Bugfix: accumulation and suggestion for learning rate finder (#1801)

* fix suggestion being too naive

* fix accumulation error and added new tests

* fix styling

* update CHANGELOG.md

* update based on review

* fix tests

* Apply suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

* Apply suggestions from code review

Co-authored-by: Nicki Skafte <[email protected]>
Co-authored-by: Jirka Borovec <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -70,8 +70,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Fixed saving native AMP scaler state (introduced in [#1561](https://github.com/PyTorchLightning/pytorch-lightning/pull/1561))
 
-- Fixed missing profiler attribute in add_argparse_args() ArgumentParser ([#1794](https://github.com/PyTorchLightning/pytorch-lightning/pull/1794))
-
+- Fixed accumulation parameter and suggestion method for learning rate finder ([#1801](https://github.com/PyTorchLightning/pytorch-lightning/pull/1801))
 
 ## [0.7.5] - 2020-04-27
 

pytorch_lightning/trainer/lr_finder.py

@@ -15,6 +15,7 @@
 from pytorch_lightning.callbacks import Callback
 from pytorch_lightning import _logger as log
 from pytorch_lightning.utilities.exceptions import MisconfigurationException
+from pytorch_lightning.utilities import rank_zero_warn
 
 
 class TrainerLRFinderMixin(ABC):
@@ -58,7 +59,8 @@ def lr_find(self,
                 max_lr: float = 1,
                 num_training: int = 100,
                 mode: str = 'exponential',
-                num_accumulation_steps: int = 1):
+                early_stop_threshold: float = 4.0,
+                num_accumulation_steps=None):
         r"""
         lr_find enables the user to do a range test of good initial learning rates,
         to reduce the amount of guesswork in picking a good starting learning rate.
@@ -81,7 +83,12 @@ def lr_find(self,
                 after each batch. If set to 'exponential', will increase learning
                 rate exponentially.
 
-            num_accumulation_steps: number of batches to calculate loss over.
+            early_stop_threshold: threshold for stopping the search. If the
+                loss at any point is larger than early_stop_threshold*best_loss
+                then the search is stopped. To disable, set to None.
+
+            num_accumulation_steps: deprepecated, number of batches to calculate loss over.
+                Set trainer argument ``accumulate_grad_batches`` instead.
 
         Example::
 
@@ -104,6 +111,12 @@ def lr_find(self,
             trainer.fit(model)
 
         """
+        if num_accumulation_steps is not None:
+            rank_zero_warn("Argument `num_accumulation_steps` has been deprepecated"
+                           " since v0.7.6 and will be removed in 0.9. Please"
+                           " set trainer argument `accumulate_grad_batches` instead.",
+                           DeprecationWarning)
+
         save_path = os.path.join(self.default_root_dir, 'lr_find_temp.ckpt')
 
         self.__lr_finder_dump_params(model)
@@ -115,7 +128,9 @@ def lr_find(self,
         lr_finder = _LRFinder(mode, min_lr, max_lr, num_training)
 
         # Use special lr logger callback
-        self.callbacks = [_LRCallback(num_training, progress_bar_refresh_rate=1)]
+        self.callbacks = [_LRCallback(num_training,
+                                      early_stop_threshold,
+                                      progress_bar_refresh_rate=1)]
 
         # No logging
         self.logger = None
@@ -127,9 +142,6 @@ def lr_find(self,
         if self.progress_bar_callback:
             self.progress_bar_callback.disable()
 
-        # Accumulation of gradients
-        self.accumulate_grad_batches = num_accumulation_steps
-
         # Disable standard checkpoint & early stopping
         self.checkpoint_callback = False
         self.early_stop_callback = None
@@ -149,7 +161,6 @@ def lr_find(self,
             raise MisconfigurationException(
                 f'`model.configure_optimizers()` returned {len(optimizers)}, but'
                 ' learning rate finder only works with single optimizer')
-        configure_optimizers = model.configure_optimizers
         model.configure_optimizers = lr_finder._get_new_optimizer(optimizers[0])
 
         # Fit, lr & loss logged in callback
@@ -164,6 +175,7 @@ def lr_find(self,
         # Transfer results from callback to lr finder object
         lr_finder.results.update({'lr': self.callbacks[0].lrs,
                                   'loss': self.callbacks[0].losses})
+        lr_finder._total_batch_idx = self.total_batch_idx  # for debug purpose
 
