics
-        assert len(logged_metrics) == 3
+        assert len(logged_metrics) == 5
 
     # -----------------------------------------
     # make sure training outputs what is expected
@@ -287,7 +287,7 @@ def test_training_step_result_log_step_and_epoch(tmpdir):
 
     out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
     assert out.signal == 0
-    assert len(out.batch_log_metrics) == 2
+    assert len(out.batch_log_metrics) == 4
 
     train_step_out = out.training_step_output_for_epoch_end
     assert len(train_step_out) == 1
@@ -328,7 +328,7 @@ def test_training_step_epoch_end_result(tmpdir):
     )
     trainer.fit(model)
 
-    assert len(trainer.logger_connector.callback_metrics) == 11
+    assert len(trainer.logger_connector.callback_metrics) == 17
 
     # make sure correct steps were called
     assert model.training_step_called
@@ -369,7 +369,7 @@ def test_training_step_epoch_end_result(tmpdir):
 
     out = trainer.train_loop.run_training_batch(batch, batch_idx, 0)
     assert out.signal == 0
-    assert len(out.batch_log_metrics) == 2
+    assert len(out.batch_log_metrics) == 4
 
     train_step_out = out.training_step_output_for_epoch_end
     assert len(train_step_out) == 1

tests/trainer/test_validation_steps_result_return.py

@@ -278,7 +278,7 @@ def test_val_step_epoch_step_metrics(tmpdir):
     )
     trainer.fit(model)
 
-    assert len(trainer.logger_connector.callback_metrics) == 7
+    assert len(trainer.logger_connector.callback_metrics) == 11
 
     # make sure correct steps were called
     assert model.validation_step_called