6

In sklearn, a serial pipeline can be defined to get the best combination of hyperparameters for all consecutive parts of the pipeline. A serial pipeline can be implemented as follows:

from sklearn.svm import SVC
from sklearn import decomposition, datasets
from sklearn.pipeline import Pipeline
from sklearn.model_selection import GridSearchCV

digits = datasets.load_digits()
X_train = digits.data
y_train = digits.target

#Use Principal Component Analysis to reduce dimensionality
# and improve generalization
pca = decomposition.PCA()
# Use a linear SVC
svm = SVC()
# Combine PCA and SVC to a pipeline
pipe = Pipeline(steps=[('pca', pca), ('svm', svm)])
# Check the training time for the SVC
n_components = [20, 40, 64]
params_grid = {
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'pca__n_components': n_components,
}

But what if I want to try different algorithms for each step of the pipeline? How can I e.g. gridsearch over

Principal Component Analysis OR Singular Value Decomposition AND Support Vector machines OR Random Forest

This would require some kind of 2nd level or "meta-gridsearch", since the type of model would be one of the hyperparameters. Is that possible in sklearn?

Vivek Kumar
  • 35,217
  • 8
  • 109
  • 132
Oblomov
  • 8,953
  • 22
  • 60
  • 106

1 Answers1

11

Pipeline supports None in its steps(list of estimators) by which certain part of the pipeline can be toggled off.

You can pass None parameter to the named_steps of the pipeline to not use that estimator by setting that in params passed to GridSearchCV.

Lets assume you want to use PCA and TruncatedSVD.

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

Add svd in pipeline

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

# Change params_grid -> Instead of dict, make it a list of dict**
# In the first element, pass `svd = None`, and in second `pca = None`
params_grid = [{
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'pca__n_components': n_components,
'svd':[None]
},
{
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'pca':[None],
'svd__n_components': n_components,
'svd__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.

Simplification if parameters have same name

If both estimators in your "OR" have same name of parameters as in this case, where PCA and TruncatedSVD has n_components (or you just want to search over this parameter, this can be simplified as:

#Here I have changed the name to `preprocessor`
pipe = Pipeline(steps=[('preprocessor', pca), ('svm', svm)])

#Now assign both estimators to `preprocessor` as below:
params_grid = {
'svm__C': [1, 10, 100, 1000],
'svm__kernel': ['linear', 'rbf'],
'svm__gamma': [0.001, 0.0001],
'preprocessor':[pca, svd],
'preprocessor__n_components': n_components,
}

Generalization of this scheme

We can make a function which can automatically populate our param_grid to be supplied to the GridSearchCV using appropriate values:-

def make_param_grids(steps, param_grids):

    final_params=[]

    # Itertools.product will do a permutation such that 
    # (pca OR svd) AND (svm OR rf) will become ->
    # (pca, svm) , (pca, rf) , (svd, svm) , (svd, rf)
    for estimator_names in itertools.product(*steps.values()):
        current_grid = {}

        # Step_name and estimator_name should correspond
        # i.e preprocessor must be from pca and select.
        for step_name, estimator_name in zip(steps.keys(), estimator_names):
            for param, value in param_grids.get(estimator_name).iteritems():
                if param == 'object':
                    # Set actual estimator in pipeline
                    current_grid[step_name]=[value]
                else:
                    # Set parameters corresponding to above estimator
                    current_grid[step_name+'__'+param]=value
        #Append this dictionary to final params            
        final_params.append(current_grid)

return final_params

And use this function on any number of transformers and estimators

# add all the estimators you want to "OR" in single key
# use OR between `pca` and `select`, 
# use OR between `svm` and `rf`
# different keys will be evaluated as serial estimator in pipeline
pipeline_steps = {'preprocessor':['pca', 'select'],
                  'classifier':['svm', 'rf']}

# fill parameters to be searched in this dict
all_param_grids = {'svm':{'object':SVC(), 
                          'C':[0.1,0.2]
                         }, 

                   'rf':{'object':RandomForestClassifier(),
                         'n_estimators':[10,20]
                        },

                   'pca':{'object':PCA(),
                          'n_components':[10,20]
                         },

                   'select':{'object':SelectKBest(),
                             'k':[5,10]
                            }
                  }  


# Call the method on the above declared variables
param_grids_list = make_param_grids(pipeline_steps, all_param_grids)

Now initialize a pipeline object with names as used in above pipeline_steps

# The PCA() and SVC() used here are just to initialize the pipeline,
# actual estimators will be used from our `param_grids_list`
pipe = Pipeline(steps=[('preprocessor',PCA()), ('classifier', SVC())])

Now, finally set out gridSearchCV object and fit data

grd = GridSearchCV(pipe, param_grid = param_grids_list)
grd.fit(X, y)
Vivek Kumar
  • 35,217
  • 8
  • 109
  • 132