Resolve string to python object

[greaber]

This seems like a dumb question, but how do I resolve a string like torch.profiler.ProfilerActivity.CPU to the corresponding python object? Like instantiation, but I don't want the _target_ to be applied to anything (because it is not even callable). I thought about using a customer OmegaConf resolver for this, but I'm thinking probably it is already builtin to hydra, since instantiation has to do these lookups anyway, and I'm just not seeing it in the docs

[Jasha10]

Hi @greaber!
There are the super-useful hydra.utils.get_class and hydra.utils.get_method functions. They are undocumented currently, but that should change soon.

>>> import hydra
>>> assert hydra.utils.get_method("functools.lru_cache") is __import__("functools").lru_cache

The get_class function allows looking up any class (i.e. any instance of type), and throws a ValueError if the looked-up object is not a type. Meanwhile, get_method does the same thing, except it allows looking up any callable object. There is a feature request open to add a hydra.utils.get_object method that would do the same thing but without the restriction that the looked-up object must be a type/callable. Please give that issue a thumbs-up if such a feature would be useful to you.

I do sometimes find myself using an omegaconf resolver that calls Python's builtins.eval. Here's the pattern I've gravitated towards in my own projects:

>>> from omegaconf import OmegaConf
>>> OmegaConf.register_new_resolver("eval", eval)
>>> yaml_data = """foo: '${eval:"bytes(123)"}'"""
>>> assert OmegaConf.create(yaml_data).foo == bytes(123)

To register this resolver would be unsafe if there's a chance that your DictConfig object will contain untrusted 3rd party data, as the 3rd party could inject some things to be eval'd on your machine. (The same caveat applies by-and-large to calling Hydra's instantiate)

[Jasha10]

Here's another thing that comes to mind:
Starting with Hydra v1.2, the instantiate function allows use of a _partial_ keyword to wrap the given _target_ in a functools.partial instance.

from hydra.utils import instantiate
from functools import partial, lru_cache
from omegaconf import OmegaConf

yaml_data = """
obj:
    _target_: functools.lru_cache
    _partial_: true
"""
cfg = OmegaConf.create(yaml_data)
instance = instantiate(cfg.obj)
assert isinstance(instance, partial)
assert instance.func is lru_cache

Incidentally, this approach does not work with torch.profiler.ProfilerActivity.CPU from your question above, as instantiate requires that the _target_ is callable (and torch.profiler.ProfilerActivity.CPU is not callable).

[Jasha10]

Since, none of the options above allow instantiating torch.profiler.ProfilerActivity.CPU, here are two ways to do so:

Using the private hydra.utils._locate method:

import torch
from hydra.utils import instantiate
from omegaconf import OmegaConf

cfg = OmegaConf.create(
    """
obj:
    _target_: hydra.utils._locate
    path: torch.profiler.ProfilerActivity.CPU
"""
)
instance = instantiate(cfg.obj)
assert instance is torch.profiler.ProfilerActivity.CPU

This is probably the cleanest pattern for achieving what you originally asked about, though it makes use of a private function (_locate).
In the future, the propsed get_object function (mentioned above) should cover this use case.

Instantiating torch.profiler.ProfilerActivity

import torch
from hydra.utils import instantiate
from omegaconf import OmegaConf

cfg = OmegaConf.create(
    """
obj:
    _target_: torch.profiler.ProfilerActivity
    _args_:
      - 0
"""
)
instance = instantiate(cfg.obj)
assert instance == torch.profiler.ProfilerActivity.CPU

[greaber]

Thanks!