add_your_own.md

Add your own objective

Three steps. Drop a file in models/, give it a name, list that name on --models.

If a method returns a number from a sequence or a structure, it can be an objective. Nothing needs to be differentiable.

---

1. Create models/my_objective.py

from typing import Dict
from core.interfaces import BaseModel
from core.registry import register_model
from evolve.individual import Individual

@register_model("my_objective")
class MyObjective(BaseModel):
    def __init__(self):
        pass

    def setup(self, config: Dict, device: str = "cpu") -> None:
        self.config = config
        self.cfg = config.models.my_objective

    def score(self, individual: Individual):
        pdb = individual.get_name()
        value = 0.0  # your number, higher = better
        individual.add_fitness({"my_objective": float(value)})

models/__init__.py imports every file in that folder. The string in @register_model(...) is the name you type on the command line.

The individual gives you:

• get_name() — path to the current PDB
• get_gen() / get_index() — generation and index
• add_fitness({...}) — one dict of name → float
• update_name(path) — only if you wrote a new PDB

---

2. Add a yaml block

Same name, under models::

models:
  my_objective:
    any_param: 1.0

Read it in setup() as config.models.my_objective.

---

3. Run it

python run.py -c configs/your_run.yml \
  --models seq_model my_objective

seq_model stays first. Put relax second if you use it. You can add several scores on that line. Each key you pass to add_fitness() becomes its own axis on the Pareto front — emit one key unless you really want more objectives.

---

Copy from these

file	what it is
models/lys_pka.py	property score (PROPKA)
models/pocket_shape.py	geometry on a PDB
models/metal3d_model.py	loaded network weights

How those look in a campaign: README.md.