Cannot instantiate augmentations with tuple parameters from Hydra config

🐛 Bug

I’d like to use Hydra to configure and instantiate augmentations. It works nicely, unless I try to use a tuple as a parameter value. Hydra translates lists and tuples to an internal type that can be casted to a tuple, but Albumentations checks the parameter types and won’t try to cast it if it’s not a list.

To Reproduce

It’s convenient to define augmentations in a Hydra configuration like this:

augmentations:
- _target_: albumentations.HorizontalFlip
  p: 0.5
- _target_: albumentations.HueSaturationValue
  sat_shift_limit:
  - -10
  - 20

One can then instantiate the augmentations like this:

from omegaconf import DictConfig
import hydra
from hydra.utils import instantiate

@hydra.main(config_path='conf', config_name='config')
def my_app(cfg: DictConfig) -> None:
    augmentations = [instantiate(a) for a in cfg.augmentations]

if __name__ == "__main__":
    my_app()

If you run the above code, it will correctly instantiate the HorizontalFlip augmentation, but fails to instantiate the HueSaturationValue. The reason is that instantiate() translates the sat_shift_limit value [-10, 20] to an omegaconf.ListConfig. It would be possible to cast omegaconf.ListConfig to a tuple, but to_tuple() checks that the type is either a list or a tuple before casting. It’s not, so instead of casting it throws an exception:

ValueError: Argument param must be either scalar (int, float) or tuple

Expected behavior

to_tuple() should cast the omegaconf.ListConfig to a tuple, so that instantiate() would instantiate the HueSaturationValue transform.

Environment

• Albumentations version (e.g., 0.1.8): 0.5.2
• Python version (e.g., 3.7): 3.8.3
• OS (e.g., Linux): Linux
• How you installed albumentations (conda, pip, source): conda

Additional context

I understand that it’s impossible to check for omegaconf.ListConfig and all other possible cases that can be casted to a tuple. Actually it’s common to use “duck typing” in Python, meaning that anything that acts like a list can be provided in place of a list, so it’s problematic to check the exact type of parameters.

[BloodAxe]

It seems hydra (Tested on 1.1) instantiate has additional parameter _convert_:

   _convert_: Conversion strategy
        none    : Passed objects are DictConfig and ListConfig, default
        partial : Passed objects are converted to dict and list, with
                  the exception of Structured Configs (and their fields).
        all     : Passed objects are dicts, lists and primitives without
                  a trace of OmegaConf containers

So adding _convert_ should do the job:

augmentations:
- _target_: albumentations.HorizontalFlip
  p: 0.5
- _target_: albumentations.HueSaturationValue
  _convert_: all
  sat_shift_limit:
  - [10, 20]

[Dipet]

It looks like we can replace check isinstance(param, (list, tuple)) with isinstance(param, Sequence) and everything will work fine. Sequence - is a standard type from typing module