create read-only property set when checkpoints loaded via Trainer.{fit,validate,test,predict}, deprecate Trainer.{validated,tested,predicted}_ckpt_path

Author: speediedan

pytorch_lightning/trainer/trainer.py

@@ -480,10 +480,14 @@ def __init__(
         # default .predict() loop
         self.predict_loop = PredictionLoop()
 
-        # .validate() and .test() set this when they load a checkpoint
-        self.validated_ckpt_path: Optional[str] = None
-        self.tested_ckpt_path: Optional[str] = None
-        self.predicted_ckpt_path: Optional[str] = None
+        # set when a checkpoint is loaded via `Trainer.{fit,validate,test,predict}`.
+        self._ckpt_path: Optional[str] = None
+
+        # .validate(), predict() and .test() set these when they load a checkpoint. They will be removed in favor of
+        #  the unified read-only `Trainer.ckpt_path` attribute in v1.8
+        self._validated_ckpt_path: Optional[str] = None  # TODO: remove in v1.8
+        self._tested_ckpt_path: Optional[str] = None  # TODO: remove in v1.8
+        self._predicted_ckpt_path: Optional[str] = None  # TODO: remove in v1.8
 
         # todo: remove in v1.7
         self._weights_summary: Optional[str] = None
@@ -758,7 +762,10 @@ def _fit_impl(
 
         # TODO: ckpt_path only in v2.0
         ckpt_path = ckpt_path or self.resume_from_checkpoint
-        results = self._run(model, ckpt_path=ckpt_path)
+        self._ckpt_path = self.__set_ckpt_path(
+            ckpt_path, model_provided=model, model_connected=self.lightning_module is not None
+        )
+        results = self._run(model, ckpt_path=self.ckpt_path)
 
         assert self.state.stopped
         self.training = False
@@ -837,12 +844,14 @@ def _validate_impl(
         # links data to the trainer
         self._data_connector.attach_data(model, val_dataloaders=dataloaders, datamodule=datamodule)
 
-        self.validated_ckpt_path = self.__set_ckpt_path(
+        self._ckpt_path = self.__set_ckpt_path(
             ckpt_path, model_provided=model_provided, model_connected=self.lightning_module is not None
         )
 
+        self._validated_ckpt_path = self.ckpt_path  # TODO: remove in v1.8
+
         # run validate
-        results = self._run(model, ckpt_path=self.validated_ckpt_path)
+        results = self._run(model, ckpt_path=self.ckpt_path)
 
         assert self.state.stopped
         self.validating = False
@@ -923,12 +932,14 @@ def _test_impl(
         # links data to the trainer
         self._data_connector.attach_data(model, test_dataloaders=dataloaders, datamodule=datamodule)
 
-        self.tested_ckpt_path = self.__set_ckpt_path(
+        self._ckpt_path = self.__set_ckpt_path(
             ckpt_path, model_provided=model_provided, model_connected=self.lightning_module is not None
         )
 
+        self._tested_ckpt_path = self.ckpt_path  # TODO: remove in v1.8
+
         # run test
-        results = self._run(model, ckpt_path=self.tested_ckpt_path)
+        results = self._run(model, ckpt_path=self.ckpt_path)
 
         assert self.state.stopped
         self.testing = False
@@ -1009,11 +1020,13 @@ def _predict_impl(
         # links data to the trainer
         self._data_connector.attach_data(model, predict_dataloaders=dataloaders, datamodule=datamodule)
 
-        self.predicted_ckpt_path = self.__set_ckpt_path(
+        self._ckpt_path = self.__set_ckpt_path(
             ckpt_path, model_provided=model_provided, model_connected=self.lightning_module is not None
         )
 
