Alternate different models in Pipeline for GridSea

Question

I want to build a Pipeline in sklearn and test different models using GridSearchCV.

Just an example (please do not pay attention on what particular models are chosen):

reg = LogisticRegression()

proj1 = PCA(n_components=2)
proj2 = MDS()
proj3 = TSNE()

pipe = [('proj', proj1), ('reg' , reg)]

pipe = Pipeline(pipe)

param_grid = {
    'reg__c': [0.01, 0.1, 1],
}

clf = GridSearchCV(pipe, param_grid = param_grid)

Here if I want to try different models for dimensionality reduction, I need to code different pipelines and compare them manually. Is there an easy way to do it?

One solution I came up with is define my own class derived from base estimator:

class Projection(BaseEstimator):
    def __init__(self, est_name):
        if est_name == "MDS":
            self.model = MDS()
        ...
    ...
    def fit_transform(self, X):
        return self.model.fit_transform(X)

I think it will work, I just create a Projection object and pass it to Pipeline, using names of the estimators as parameters for it.

But to me this way is a bit chaotic and not scalable: it makes me to define new class each time I want to compare different models. Also to continue this solution, one could implement a class that does the same job, but with arbitrary set of models. It seems overcomplicated to me.

What is the most natural and pythonic way to compare different models?

Answer 1

Lets assume you want to use PCA and TruncatedSVD as your dimesionality reduction step.

pca = decomposition.PCA()
svd = decomposition.TruncatedSVD()
svm = SVC()
n_components = [20, 40, 64]

You can do this:

pipe = Pipeline(steps=[('reduction', pca), ('svm', svm)])

# Change params_grid -> Instead of dict, make it a list of dict
# In the first element, pass parameters related to pca, and in second related to svd

params_grid = [{
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'reduction':pca,
'reduction__n_components': n_components,
},
{
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'reduction':svd,
'reduction__n_components': n_components,
'reduction__algorithm':['randomized']
}]

and now just pass the pipeline object to gridsearchCV

grd = GridSearchCV(pipe, param_grid = params_grid)

Calling grd.fit() will search the parameters over both the elements of the params_grid list, using all values from one at a time.

Please look at my other answer for more details: "Parallel" pipeline to get best model using gridsearch

Answer 2

An alternative solution that does not require to prefix the estimators names in the parameter grid is the following:

from sklearn.ensemble import RandomForestClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.linear_model import LogisticRegression

# the models that you want to compare
models = {
    'RandomForestClassifier': RandomForestClassifier(),
    'KNeighboursClassifier': KNeighborsClassifier(),
    'LogisticRegression': LogisticRegression()
}

# the optimisation parameters for each of the above models
params = {
    'RandomForestClassifier':{ 
            "n_estimators"      : [100, 200, 500, 1000],
            "max_features"      : ["auto", "sqrt", "log2"],
            "bootstrap": [True],
            "criterion": ['gini', 'entropy'],
            "oob_score": [True, False]
            },
    'KNeighboursClassifier': {
        'n_neighbors': np.arange(3, 15),
        'weights': ['uniform', 'distance'],
        'algorithm': ['ball_tree', 'kd_tree', 'brute']
        },
    'LogisticRegression': {
        'solver': ['newton-cg', 'sag', 'lbfgs'],
        'multi_class': ['ovr', 'multinomial']
        }  
}

and you can define:

from sklearn.model_selection import GridSearchCV

def fit(train_features, train_actuals):
        """
        fits the list of models to the training data, thereby obtaining in each 
        case an evaluation score after GridSearchCV cross-validation
        """
        for name in models.keys():
            est = models[name]
            est_params = params[name]
            gscv = GridSearchCV(estimator=est, param_grid=est_params, cv=5)
            gscv.fit(train_features, train_actuals)
            print("best parameters are: {}".format(gscv.best_estimator_))

basically running through the different models, each model referring to its own set of optimisation parameters through a dictionary. Of course do not forget to pass the models and the parameters dictionary to the fit function, in case you do not have them as global variables. Have a look at this GitHub project for a more complete overview.