Many a times while working on a dataset and using a Machine Learning model we don't know which set of hyperparameters will give us the best result. Passing all sets of hyperparameters manually through the model and checking the result might be a hectic work and may not be possible to do.
This data science python source code does the following:
1. Hyper-parameters of logistic regression.
2. Implements Standard Scaler function on the dataset.
3. Performs train_test_split on your dataset.
4. Uses Cross Validation to prevent overfitting.
To get the best set of hyperparameters we can use Grid Search. Grid Search passes all combinations of hyperparameters one by one into the model and check the result. Finally it gives us the set of hyperparemeters which gives the best result after passing in the model.
So this recipe is a short example of how to use Grid Search and get the best set of hyperparameters.
import numpy as np from sklearn import linear_model, decomposition, datasets from sklearn.pipeline import Pipeline from sklearn.model_selection import GridSearchCV from sklearn.preprocessing import StandardScaler
Here we have imported various modules like decomposition, datasets, linear_model, Pipeline, StandardScaler and GridSearchCV from differnt libraries. We will understand the use of these later while using it in the in the code snipet.
For now just have a look on these imports.
Here we have used datasets to load the inbuilt wine dataset and we have created objects X and y to store the data and the target value respectively.
dataset = datasets.load_wine()
X = dataset.data
y = dataset.target
StandardScaler is used to remove the outliners and scale the data by making the mean of the data 0 and standard deviation as 1. So we are creating an object std_scl to use standardScaler.
std_slc = StandardScaler()
We are also using Principal Component Analysis(PCA) which will reduce the dimension of features by creating new features which have most of the varience of the original data.
pca = decomposition.PCA()
Here, we are using Logistic Regression as a Machine Learning model to use GridSearchCV. So we have created an object Logistic_Reg.
logistic_Reg = linear_model.LogisticRegression()
Pipeline will helps us by passing modules one by one through GridSearchCV for which we want to get the best parameters. So we are making an object pipe to create a pipeline for all the three objects std_scl, pca and logistic_Reg.
pipe = Pipeline(steps=[('std_slc', std_slc),
Now we have to define the parameters that we want to optimise for these three objects.
StandardScaler doesnot requires any parameters to be optimised by GridSearchCV.
Principal Component Analysis requires a parameter 'n_components' to be optimised. 'n_components' signifies the number of components to keep after reducing the dimension.
n_components = list(range(1,X.shape+1,1))
Logistic Regression requires two parameters 'C' and 'penalty' to be optimised by GridSearchCV. So we have set these two parameters as a list of values form which GridSearchCV will select the best value of parameter.
C = np.logspace(-4, 4, 50)
penalty = ['l1', 'l2']
Now we are creating a dictionary to set all the parameters options for different modules.
parameters = dict(pca__n_components=n_components,
Before using GridSearchCV, lets have a look on the important parameters.
clf_GS = GridSearchCV(pipe, parameters) clf_GS.fit(X, y)Now we are using print statements to print the results. It will give the values of hyperparameters as a result.
print('Best Penalty:', clf_GS.best_estimator_.get_params()['logistic_Reg__penalty']) print('Best C:', clf_GS.best_estimator_.get_params()['logistic_Reg__C']) print('Best Number Of Components:', clf_GS.best_estimator_.get_params()['pca__n_components']) print(); print(clf_GS.best_estimator_.get_params()['logistic_Reg'])As an output we get:
Best Penalty: l1 Best C: 109.85411419875572 Best Number Of Components: 13 LogisticRegression(C=109.85411419875572, class_weight=None, dual=False, fit_intercept=True, intercept_scaling=1, max_iter=100, multi_class='warn', n_jobs=None, penalty='l1', random_state=None, solver='warn', tol=0.0001, verbose=0, warm_start=False)