LightningCLI instantiator receives values applied by instantiation links to set in hparams

requirements/pytorch/extra.txt

@@ -5,7 +5,7 @@
 matplotlib>3.1, <3.9.0
 omegaconf >=2.2.3, <2.4.0
 hydra-core >=1.2.0, <1.4.0
-jsonargparse[signatures] >=4.28.0, <=4.40.0
+jsonargparse[signatures] >=4.39.0, <4.40.0
 rich >=12.3.0, <13.6.0
 tensorboardX >=2.2, <2.7.0  # min version is set by torch.onnx missing attribute
 bitsandbytes >=0.45.2,<0.45.3; platform_system != "Darwin"

src/lightning/pytorch/CHANGELOG.md

@@ -24,7 +24,7 @@ The format is based on [Keep a Changelog](http://keepachangelog.com/en/1.0.0/).
 
 ### Fixed
 
--
+- Fixed `save_hyperparameters` not working correctly with `LightningCLI` when there are parsing links applied on instantiation ([#20777](https://github.com/Lightning-AI/pytorch-lightning/pull/20777))
 
 
 ---

src/lightning/pytorch/cli.py

@@ -320,6 +320,7 @@ def __init__(
         args: ArgsType = None,
         run: bool = True,
         auto_configure_optimizers: bool = True,
+        load_from_checkpoint_support: bool = True,
     ) -> None:
         """Receives as input pytorch-lightning classes (or callables which return pytorch-lightning classes), which are
         called / instantiated using a parsed configuration file and / or command line args.
@@ -360,6 +361,11 @@ def __init__(
                 ``dict`` or ``jsonargparse.Namespace``.
             run: Whether subcommands should be added to run a :class:`~lightning.pytorch.trainer.trainer.Trainer`
                 method. If set to ``False``, the trainer and model classes will be instantiated only.
+            auto_configure_optimizers: Whether to automatically add default optimizer and lr_scheduler arguments.
+            load_from_checkpoint_support: Whether ``save_hyperparameters`` should save the original parsed
+                hyperparameters (instead of what ``__init__`` receives), such that it is possible for
+                ``load_from_checkpoint`` to correctly instantiate classes even when using complex nesting and
+                dependency injection.
 
         """
         self.save_config_callback = save_config_callback
@@ -389,7 +395,8 @@ def __init__(
 
         self._set_seed()
 
-        self._add_instantiators()
+        if load_from_checkpoint_support:
+            self._add_instantiators()
         self.before_instantiate_classes()
         self.instantiate_classes()
         self.after_instantiate_classes()
@@ -537,11 +544,14 @@ def parse_arguments(self, parser: LightningArgumentParser, args: ArgsType) -> No
         else:
             self.config = parser.parse_args(args)
 
-    def _add_instantiators(self) -> None:
+    def _dump_config(self) -> None:
+        if hasattr(self, "config_dump"):
+            return
         self.config_dump = yaml.safe_load(self.parser.dump(self.config, skip_link_targets=False, skip_none=False))
         if "subcommand" in self.config:
             self.config_dump = self.config_dump[self.config.subcommand]
 
+    def _add_instantiators(self) -> None:
         self.parser.add_instantiator(
             _InstantiatorFn(cli=self, key="model"),
             _get_module_type(self._model_class),
@@ -792,12 +802,27 @@ def _get_module_type(value: Union[Callable, type]) -> type:
     return value
 
 
+def _set_dict_nested(data: dict, key: str, value: Any) -> None:
+    keys = key.split(".")
+    for k in keys[:-1]:
+        assert k in data, f"Expected key {key} to be in data"
+        data = data[k]
+    data[keys[-1]] = value
+
+
 class _InstantiatorFn:
     def __init__(self, cli: LightningCLI, key: str) -> None:
         self.cli = cli
         self.key = key
 
-    def __call__(self, class_type: type[ModuleType], *args: Any, **kwargs: Any) -> ModuleType:
+    def __call__(
+        self,
+        class_type: type[ModuleType],
+        *args: Any,
+        applied_instantiation_links: dict,
+        **kwargs: Any,
+    ) -> ModuleType:
+        self.cli._dump_config()
         hparams = self.cli.config_dump.get(self.key, {})
         if "class_path" in hparams:
             # To make hparams backwards compatible, and so that it is the same irrespective of subclass_mode, the
@@ -808,6 +833,15 @@ def __call__(self, class_type: type[ModuleType], *args: Any, **kwargs: Any) -> M
                 **hparams.get("init_args", {}),
                 **hparams.get("dict_kwargs", {}),
             }
+        # get instantiation link target values from kwargs
+        for key, value in applied_instantiation_links.items():
+            if not key.startswith(f"{self.key}."):
+                continue
+            key = key[len(f"{self.key}.") :]
+            if key.startswith("init_args."):
+                key = key[len("init_args.") :]
+            _set_dict_nested(hparams, key, value)
+
         with _given_hyperparameters_context(
             hparams=hparams,
             instantiator="lightning.pytorch.cli.instantiate_module",

tests/tests_pytorch/test_cli.py

@@ -550,6 +550,7 @@ def __init__(self, activation: torch.nn.Module = None, transform: Optional[list[
 class BoringModelRequiredClasses(BoringModel):
     def __init__(self, num_classes: int, batch_size: int = 8):
         super().__init__()
+        self.save_hyperparameters()
         self.num_classes = num_classes
         self.batch_size = batch_size
 
