How to load 'best' model/emulator after training

[jvwilliams23]

Hi,

I loosely followed the quickstart example where we have some training data and find the best emulator doing something like:

from autoemulate.core.compare import AutoEmulate
from autoemulate.emulators import GaussianProcess
ae = AutoEmulate(x_train, y_train, log_level="progress_bar", models=[GaussianProcess])
print(ae.summarise())
best = ae.best_result()
best_result_filepath = ae.save(best, "ae_best", use_timestamp=True)

I then want to use the best emulator in another downstream application, but it seems that I cannot do that without retraining the model. E.g., the compare() function is called in AutoEmulate.__init__(). How do I reuse the best model without having to retrain?

[radka-j]

Hi @jvwilliams23! You do not need to retrain the model. The trained model has been saved in the results object and can be accessed as best.model. There is no need to re-instantiate the AutoEmulate object. I hope this helps.

[jvwilliams23]

Hi @radka-j

Yes, but I mean how to use the model in a different script at inference time. I think I found the answer:

import torch
from autoemulate.core.save import ModelSerialiser
from autoemulate.core.logging_config import get_configured_logger

# load model, we need to do `get_configured_logger` otherwise the exceptions are not properly read and give a different but
model = ModelSerialiser(get_configured_logger("info")[0])._load_result("ae_best").model
# test it works
print(model.predict(torch.tensor([0,0,0,0,0,0])).mean)

I am not sure if this is the best/intended usage, since the function is prefixed with _ (_load_result) which usually means it is a private function.

If I give ModelSerialiser(None) and provide an incorrect path, I get the following:

File ~/miniconda3/envs/autoemulate/lib/python3.11/site-packages/autoemulate/core/save.py:103, in ModelSerialiser._load_model(self, path)
    102 model = joblib.load(path)
--> 103 self.logger.info("Model loaded from %s", path)
    104 return model

AttributeError: 'NoneType' object has no attribute 'info'

I think it would be more user-friendly to have filename:str = None and maybe load:bool = False as kwargs to AutoEmulate class, then if they are true, the appropriate model is read using ModelSerialiser._load_result instead of training a new model with AutoEmulate.compare(). I could raise a PR if that helps.

Best,
Josh

[radka-j]

You're totally right, the API needs updating. Thank you so much for raising this issue!

In the meantime, you can load the model using the below code:

import joblib
model = joblib.load(f"{best_result_filepath}.joblib")

Please let us know if you run into any more issues and we'll aim to get the API fixed ASAP.

[jvwilliams23]

No worries, thanks for the solution. I will convert this discussion to an issue so it is properly tracked.

[radka-j]

That would be amazing, thank you!