Do not add return dict items to callback_metrics (#6682)

CHANGELOG.md

@@ -10,6 +10,10 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Changed the behavior of `on_epoch_start` to run at the beginning of validation & test epoch ([#6498](https://github.com/PyTorchLightning/pytorch-lightning/pull/6498))
 
+### Removed
+
+- Removed legacy code to include `step` dictionary returns in `callback_metrics`. Use `self.log_dict` instead. ([#6682](https://github.com/PyTorchLightning/pytorch-lightning/pull/6682))
+
 ### Fixed
 
 - Fixed `DummyLogger.log_hyperparams` raising a `TypeError` when running with `fast_dev_run=True` ([#6398](https://github.com/PyTorchLightning/pytorch-lightning/pull/6398))

docs/source/ecosystem/asr_nlp_tts.rst

@@ -270,12 +270,12 @@ with PyTorch Lightning since every NeMo model is a Lightning Module.
                 log_probs=log_probs, targets=transcript, input_lengths=encoded_len, target_lengths=transcript_len
             )
             wer_num, wer_denom = self._wer(predictions, transcript, transcript_len)
-            tensorboard_logs = {
+            self.log_dict({
                 'train_loss': loss_value,
                 'training_batch_wer': wer_num / wer_denom,
                 'learning_rate': self._optimizer.param_groups[0]['lr'],
-            }
-            return {'loss': loss_value, 'log': tensorboard_logs}
+            })
+            return loss_value
 
 Neural Types in NeMo ASR
 ------------------------
@@ -539,8 +539,8 @@ since every NeMo model is a Lightning Module.
             logits = self(input_ids=input_ids, token_type_ids=input_type_ids, attention_mask=input_mask)
 
             loss = self.loss(logits=logits, labels=labels, loss_mask=loss_mask)
-            tensorboard_logs = {'train_loss': loss, 'lr': self._optimizer.param_groups[0]['lr']}
-            return {'loss': loss, 'log': tensorboard_logs}
+            self.log_dict({'train_loss': loss, 'lr': self._optimizer.param_groups[0]['lr']})
+            return loss
         ...
 
 Neural Types in NeMo NLP

docs/source/ecosystem/bolts.rst

@@ -68,8 +68,8 @@ you can trust the implementations and use them to bootstrap your research much f
 
             loss = self.criterion(logits.view(-1, logits.size(-1)), x.view(-1).long())
 
-            logs = {"loss": loss}
-            return {"loss": loss, "log": logs}
+            self.log("loss", loss)
+            return loss
 
 ----------
 

pytorch_lightning/callbacks/model_checkpoint.py

@@ -490,7 +490,7 @@ def _validate_monitor_key(self, trainer):
             m = (
                 f"ModelCheckpoint(monitor='{self.monitor}') not found in the returned metrics:"
                 f" {list(metrics.keys())}. "
-                f"HINT: Did you call self.log('{self.monitor}', tensor) in the LightningModule?"
+                f"HINT: Did you call self.log('{self.monitor}', value) in the LightningModule?"
             )
             raise MisconfigurationException(m)
 

pytorch_lightning/trainer/connectors/logger_connector/epoch_result_store.py

@@ -346,7 +346,6 @@ def update_logger_connector(self) -> Tuple[Dict, Dict]:
 
         # update callback_metrics
         logger_connector._callback_metrics.update(callback_metrics)
-        logger_connector._callback_metrics.pop("epoch", None)
 
         batch_pbar_metrics.pop("debug_epoch", None)
         return batch_pbar_metrics, batch_log_metrics

pytorch_lightning/trainer/connectors/logger_connector/logger_connector.py

@@ -78,7 +78,7 @@ def progress_bar_metrics(self, progress_bar_metrics: Dict) -> None:
 
     @property
     def cached_results(self) -> Union[EpochResultStore, None]:
-        return self._cached_results.get(self.trainer._running_stage)  # type: ignore
+        return self._cached_results.get(self.trainer._running_stage)
 
     def get_metrics(self, key: str) -> Dict:
         metrics_holder = getattr(self, f"_{key}", None)
@@ -125,8 +125,6 @@ def cache_logged_metrics(self):
     def on_trainer_init(self, logger, flush_logs_every_n_steps: int, log_every_n_steps: int, move_metrics_to_cpu: bool):
         # logging
         self.configure_logger(logger)
-        # todo: IDE is complaining, these shall be initialized in the Trainer init at leas as placeholders
-        # and assign here the desired value
         self.trainer.flush_logs_every_n_steps = flush_logs_every_n_steps
         self.trainer.log_every_n_steps = log_every_n_steps
         self.trainer.move_metrics_to_cpu = move_metrics_to_cpu
@@ -189,9 +187,6 @@ def cache_training_step_metrics(self, opt_closure_result):
             batch_log_metrics = opt_closure_result.training_step_output.log_metrics
             logged_metrics_tmp.update(batch_log_metrics)
 
