[Bug] Cannot use plugin's conf dataclasses in structured config

[mayeroa]

🐛 Bug

Description

Hi,

First of all, many thanks for that wonderful tool you've made. It is, for sure, extremely helpful to have it when production environment is necessary. I have met an error/bug when I wanted to use the OptunaSweeperConf using the Python API (instead of using the traditional YAML files).

Checklist

• [x] I checked on the latest version of Hydra
• [x] I created a minimal repro (See this for tips).

To reproduce

Minimal Code/Config snippet to reproduce

# Standard libraries
from dataclasses import dataclass

# Third-party libraries
import hydra
from hydra.conf import HydraConf
from hydra.core.config_store import ConfigStore
from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf
from omegaconf import DictConfig

@dataclass
class ExampleConfig():
    hydra: HydraConf = HydraConf(sweeper=OptunaSweeperConf())

ConfigStore.instance().store(name='config', node=ExampleConfig)

@hydra.main(config_path=None, config_name='config')
def main(config: DictConfig) -> None:
    """Main function."""
    print(config)

if __name__ == '__main__':
    main()

And the stack trace with the returned error:

Stack trace/error message

In 'config': Validation error while composing config:
Merge error: OptunaSweeperConf is not a subclass of BasicSweeperConf. value: {'_target_': 'hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper', 'defaults': [{'sampler': 'tpe'}], 'sampler': '???', 'direction': <Direction.minimize: 1>, 'storage': None, 'study_name': None, 'n_trials': 20, 'n_jobs': 2, 'search_space': {}}
    full_key: 
    object_type=dict

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

Expected Behavior

I would like to override the hydra config group, directly into the structure config. To do so, I import the HydraConf from hydra and install the hydra-optuna-sweeper plugin. The error seems to indicate that the OptunaSweeperConf must inherit from the BasicSweeperConf structured config (current implementation: https://github.com/facebookresearch/hydra/blob/v1.1.1/plugins/hydra_optuna_sweeper/hydra_plugins/hydra_optuna_sweeper/config.py#L133). The idea is to be able to override hydra config group directly from a python script. Does that make sense?

System information

• Hydra Version : 1.1.1
• Hydra Optuna Sweeper Version : 1.1.1
• Python version : 3.8.11
• Virtual environment type and version : MiniConda environment
• Operating system : LinuxMint

Additional context

Add any other context about the problem here.

[jieru-hu]

hi @mayeroa pls try using the defaults list instead. something like the following should work

@dataclass
class ExampleConfig():
    defaults: List[Any] = field(
        default_factory=lambda: [
            {"override hydra/sweeper": "optuna"},
        ]
    )

Run the application, you should see optuna's config

$python my_app.py --cfg hydra -p hydra.sweeper
# @package hydra.sweeper
sampler:
  _target_: optuna.samplers.TPESampler
  seed: null
  consider_prior: true
  prior_weight: 1.0
  consider_magic_clip: true
  consider_endpoints: false
  n_startup_trials: 10
  n_ei_candidates: 24
  multivariate: false
  warn_independent_sampling: true
_target_: hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper
direction: minimize
storage: null
study_name: null
n_trials: 20
n_jobs: 2
search_space: {}

Also the cfg object you get via hydra.main does not contain Hydra's config, to access the configs, try something like the following

@hydra.main(config_path=None, config_name='config')
def main(config: DictConfig) -> None:
    """Main function."""
    print(HydraConfig.get().sweeper)

[mayeroa]

Thank you for your answer @jieru-hu, it helps. And if I want to override some parameters of the OptunaSweeper directly into the ExampleConfig, how should I proceed?

For example :

@dataclass
class ExampleConfig():
    defaults: List[Any] = field(
        default_factory=lambda: [
            {"override hydra/sweeper": "optuna"},
        ]
    )
# How to override n_trials?
# How to define search space?

An other question: would it be possible to extend Hydra in order to support the following scenario ?