-        results = self._run(model, ckpt_path=self.predicted_ckpt_path)
+        self._predicted_ckpt_path = self.ckpt_path  # TODO: remove in v1.8
+
+        results = self._run(model, ckpt_path=self.ckpt_path)
 
         assert self.state.stopped
         self.predicting = False
@@ -2217,6 +2230,74 @@ def resume_from_checkpoint(self) -> Optional[Union[str, Path]]:
 
         return resume_from_checkpoint
 
+    @property
+    def ckpt_path(self) -> Optional[str]:
+        """Set to the path/URL of checkpoints loaded via :meth:`~pytorch_lightning.trainer.trainer.Trainer.fit`,
+        :meth:`~pytorch_lightning.trainer.trainer.Trainer.validate`,
+        :meth:`~pytorch_lightning.trainer.trainer.Trainer.test`, or
+        :meth:`~pytorch_lightning.trainer.trainer.Trainer.predict`. ``None`` otherwise."""
+        return self._ckpt_path
+
+    @property
+    def validated_ckpt_path(self) -> Optional[str]:
+        rank_zero_deprecation(
+            "The `Trainer.validated_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via"
+            " `Trainer.ckpt_path` instead.",
+            stacklevel=5,
+        )
+        return self._validated_ckpt_path
+
+    @validated_ckpt_path.setter
+    def validated_ckpt_path(self, ckpt_path: Optional[str]) -> None:
+        rank_zero_deprecation(
+            "The `Trainer.validated_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via the read-only"
+            " `Trainer.ckpt_path`.",
+            stacklevel=5,
+        )
+        self._validated_ckpt_path = ckpt_path
+
+    @property
+    def tested_ckpt_path(self) -> Optional[str]:
+        rank_zero_deprecation(
+            "The `Trainer.tested_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via"
+            " `Trainer.ckpt_path` instead.",
+            stacklevel=5,
+        )
+        return self._tested_ckpt_path
+
+    @tested_ckpt_path.setter
+    def tested_ckpt_path(self, ckpt_path: Optional[str]) -> None:
+        rank_zero_deprecation(
+            "The `Trainer.tested_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via the read-only"
+            " `Trainer.ckpt_path` instead.",
+            stacklevel=5,
+        )
+        self._tested_ckpt_path = ckpt_path
+
+    @property
+    def predicted_ckpt_path(self) -> Optional[str]:
+        rank_zero_deprecation(
+            "The `Trainer.predicted_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via"
+            " `Trainer.ckpt_path` instead.",
+            stacklevel=5,
+        )
+        return self._predicted_ckpt_path
+
+    @predicted_ckpt_path.setter
+    def predicted_ckpt_path(self, ckpt_path: Optional[str]) -> None:
+        rank_zero_deprecation(
+            "The `Trainer.predicted_ckpt_path` attribute was deprecated in v1.6 and will be removed in v1.8. The"
+            " path of checkpoints loaded via `Trainer.{fit,validate,test,predict}` should be accessed via the read-only"
+            " `Trainer.ckpt_path` instead.",
+            stacklevel=5,
+        )
+        self._predicted_ckpt_path = ckpt_path
+
     def save_checkpoint(self, filepath: _PATH, weights_only: bool = False) -> None:
         r"""
         Runs routine to create a checkpoint.

tests/deprecated_api/test_remove_1-8.py

@@ -146,6 +146,16 @@ def test_v1_8_0_trainer_verbose_evaluate():
         trainer.verbose_evaluate = False
 
 
+@pytest.mark.parametrize("fn_prefix", ["validated", "tested", "predicted"], ids=["validated", "tested", "predicted"])
+def test_v1_8_0_trainer_ckpt_path_attributes(fn_prefix: str):
+    test_attr = f"{fn_prefix}_ckpt_path"
+    trainer = Trainer()
+    with pytest.deprecated_call(match=f"{test_attr}` attribute was deprecated in v1.6 and will be removed in v1.8"):
+        _ = getattr(trainer, test_attr)
+    with pytest.deprecated_call(match=f"{test_attr}` attribute was deprecated in v1.6 and will be removed in v1.8"):
+        setattr(trainer, test_attr, "v")
+
+
 def test_v1_8_0_deprecated_trainer_should_rank_save_checkpoint(tmpdir):
     trainer = Trainer()
     with pytest.deprecated_call(

