Have bytes as supported value type.

[pedropgusmao]

Is your feature request related to a problem? Please describe.
I need to instantiate an object with another one which is basically a dataclass instance like:

class Parameters:
    """Model parameters."""
    tensors: List[bytes] 
    tensor_type: str 

when trying to use this I get:
omegaconf.errors.ValidationError: Unsupported value type: <class 'bytes'>
Example:

from typing import List

from omegaconf import OmegaConf

@dataclass
class Parameters:
    """Model parameters."""
    tensors: List[bytes] 
    tensor_type: str 

# omegaconf.errors.ValidationError: Unsupported value type: <class 'bytes'>
OmegaConf.structured(Parameters)

Describe the solution you'd like
Accepting bytes.

Describe alternatives you've considered
Internally interpreting bytes as str (when parsing) then converting it to bytes.

Additional context
...

[Jasha10]

Thanks for suggesting this @pedropgusmao.
Supporting bytes seems reasonable to me since bytes is a built-in type and since yaml supports the bytes type binary data natively.

[pixelb]

Thanks for the suggestion and implementation!

I agree this has some merit, as binary is a yaml primitive: https://yaml.org/type/binary.html
Playing devil's advocate, some thoughts against implementing this:

• "binary" is just a base64 string, so why not do a little manual conversion in code to/from storing b64 in existing strings?
  • It would be good to have a concrete example of before/after code for this use case. BTW I vaguely remember a suggestion to support tensor as a primitive type, which had less merit, but some of the same considerations as this, as that's what's presented as the use case. Also slightly related to leaning on explicit conversion more is issue Value 'float64' is not a supported primitive type  #805 discussing numpy float handling.
• We have to be careful to avoid becoming a data serialization mechanism. We're tuned for smaller amounts of config, rather than larger amounts of data

If we do support bytes, I'm not sure about the support for converting to/from strings. That would have to be restricted to utf-8, so I don't see the use case for this conversion support

[Jasha10]

I've agree about the conversion support -- it's cleaner in my opinion if assigning str to a bytes field fails and assigning bytes to a str field fails.
cc @jieru-hu

[Jasha10]

* "binary" is just a base64 string, so why not do a little manual conversion in code to/from storing b64 in existing strings?
  
  * It would be good to have a concrete example of before/after code for this use case. 

@pedropgusmao, would you mind giving a motivating example?

[pedropgusmao]

Hello @Jasha10 and @pixelb ,
Thank you both for the comments.

Tbh, my use case is not really related to conversion the str <-> bytes , nor is it directly related to loading bytes from .yaml files.
Instead, I'm looking for simple/clean way to set all my regular parameters (str, float, etc...) and to instantiate different implementations of an given interface inside a single .yaml config file.

As you must definitely know, both things can be achieved using Hydra, where calling hydra.utils.instantiate(cfg.obj,named_arg=parameters) would solve for the instantiation part of the problem. However, even if parameters is created inside my python script (i.e. not inside the '.yaml' file), I get the error seen at the beginning of this Issue. Allowing bytes to be supported in OmegaConf would allow Hydra to accept bytes-containing data classes as input to its hydra.utils.instantiate.

Forgot to mention: This will also help generations to come... eventually.

[Jasha10]

Support for bytes is now implemented on OmegaConf's master branch and can be tested by installing the OmegaConf 2.2.0.dev1 development release from pypi.