# Standard libraries
from dataclasses import dataclass

# Third-party libraries
import hydra
from hydra.conf import HydraConf
from hydra.core.config_store import ConfigStore
from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf
from omegaconf import DictConfig

@dataclass
class ExampleConfig():
    hydra: HydraConf = HydraConf(sweeper=OptunaSweeperConf())

ConfigStore.instance().store(name='config', node=ExampleConfig)

@hydra.main(config_path=None, config_name='config')
def main(config: DictConfig) -> None:
    """Main function."""
    print(config)

if __name__ == '__main__':
    main()

While it does not work for configuring a sweeper, it effectively works, for instance, if a user wants to override the RunDir:

# Standard libraries
from dataclasses import dataclass

# Third-party libraries
import hydra
from hydra.conf import HydraConf, Rundir
from hydra.core.config_store import ConfigStore
from omegaconf import DictConfig

@dataclass
class ExampleConfig():
    hydra: HydraConf = HydraConf(run=RunDir('./test'))

ConfigStore.instance().store(name='config', node=ExampleConfig)

@hydra.main(config_path=None, config_name='config')
def main(config: DictConfig) -> None:
    """Main function."""
    print(config)

if __name__ == '__main__':
    main()

I think, it would be of great value to support such sweeper instantiation directly in python code. What do you think?

[jieru-hu]

hi @mayeroa

And if I want to override some parameters of the OptunaSweeper directly into the ExampleConfig, how should I proceed?

We do have in detailed documentation on how you'd configure a plugin, pls check here https://hydra.cc/docs/patterns/configuring_plugins

I think, it would be of great value to support such sweeper instantiation directly in python code. What do you think?

Essentially you want to override Hydra configuration here right? You can configure Hydra just like how you'd configure your application configuration. For example, you can write it in your config files, or override it from the commandline.

# in your primary config
hydra:
  run:
     dir:
          ./my_dir

or from the commandline

hydra.run.dir=./my_dir

Hydra takes care of the config composition for you so you do not need to manage them yourself. I'd suggest go through our basic tutorial and the configure Hydra section of the doc if you have not already https://hydra.cc/docs/next/tutorials/intro/ https://hydra.cc/docs/next/configure_hydra/intro

[mayeroa]

Hi @jieru-hu, Sorry for the explanation, I think I am not clear enough. I am perfectly aware of how to configure a sweeper using YAML files (this is well documented) or directly through the command line. My question was about configuring a Sweeper directly using the Python API (structured config)? But your answers suggest that it is not possible or not proposed by hydra.

To put it in some context, I'm using Hydra to run Deep Learning trainings using PyTorch-Lightning like that:

@dataclass
class ExampleConfig:

    datamodule = DataModuleConf(
        train_dataset=RandomDatasetConf(),
        val_dataset=RandomDatasetConf(),
        test_dataset=RandomDatasetConf(),
        num_workers=0,
        batch_size=8
    )

    module = ModuleConf(
        model=RandomModelConf(),
        loss=CrossEntropyLossConf(),
        optimizer=SGDConf(lr=0.1),
        scheduler=StepLRConf(step_size=1),
    )

    trainer = TrainerConf(
        callbacks=[LearningRateMonitorConf()],
        max_epochs=20
    )

ConfigStore.instance().store(name='config', node=ExampleConfig)

@hydra.main(config_path=None, config_name='config')
def main(config) -> None:
    """Main function."""
    # Instantiate datamodule
    datamodule = instantiate(config=config.datamodule)

    # Instantiate module
    module = instantiate(config=config.module)

    # Instantiate trainer
    trainer = instantiate(config=config.trainer)

    # Run training
    trainer.fit(module, datamodule=datamodule)

if __name__ == '__main__':
    main()

I really like to use the Python API because it offers the benefits of using an IDE (like VSCode) to easily manipulate structured configurations with auto-completion. And I would like to be happy to override HydraConf directly in that file (having a single file to centralize all the config used for training). I managed to override the run dir like so, with adding the hydra config group:

    hydra: HydraConf = HydraConf(run=RunDir('./test'))

But if I want to do the same thing with a sweeper (like Optuna), it fails (see the error at the beginning of the thread). I understand that Hydra relies essentially on manipulating and composing YAML files but I also see an interest of manipulating the structured config (the schemas) directly through a Python script. Maybe I am wrong 😃 .

