Scikit Learn GridSearchCV without cross validation (unsupervised learning)

Question

Is it possible to use GridSearchCV without cross validation? I am trying to optimize the number of clusters in KMeans clustering via grid search, and thus I don't need or want cross validation.

The documentation is also confusing me because under the fit() method, it has an option for unsupervised learning (says to use None for unsupervised learning). But if you want to do unsupervised learning, you need to do it without cross validation and there appears to be no option to get rid of cross validation.

Metempsychosis

You can implement a custom cv which will put all data into training and test. – Trimming

Answer 1

After much searching, I was able to find this thread. It appears that you can get rid of cross validation in GridSearchCV if you use:

cv=[(slice(None), slice(None))]

I have tested this against my own coded version of grid search without cross validation and I get the same results from both methods. I am posting this answer to my own question in case others have the same issue.

Edit: to answer jjrr's question in the comments, here is an example use case:

from sklearn.metrics import silhouette_score as sc

def cv_silhouette_scorer(estimator, X):
    estimator.fit(X)
    cluster_labels = estimator.labels_
    num_labels = len(set(cluster_labels))
    num_samples = len(X.index)
    if num_labels == 1 or num_labels == num_samples:
        return -1
    else:
        return sc(X, cluster_labels)

cv = [(slice(None), slice(None))]
gs = GridSearchCV(estimator=sklearn.cluster.MeanShift(), param_grid=param_dict, 
                  scoring=cv_silhouette_scorer, cv=cv, n_jobs=-1)
gs.fit(df[cols_of_interest])

Metempsychosis

I get the error: AttributeError: 'slice' object has no attribute 'flags'. Python 3.6, sklearn 0.20.3 – Eschar

AttributeError: 'slice' object has no attribute 'flags' – Warbler

Python 3.8: AttributeError: 'memmap' object has no attribute 'index' – Auxochrome

why do you call fit in the scoring function? – Lackaday

I am using hdbscan and would like to implement this method, but when I run GridSearchCV in verbose =10 it says score=nan in the outputs. I have scorer= make_scorer(hdbscan.validity.validity_index,greater_is_better=True). Any help appreciated – Induration

How does cv=[(slice(None), slice(None))] help me when I use train_test_split and want to perform hyperparameter tuning on the test set? I want to fit the model on the training set but score on the test set. – Lancastrian

Answer 2

I'm going to answer your question since it seems like it has been unanswered still. Using the parallelism method with the for loop, you can use the multiprocessing module.

from multiprocessing.dummy import Pool
from sklearn.cluster import KMeans
import functools

kmeans = KMeans()

# define your custom function for passing into each thread
def find_cluster(n_clusters, kmeans, X):
    from sklearn.metrics import silhouette_score  # you want to import in the scorer in your function

    kmeans.set_params(n_clusters=n_clusters)  # set n_cluster
    labels = kmeans.fit_predict(X)  # fit & predict
    score = silhouette_score(X, labels)  # get the score

    return score

# Now's the parallel implementation
clusters = [3, 4, 5]
pool = Pool()
results = pool.map(functools.partial(find_cluster, kmeans=kmeans, X=X), clusters)
pool.close()
pool.join()

# print the results
print(results)  # will print a list of scores that corresponds to the clusters list

Samira

Answer 3

I think that using cv=ShuffleSplit(test_size=0.20, n_splits=1) with n_splits=1 is a better solution like this post suggested

Hitchhike

This yields another result as cv = [(slice(None), slice(None))] (the resulting scores are different) – Pelagian

Answer 4

I recently came out with the following custom cross-validator, based on this answer. I passed it to GridSearchCV and it properly disabled the cross-validation for me:

import numpy as np

class DisabledCV:
    def __init__(self):
        self.n_splits = 1

    def split(self, X, y, groups=None):
        yield (np.arange(len(X)), np.arange(len(y)))

    def get_n_splits(self, X, y, groups=None):
        return self.n_splits

I hope it can help.

Extranuclear

I test your solution, I got this error: "return self.n_splits AttributeError: 'numpy.ndarray' object has no attribute 'n_splits' ". Do you know how to fix it? – Suffuse

Answer 5

You can create your own GridSearch using ParameterGrid.

For example:

from sklearn.model_selection import ParameterGrid

param_grid = {'a': [1, 2], 'b': [True, False]}

param_candidates = ParameterGrid(param_grid)
print(f'{len(param_candidates)} candidates')
results = []
for i, params in enumerate(param_candidates):
    model = estimator.set_params(**params)
    model.fit(X_train, y_train)
    score = model.score(X_val, y_val)
    results.append([params, score])
    print(f'{i+1}/{len(param_candidates)}: ', params, score)

print(max(results, key=lambda x: x[1]))

To increase performance I would suggest parallelizing the loop:

from joblib import Parallel, delayed

param_grid = {'a': [1, 2], 'b': [True, False]}
param_candidates = ParameterGrid(param_grid)
print(f'{len(param_candidates)} candidates')

def fit_model(params):
    model = estimator.set_params(**params)
    model.fit(X_train, y_train)
    score = model.score(X_val, y_val)
    return [params, score]

results = Parallel(n_jobs=-1)(delayed(fit_model)(params) for params in param_candidates)
print(max(results, key=lambda x: x[1]))

Lancastrian