@@ -561,35 +562,103 @@ def __init__(self, batch_size: int = 8):
         self.num_classes = 5  # only available after instantiation
 
 
-def test_lightning_cli_link_arguments():
+def test_lightning_cli_link_arguments(cleandir):
     class MyLightningCLI(LightningCLI):
         def add_arguments_to_parser(self, parser):
             parser.link_arguments("data.batch_size", "model.batch_size")
             parser.link_arguments("data.num_classes", "model.num_classes", apply_on="instantiate")
 
-    cli_args = ["--data.batch_size=12"]
+    cli_args = ["--data.batch_size=12", "--trainer.max_epochs=1"]
 
     with mock.patch("sys.argv", ["any.py"] + cli_args):
         cli = MyLightningCLI(BoringModelRequiredClasses, BoringDataModuleBatchSizeAndClasses, run=False)
 
     assert cli.model.batch_size == 12
     assert cli.model.num_classes == 5
 
-    class MyLightningCLI(LightningCLI):
+    cli.trainer.fit(cli.model)
+    hparams_path = Path(cli.trainer.log_dir) / "hparams.yaml"
+    assert hparams_path.is_file()
+    hparams = yaml.safe_load(hparams_path.read_text())
+
+    hparams.pop("_instantiator")
+    assert hparams == {"batch_size": 12, "num_classes": 5}
+
+    class MyLightningCLI2(LightningCLI):
         def add_arguments_to_parser(self, parser):
             parser.link_arguments("data.batch_size", "model.init_args.batch_size")
             parser.link_arguments("data.num_classes", "model.init_args.num_classes", apply_on="instantiate")
 
-    cli_args[-1] = "--model=tests_pytorch.test_cli.BoringModelRequiredClasses"
+    cli_args[0] = "--model=tests_pytorch.test_cli.BoringModelRequiredClasses"
 
     with mock.patch("sys.argv", ["any.py"] + cli_args):
-        cli = MyLightningCLI(
+        cli = MyLightningCLI2(
             BoringModelRequiredClasses, BoringDataModuleBatchSizeAndClasses, subclass_mode_model=True, run=False
         )
 
     assert cli.model.batch_size == 8
     assert cli.model.num_classes == 5
 
+    cli.trainer.fit(cli.model)
+    hparams_path = Path(cli.trainer.log_dir) / "hparams.yaml"
+    assert hparams_path.is_file()
+    hparams = yaml.safe_load(hparams_path.read_text())
+
+    hparams.pop("_instantiator")
+    assert hparams == {"batch_size": 8, "num_classes": 5}
+
+
+class CustomAdam(torch.optim.Adam):
+    def __init__(self, params, num_classes: Optional[int] = None, **kwargs):
+        super().__init__(params, **kwargs)
+
+
+class DeepLinkTargetModel(BoringModel):
+    def __init__(
+        self,
+        optimizer: OptimizerCallable = torch.optim.Adam,
+    ):
+        super().__init__()
+        self.save_hyperparameters()
+        self.optimizer = optimizer
+
+    def configure_optimizers(self):
+        optimizer = self.optimizer(self.parameters())
+        return {"optimizer": optimizer}
+
+
+def test_lightning_cli_link_arguments_subcommands_nested_target(cleandir):
+    class MyLightningCLI(LightningCLI):
+        def add_arguments_to_parser(self, parser):
+            parser.link_arguments(
+                "data.num_classes",
+                "model.init_args.optimizer.init_args.num_classes",
+                apply_on="instantiate",
+            )
+
+    cli_args = [
+        "fit",
+        "--data.batch_size=12",
+        "--trainer.max_epochs=1",
+        "--model=tests_pytorch.test_cli.DeepLinkTargetModel",
+        "--model.optimizer=tests_pytorch.test_cli.CustomAdam",
+    ]
+
+    with mock.patch("sys.argv", ["any.py"] + cli_args):
+        cli = MyLightningCLI(
+            DeepLinkTargetModel,
+            BoringDataModuleBatchSizeAndClasses,
+            subclass_mode_model=True,
+            auto_configure_optimizers=False,
+        )
+
+    hparams_path = Path(cli.trainer.log_dir) / "hparams.yaml"
+    assert hparams_path.is_file()
+    hparams = yaml.safe_load(hparams_path.read_text())
+
+    assert hparams["optimizer"]["class_path"] == "tests_pytorch.test_cli.CustomAdam"
+    assert hparams["optimizer"]["init_args"]["num_classes"] == 5
+
 
 class EarlyExitTestModel(BoringModel):
     def on_fit_start(self):