-            callback_metrics = opt_closure_result.training_step_output.callback_metrics
-            callback_metrics_tmp.update(callback_metrics)
-
             batch_pbar_metrics = opt_closure_result.training_step_output.pbar_on_batch_end
             pbar_metrics_tmp.update(batch_pbar_metrics)
 
@@ -214,9 +209,6 @@ def log_metrics(self, metrics, grad_norm_dic, step=None):
             metrics (dict): Metric values
             grad_norm_dic (dict): Gradient norms
             step (int): Step for which metrics should be logged. Default value corresponds to `self.global_step`
-            log_train_step_metrics (bool): Used to track if `log_metrics` function is being called in during training
-                steps. In training steps, we will log metrics on step: `total_nb_idx` (for accumulated gradients)
-                and global_step for the rest.
         """
         # add gpu memory
         if self.trainer._device_type == DeviceType.GPU and self.log_gpu_memory:
@@ -350,27 +342,6 @@ def _track_callback_metrics(self, eval_results):
             if self.trainer.testing:
                 self.trainer.logger_connector.evaluation_callback_metrics.update(flat)
 
-    def __process_eval_epoch_end_results_and_log_legacy_update(self, prog_bar_metrics, log_metrics, callback_metrics):
-        # eval loop returns all metrics
-        dataloader_result_metrics = {**prog_bar_metrics, **log_metrics, **callback_metrics}
-
-        # add metrics to prog bar
-        self.trainer.logger_connector.add_progress_bar_metrics(prog_bar_metrics)
-
-        # log metrics
-        if len(log_metrics) > 0:
-            self.trainer.logger_connector.log_metrics(log_metrics, {})
-
-        # track metrics for callbacks (all prog bar, logged and callback metrics)
-        callback_metrics.update(log_metrics)
-        callback_metrics.update(prog_bar_metrics)
-        self.trainer.logger_connector.callback_metrics.update(callback_metrics)
-        if self.trainer.testing:
-            self.trainer.logger_connector.evaluation_callback_metrics.update(callback_metrics)
-
-        if len(dataloader_result_metrics) > 0:
-            self.eval_loop_results.append(dataloader_result_metrics)
-
     def __process_eval_epoch_end_results_and_log_legacy(self, eval_results):
         if self.trainer.running_sanity_check:
             return
@@ -381,21 +352,21 @@ def __process_eval_epoch_end_results_and_log_legacy(self, eval_results):
             if not isinstance(eval_results, list):
                 eval_results = [eval_results]
 
-            num_loaders: int = self.trainer.evaluation_loop.num_dataloaders
-            prog_bar_metrics, log_metrics, callback_metrics = {}, {}, {}
-
             for result_idx, result in enumerate(eval_results):
-                _, prog_bar_metrics, log_metrics, callback_metrics, _ = self.trainer.process_dict_result(result)
+                _, prog_bar_metrics, log_metrics, _ = self.trainer.process_dict_result(result)
+
+                # eval loop returns all metrics
+                dataloader_result_metrics = {**prog_bar_metrics, **log_metrics}
+
+                # add metrics to prog bar
+                self.trainer.logger_connector.add_progress_bar_metrics(prog_bar_metrics)
 
-                if num_loaders > 1:
-                    self.__process_eval_epoch_end_results_and_log_legacy_update(
-                        prog_bar_metrics, log_metrics, callback_metrics
-                    )
+                # log metrics
+                if len(log_metrics) > 0:
+                    self.trainer.logger_connector.log_metrics(log_metrics, {})
 
-            if num_loaders == 1:
-                self.__process_eval_epoch_end_results_and_log_legacy_update(
-                    prog_bar_metrics, log_metrics, callback_metrics
-                )
+                if len(dataloader_result_metrics) > 0:
+                    self.eval_loop_results.append(dataloader_result_metrics)
 
     def on_train_epoch_end(self):
         # inform cached logger connector epoch finished
@@ -448,10 +419,9 @@ def log_train_epoch_end_metrics(
 
         # TODO: deprecate 1.0
         else:
-            out = self.__run_legacy_training_epoch_end(
-                num_optimizers, epoch_output, model, is_result_obj, epoch_callback_metrics
+            epoch_log_metrics, epoch_progress_bar_metrics = self.__run_legacy_training_epoch_end(
+                num_optimizers, epoch_output, model, is_result_obj
             )
-            epoch_log_metrics, epoch_progress_bar_metrics, epoch_callback_metrics = out
 