         # Reset model state
         self.restore(str(save_path), on_gpu=self.on_gpu)
@@ -184,7 +196,6 @@ def __lr_finder_dump_params(self, model):
             'logger': self.logger,
             'max_steps': self.max_steps,
             'progress_bar_refresh_rate': self.progress_bar_refresh_rate,
-            'accumulate_grad_batches': self.accumulate_grad_batches,
             'checkpoint_callback': self.checkpoint_callback,
             'early_stop_callback': self.early_stop_callback,
             'enable_early_stop': self.enable_early_stop,
@@ -198,7 +209,6 @@ def __lr_finder_restore_params(self, model):
         self.callbacks = self.__dumped_params['callbacks']
         self.max_steps = self.__dumped_params['max_steps']
         self.progress_bar_refresh_rate = self.__dumped_params['progress_bar_refresh_rate']
-        self.accumulate_grad_batches = self.__dumped_params['accumulate_grad_batches']
         self.checkpoint_callback = self.__dumped_params['checkpoint_callback']
         self.early_stop_callback = self.__dumped_params['early_stop_callback']
         self.enable_early_stop = self.__dumped_params['enable_early_stop']
@@ -242,6 +252,7 @@ def __init__(self, mode: str, lr_min: float, lr_max: float, num_training: int):
         self.num_training = num_training
 
         self.results = {}
+        self._total_batch_idx = 0  # for debug purpose
 
     def _get_new_optimizer(self, optimizer: torch.optim.Optimizer):
         """ Construct a new `configure_optimizers()` method, that has a optimizer
@@ -298,30 +309,49 @@ def plot(self, suggest: bool = False, show: bool = False):
 
         return fig
 
-    def suggestion(self):
+    def suggestion(self, skip_begin: int = 10, skip_end: int = 1):
         """ This will propose a suggestion for choice of initial learning rate
         as the point with the steepest negative gradient.
 
         Returns:
             lr: suggested initial learning rate to use
+            skip_begin: how many samples to skip in the beginning. Prevent too naive estimates
+            skip_end: how many samples to skip in the end. Prevent too optimistic estimates
 
         """
         try:
-            min_grad = (np.gradient(np.array(self.results["loss"]))).argmin()
-            self._optimal_idx = min_grad
-            return self.results["lr"][min_grad]
+            loss = self.results["loss"][skip_begin:-skip_end]
+            min_grad = (np.gradient(np.array(loss))).argmin()
+            self._optimal_idx = min_grad + skip_begin
+            return self.results["lr"][self._optimal_idx]
         except Exception:
-            log.warning('Failed to compute suggesting for `lr`.'
-                        ' There might not be enough points.')
+            log.exception('Failed to compute suggesting for `lr`. There might not be enough points.')
             self._optimal_idx = None
 
 
 class _LRCallback(Callback):
     """ Special callback used by the learning rate finder. This callbacks log
     the learning rate before each batch and log the corresponding loss after
-    each batch. """
-    def __init__(self, num_training: int, progress_bar_refresh_rate: bool = False, beta: float = 0.98):
+    each batch.
+
+    Args:
+        num_training: number of iterations done by the learning rate finder
+        early_stop_threshold: threshold for stopping the search. If the
+            loss at any point is larger than ``early_stop_threshold*best_loss``
+            then the search is stopped. To disable, set to ``None``.
+        progress_bar_refresh_rate: rate to refresh the progress bar for
+            the learning rate finder
+        beta: smoothing value, the loss being logged is a running average of
+            loss values logged until now. ``beta`` controls the forget rate i.e.
+            if ``beta=0`` all past information is ignored.
+
+    """
+    def __init__(self, num_training: int,
+                 early_stop_threshold: float = 4.0,
+                 progress_bar_refresh_rate: bool = False,
+                 beta: float = 0.98):
         self.num_training = num_training
+        self.early_stop_threshold = early_stop_threshold
         self.beta = beta
         self.losses = []
         self.lrs = []
@@ -332,13 +362,19 @@ def __init__(self, num_training: int, progress_bar_refresh_rate: bool = False, b
 
     def on_batch_start(self, trainer, pl_module):
         """ Called before each training batch, logs the lr that will be used """
+        if (trainer.batch_idx + 1) % trainer.accumulate_grad_batches != 0:
+            return
+
         if self.progress_bar_refresh_rate and self.progress_bar is None:
             self.progress_bar = tqdm(desc='Finding best initial lr', total=self.num_training)
 
         self.lrs.append(trainer.lr_schedulers[0]['scheduler'].lr[0])
 
     def on_batch_end(self, trainer, pl_module):
         """ Called when the training batch ends, logs the calculated loss """
+        if (trainer.batch_idx + 1) % trainer.accumulate_grad_batches != 0:
+            return
+
         if self.progress_bar:
             self.progress_bar.update()
 
