Separate epoch validation from step validation (#5208)

* Seperate epoch validaton from step validation

* update system

* test

* baked logic in callbacks

* unbake logic in callbacks

* fix the call for scheduler

* use property

* pep

* correct rebase

* gitignore

* ref

* add tests

* fix

* add early stopping test

* trigger

* chlog

* rev

* 1.3

* log

* Apply suggestions from code review

Co-authored-by: Carlos Mocholí <[email protected]>

* Update pytorch_lightning/trainer/training_loop.py

* Update CHANGELOG.md

* Apply suggestions from code review

Co-authored-by: chaton <[email protected]>
Co-authored-by: Carlos Mocholí <[email protected]>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Co-authored-by: Jirka Borovec <[email protected]>

(cherry picked from commit e429f97)

Author: rohitgr7

.gitignore

@@ -151,3 +151,6 @@ wandb
 
 # dataset generated from bolts in examples.
 cifar-10-batches-py
+
+# ctags
+tags

CHANGELOG.md

@@ -175,10 +175,11 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 - Fixed support custom DataLoader with DDP if they can be re-instantiated ([#5745](https://github.com/PyTorchLightning/pytorch-lightning/pull/5745))
 
 
-## [1.1.8] - 2021-02-06
+## [1.1.8] - 2021-02-08
 
 ### Fixed
 
+- Separate epoch validation from step validation ([#5208](https://github.com/PyTorchLightning/pytorch-lightning/pull/5208))
 - Fixed `toggle_optimizers` not handling all optimizer parameters ([#5775](https://github.com/PyTorchLightning/pytorch-lightning/pull/5775))
 
 

pytorch_lightning/callbacks/early_stopping.py

@@ -86,9 +86,6 @@ def __init__(
         self.stopped_epoch = 0
         self.mode = mode
         self.warned_result_obj = False
-        # Indicates, if eval results are used as basis for early stopping
-        # It is set to False initially and overwritten, if eval results have been validated
-        self.based_on_eval_results = False
 
         self.__init_monitor_mode()
 
@@ -159,21 +156,6 @@ def on_validation_end(self, trainer, pl_module):
 
         self._run_early_stopping_check(trainer, pl_module)
 
-    def on_validation_epoch_end(self, trainer, pl_module):
-        if trainer.fast_dev_run or trainer.running_sanity_check:
-            return
-
-        if self._validate_condition_metric(trainer.callback_metrics):
-            # turn off early stopping in on_train_epoch_end
-            self.based_on_eval_results = True
-
-    def on_train_epoch_end(self, trainer, pl_module, outputs):
-        # disable early stopping in train loop when there's a val loop
-        if self.based_on_eval_results:
-            return
-
-        self._run_early_stopping_check(trainer, pl_module)
-
     def _run_early_stopping_check(self, trainer, pl_module):
         """
         Checks whether the early stopping condition is met

pytorch_lightning/callbacks/model_checkpoint.py

@@ -167,7 +167,7 @@ def __init__(
         self.save_top_k = save_top_k
         self.save_weights_only = save_weights_only
         self.period = period
-        self.last_global_step_saved = -1
+        self._last_global_step_saved = -1
         self.prefix = prefix
         self.current_score = None
         self.best_k_models = {}
@@ -232,15 +232,15 @@ def save_checkpoint(self, trainer, pl_module):
             or self.period < 1  # no models are saved
             or (epoch + 1) % self.period  # skip epoch
             or trainer.running_sanity_check  # don't save anything during sanity check
-            or self.last_global_step_saved == global_step  # already saved at the last step
+            or self._last_global_step_saved == global_step  # already saved at the last step
         ):
             return
 
         self._add_backward_monitor_support(trainer)
         self._validate_monitor_key(trainer)
 
         # track epoch when ckpt was last checked
-        self.last_global_step_saved = global_step
+        self._last_global_step_saved = global_step
 