         # it will perform reduction over epoch and return log metrics
         cached_epoch_log_metrics = self.cached_results.get_epoch_log_metrics()
@@ -503,9 +473,7 @@ def training_epoch_end(self, model, epoch_output, num_optimizers):
         # capture logging
         self.trainer.logger_connector.cache_logged_metrics()
 
-    def __run_legacy_training_epoch_end(
-        self, num_optimizers, epoch_output, model, is_result_obj, epoch_callback_metrics
-    ):
+    def __run_legacy_training_epoch_end(self, num_optimizers, epoch_output, model, is_result_obj):
 
         epoch_log_metrics = {}
         epoch_progress_bar_metrics = {}
@@ -536,15 +504,14 @@ def __run_legacy_training_epoch_end(
                 _processed_outputs = self.trainer.process_dict_result(epoch_output)
                 epoch_progress_bar_metrics = _processed_outputs[1]
                 epoch_log_metrics = _processed_outputs[2]
-                epoch_callback_metrics = _processed_outputs[3]
 
         # --------------------------
         # Structured Result (auto epoch end)
         # --------------------------
         elif is_result_obj:
             epoch_log_metrics, epoch_progress_bar_metrics = self.__auto_reduce_results_on_epoch_end(epoch_output)
 
-        return epoch_log_metrics, epoch_progress_bar_metrics, epoch_callback_metrics
+        return epoch_log_metrics, epoch_progress_bar_metrics
 
     def __auto_reduce_results_on_epoch_end(self, epoch_output):
         epoch_log_metrics = {}

pytorch_lightning/trainer/logging.py

@@ -21,6 +21,7 @@
 from pytorch_lightning.loggers import LightningLoggerBase
 from pytorch_lightning.utilities import DeviceType, DistributedType
 from pytorch_lightning.utilities.distributed import rank_zero_warn
+from pytorch_lightning.utilities.exceptions import MisconfigurationException
 from pytorch_lightning.utilities.memory import recursive_detach
 
 
@@ -42,8 +43,14 @@ class TrainerLoggingMixin(ABC):
 
     def metrics_to_scalars(self, metrics):
         new_metrics = {}
+        # TODO: this is duplicated in MetricsHolder. should be unified
         for k, v in metrics.items():
             if isinstance(v, torch.Tensor):
+                if v.numel() != 1:
+                    raise MisconfigurationException(
+                        f"The metric `{k}` does not contain a single element"
+                        f" thus it cannot be converted to float. Found `{v}`"
+                    )
                 v = v.item()
 
             if isinstance(v, dict):
@@ -81,23 +88,8 @@ def process_dict_result(self, output, train=False):
         if isinstance(output, torch.Tensor):
             progress_bar_metrics = {}
             log_metrics = {}
-            callback_metrics = {}
             hiddens = None
-            return output, progress_bar_metrics, log_metrics, callback_metrics, hiddens
-
-        # ---------------
-        # EXTRACT CALLBACK KEYS
-        # ---------------
-        # all keys not progress_bar or log are candidates for callbacks
-        callback_metrics = {}
-        if isinstance(output, Mapping):
-            for k, v in output.items():
-                if k not in ['progress_bar', 'log', 'hiddens']:
-                    callback_metrics[k] = v
-
-        if train and self._distrib_type in (DistributedType.DP, DistributedType.DDP2):
-            num_gpus = self.num_gpus
-            callback_metrics = self.reduce_distributed_output(callback_metrics, num_gpus)
+            return output, progress_bar_metrics, log_metrics, hiddens
 
         # ---------------
         # EXTRACT PROGRESS BAR KEYS
@@ -159,17 +151,12 @@ def process_dict_result(self, output, train=False):
         # ---------------
         hiddens = output.get('hiddens', None) if isinstance(output, Mapping) else None
 
-        # use every metric passed in as a candidate for callback
-        callback_metrics.update(progress_bar_metrics)
-        callback_metrics.update(log_metrics)
-
         # detach all metrics for callbacks to prevent memory leaks
         # no .item() because it will slow things down
-        callback_metrics = recursive_detach(callback_metrics)
         progress_bar_metrics = recursive_detach(progress_bar_metrics)
         log_metrics = recursive_detach(log_metrics)
 
-        return loss, progress_bar_metrics, log_metrics, callback_metrics, hiddens
+        return loss, progress_bar_metrics, log_metrics, hiddens
 
     def reduce_distributed_output(self, output, num_gpus):
         if num_gpus <= 1:

pytorch_lightning/trainer/trainer.py

@@ -859,15 +859,6 @@ def run_sanity_check(self, ref_model):
             # run eval step
             _, eval_results = self.run_evaluation(max_batches=self.num_sanity_val_batches)
 
