feature: use load_model() function

[YuriCat]

load_model() is more natural than get_model().

Especially for ONNX models, the model can be loaded with just the model_path.
Therefore, it makes more sense to have the model_path as the first argument, as in build_agent(), but I'm not sure.

[ikki407]

IMHO, it's natural to pass the model first and then the weights (load_model(model, model_path)).
How about using a model as an argument?

E.g.

model = env.net()
model = load_model(model_path, model)
# or
model = load_model(env.net(), model_path)

[YuriCat]

I was worried that unnecessary torch would be imported when we loaded ONNX models.
However, as long as the neural net is defined in the environment file, torch is already imported when an environment instance is created.

[ikki407]

Ah, I got it!!

[ikki407]

So, if the definition of the neural network is separated from the environment module, load_model() can return an ONNX model without importing PyTorch when the extension is .onnx.

[YuriCat]

That's right. This torch lazy import will be written by whoever needs it, and I think your idea is cool.

[ikki407]

👍

[ikki407]

LGTM