         # what can be monitored
         monitor_candidates = self._monitor_candidates(trainer)

pytorch_lightning/trainer/evaluation_loop.py

@@ -70,17 +70,8 @@ def get_evaluation_dataloaders(self, max_batches):
 
         return dataloaders, max_batches
 
-    def should_skip_evaluation(self, dataloaders, max_batches):
-        # skip when dataloaders aren't defined
-        if dataloaders is None:
-            return True
-
-        # enable disabling validation step with limit_val_batches = 0
-        should_skip = sum(max_batches) == 0
-        if should_skip:
-            return True
-
-        return False
+    def should_skip_evaluation(self, max_batches):
+        return sum(max_batches) == 0
 
     def on_evaluation_start(self, *args, **kwargs):
         if self.trainer.testing:

pytorch_lightning/trainer/trainer.py

@@ -598,9 +598,6 @@ def train(self):
                 if self.max_steps and self.max_steps <= self.global_step:
                     return
 
-                # update LR schedulers
-                self.optimizer_connector.update_learning_rates(interval='epoch')
-
                 # early stopping
                 met_min_epochs = epoch >= self.min_epochs - 1
                 met_min_steps = self.global_step >= self.min_steps if self.min_steps else True
@@ -626,7 +623,7 @@ def train(self):
             # hook
             self.train_loop.on_train_end()
 
-    def run_evaluation(self, max_batches=None):
+    def run_evaluation(self, max_batches=None, on_epoch=False):
 
         # used to know if we are logging for val, test + reset cached results
         self._set_wide_running_stage(RunningStage.TESTING if self.testing else RunningStage.EVALUATING)
@@ -639,7 +636,7 @@ def run_evaluation(self, max_batches=None):
         dataloaders, max_batches = self.evaluation_loop.get_evaluation_dataloaders(max_batches)
 
         # check if we want to skip this evaluation
-        if self.evaluation_loop.should_skip_evaluation(dataloaders, max_batches):
+        if self.evaluation_loop.should_skip_evaluation(max_batches):
             return [], []
 
         # ref model
@@ -705,6 +702,10 @@ def run_evaluation(self, max_batches=None):
         # hook
         self.evaluation_loop.on_evaluation_epoch_end()
 
+        # update epoch-level lr_schedulers
+        if on_epoch:
+            self.optimizer_connector.update_learning_rates(interval='epoch')
+
         # hook
         self.evaluation_loop.on_evaluation_end()
 

pytorch_lightning/trainer/training_loop.py

@@ -18,7 +18,7 @@
 import numpy as np
 import torch
 
-from pytorch_lightning.callbacks import ModelCheckpoint
+from pytorch_lightning.callbacks import EarlyStopping
 from pytorch_lightning.core.memory import ModelSummary
 from pytorch_lightning.core.optimizer import LightningOptimizer
 from pytorch_lightning.core.step_result import Result
@@ -161,7 +161,7 @@ def on_train_end(self):
         # trigger checkpoint check. need to temporarily decrease the global step to avoid saving duplicates
         # when a checkpoint was saved at the last step
         self.trainer.global_step -= 1
-        self.check_checkpoint_callback(should_save=True, is_last=True)
+        self.check_checkpoint_callback(should_update=True, is_last=True)
         self.trainer.global_step += 1
 
         # hook
@@ -184,18 +184,27 @@ def on_train_end(self):
             model.cpu()
             torch.cuda.empty_cache()
 
-    def check_checkpoint_callback(self, should_save, is_last=False):
-        # TODO bake this logic into the checkpoint callback
-        if should_save and self.trainer.checkpoint_connector.has_trained:
-            checkpoint_callbacks = [c for c in self.trainer.callbacks if isinstance(c, ModelCheckpoint)]
+    def check_checkpoint_callback(self, should_update, is_last=False):
+        # TODO bake this logic into the ModelCheckpoint callback
+        if should_update and self.trainer.checkpoint_connector.has_trained:
+            callbacks = self.trainer.checkpoint_callbacks
 
