README.rst

Stream Likelihoods with ML

This is the PyTorch implementation of the StreamMapper code, which can be used to model stellar streams. StreamMapper-PyTorch is a PyTorch framework for building Bayesian Mixture Density Networks, which can then be trained using the standard PyTorch tooling. Detailed explanations can be found in our paper (https://ui.adsabs.harvard.edu/abs/2023arXiv231116960S/abstract) and especially in the code repository for the paper (https://github.com/nstarman/stellar_stream_density_ml_paper).

As an illustruative example:

bkg_phi2_model = sml.builtin.Uniform(
    data_scaler=scaler,
    indep_coord_names=("phi1",),
    coord_names=("phi2",),
    coord_bounds={"phi2": (lower, upper)},
    params=ModelParameters(),
)

bkg_plx_model = sml.builtin.Exponential(
    net=sml.nn.sequential(
        data=1, hidden_features=32, layers=3, features=1, dropout=0.15
    ),
    data_scaler=scaler,
    indep_coord_names=("phi1",),
    coord_names=("parallax",),
    coord_bounds={"parallax": (lower, upper)},
    params=ModelParameters(
        {"parallax": {"slope": ModelParameter(bounds=SigmoidBounds(15.0, 25.0))}}
    ),
)


bkg_flow = sml.builtin.compat.ZukoFlowModel(
    net=zuko.flows.MAF(features=2, context=1, transforms=4, hidden_features=[4] * 4),
    jacobian_logdet=-xp.log(xp.prod(...)),
    data_scaler=scaler[("phi1", "g", "r")],
    coord_names=phot_names,
    coord_bounds=phot_bounds,
    params=ModelParameters(),
)

background_model = sml.IndependentModels(
    {
        "astrometric": sml.IndependentModels(
            {"phi2": bkg_phi2_model, "parallax": bkg_plx_model}
        ),
        "photometric": bkg_flow,
    }
)


stream_astrometric_model = sml.builtin.Normal(
    net=...,  # PyTorch NN
    data_scaler=scaler,
    coord_names=coord_astrometric_names,
    coord_bounds=coord_astrometric_bounds,
    params=ModelParameters(
        {
            "phi2": {
                "mu": ModelParameter(bounds=..., scaler=...),
                "ln-sigma": ModelParameter(bounds=..., scaler=...),
            },
            "parallax": {
                "mu": ModelParameter(bounds=..., scaler=...),
                "ln-sigma": ModelParameter(bounds=..., scaler=...),
            },
        }
    ),
)

stream_isochrone_model = sml.builtin.IsochroneMVNorm(...)

stream_model = sml.IndependentModels(
    {"astrometric": stream_astrometric_model, "photometric": stream_isochrone_model},
    unpack_params_hooks=(
        Parallax2DistMod(
            astrometric_coord="astrometric.parallax",
            photometric_coord="photometric.distmod",
        ),
    ),
)

model = sml.MixtureModel(
    {"stream": stream_model, "background": background_model},
    net=...,
    data_scaler=scaler,
    params=ModelParameters(
        {
            f"stream.ln-weight": ModelParameter(...),
            f"background.ln-weight": ModelParameter(...),
        }
    ),
)