Is it possible to override parameters from a .yaml file via defaults list?

[YerePhy]

Hi,

Is it possible to override parameters from the defaults list in a .yaml file? I mean something like that:

defaults:
  - _self_
  - datamodule/monai_holdout_datamodule
  - lightningmodule/unetlike
  - override datamodule/monai_holdout_datamodule/batch_size: 2

I know its possible from the CLI with python train.py -m datamodule.batch_size=2, but I don't know if there is any option of doing this from the .yaml file itself.

[jieru-hu]

yes - since this is a config value you can override it in the yaml. note: defaults lists is used for config groups, not values.

datamodule:
  monai_holdout_datamodule:
    batch_size: 2

[YerePhy]

Thanks for your help but your proposal doesn't work for me, do you see something wrong here?

Folder structure:

train/
├── config.yaml
├── datamodule
│   ├── monai_holdout_datamodule.yaml
│   ├── patch_func
│   │   └── label.yaml
│   ├── splitter
│   │   └── hold_out.yaml
│   ├── test_stage_transformations
│   │   └── dummy.yaml
│   └── train_stage_transformations
│       └── dummy.yaml
└── lightningmodule
    ├── architecture
    │   └── unet.yaml
    ├── loss
    │   └── generalized_dice.yaml
    ├── metrics
    │   └── dice.yaml
    ├── optimizer_configuration
    │   └── adam.yaml
    ├── scheduler_configuration
    │   └── step_lr.yaml
    └── unet_like.yaml

My config.yaml:

defaults:
  - _self_
  - datamodule/monai_holdout_datamodule
  - lightningmodule/unet_like

datamodule_type: monai_hold_out
lightningmodule_type: unet_like
datamodule:
  monai_holdout_datamodule:
    batch_size: 2

My script:

import hydra
from omegaconf import DictConfig, OmegaConf, OmegaConf
from utils import datamodule_factory, lightningmodule_factory

@hydra.main(config_path="./configs/train", config_name="config")
def main(cfg: DictConfig):
    """
    Perform training.
    """
    OmegaConf.register_new_resolver("get_method", hydra.utils.get_method)
    cfg = hydra.utils.instantiate(cfg, _recursive_=True)
    
    print(cfg["datamodule"]["batch_size"])

    return None

main()

Output: 1.

[Jasha10]

What if you put _self_ at the end of your defaults list in config.yaml?

defaults:
  - datamodule/monai_holdout_datamodule
  - lightningmodule/unet_like
  - _self_

...
datamodule:
  monai_holdout_datamodule:
    batch_size: 2

The order of the defaults list controls the order in which the configs will be merged together.

[YerePhy]

With this folder structure:

train/
├── config.yaml
├── datamodule
│   ├── monai_holdout.yaml
│   ├── patch_func
│   │   └── label.yaml
│   ├── splitter
│   │   └── hold_out.yaml
│   ├── test_stage_transformations
│   │   └── dummy.yaml
│   └── train_stage_transformations
│       └── dummy.yaml
├── lightningmodule
│   ├── architecture
│   │   └── unet.yaml
│   ├── loss
│   │   └── generalized_dice.yaml
│   ├── metrics
│   │   └── dice.yaml
│   ├── optimizer_configuration
│   │   └── adam.yaml
│   ├── scheduler_configuration
│   │   └── step_lr.yaml
│   └── unet_like.yaml
└── trainer
    └── standard.yaml

The change that @Jasha10 proposed:

defaults:
  - datamodule: monai_holdout
  - lightningmodule: unet_like
  - trainer: standard
  - _self_
datamodule_type: monai_hold_out
lightningmodule_type: unet_like
datamodule:
  batch_size: 5

Overrides properly the batch_size in the main script.

Thanks @Jasha10 and @jieru-hu, could you explain a bit more about the _self_ item in the defaults lits? I don really understand why this is happening. My hypothesis, according to what you've pointed, is that _self_ refers to the configuration explicitly written in config.yaml, i.e.:

datamodule:
  monai_holdout_datamodule:
    batch_size: 2

in this case, am I right? So if you put it at the end of the defaults list will override all the previous values given by the groups, does it make sense?

[Jasha10]

My hypothesis, according to what you've pointed, is that _self_ refers to the configuration explicitly written in config.yaml.

Yes, that is exactly correct. You can read more about _self_ here and here.