-            if is_last and any(c.save_last for c in checkpoint_callbacks):
+            if is_last and any(cb.save_last for cb in callbacks):
                 rank_zero_info("Saving latest checkpoint...")
 
             model = self.trainer.get_model()
 
-            for callback in checkpoint_callbacks:
-                callback.on_validation_end(self.trainer, model)
+            for cb in callbacks:
+                cb.on_validation_end(self.trainer, model)
+
+    def check_early_stopping_callback(self, should_update):
+        # TODO bake this logic into the EarlyStopping callback
+        if should_update and self.trainer.checkpoint_connector.has_trained:
+            callbacks = [c for c in self.trainer.callbacks if isinstance(c, EarlyStopping)]
+            model = self.trainer.get_model()
+
+            for cb in callbacks:
+                cb.on_validation_end(self.trainer, model)
 
     def on_train_epoch_start(self, epoch):
 
@@ -521,7 +530,6 @@ def tbptt_split_batch(self, batch):
         return splits
 
     def run_training_epoch(self):
-
         # get model
         model = self.trainer.get_model()
 
@@ -584,11 +592,12 @@ def run_training_epoch(self):
             self.trainer.checkpoint_connector.has_trained = True
 
             # max steps reached, end training
-            if self.trainer.max_steps is not None and self.trainer.max_steps == self.trainer.global_step + 1:
-                accumulation_done = self._accumulated_batches_reached()
-                # Ensure accumulation across batches has completed before breaking loop
-                if accumulation_done:
-                    break
+            if (
+                self.trainer.max_steps is not None
+                and self.trainer.max_steps == self.trainer.global_step + 1
+                and self._accumulated_batches_reached()
+            ):
+                break
 
             # end epoch early
             # stop when the flag is changed or we've gone past the amount
@@ -599,7 +608,7 @@ def run_training_epoch(self):
             self.trainer.total_batch_idx += 1
 
             # stop epoch if we limited the number of training batches
-            if (batch_idx + 1) >= self.trainer.num_training_batches:
+            if self._num_training_batches_reached(is_last_batch):
                 break
 
             # progress global step according to grads progress
@@ -613,8 +622,20 @@ def run_training_epoch(self):
             epoch_output, self.checkpoint_accumulator, self.early_stopping_accumulator, self.num_optimizers
         )
 
-        # when no val loop is present or fast-dev-run still need to call checkpoints
-        self.check_checkpoint_callback(not (should_check_val or is_overridden('validation_step', model)))
+        should_check_val = self.should_check_val_fx(batch_idx, is_last_batch, on_epoch=True)
+        if should_check_val:
+            self.trainer.run_evaluation(on_epoch=True)
+            # reset stage to train
+            self.trainer.logger_connector.set_stage("train")
+
+        should_skip_eval = self.trainer.evaluation_loop.should_skip_evaluation(self.trainer.num_val_batches)
+        should_train_only = self.trainer.disable_validation or should_skip_eval
+
+        if should_train_only:
+            # update epoch level lr_schedulers
+            self.trainer.optimizer_connector.update_learning_rates(interval='epoch')
+            self.check_checkpoint_callback(True)
+            self.check_early_stopping_callback(True)
 
         # increment the global step once
         # progress global step according to grads progress
@@ -840,25 +861,33 @@ def increment_accumulated_grad_global_step(self):
     def _accumulated_batches_reached(self):
         return (self.trainer.batch_idx + 1) % self.trainer.accumulate_grad_batches == 0
 
-    def _num_training_batches_reached(self):
-        return (self.trainer.batch_idx + 1) == self.trainer.num_training_batches
+    def _num_training_batches_reached(self, is_last_batch=False):
+        return (self.trainer.batch_idx + 1) == self.trainer.num_training_batches or is_last_batch
 
     def should_accumulate(self):
         # checks if backward or backward + optimizer step (via closure)
         accumulation_done = self._accumulated_batches_reached()
         is_final_batch = self._num_training_batches_reached()
         return not (accumulation_done or is_final_batch)
 
