Bring your own

Using your own trained model inside a pipeline.

Wrap it in a card

Signal Studio does not train models. It runs them. Put your weights file alongside a card that loads it and exposes whatever you want adjustable as parameters.

import torch
from pathlib import Path

_model = None

def run(ctx):
    global _model
    if _model is None:
        # Loaded once: the execution cache calls run() again
        # on every upstream parameter change.
        _model = torch.load(Path(__file__).parent / "weights.pt")
        _model.eval()

    x = ctx.inputs["in"].get_data()
    with torch.no_grad():
        y = _model(torch.from_numpy(x).float()).numpy()
    return { "out": ctx.table(prediction=y) }

Weights are large

The marketplace accepts files up to 200 MB. Above that, host the weights elsewhere and have the card download them on first use, stating clearly in the description that it will.

Declare what it expects

Sampling rate, channel count, montage, preprocessing. Then check them in run() and raise a clear error when they do not match. A model that accepts anything is a model that will be misused.