PS: there was an attempt, in https://github.com/pytorch/hydra-torch/blob/main/examples/mnist_00.py to use directly structured configs through a script. And I think we can go a step further with the example above. What do you think?

[Jasha10]

# How to override n_trials?
# How to define search space?

I have tried the following, which yields a surprising error:

from dataclasses import dataclass, field

import hydra
from hydra.conf import HydraConf
from hydra.core.config_store import ConfigStore
from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf
from omegaconf import DictConfig
from typing import List
from typing import Any

@dataclass
class ExampleConfig:
    defaults: List[Any] = field(
        default_factory=lambda: [
            {"override hydra/sweeper": "optuna"},
        ]
    )
    hydra: HydraConf = HydraConf(sweeper=OptunaSweeperConf(n_trials=100))

ConfigStore.instance().store(name="config", node=ExampleConfig)

@hydra.main(config_path=None, config_name="config")
def main(config: DictConfig) -> None:
    """Main function."""
    print(config)

if __name__ == "__main__":
    main()

Here's the output:

$ python my_app.py
In 'config': ValidationError raised while composing config:
Merge error: OptunaSweeperConf is not a subclass of OptunaSweeperConf. value: {'_target_': 'hydra_plugins.hydra_optuna_sweeper.optuna_sweeper.OptunaSweeper', 'defaults': [{'sampler': 'tpe'}], 'sampler': '???', 'direction': <Direction.minimize: 1>, 'storage': None, 'study_name': None, 'n_trials': 100, 'n_jobs': 2, 'search_space': {}}
    full_key:
    object_type=dict

Set the environment variable HYDRA_FULL_ERROR=1 for a complete stack trace.

Note in particular that the error message says "OptunaSweeperConf is not a subclass of OptunaSweeperConf." This smells fishy to me... I think there might be a bug here.

[omry]

This is probably related to how plugins are loaded discovered and loaded. There are two different OptunaSweeperConf classes loaded.

[mayeroa]

This is probably related to how plugins are loaded discovered and loaded. There are two different OptunaSweeperConf classes loaded.

Do you think I can have a look on it or does it seem like a complex fix to implement? If first answer, on which part of the source code should I focus?

[jieru-hu]

sorry @mayeroa for getting back to you late, thanks for further explaining your use case.

like @omry suspected earlier, it does seem to be related to how the two classes are imported, in particular, the application fails at merging the configs here, although the two Conf classes have the same type, the issubclass return False.

I tried to move the ConfigStore call to the __init__.py like the following

a/plugins/hydra_optuna_sweeper/hydra_plugins/hydra_optuna_sweeper/__init__.py
+++ b/plugins/hydra_optuna_sweeper/hydra_plugins/hydra_optuna_sweeper/__init__.py
@@ -1,3 +1,14 @@
 # Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved

 __version__ = "1.2.0dev1"
+
+from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf
+
+from hydra.core.config_store import ConfigStore
+
+ConfigStore.instance().store(
+    group="hydra/sweeper",
+    name="optuna",
+    node=OptunaSweeperConf,
+    provider="optuna_sweeper",
+)

which seems to resolve the issue, and the config override worked running @Jasha10's example above

$ python /Users/jieru/workspace/hydra-fork/hydra/examples/tutorials/structured_configs/1_minimal/my_app.py --cfg hydra -p hydra.sweeper.n_trials
100

I'm not sure if this is the best way to fix this though, @omry thoughts?

[omry]

Worth trying to figure out the root cause.

[mayeroa]

Hello :) Do you have any update on that topic? Or need some help to further investigate on that?

