Implementation:Scikit learn Scikit learn ParameterGrid Init

Template:Implementation Metadata

Overview

Concrete tool for generating exhaustive parameter combinations from a grid specification provided by scikit-learn.

The ParameterGrid class takes a dictionary (or list of dictionaries) mapping parameter names to sequences of allowed values, and produces an iterable over all combinations via the Cartesian product. It supports indexing, length queries, and deterministic iteration order.

Code Reference

Class Signature

class ParameterGrid:
    def __init__(self, param_grid):
        ...

Constructor Parameter

Parameter	Type	Description
param_grid	dict of str to sequence, or sequence of such dicts	The parameter grid to explore. Each dictionary maps estimator parameter names (strings) to sequences of allowed values. A sequence of dicts specifies multiple sub-grids to search, which is useful for avoiding nonsensical parameter combinations.

I/O Contract

Input:

• A single dictionary {'param_name': [val1, val2, ...], ...} or a list of such dictionaries.
• Each value sequence must be a non-empty list or numpy array (not a string, not a scalar).
• Numpy arrays must be one-dimensional.

Output (via iteration):

• Yields dictionaries mapping each parameter name to one of its allowed values.
• The total number of yielded dictionaries equals the sum of the Cartesian product sizes for each sub-grid.
• Keys are sorted alphabetically within each yielded dictionary for reproducibility.

Protocols supported:

• __iter__ -- iterate over all parameter combinations
• __len__ -- return the total number of combinations
• __getitem__(ind) -- return the ind-th combination (enables memory-efficient random sampling without replacement)

Validation Rules

The constructor validates the input and raises:

• TypeError if param_grid is not a dict or list of dicts.
• TypeError if any value is a string or is not a list/numpy array.
• ValueError if any numpy array has more than one dimension.
• ValueError if any value sequence is empty.

Usage Examples

Single Grid

from sklearn.model_selection import ParameterGrid

param_grid = {'C': [0.1, 1, 10], 'kernel': ['linear', 'rbf']}
grid = ParameterGrid(param_grid)

print(len(grid))  # 6 (3 * 2)

for params in grid:
    print(params)
# {'C': 0.1, 'kernel': 'linear'}
# {'C': 0.1, 'kernel': 'rbf'}
# {'C': 1, 'kernel': 'linear'}
# {'C': 1, 'kernel': 'rbf'}
# {'C': 10, 'kernel': 'linear'}
# {'C': 10, 'kernel': 'rbf'}

Multiple Sub-grids

from sklearn.model_selection import ParameterGrid

# Avoid nonsensical combinations: gamma only applies to rbf kernel
param_grid = [
    {'kernel': ['linear'], 'C': [1, 10]},
    {'kernel': ['rbf'], 'C': [1, 10], 'gamma': [0.1, 1]}
]
grid = ParameterGrid(param_grid)

print(len(grid))  # 2 + 4 = 6
print(grid[3])    # Indexing: {'C': 10, 'gamma': 0.1, 'kernel': 'rbf'}

Randomized Alternative: ParameterSampler

For randomized search over parameter spaces, scikit-learn provides ParameterSampler as the counterpart to ParameterGrid.

class ParameterSampler:
    def __init__(self, param_distributions, n_iter, *, random_state=None):
        ...

Parameter	Type	Description
param_distributions	dict or list of dicts	Maps parameter names to distributions (objects with an rvs method, such as scipy.stats distributions) or lists of discrete values.
n_iter	int	Number of parameter settings to sample.
random_state	int, RandomState, or None	Controls reproducibility.

Key behaviors:

• If all parameters are lists, sampling is done without replacement from the full grid.
• If at least one parameter has a distribution with an rvs method, sampling is done with replacement.
• A warning is issued if n_iter exceeds the total grid size when all parameters are lists.

Related Pages

• Principle:Scikit_learn_Scikit_learn_Parameter_Space_Definition