-    def should_check_val_fx(self, batch_idx, is_last_batch):
+    def should_check_val_fx(self, batch_idx, is_last_batch, on_epoch=False):
         # decide if we should run validation
         is_val_check_batch = (batch_idx + 1) % self.trainer.val_check_batch == 0
         is_val_check_epoch = (self.trainer.current_epoch + 1) % self.trainer.check_val_every_n_epoch == 0
         can_check_val = self.trainer.enable_validation and is_val_check_epoch
-        should_check_val = is_val_check_batch or self.trainer.should_stop
         is_last_batch_for_infinite_dataset = is_last_batch and self.trainer.val_check_batch == float("inf")
-        should_check_val = can_check_val and (should_check_val or is_last_batch_for_infinite_dataset)
+        epoch_end_val_check = self.trainer.val_check_batch == self.trainer.num_training_batches
+
+        should_check_val = (
+            (is_val_check_batch and epoch_end_val_check)
+            or self.trainer.should_stop
+            or is_last_batch_for_infinite_dataset
+        ) if on_epoch else (
+            is_val_check_batch
+            and not epoch_end_val_check
+        )
 
-        return should_check_val
+        return should_check_val and can_check_val
 
     def build_train_args(self, batch, batch_idx, opt_idx, hiddens):
         # enable not needing to add opt_idx to training_step

tests/callbacks/test_early_stopping.py

@@ -115,11 +115,9 @@ def test_early_stopping_patience(tmpdir, loss_values, patience, expected_stop_ep
 
     class ModelOverrideValidationReturn(EvalModelTemplate):
         validation_return_values = torch.Tensor(loss_values)
-        count = 0
 
         def validation_epoch_end(self, outputs):
-            loss = self.validation_return_values[self.count]
-            self.count += 1
+            loss = self.validation_return_values[self.current_epoch]
             return {"test_val_loss": loss}
 
     model = ModelOverrideValidationReturn()
@@ -135,6 +133,41 @@ def validation_epoch_end(self, outputs):
     assert trainer.current_epoch == expected_stop_epoch
 
 
+@pytest.mark.parametrize('validation_step', ['base', None])
+@pytest.mark.parametrize(
+    "loss_values, patience, expected_stop_epoch",
+    [
+        ([6, 5, 5, 5, 5, 5], 3, 4),
+        ([6, 5, 4, 4, 3, 3], 1, 3),
+        ([6, 5, 6, 5, 5, 5], 3, 4),
+    ],
+)
+def test_early_stopping_patience_train(tmpdir, validation_step, loss_values, patience, expected_stop_epoch):
+    """Test to ensure that early stopping is not triggered before patience is exhausted."""
+
+    class ModelOverrideTrainReturn(EvalModelTemplate):
+        train_return_values = torch.Tensor(loss_values)
+
+        def training_epoch_end(self, outputs):
+            loss = self.train_return_values[self.current_epoch]
+            self.log('train_loss', loss)
+
+    model = ModelOverrideTrainReturn()
+
+    if validation_step is None:
+        model.validation_step = None
+
+    early_stop_callback = EarlyStopping(monitor="train_loss", patience=patience, verbose=True)
+    trainer = Trainer(
+        default_root_dir=tmpdir,
+        callbacks=[early_stop_callback],
+        num_sanity_val_steps=0,
+        max_epochs=10,
+    )
+    trainer.fit(model)
+    assert trainer.current_epoch == expected_stop_epoch
+
+
 def test_pickling(tmpdir):
     early_stopping = EarlyStopping()
 

tests/checkpointing/test_checkpoint_callback_frequency.py

@@ -59,6 +59,7 @@ def test_default_checkpoint_freq(save_mock, tmpdir, epochs, val_check_interval,
         max_epochs=epochs,
         weights_summary=None,
         val_check_interval=val_check_interval,
+        progress_bar_refresh_rate=0,
     )
     trainer.fit(model)