[jieru-hu]

hi @mayeroa thanks for the ping, we will get to this for the next release. in the meanwhile, feel free to further investigate and maybe even suggest a fix :)

[jieru-hu]

Ok - I took another look at this and here is what is happening: the tl;dr is the plugins are being imported twice and as a result OmegaConf fails the validation here

1. OmegaConf fails here - complaining OptunaSweeperConf is not a subclass of OptunaSweeperConf. In debug mode, comparing the dest_obj_type and src_obj_type, i can see that although they are the same class, they have different ids (meaning they are different types due to the double imports)
2. This seems to be only happening to plugins, since they are being imported twice: 1) first time in the user application 2) second time here
3. This is probably not an optuna only issue but affects all plugins.

I also found a so question that describes a similar issue here

[Jasha10]

Nice detective work!!

Here is a minimal reproducer for the issue based on the comment above:

# min_repo1.py
import importlib
import pkgutil

from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf

mdl = importlib.import_module("hydra_plugins")
for importer, modname, ispkg in pkgutil.walk_packages(
    path=mdl.__path__, prefix=mdl.__name__ + ".", onerror=lambda x: None
):
    m = importer.find_module(modname)
    loaded_mod = m.load_module(modname)
    if loaded_mod.__name__ == "hydra_plugins.hydra_optuna_sweeper.config":
        assert loaded_mod.OptunaSweeperConf is OptunaSweeperConf, "comparison failed"
        print("assertion passed!")

$ python min_repo1.py
Traceback (most recent call last):
  File "min_repo1.py", line 14, in <module>
    assert loaded_mod.OptunaSweeperConf is OptunaSweeperConf, "comparison failed"
AssertionError: comparison failed

I have an idea for a diff to solve the issue.

$ diff min_repo1.py min_repo2.py
1c1
< # min_repo1.py
---
> # min_repo2.py
11,12c11
<     m = importer.find_module(modname)
<     loaded_mod = m.load_module(modname)
---
>     loaded_mod = importlib.import_module(modname)

$ python min_repo2.py
assertion passed!

By doing the import with importlib.import_module instead of pkgutil, we get the same module as is loaded by the import statement at the top of the file.

[mayeroa]

Awesome investigation! That'a a bit weird by the way because I thought pkgutil was using importlib under the hood. Obviously not. So changing the way we import module should be enough to fix the issue? Do you want me to send a PR or does it need further investigation (and another solution)?

[jieru-hu]

Thanks @Jasha10 . It's interesting to see all these tiny (and undocumented 🙃 ) subtleties between the standard libs.

~Another part of the root cause is we call ConfigStore.store twice: 1) first time when users import the plugin conf dataclasses. 2) when scanning all plugins we try to import module under the hydra_plugin namespace except for the ones starts with __ and _, which explains why in https://github.com/facebookresearch/hydra/issues/1830#issuecomment-933914988 when I put the ConfigStore code in __init__.py, everything works fine. (__init__.py is not imported again in Plugins)~

updating loaded_mod = importlib.import_module(modname) looks good to me.

[Jasha10]

__init__.py is not imported again in Plugins

Single underscores are ignored but double underscores are not ignored.

if module_name.startswith("_") and not module_name.startswith("__"):

(ref)

So __init__.py should still be imported. I am confused xD

[jieru-hu]

@Jasha10 you are right! 🤣 i updated the previous comment.

[Jasha10]

So changing the way we import module should be enough to fix the issue?

I think changing the way we import the module is necessary, and hopefully it will be sufficient too.

Do you want me to send a PR or does it need further investigation (and another solution)?

Thanks for the offer :) I've just opened PR #2019. @mayeroa, would you mind testing it out to see if that diff is sufficient to solve the problem that was coming up in your use-case?

