Using defaults list in sub parameters

[MaugrimEP]

Hello, I have this example where I have this structure:

Params1
├── v: Params2
|       └── val: 123 
└── t : 
         ├── name: a
         └──  var: 1

And I can use the command line to swap between t=a and t=b using python TestHydra.py t=a or python TestHydra.py t=b

Now I would like to have this t as a field in the dataclass Params2 with a default value and be able to swap between configuration a or b for it.
How can I do that, just swapping the variable and the default list does not works for me?

from dataclasses import dataclass, field
from typing import Any, List

import hydra
from hydra.core.config_store import ConfigStore
from omegaconf import MISSING, OmegaConf

@dataclass
class a:
    name: str = 'a'
    var: int = 1


@dataclass
class b:
    name: str = 'b'
    var: int = 2


defaults = [
    '_self_',
    {'t': 'a'}
]
@dataclass
class Params2:
    val: int = 123



@dataclass
class Params1:
    v: Params2 = Params2()
    defaults: List[Any] = field(default_factory=lambda: defaults)
    t: Any = MISSING


cs = ConfigStore.instance()
cs.store(group='t', name='a', node=a)
cs.store(group='t', name='b', node=b)
cs.store(name='params1', node=Params1)

@hydra.main(version_base=None, config_name='params1')
def main(cfg: Params1) -> None:
    print(OmegaConf.to_yaml(cfg))


if __name__ == '__main__':
    main()

[Jasha10]

Hello, I'm not sure what you are trying to do.

Now I would like to have this t as a field in the dataclass Params2 with a default value and be able to swap between configuration a or b for it.

Can you please give an example output? What is the yaml output that you want to achieve?

[MaugrimEP]

Thanks for your answer.
Yes, this is pretty much what I would like to have, except for the t in Params1 which should not be here.

v:
  t:
    name: b
    var: 2
  val: 123

I would like to declare the defaults and t in the Params2 as something like:

@dataclass
class Params2:
    defaults: List[Any] = field(default_factory=lambda: defaults_for_params2)
    t: Any = MISSING
    val: int = 123

which output instead

v:
  defaults:
  - _self_
  - t: a
  t: ???
  val: 123

I read the documentation about Structured Configuration and their Default List without finding what I want.

[Jasha10]

Like this?

# TestHydra.py
from dataclasses import dataclass, field
from typing import Any, List
import hydra
from hydra.core.config_store import ConfigStore
from omegaconf import MISSING, OmegaConf

@dataclass
class a:
    name: str = "a"
    var: int = 1

@dataclass
class b:
    name: str = "b"
    var: int = 2

defaults = ["_self_", {"v/t": "a"}]

@dataclass
class Params2:
    t: Any = MISSING
    val: int = 123

@dataclass
class Params1:
    v: Params2 = Params2()
    defaults: List[Any] = field(default_factory=lambda: defaults)

cs = ConfigStore.instance()
cs.store(group="v/t", name="a", node=a)
cs.store(group="v/t", name="b", node=b)
cs.store(name="params1", node=Params1)

@hydra.main(version_base=None, config_name="params1")
def main(cfg: Params1) -> None:
    OmegaConf.resolve(cfg)
    print(OmegaConf.to_yaml(cfg))

if __name__ == "__main__":
    main()

$ python TestHydra.py
v:
  t:
    name: a
    var: 1
  val: 123

$ python TestHydra.py v/t=b
v:
  t:
    name: b
    var: 2
  val: 123

[Jasha10]

Another way to do it is like this:

defaults = ["_self_", {"[email protected]": "a"}]

...

cs = ConfigStore.instance()
cs.store(group="t", name="a", node=a)
cs.store(group="t", name="b", node=b)
cs.store(name="params1", node=Params1)

$ python TestHydra.py
v:
  t:
    name: a
    var: 1
  val: 123

$ python TestHydra.py [email protected]=b
v:
  t:
    name: b
    var: 2
  val: 123

[MaugrimEP]

Thanks you, your answers are really helpful.

I am trying to keep everything related to the default configuration of Params2 in Params2.
In practice, my Params1 and Params2 are in different files, and the default configuration of Params2 has nothing to do with Params1.

And with this method, the issue is that:

1. I would like the defaults settings relative to Params2 only, to stay in the class Params2 where they belong
2. do not use paths like v/t. In practice, I have several levels which means that if I change a parameter name, I have to change the associated 'path string' in the default configuration etc.

This is what I did not find in the documentation. But I will set up with your proposition if it's not possible.

[Jasha10]

Another idea is to use nested defaults lists. This way you can have separate defaults for Params2 vs for Params1:

Using absolute path in defaults list:

...

defaults_params2 = ["_self_", {"/t": "a"}]  # absolute path in defaults list for Params2

@dataclass
class Params2:
    t: Any = MISSING
    val: int = 123
    defaults: List[Any] = field(default_factory=lambda: defaults_params2)

cs = ConfigStore.instance()
cs.store(group="t", name="a", node=a)
cs.store(group="t", name="b", node=b)
cs.store(group="v", name="params2", node=Params2)

defaults_params1 = ["_self_", {"v": "params2"}]

@dataclass
class Params1:
    v: Params2 = Params2()
    defaults: List[Any] = field(default_factory=lambda: defaults_params1)

cs.store(name="params1", node=Params1)

...

$ python TestHydra.py [email protected]=b
v:
  t:
    name: b
    var: 2
  val: 123

Using relative path in defaults list:

...

defaults_params2 = ["_self_", {"t": "a"}]  # relative path in defaults list for Params2

@dataclass
class Params2:
    t: Any = MISSING
    val: int = 123
    defaults: List[Any] = field(default_factory=lambda: defaults_params2)

cs = ConfigStore.instance()
cs.store(group="v/t", name="a", node=a)
cs.store(group="v/t", name="b", node=b)
cs.store(group="v", name="params2", node=Params2)

defaults_params1 = ["_self_", {"v": "params2"}]

@dataclass
class Params1:
    v: Params2 = Params2()
    defaults: List[Any] = field(default_factory=lambda: defaults_params1)

cs.store(name="params1", node=Params1)

...

$ python TestHydra.py v/t=b
v:
  t:
    name: b
    var: 2
  val: 123

[omry]

Do read about the Defaults List in the docs.

[MaugrimEP]

I did, but I think I miss the part where they indicate how to do what I would like to do using data classes. I find the part for the YAML file but I did not manage to make it works for data class.