fix result obj dp auto reduce (#3013)

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* fix result for dp

* added warning when changing monitor and using results obj

pytorch_lightning/core/step_result.py

@@ -361,6 +361,14 @@ class Result(Dict):
         result['meta'] = meta
         return result
 
+    def dp_reduce(self):
+        for k, value in self.items():
+            if k == 'meta':
+                continue
+            if isinstance(value, list):
+                value = torch.tensor(value)
+            self[k] = value.mean(dim=-1)
+
     @property
     def should_reduce_on_epoch_end(self) -> bool:
         return self['meta']['_internal']['_reduce_on_epoch']

pytorch_lightning/trainer/evaluation_loop.py

@@ -343,17 +343,20 @@ class TrainerEvaluationLoopMixin(ABC):
                     m = 'only EvalResults or dicts are allowed from validation_step'
                     raise MisconfigurationException(m)
 
+                # ------------------
+                # EVAL STEP END
+                # ------------------
                 # on dp / ddp2 might still want to do something with the batch parts
-                if test_mode:
-                    if self.is_overridden('test_step_end'):
-                        model_ref = self.get_model()
-                        with self.profiler.profile('test_step_end'):
-                            output = model_ref.test_step_end(output)
-                else:
-                    if self.is_overridden('validation_step_end'):
-                        model_ref = self.get_model()
-                        with self.profiler.profile('validation_step_end'):
-                            output = model_ref.validation_step_end(output)
+                eval_step_end_hook_name = 'test_step_end' if test_mode else 'validation_step_end'
+                if self.is_overridden(eval_step_end_hook_name):
+                    model_ref = self.get_model()
+                    with self.profiler.profile(eval_step_end_hook_name):
+                        eval_step_end = getattr(model_ref, eval_step_end_hook_name)
+                        output = eval_step_end(output)
+
+                elif is_result_obj and (self.use_dp or self.use_ddp2):
+                    # result auto reduce
+                    output.dp_reduce()
 
                 # callbacks (on __batch_end)
                 if test_mode:

pytorch_lightning/trainer/training_loop.py

@@ -1221,6 +1221,8 @@ class TrainerTrainLoopMixin(ABC):
         else:
             output = self.model.training_step(*args)
 
+        is_result_obj = isinstance(output, Result)
+
         # allow any mode to define training_step_end
         # do something will all the dp outputs (like softmax)
         if self.is_overridden('training_step_end'):
@@ -1229,6 +1231,9 @@ class TrainerTrainLoopMixin(ABC):
                 # TODO: modify when using result obj
                 output = model_ref.training_step_end(output)
 
+        elif is_result_obj and (self.use_dp or self.use_ddp2):
+            output.dp_reduce()
+
         # allow any mode to define training_end
         # TODO: remove in 1.0.0
         if self.is_overridden('training_end'):

tests/base/model_train_steps.py

@@ -79,6 +79,28 @@ class TrainingStepVariations(ABC):
         self.training_step_called = True
         return result
 
+    def training_step_result_obj_dp(self, batch, batch_idx, optimizer_idx=None):
+        # forward pass
+        x, y = batch
+        x = x.view(x.size(0), -1)
+        y_hat = self(x.to(self.device))
+
+        # calculate loss
+        loss_val = self.loss(y.to(y_hat.device), y_hat)
+        log_val = loss_val
+
+        # alternate between tensors and scalars for "log" and "progress_bar"
+        if batch_idx % 2 == 0:
+            log_val = log_val.item()
+
+        result = TrainResult(loss_val)
+        result.log('some_val', log_val * log_val, prog_bar=True, logger=False)
+        result.log('train_some_val', log_val * log_val)
+
+        self.training_step_called = True
+
+        return result
+
     def training_step_end_full_loop_result_obj_dp(self, result):
         """
         Full loop flow train step (result obj + dp)

tests/base/model_valid_steps.py

@@ -52,6 +52,28 @@ class ValidationStepVariations(ABC):
         })
         return result
 
+    def validation_step_result_obj_dp(self, batch, batch_idx, *args, **kwargs):
+        x, y = batch
+        x = x.view(x.size(0), -1)
+        y_hat = self(x.to(self.device))
+
+        y = y.to(y_hat.device)
+        loss_val = self.loss(y, y_hat)
+
+        # acc
+        labels_hat = torch.argmax(y_hat, dim=1)
+        val_acc = torch.sum(y == labels_hat).item() / (len(y) * 1.0)
+        val_acc = torch.tensor(val_acc).type_as(x)
+
+        result = EvalResult(checkpoint_on=loss_val, early_stop_on=loss_val)
+        result.log_dict({
+            'val_loss': loss_val,
+            'val_acc': val_acc,
+        })
+
+        self.validation_step_called = True
+        return result
+
     def validation_step__multiple_dataloaders(self, batch, batch_idx, dataloader_idx, **kwargs):
         """
         Lightning calls this inside the validation loop

tests/trainer/test_trainer_steps_result_return.py

@@ -535,6 +535,41 @@ def test_full_train_loop_with_results_obj_dp(tmpdir):
     assert 'epoch_train_epoch_end_metric' in seen_keys
 
 
+@pytest.mark.skipif(torch.cuda.device_count() < 2, reason="test requires multi-GPU machine")
+def test_loop_steps_only_dp(tmpdir):
+    os.environ['PL_DEV_DEBUG'] = '1'
+
+    batches = 10
+    epochs = 3
+
+    model = EvalModelTemplate()
+    model.validation_step = None
+    model.test_step = None
+    model.training_step = model.training_step_result_obj_dp
+    model.training_step_end = None
+    model.training_epoch_end = None
+    model.validation_step = model.validation_step_result_obj_dp
+    model.validation_step_end = None
+    model.validation_epoch_end = None
+    model.test_dataloader = None
+
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        distributed_backend='dp',
+        gpus=[0, 1],
+        max_epochs=epochs,
+        early_stop_callback=True,
+        row_log_interval=2,
+        limit_train_batches=batches,
+        weights_summary=None,
+    )
+
+    trainer.fit(model)
+
+    assert model.training_step_called
+    assert model.validation_step_called
+
+
 def test_result_map(tmpdir):
     result = TrainResult()
     result.log_dict({'x1': torch.tensor(1), 'x2': torch.tensor(2)})