Remove NaNs from loss in LRFinder (#1862)

* Remove NaNs from loss in LRFinder

* np.isfinite

* chlog

* add test

* chlog

Co-authored-by: Jirka <jirka@pytorchlightning.ai>

CHANGELOG.md

@@ -14,6 +14,8 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 - Allow user to select individual TPU core to train on ([#1729](https://github.com/PyTorchLightning/pytorch-lightning/pull/1729))
 
+- Removed non-finite values from loss in `LRFinder` ([#1862](https://github.com/PyTorchLightning/pytorch-lightning/pull/1862))
+
 ### Deprecated
 
 ### Removed

pytorch_lightning/trainer/lr_finder.py

@@ -321,8 +321,9 @@ class _LRFinder(object):
 
         """
         try:
-            loss = self.results["loss"][skip_begin:-skip_end]
+            loss = np.array(self.results["loss"][skip_begin:-skip_end])
-            min_grad = (np.gradient(np.array(loss))).argmin()
+            loss = loss[np.isfinite(loss)]
+            min_grad = np.gradient(loss).argmin()
             self._optimal_idx = min_grad + skip_begin
             return self.results["lr"][self._optimal_idx]
         except Exception:

tests/trainer/test_lr_finder.py

@@ -124,7 +124,7 @@ def test_call_to_trainer_method(tmpdir):
     # logger file to get meta
     trainer = Trainer(
         default_save_path=tmpdir,
-        max_epochs=5,
+        max_epochs=5
     )
 
     lrfinder = trainer.lr_find(model, mode='linear')
@@ -170,7 +170,7 @@ def test_suggestion_parameters_work(tmpdir):
     # logger file to get meta
     trainer = Trainer(
         default_save_path=tmpdir,
-        max_epochs=10,
+        max_epochs=10
     )
 
     lrfinder = trainer.lr_find(model)
@@ -179,3 +179,24 @@ def test_suggestion_parameters_work(tmpdir):
 
     assert lr1 != lr2, \
         'Skipping parameter did not influence learning rate'
+
+
+def test_suggestion_with_non_finite_values(tmpdir):
+    """ Test that non-finite values does not alter results """
+
+    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)
+    before_lr = lrfinder.suggestion()
+    lrfinder.results['loss'][-1] = float('nan')
+    after_lr = lrfinder.suggestion()
+
+    assert before_lr == after_lr, \
+        'Learning rate was altered because of non-finite loss values'