sample.optimize module

Automatic optimization of SAMPLE hyperparameters

class sample.optimize.SAMPLEOptimizer(model: ~sample.sample.SAMPLE = None, loss_fn: ~typing.Callable[[~numpy.ndarray, ~numpy.ndarray], float] = <function multiscale_spectral_loss>, loss_kw: ~typing.Optional[~typing.Dict[str, ~typing.Any]] = None, remap: ~typing.Optional[~typing.Callable[[...], ~typing.Dict[str, ~typing.Any]]] = <function sample_kwargs_remapper>, clip: bool = True, **kwargs)

Bases: object

Hyperparameter optimizer for a SAMPLE model, based on Gaussian Process minimization

Parameters:

  • sample_fn (callable) – Constructor for a SAMPLE model

  • sample_kw (dict) – Keyword arguments for sample_fn. These parameters will not be optimized. These parameters will potentially be remapped by remap

  • loss_fn (callable) – Loss function. It should take, as two positional arguments, the arrays of original and resynthesised audio samples

  • loss_kw (dict) – Keyword arguments for the loss function

  • remap (callable) – Function that accepts keyword arguments and returns a new dictionary of keyword arguments for sample_fn. Default is sample_kwargs_remapper()

  • clip (bool) – If True (default), then clip resynthesised audio to the same peak of the original audio

  • **kwargs – Parameters to optimize. See skopt.gp_minimize() dimensions for definition options

gp_minimize(x: ndarray, fs: float = 44100, state: Optional[OptimizeResult] = None, ignore_warnings: bool = True, fit_kws: Optional[Dict[str, Any]] = None, **kwargs) → Tuple[SAMPLE, OptimizeResult]

Use skopt.gp_minimize() to tune the hyperparameters

Parameters:

  • x (array) – Audio samples

  • fs (float) – Sample rate

  • ignore_warnings (bool) – If True (default), then ignore warnings while optimizing

  • fit_kws (dict) – Arguments for the self.model.fit() method

  • **kwargs – Keyword arguments for skopt.gp_minimize()

Returns:

Best model, and optimization summary

Return type:

SAMPLE, OptimizeResult

loss(x: ndarray, fs: float) → Callable[[...], float]

Define a loss function for the target audio based on computing loss_fn on the target and resynthesised audio

Parameters:

  • x (array) – Audio samples

  • fs (float) – Sample rate

Returns:

Loss function

Return type:

callable

property model

Default fget function

class sample.optimize.TqdmCallback(sample_opt: ~sample.optimize.SAMPLEOptimizer, n_calls: int, n_initial_points: int = 0, tqdm_fn: ~typing.Callable[[...], ~tqdm.std.tqdm] = <class 'tqdm.std.tqdm'>, minimum: bool = True)

Bases: object

Callback for using tqdm with SAMPLEOptimizer

Parameters:

  • sample_opt (SAMPLEOptimizer) – Optimizer instance

  • n_calls (int) – Number of total calls

  • n_initial_points (int) – Number of initial (random) points

  • tqdm_fn (callable) – Constructor for a tqdm object

  • minimum (bool) – If True (default), show current minimum in postfix

reset() → TqdmCallback

Reset the state of the callback, e.g. for using it again

start() → TqdmCallback

Start the callback. Calls to this method initialize internal objects

property started: bool

If True, the callback has already been started

sample.optimize.sample_kwargs_remapper(sinusoidal__log_n: Optional[int] = None, sinusoidal__wtype: str = 'hamming', sinusoidal__wsize: float = 1.0, sinusoidal__overlap: float = 0.5, **kwargs) → Dict[str, Any]

Default argument remapper for SAMPLEOptimizer. It remaps stft window paramaters and lets every other parameter pass through

Parameters:

  • sinusoidal__log_n (int) – Logarithm of fft size. Will be remapped to sinusoidal__n if not in kwargs

  • sinusoidal__wtype (str) – Name of the window to use. Default is "hamming". It is used to compute the window sinusoidal__w if not in kwargs

  • sinusoidal__wsize (float) – Window size as a fraction of fft size. Default is 1.0. It is used to compute the window sinusoidal__w if not in kwargs

  • sinusoidal__overlap (float) – Window overlap as a fraction of the window size. Default is 0.5. It is used to compute the hop size sinusoidal__tracker__h if not in kwargs

  • **kwargs – Pass-through keyword arguments

Returns:

Remapped keyword arguments

Return type:

dict