@@ -350,10 +386,11 @@ def on_batch_end(self, trainer, pl_module):
         smoothed_loss = self.avg_loss / (1 - self.beta**current_step)
 
         # Check if we diverging
-        if current_step > 1 and smoothed_loss > 4 * self.best_loss:
-            trainer.max_steps = current_step  # stop signal
-            if self.progress_bar:
-                self.progress_bar.close()
+        if self.early_stop_threshold is not None:
+            if current_step > 1 and smoothed_loss > self.early_stop_threshold * self.best_loss:
+                trainer.max_steps = current_step  # stop signal
+                if self.progress_bar:
+                    self.progress_bar.close()
 
         # Save best loss for diverging checking
         if smoothed_loss < self.best_loss or current_step == 1:

tests/trainer/test_lr_finder.py

@@ -76,15 +76,15 @@ def test_trainer_reset_correctly(tmpdir):
 
 
 def test_trainer_arg_bool(tmpdir):
-
+    """ Test that setting trainer arg to bool works """
     hparams = EvalModelTemplate.get_default_hparams()
     model = EvalModelTemplate(hparams)
     before_lr = hparams.learning_rate
 
     # logger file to get meta
     trainer = Trainer(
         default_save_path=tmpdir,
-        max_epochs=1,
+        max_epochs=5,
         auto_lr_find=True
     )
 
@@ -95,7 +95,7 @@ def test_trainer_arg_bool(tmpdir):
 
 
 def test_trainer_arg_str(tmpdir):
-
+    """ Test that setting trainer arg to string works """
     hparams = EvalModelTemplate.get_default_hparams()
     hparams.__dict__['my_fancy_lr'] = 1.0  # update with non-standard field
     model = EvalModelTemplate(hparams)
@@ -104,7 +104,7 @@ def test_trainer_arg_str(tmpdir):
     # logger file to get meta
     trainer = Trainer(
         default_save_path=tmpdir,
-        max_epochs=1,
+        max_epochs=5,
         auto_lr_find='my_fancy_lr'
     )
 
@@ -115,6 +115,7 @@ def test_trainer_arg_str(tmpdir):
 
 
 def test_call_to_trainer_method(tmpdir):
+    """ Test that directly calling the trainer method works """
 
     hparams = EvalModelTemplate.get_default_hparams()
     model = EvalModelTemplate(hparams)
@@ -123,7 +124,7 @@ def test_call_to_trainer_method(tmpdir):
     # logger file to get meta
     trainer = Trainer(
         default_save_path=tmpdir,
-        max_epochs=1,
+        max_epochs=5,
     )
 
     lrfinder = trainer.lr_find(model, mode='linear')
@@ -133,3 +134,48 @@ def test_call_to_trainer_method(tmpdir):
 
     assert before_lr != after_lr, \
         'Learning rate was not altered after running learning rate finder'
+
+
+def test_accumulation_and_early_stopping(tmpdir):
+    """ Test that early stopping of learning rate finder works, and that
+        accumulation also works for this feature """
+
+    hparams = EvalModelTemplate.get_default_hparams()
+    model = EvalModelTemplate(hparams)
+
+    before_lr = hparams.learning_rate
+    # logger file to get meta
+    trainer = Trainer(
+        default_save_path=tmpdir,
+        accumulate_grad_batches=2
+    )
+
+    lrfinder = trainer.lr_find(model, early_stop_threshold=None)
+    after_lr = lrfinder.suggestion()
+
+    assert before_lr != after_lr, \
+        'Learning rate was not altered after running learning rate finder'
+    assert len(lrfinder.results['lr']) == 100, \
+        'Early stopping for learning rate finder did not work'
+    assert lrfinder._total_batch_idx == 100 * 2, \
+        'Accumulation parameter did not work'
+
+
+def test_suggestion_parameters_work(tmpdir):
+    """ Test that default skipping does not alter results in basic case """
+
+    hparams = EvalModelTemplate.get_default_hparams()
+    model = EvalModelTemplate(hparams)
+
+    # logger file to get meta
+    trainer = Trainer(
+        default_save_path=tmpdir,
+        max_epochs=10,
+    )
+
+    lrfinder = trainer.lr_find(model)
+    lr1 = lrfinder.suggestion(skip_begin=10)  # default
+    lr2 = lrfinder.suggestion(skip_begin=80)  # way too high, should have an impact
+
+    assert lr1 != lr2, \
+        'Skipping parameter did not influence learning rate'