Supported models

PoreMind provides classifiers for event-feature tables and waveform-based deep-learning models.

Classical machine learning

build_best_model evaluates the built-in scikit-learn model set using event features and cross-validation. By default, it uses duration_s, blockade_ratio, segment_std, segment_skew, and segment_kurt.

best_pkg = analysis.build_best_model(cv=5, scoring="accuracy")

Waveform deep learning

build_DL_model supports 1D-CNN, residual-CNN, MAGJAM, and MAGJAM_d4. MAGJAM reads the event waveform directly and disables handcrafted feature concatenation.

# MAGJAM d8 (default)
magjam_d8 = analysis.build_DL_model(
    model_name="MAGJAM",
    interp_length=500,
    interp_method="interp",
    device="cuda",
    cv=5,
)

# MAGJAM d4
magjam_d4 = analysis.build_DL_model(
    model_name="MAGJAM_d4",
    interp_length=500,
    interp_method="interp",
    device="cuda",
    cv=5,
)

interp_length defaults to 500. interp_method="interp" linearly resamples each event waveform; interp_method="padding" pads short waveforms with zeros and center-crops long waveforms. MAGJAM's d8 architecture is the default; d4 uses the shallower variant.

Overview of the MAGJAM framework

For a custom DL extractor, pass an instantiated torch.nn.Module as model; a string naming a class is not a module instance. See the DL model API reference.

Save, reload, and predict

The model-package helpers serialize and load supported model dictionaries. See save_model_package and load_model_package. For predictions on new traces, use classify_new_samples.

Self-supervised waveform embeddings

The repository's current Python package does not export a callable PoreMindWaveformEncoder API. This book therefore does not provide executable embedding examples; use the implemented feature-based or waveform-classification workflows above.