Get descriptor without training

[braxtonowens]

Is there any way to retrieve the descriptor without also training?

Ideally, with some input, like an ASE object or LAMMPS dump I could just retrieve the descriptor?

Essentially, something like the Dscribe python library? https://singroup.github.io/dscribe/latest/

[njzjz]

DeePMD-kit's descriptors are trainable, unlike others.

[braxtonowens]

I don't quite understand your response. What do you mean by trainable?

I just want the described representation of my data. Is there a way I can do this?

[njzjz]

The parameters of the descriptor are trainable. The initial parameters are just random values. So there is no way to retrieve the descriptor without training. You need to train a model firstly, and then infer the descriptor.