sample.beatsdrop.regression module

Regression models for beats

class sample.beatsdrop.regression.BeatRegression(fs: float = 100, lpf: float = 50, linear_regressor=None, linear_regressor_k: str = 'coef_', linear_regressor_q: str = 'intercept_', params_init: Optional[Callable[[ndarray, ndarray, ndarray, BeatRegression], Tuple[float, float, float, float, float, float, float, float]]] = None, bounds: Optional[Callable[[ndarray, ndarray, ndarray, Tuple[float, float, float, float, float, float, float, float], BeatRegression], Tuple[Tuple[float, float, float, float, float, float, float, float], Tuple[float, float, float, float, float, float, float, float]]]] = None)

Bases: RegressorMixin, BaseEstimator

Regressor for fitting to a beat pattern

Parameters:

  • fs (float) – Sampling frequency for autocorrelogram computation

  • lpf (float) – Corner frequency for LPF of spectrum in autocorrelation computation

  • linear_regressor (sklearn.base.BaseEstimator) – Linear regression model instance. Must be sklearn-compatible

  • linear_regressor_k (str) – Attribute name for the estimated slope coefficient of the linear regression

  • linear_regressor_q (str) – Attribute name for the estimated intercept coefficient of the linear regression

  • params_init (callable) – Initializer for beat model parameters. Signature should be f(t, a, f, res_fn, model) -> a0, a1, f0, f1, d0, d1, p0, p1. It should return initial parameters for nonlinear least squares using input data t, a, and f, the residual function res_fn, and the BeatRegression instance model. If None, use default

  • bounds (callable) – Callable for computing beat model coefficient boundaries. Signature should be bounds(t, a, f, p, model) -> (bounds_min, bounds_max). It should return lower and upper boundaries for all eight parameters using input data t, a, and f, initial parameter estimates p, and the BeatRegression instance model. If None, use default

property bounds

Default fget function

fit(t: ndarray, a: ndarray, f: ndarray, tr_solver: Optional[str] = 'lsmr', method: str = 'dogbox', loss: str = 'cauchy', **kwargs)

Fit beat pattern

Parameters:

  • t (array) – Time. Must be a column vector (shape like (-1, 1))

  • a (array) – Amplitude at time t (in dB)

  • f (array) – Frequency at time t

  • **kwargs – Keyword arguments for scipy.optimize.least_squares()

Returns:

self

Return type:

BeatRegression

property k_: float

Linear regression slope

property params_init

Default fget function

predict(t: ndarray, *args) → List[ndarray]

Predict beat pattern

Parameters:

  • t (array) – Time

  • args – The names of the outputs to predict. If none is specified, then the audio signal ("x") is predicted. For a list of available output names, see BeatRegression.beat_.variables

Returns:

One array per output

Return type:

list of arrays

property q_: float

Linear regression intercept

class sample.beatsdrop.regression.DualBeatRegression(fs: float = 100, lpf: float = 50, linear_regressor=None, linear_regressor_k: str = 'coef_', linear_regressor_q: str = 'intercept_', params_init: Optional[Callable[[ndarray, ndarray, ndarray, BeatRegression], Tuple[float, float, float, float, float, float, float, float]]] = None, bounds: Optional[Callable[[ndarray, ndarray, ndarray, Tuple[float, float, float, float, float, float, float, float], BeatRegression], Tuple[Tuple[float, float, float, float, float, float, float, float], Tuple[float, float, float, float, float, float, float, float]]]] = None, freq_loss: Optional[Callable[[ndarray, ndarray], ndarray]] = None, amp_loss: Optional[Callable[[ndarray, ndarray], ndarray]] = None, freq_w: float = 0.0002631578947368421, freq_clip: Optional[float] = 2.5, disambiguate: bool = True)

Bases: BeatRegression

Regressor for fitting to a beat pattern, using information from both the amplitude modulation and the frequency modulation

Parameters:

  • fs (float) – Sampling frequency for autocorrelogram computation

  • lpf (float) – Corner frequency for LPF of spectrum in autocorrelation computation

  • linear_regressor (sklearn.base.BaseEstimator) – Linear regression model instance. Must be sklearn-compatible

  • linear_regressor_k (str) – Attribute name for the estimated slope coefficient of the linear regression

  • linear_regressor_q (str) – Attribute name for the estimated intercept coefficient of the linear regression

  • params_init (callable) – Initializer for beat model parameters. Signature should be f(t, a, f, res_fn, model) -> a0, a1, f0, f1, d0, d1, p0, p1. It should return initial parameters for nonlinear least squares using input data t, a, and f, the residual function res_fn, and the DualBeatRegression instance model. If None, use default

  • bounds (callable) – Callable for computing beat model coefficient boundaries. Signature should be bounds(t, a, f, p, model) -> (bounds_min, bounds_max). It should return lower and upper boundaries for all eight parameters using input data t, a, and f, initial parameter estimates p, and the DualBeatRegression instance model. If None, use default

  • freq_loss (callable) – Loss function for frequencies. Signature should be freq_loss(f_true, f_est) -> f_loss. If None, use default

  • amp_loss (callable) – Loss function for amplitudes. Signature should be amp_loss(a_true, a_est) -> a_loss. If None, use default

  • freq_w (float) – Weighting coefficient for the frequency loss function

  • freq_clip (float) – Maximum estimatable frequency deviation (as a multiple of the iterquartile range)

  • disambiguate (bool) – If True (default), then disambiguate association between frequencies and amplitudes

property amp_loss

Default fget function

fit(t: ndarray, a: ndarray, f: ndarray, **kwargs)

Fit beat pattern

Parameters:

  • t (array) – Time. Must be a column vector (shape like (-1, 1))

  • a (array) – Amplitude at time t (in dB)

  • f (array) – Frequency at time t

  • **kwargs – Keyword arguments for scipy.optimize.least_squares()

Returns:

self

Return type:

DualBeatRegression

property freq_loss

Default fget function

sample.beatsdrop.regression.sort_params(p: ~typing.Tuple[float, float, float, float, float, float, float, float], key: ~typing.Callable[[~typing.Tuple[float, float, float, float, float, float, float, float]], ~typing.Tuple[float, float]] = <function _param_energies>, descending: bool = False) → Tuple[float, float, float, float, float, float, float, float]

Sort beat model parameters

Parameters:

  • p (tuple) – Beat model parameters

  • key (callable) – Function with respect to which to sort. Defaults to modal energy

  • descending (bool) – If True, then sort in descending order

Returns:

Sorted beat model parameters (maintaining association)

Return type:

tuple