[mayeroa]

Hi @Jasha10,

I have made some tests using your PR with the following example (simply implementing the sphere example but using the python API:

# Standard libraries
from dataclasses import dataclass, field
from typing import Any, List

# Third-party libraries
import hydra
from hydra.conf import HydraConf
from hydra.core.config_store import ConfigStore
from hydra_plugins.hydra_optuna_sweeper.config import OptunaSweeperConf, TPESamplerConfig
from omegaconf import DictConfig

@dataclass
class ExampleConfig:
    defaults: List[Any] = field(
        default_factory=lambda: [
            {"override hydra/sweeper": "optuna"},
        ]
    )

    hydra: HydraConf = HydraConf(
        sweeper=OptunaSweeperConf(
            n_trials=20,
            n_jobs=1,
            storage=None,
            study_name='sphere',
            direction='minimize',
            sampler=TPESamplerConfig(seed=123),
            search_space={
                'x': {'type': 'float', 'low': -5.5, 'high': 5.5, 'step': 0.5},
                'y': {'type': 'categorical', 'choices': [-5, 0, 5]},
            }
        )
    )

    x: float = 1.0
    y: float = 1.0
    z: float = 1.0

ConfigStore.instance().store(name="config", node=ExampleConfig)

@hydra.main(config_path=None, config_name="config")
def main(config: DictConfig) -> None:
    """Main function."""
    x: float = config.x
    y: float = config.y
    return x**2 + y**2

if __name__ == "__main__":
    main()

Running with following command python sphere.py --multirun, I get this output:

[I 2022-02-08 10:04:15,274] A new study created in memory with name: sphere
[2022-02-08 10:04:15,274][HYDRA] Study name: sphere
[2022-02-08 10:04:15,274][HYDRA] Storage: None
[2022-02-08 10:04:15,274][HYDRA] Sampler: TPESampler
[2022-02-08 10:04:15,274][HYDRA] Directions: ['minimize']
[2022-02-08 10:04:15,277][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,277][HYDRA]        #0 : x=2.5 y=5
[2022-02-08 10:04:15,446][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,446][HYDRA]        #1 : x=2.5 y=0
[2022-02-08 10:04:15,559][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,559][HYDRA]        #2 : x=0.0 y=5
[2022-02-08 10:04:15,671][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,671][HYDRA]        #3 : x=-0.5 y=5
[2022-02-08 10:04:15,782][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,782][HYDRA]        #4 : x=-3.5 y=5
[2022-02-08 10:04:15,894][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:15,894][HYDRA]        #5 : x=1.5 y=-5
[2022-02-08 10:04:16,008][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,008][HYDRA]        #6 : x=2.5 y=0
[2022-02-08 10:04:16,118][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,118][HYDRA]        #7 : x=-2.5 y=-5
[2022-02-08 10:04:16,230][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,230][HYDRA]        #8 : x=-1.0 y=-5
[2022-02-08 10:04:16,342][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,342][HYDRA]        #9 : x=-1.0 y=0
[2022-02-08 10:04:16,455][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,455][HYDRA]        #10 : x=5.5 y=0
[2022-02-08 10:04:16,568][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,568][HYDRA]        #11 : x=5.0 y=0
[2022-02-08 10:04:16,689][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,689][HYDRA]        #12 : x=-5.0 y=0
[2022-02-08 10:04:16,802][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,802][HYDRA]        #13 : x=4.0 y=0
[2022-02-08 10:04:16,918][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:16,918][HYDRA]        #14 : x=1.5 y=0
[2022-02-08 10:04:17,094][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:17,094][HYDRA]        #15 : x=-2.0 y=0
[2022-02-08 10:04:17,208][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:17,208][HYDRA]        #16 : x=1.5 y=0
[2022-02-08 10:04:17,322][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:17,322][HYDRA]        #17 : x=1.0 y=0
[2022-02-08 10:04:17,437][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:17,437][HYDRA]        #18 : x=-5.5 y=0
[2022-02-08 10:04:17,551][HYDRA] Launching 1 jobs locally
[2022-02-08 10:04:17,551][HYDRA]        #19 : x=-3.5 y=-5
[2022-02-08 10:04:17,662][HYDRA] Best parameters: {'x': -1.0, 'y': 0}
[2022-02-08 10:04:17,662][HYDRA] Best value: 1.0

Which seems to be fine for me. :)

[Jasha10]

Which seems to be fine for me. :)

Great!

[Jasha10]

We are deferring work on this due to unforeseen complexity (see comments on PR #2019).

[mayeroa]

Related to this new issue: #2697