tests/trainer/test_trainer.py

@@ -686,8 +686,7 @@ def predict_step(self, batch, *_):
     trainer.fit(model)
 
     trainer_fn = getattr(trainer, fn)
-    path_attr = f"{fn}{'d' if fn == 'validate' else 'ed'}_ckpt_path"
-    assert getattr(trainer, path_attr) is None
+    assert getattr(trainer, "ckpt_path") is None
 
     if ckpt_path == "best":
         # ckpt_path is 'best', meaning we load the best weights
@@ -698,20 +697,20 @@ def predict_step(self, batch, *_):
                 trainer_fn(model, ckpt_path=ckpt_path)
         else:
             trainer_fn(ckpt_path=ckpt_path)
-            assert getattr(trainer, path_attr) == trainer.checkpoint_callback.best_model_path
+            assert getattr(trainer, "ckpt_path") == trainer.checkpoint_callback.best_model_path
 
             trainer_fn(model, ckpt_path=ckpt_path)
-            assert getattr(trainer, path_attr) == trainer.checkpoint_callback.best_model_path
+            assert getattr(trainer, "ckpt_path") == trainer.checkpoint_callback.best_model_path
     elif ckpt_path is None:
         # ckpt_path is None, meaning we don't load any checkpoints and use the provided model
         trainer_fn(model, ckpt_path=ckpt_path)
-        assert getattr(trainer, path_attr) is None
+        assert getattr(trainer, "ckpt_path") is None
 
         if save_top_k > 0:
             # ckpt_path is None with no model provided means load the best weights
             with pytest.warns(UserWarning, match="The best model of the previous `fit` call will be used"):
                 trainer_fn(ckpt_path=ckpt_path)
-                assert getattr(trainer, path_attr) == trainer.checkpoint_callback.best_model_path
+                assert getattr(trainer, "ckpt_path") == trainer.checkpoint_callback.best_model_path
     else:
         # specific checkpoint, pick one from saved ones
         if save_top_k == 0:
@@ -724,10 +723,10 @@ def predict_step(self, batch, *_):
                 ].absolute()
             )
             trainer_fn(ckpt_path=ckpt_path)
-            assert getattr(trainer, path_attr) == ckpt_path
+            assert getattr(trainer, "ckpt_path") == ckpt_path
 
             trainer_fn(model, ckpt_path=ckpt_path)
-            assert getattr(trainer, path_attr) == ckpt_path
+            assert getattr(trainer, "ckpt_path") == ckpt_path
 
 
 @pytest.mark.parametrize("enable_checkpointing", (False, True))
@@ -758,15 +757,14 @@ def predict_step(self, batch, *_):
     trainer.fit(model)
 
     trainer_fn = getattr(trainer, fn)
-    path_attr = f"{fn}{'d' if fn == 'validate' else 'ed'}_ckpt_path"
-    assert getattr(trainer, path_attr) is None
+    assert getattr(trainer, "ckpt_path") is None
 
     if enable_checkpointing:
         trainer_fn(ckpt_path="best")
-        assert getattr(trainer, path_attr) == trainer.checkpoint_callback.best_model_path
+        assert getattr(trainer, "ckpt_path") == trainer.checkpoint_callback.best_model_path
 
         trainer_fn(model, ckpt_path="best")
-        assert getattr(trainer, path_attr) == trainer.checkpoint_callback.best_model_path
+        assert getattr(trainer, "ckpt_path") == trainer.checkpoint_callback.best_model_path
     else:
         with pytest.raises(MisconfigurationException, match="`ModelCheckpoint` is not configured."):
             trainer_fn(ckpt_path="best")