Configurate using code

[NightWinkle]

Is there a way to configurate using python code input ?
I'm trying to configurate experiments by defining an input function f(x).
I've been thinking of the following solution :

1. Create a input.py file containing a class :

class Input:
    def __init__(self, param1):
        self.param1 = param1
    def f(self, x):
        return x*self.param1 / 2

2. Instantiate using yaml config :

input:
    _target_: input.Input
    param1: 0.3

3. Add config path to pythonpath

hcfg = HydraConfig.get()
for source in hcfg.runtime.config_sources:
    os.environ["PYTHONPATH"] = os.environ.get("PYTHONPATH", "") + ":" + source.path

However, this looks like a quite dirty solution. Is there any better way ?

[odelalleau]

What do you mean "configurate"? Is the problem that your input.py is not in PYTHONPATH? If yes AFAIK it's your responsibility to ensure that whatever you instantiate using the _target_ mechanism is in PYTHONPATH. In general Hydra has no way to know where those are located.

[NightWinkle]

I want to specify functions as input parameters. As an example usecase, if you are using Hydra to configurate PDEs solving, you might want to be able to specify boundary conditions (initial value function).

[odelalleau]

Ok -- it sounds like Hydra's instantiation mechanism is indeed what you should be using then. What is it exactly you would like to do differently? If what you want is to get rid of the .yaml, you can use structured configs to do everything in Python. You can also look into hydra-zen, which offers additional functionalities to automate the generation of structured configs from Python code.

[NightWinkle]

Sorry for the late response.
What I mean is I want to have "code" as parameter. For instance I would like one of the input of my program to be a function that the user can define (and I'm perfectly fine if it is using yaml).
For instance, my yaml would contain :

source:
    f: x**2 + 2*t

Which then allows me to configurate simulations whose input is a matrix f(x[:, None], t[None, :]). From my understanding of previous issues and discussions here, this is however not possible. This is why I was suggesting as an alternative that I might define a function f in a python file next to my yaml config, and then instantiate this python file using the yaml config.
However, this is actually not working because the python file is not in python path.
As I understand there is no dedicated solution for this kind of problem ?

[odelalleau]

Thanks for the clarification. One way to achieve this is through resolvers. This would become for instance

source:
  f: "${func: x**2 + 2*t}"

with

OmegaConf.register_new_resolver("func", lambda txt: eval(f"lambda x, t: {txt}"))

Just be aware of the security risks associated with the usage of eval() in Python.