utils.sweep¶

Sweep hyperparameter or argument search spaces.

class utils.sweep.Sweep(search_space: dict, num_combinations: int = None)¶

Reads a search_space dictionary and generates appropriate parameter combinations.

Parameters:: search_space (dict) – Dictionaries of argument keys and their possible values (or scipy distributions and parameters), grouped by top-level keys specifying the sweep strategy to be used with each. Search spaces are typically stored in the sweep key of a Component configuration dictionary.

Search space parameters are expected to be found under one of four top-level keys:

1. Fixed: The argument takes the same value across all branches. E.g., if settings[“fixed”][“resample”]=100, resample will always take the value 100.

2. Zipped: The values of the arguments listed are iterated over such that the Nth value of any argument is matched with the Nth value of other arguments. Combinations are repeated as many times as it takes to reach the number of combinations dictated by construction or by specification. E.g., if settings[“zipped”][“freq_band”] = [[5, 10], [30, 50]] and … [“resample”] = [50, 100], then self.combinations will have freq_band = [5, 10] with resample = 50 in half the branches, and freq_band = [30, 50] with resample = 100 in the other half.

3. Grid: The values of the keys listed are iterated over such that any value combination is equally represented in the resulting dictionary. Combinations are repeated as many times as it takes to reach the number of combinations dictated by construction or by specification. E.g., if settings[“zipped”][“freq_band”] = [[5, 10], [30, 50]] and …[“resample”] = [50, 100], then self.combinations will have four qualitatively distinct groups, with (freq_band, resample) taking one of the following combined values: ([5, 10], 50), ([5, 10], 100), ([30, 50], 50), ([30, 50], 100).

4. Random: The values of the parameter keys listed are drawn from a scipy.stats distribution given under the key “method”, with the arguments “args”. E.g., if settings[“random”][“resample”][“method”]=”uniform” and …[“args”] = [x, y], then resample values will be drawn from U(x, x+y).

Example

Define a search space and generate a list of parameter combination dictionaries:

>>> s = Sweep(search_space={"fixed": {"a": 4}, "grid": {"b": [6, 7], "c": [8, 9]}}, num_combinations=4)
>>> print(s())
[{'a': 4, 'b': 6, 'c': 8}, {'a': 4, 'b': 6, 'c': 9}, {'a': 4, 'b': 7, 'c': 8}, {'a': 4, 'b': 7, 'c': 9}]