Undo test changes

Author: carmocca

pytorch_lightning/loops/base.py

@@ -205,8 +205,7 @@ def run(self, *args, **kwargs):
                 self._restarting = False
             except StopIteration:
                 break
-        else:
-            self._restarting = False
+        self._restarting = False
 
         output = self.on_run_end()
         return output

tests/checkpointing/test_trainer_checkpoint.py

@@ -20,35 +20,62 @@
 
 
 def test_finetuning_with_ckpt_path(tmpdir):
-    model = BoringModel()
+    """This test validates that generated ModelCheckpoint is pointing to the right best_model_path during test."""
+
+    checkpoint_callback = ModelCheckpoint(monitor="val_loss", dirpath=tmpdir, filename="{epoch:02d}", save_top_k=-1)
+
+    class ExtendedBoringModel(BoringModel):
+        def configure_optimizers(self):
+            import torch
+
+            optimizer = torch.optim.SGD(self.layer.parameters(), lr=0.001)
+            lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1)
+            return [optimizer], [lr_scheduler]
+
+        def validation_step(self, batch, batch_idx):
+            output = self.layer(batch)
+            loss = self.loss(batch, output)
+            self.log("val_loss", loss, on_epoch=True, prog_bar=True)
+
+    model = ExtendedBoringModel()
+    model.validation_epoch_end = None
     trainer = Trainer(
         default_root_dir=tmpdir,
         max_epochs=1,
-        limit_train_batches=1,
-        callbacks=ModelCheckpoint(dirpath=tmpdir, filename="{epoch:02d}", save_top_k=-1),
+        limit_train_batches=12,
+        limit_val_batches=6,
+        limit_test_batches=12,
+        callbacks=[checkpoint_callback],
         logger=False,
     )
     trainer.fit(model)
     assert os.listdir(tmpdir) == ["epoch=00.ckpt"]
 
-    best_model_paths = [trainer.checkpoint_callback.best_model_path]
+    best_model_paths = [checkpoint_callback.best_model_path]
+    results = []
 
     for idx in range(3, 6):
         # load from checkpoint
         trainer = pl.Trainer(
             default_root_dir=tmpdir,
             max_epochs=idx,
-            limit_train_batches=1,
+            limit_train_batches=12,
+            limit_val_batches=12,
+            limit_test_batches=12,
             enable_progress_bar=False,
-            callbacks=ModelCheckpoint(dirpath=tmpdir, filename="{epoch:02d}", save_top_k=-1),
         )
         trainer.fit(model, ckpt_path=best_model_paths[-1])
-        trainer.test(model)
+        trainer.test()
+        from copy import deepcopy
+
+        results.append(deepcopy(trainer.callback_metrics))
         best_model_paths.append(trainer.checkpoint_callback.best_model_path)
 
-    assert len(best_model_paths) == 4
     for idx, best_model_path in enumerate(best_model_paths):
-        assert best_model_path.endswith(f"epoch=0{idx}.ckpt")
+        if idx == 0:
+            assert best_model_path.endswith(f"epoch=0{idx}.ckpt")
+        else:
+            assert f"epoch={idx + 1}" in best_model_path
 
 
 def test_accumulated_gradient_batches_with_ckpt_path(tmpdir):