fixed_sample_integrator module

Integrator that does not sample points during the training phase but uses a fixed dataset of points

class FixedSampleSurveyIntegrator(*args, **kwargs)[source]

Bases: zunis.integration.base_integrator.BaseIntegrator

Integrator that trains its model during the survey phase using a pre-computed sample provided externally

Parameters

  • f (callable) – ZuNIS-compatible function

  • trainer (BasicTrainer) – trainer object used to perform the survey

  • sample (tuple of torch.Tensor) – (x, fx, px): target-space point batch drawn from some PDF p, function value batch, PDF value batch p(x)

  • n_iter (int) – number of iterations (used for both survey and refine unless specified)

  • n_iter_survey (int) – number of iterations for survey

  • n_iter_refine (int) – number of iterations for refine

  • n_points (int) – number of points for both survey and refine unless specified

  • n_points_survey (int) – number of points for survey

  • n_points_refine (int) – number of points for refine

  • use_survey (bool) – whether to use the integral estimations from the survey phase. This makes error estimation formally incorrect since samples from the refine depend on the survey training, but these correlation can be negligible in some cases.

  • verbosity (int) – level of verbosity for the integrator-level logger

  • trainer_verbosity (int) – level of verbosity for the trainer-level logger

  • kwargs –

sample_survey(n_points=None, **kwargs)[source]

Sample points from the internally stored sample

Parameters

  • n_points (int, None) – size of the batch to select from the sample

  • kwargs –

Returns

(x,px,fx): sampled points, sampling distribution PDF values, function values

Return type

tuple of torch.Tensor

set_sample(sample)[source]

Assign a sample to be trained on

Parameters

sample (tuple of torch.Tensor) – (x,px,fx): sampled points, sampling distribution PDF values, function values

set_sample_csv(csv_path, device=None, delimiter=', ', dtype=<class 'float'>)[source]

Assign a sample to be trained on from a csv file The file must contain equal length rows with at least four columns, all numerical. All columns but the last two are interpreted as point coordinates, the next-to-last is the point PDF and the last is the function value.

Parameters

  • csv_path (str) – path to the csv file

  • device (torch.device) – device to which to send the sample

  • delimiter (str) – delimiter of the csv file

set_sample_pickle(pickle_path, device=None)[source]

Assign a sample to be trained on from a pickle file The pickle file must either contain a tuple (x,px,fx) of point batch, PDF value batch, function batch or a mapping with keys “x”, “px”, “fx”. In either case, these batches must be valid inputs for torch.tensor

Parameters

  • pickle_path (str) – path to the pickle file.

  • device (torch.device, None) – device on which to send the sample. If none is provided, flow parameter device will be used