-            # allow no returns from eval
-            if eval_results is not None and len(eval_results) > 0:
-                # when we get a list back, used only the last item
-                if isinstance(eval_results, list):
-                    eval_results = eval_results[-1]
-
-                _, _, _, callback_metrics, _ = self.process_dict_result(eval_results)
-                self.logger_connector.callback_metrics = callback_metrics
-
             self.on_sanity_check_end()
             self.running_sanity_check = False
 

pytorch_lightning/trainer/training_loop.py

@@ -360,8 +360,7 @@ def _process_training_step_output(self, training_step_output, split_batch):
             batch_loss=training_step_output[0],
             pbar_on_batch_end=training_step_output[1],
             log_metrics=training_step_output[2],
-            callback_metrics=training_step_output[3],
-            hiddens=training_step_output[4],
+            hiddens=training_step_output[3],
         )
         # if the user decides to finally reduce things in epoch_end, save raw output without graphs
         if isinstance(training_step_output_for_epoch_end, torch.Tensor):

tests/base/model_valid_epoch_ends.py

@@ -43,9 +43,8 @@ def _mean(res, key):
             val_loss_mean = val_loss_mean.item()
             val_acc_mean = val_acc_mean.item()
 
-        metrics_dict = {'early_stop_on': val_loss_mean, 'val_acc': val_acc_mean}
-        results = {'progress_bar': metrics_dict, 'log': metrics_dict}
-        return results
+        self.log('early_stop_on', val_loss_mean, prog_bar=True)
+        self.log('val_acc', val_acc_mean, prog_bar=True)
 
     def validation_epoch_end__multiple_dataloaders(self, outputs):
         """

tests/callbacks/test_early_stopping.py

@@ -127,7 +127,7 @@ class ModelOverrideValidationReturn(BoringModel):
 
         def validation_epoch_end(self, outputs):
             loss = self.validation_return_values[self.current_epoch]
-            return {"test_val_loss": loss}
+            self.log("test_val_loss", loss)
 
     model = ModelOverrideValidationReturn()
     early_stop_callback = EarlyStopping(monitor="test_val_loss", patience=patience, verbose=True)
@@ -217,7 +217,7 @@ class CurrentModel(BoringModel):
         def validation_epoch_end(self, outputs):
             losses = [8, 4, 2, 3, 4, 5, 8, 10]
             val_loss = losses[self.current_epoch]
-            self.log('abc', torch.tensor(val_loss))
+            self.log('abc', val_loss)
 
     model = CurrentModel()
 
@@ -231,28 +231,6 @@ def validation_epoch_end(self, outputs):
     assert trainer.current_epoch == 5, 'early_stopping failed'
 
 
-def test_early_stopping_functionality_arbitrary_key(tmpdir):
-    """Tests whether early stopping works with a custom key and dictionary results on val step."""
-
-    class CurrentModel(BoringModel):
-
-        def validation_epoch_end(self, outputs):
-            losses = [8, 4, 2, 3, 4, 5, 8, 10]
-            val_loss = losses[self.current_epoch]
-            return {'jiraffe': torch.tensor(val_loss)}
-
-    model = CurrentModel()
-
-    trainer = Trainer(
-        default_root_dir=tmpdir,
-        callbacks=[EarlyStopping(monitor='jiraffe')],
-        overfit_batches=0.20,
-        max_epochs=20,
-    )
-    trainer.fit(model)
-    assert trainer.current_epoch >= 5, 'early_stopping failed'
-
-
 @pytest.mark.parametrize('step_freeze, min_steps, min_epochs', [(5, 1, 1), (5, 1, 3), (3, 15, 1)])
 def test_min_steps_override_early_stopping_functionality(tmpdir, step_freeze, min_steps, min_epochs):
     """Excepted Behaviour:
@@ -269,7 +247,7 @@ def test_min_steps_override_early_stopping_functionality(tmpdir, step_freeze, mi
         when `early_stopping` is being triggered,
     THEN the highest between `min_epochs * len(train_dataloader)` and `min_steps` would be reached.
 
-    Caviat: IF min_steps is divisible by len(train_dataloader), then it will do min_steps + len(train_dataloader)
+    Caveat: IF min_steps is divisible by len(train_dataloader), then it will do min_steps + len(train_dataloader)
 
     This test validate those expected behaviours
     """
@@ -306,7 +284,7 @@ def validation_epoch_end(self, outputs):
                 self._count_decrease += 1
                 self._loss_value -= self._eps
             self._values.append(_mean)
-            return {"test_val_loss": _mean}
+            self.log('test_val_loss', _mean)
 
     model = Model(step_freeze)
     model.training_step_end = None