I was writing a gridsearch experiment for interpolation/smoothing. I train on some regular interval of data points, and validate on some intermediate points, so here's my splitter:

class InterpSplitter(BaseCrossValidator):
    def __init__(self, thin_step: int = 2):
        if thin_step < 2:
            raise ValueError("thin_step must be >= 2")
        self.thin_step = thin_step

    def split(self, t_obs: np.ndarray, x: None=None, groups: None=None):
        n_obs = len(t_obs)
        train_set = tuple(range(n_obs))[:: self.thin_step]
        test_set = tuple(sorted(set(range(n_obs)) - set(train_set)))
        yield np.array(train_set), np.array(test_set)

    def get_n_splits(self, X=None, y=None, groups=None):
        return 1

If I have a list of trajectories of the same length, it should work to stack the data in shape (n_obs, n_trajectories). However, if the trajectories are different lengths, I can't stack them. The only solution I can think of is to pad with nans and stack, but that's a bad idea: it requires changes to the library's model.fit(), predict() in order to implement a convention that only matters in client code.

Would love any advice on how to write a splitter for ragged (list of list) data. Or more generally, how to write a splitter for custom containers.