Inference of LMFIT model using saved parameters

[dean-sh]

Hello!

We are facing some issue trying to use LMFIT for inference.
Our training pipeline uses ExpressionModel, fitting it, then saving the train params as well as the confidence in a database.
In our inference pipeline, we would like to initialize the same LMFIT model with the parameters found during training, similarly to loading from dumped model.
Something of this sort:

train_params = load_from_db()
model = ExpressionModel(...)
model.params = train_params 
model.eval(x = ...)

However, we cannot run eval without running fit first, or without dumping the trained model with dill.

This sounds to us like a common use case, is there any way how to do it?

[newville]

On Wed, Aug 31, 2022 at 9:36 AM Dean Shabi ***@***.***> wrote: Hello! We are facing some issue trying to use LMFIT for inference. Our training pipeline uses ExpressionModel, fitting it, then saving the train params as well as the confidence in a database. In our inference pipeline, we would like to initialize the same LMFIT model with the parameters found during training, similarly to loading from dumped model. Something of this sort: train_params = load_from_db() model = ExpressionModel(...) model.params = train_params model.eval(x = ...) However, we cannot run eval without running fit first, or without dumping the trained model with dill. This sounds to us like a common use case, is there any way how to do it?

pass your `params` to `Model.eval()`: import numpy as np from lmfit import Parameters from lmfit.models import ExpressionModel params = Parameters() params.add_many(('off', 1.0), ('slope', 0.220)) model = ExpressionModel('off + slope*x') x = np.linspace(0, 10, 21) y = model.eval(params, x=x) That is, a Model doesn't "have" Parameters, a Model represents a functional form. It can be evaluated with Parameters and the values for the expected independent variables. A ModelResult does have Parameters.

…

--Matt