Use scikit-learn¶
This section describes how to use scikit-learn functionalities via pandas-ml.
Basics¶
You can create ModelFrame instance from scikit-learn datasets directly.
>>> import pandas_ml as pdml
>>> import sklearn.datasets as datasets
>>> df = pdml.ModelFrame(datasets.load_iris())
>>> df.head()
.target sepal length (cm) sepal width (cm) petal length (cm) \
0 0 5.1 3.5 1.4
1 0 4.9 3.0 1.4
2 0 4.7 3.2 1.3
3 0 4.6 3.1 1.5
4 0 5.0 3.6 1.4
petal width (cm)
0 0.2
1 0.2
2 0.2
3 0.2
4 0.2
# make columns be readable
>>> df.columns = ['.target', 'sepal length', 'sepal width', 'petal length', 'petal width']
ModelFrame has accessor methods which makes easier access to scikit-learn namespace.
>>> df.cluster.KMeans
<class 'sklearn.cluster.k_means_.KMeans'>
Following table shows scikit-learn module and corresponding ModelFrame module. Some accessors has its abbreviated versions.
Thus, you can instanciate each estimator via ModelFrame accessors. Once create an estimator, you can pass it to ModelFrame.fit then predict. ModelFrame automatically uses its data and target properties for each operations.
>>> estimator = df.cluster.KMeans(n_clusters=3)
>>> df.fit(estimator)
>>> predicted = df.predict(estimator)
>>> predicted
0 1
1 1
2 1
...
147 2
148 2
149 0
Length: 150, dtype: int32
ModelFrame preserves the most recently used estimator in estimator atribute, and predicted results in predicted attibute.
>>> df.estimator
KMeans(copy_x=True, init='k-means++', max_iter=300, n_clusters=3, n_init=10,
n_jobs=1, precompute_distances=True, random_state=None, tol=0.0001,
verbose=0)
>>> df.predicted
0 1
1 1
2 1
...
147 2
148 2
149 0
Length: 150, dtype: int32
ModelFrame has following methods corresponding to various scikit-learn estimators. The last results are saved as corresponding ModelFrame properties.
ModelFrame method |
ModelFrame property |
|---|---|
ModelFrame.fit |
(None) |
ModelFrame.transform |
(None) |
ModelFrame.fit_transform |
(None) |
ModelFrame.inverse_transform |
(None) |
ModelFrame.predict |
ModelFrame.predicted |
ModelFrame.fit_predict |
ModelFrame.predicted |
ModelFrame.score |
(None) |
ModelFrame.predict_proba |
ModelFrame.proba |
ModelFrame.predict_log_proba |
ModelFrame.log_proba |
ModelFrame.decision_function |
ModelFrame.decision |
Note
If you access to a property before calling ModelFrame methods, ModelFrame automatically calls corresponding method of the latest estimator and return the result.
Following example shows to perform PCA, then revert principal components back to original space. inverse_transform should revert the original columns.
>>> estimator = df.decomposition.PCA()
>>> df.fit(estimator)
>>> transformed = df.transform(estimator)
>>> transformed.head()
.target 0 1 2 3
0 0 -2.684207 -0.326607 0.021512 0.001006
1 0 -2.715391 0.169557 0.203521 0.099602
2 0 -2.889820 0.137346 -0.024709 0.019305
3 0 -2.746437 0.311124 -0.037672 -0.075955
4 0 -2.728593 -0.333925 -0.096230 -0.063129
>>> type(transformed)
<class 'pandas_ml.core.frame.ModelFrame'>
>>> transformed.inverse_transform(estimator)
.target sepal length sepal width petal length petal width
0 0 5.1 3.5 1.4 0.2
1 0 4.9 3.0 1.4 0.2
2 0 4.7 3.2 1.3 0.2
3 0 4.6 3.1 1.5 0.2
4 0 5.0 3.6 1.4 0.2
.. ... ... ... ... ...
145 2 6.7 3.0 5.2 2.3
146 2 6.3 2.5 5.0 1.9
147 2 6.5 3.0 5.2 2.0
148 2 6.2 3.4 5.4 2.3
149 2 5.9 3.0 5.1 1.8
[150 rows x 5 columns]
If ModelFrame both has target and predicted values, the model evaluation can be performed using functions available in ModelFrame.metrics.
>>> estimator = df.svm.SVC()
>>> df.fit(estimator)
>>> df.predict(estimator)
0 0
1 0
2 0
...
147 2
148 2
149 2
Length: 150, dtype: int64
>>> df.predicted
0 0
1 0
2 0
...
147 2
148 2
149 2
Length: 150, dtype: int64
>>> df.metrics.confusion_matrix()
Predicted 0 1 2
Target
0 50 0 0
1 0 48 2
2 0 0 50
Use Module Level Functions¶
Some scikit-learn modules define functions which handle data without instanciating estimators. You can call these functions from accessor methods directly, and ModelFrame will pass corresponding data on background. Following example shows to use sklearn.cluster.k_means function to perform K-means.
Important
When you use module level function, ModelFrame.predicted WILL NOT be updated. Thus, using estimator is recommended.
# no need to pass data explicitly
# sklearn.cluster.kmeans returns centroids, cluster labels and inertia
>>> c, l, i = df.cluster.k_means(n_clusters=3)
>>> l
0 1
1 1
2 1
...
147 2
148 2
149 0
Length: 150, dtype: int32
Pipeline¶
ModelFrame can handle pipeline as the same as normal estimators.
>>> estimators = [('reduce_dim', df.decomposition.PCA()),
... ('svm', df.svm.SVC())]
>>> pipe = df.pipeline.Pipeline(estimators)
>>> df.fit(pipe)
>>> df.predict(pipe)
0 0
1 0
2 0
...
147 2
148 2
149 2
Length: 150, dtype: int64
Above expression is the same as below:
>>> df2 = df.copy()
>>> df2 = df2.fit_transform(df2.decomposition.PCA())
>>> svm = df2.svm.SVC()
>>> df2.fit(svm)
SVC(C=1.0, cache_size=200, class_weight=None, coef0=0.0, degree=3, gamma=0.0,
kernel='rbf', max_iter=-1, probability=False, random_state=None,
shrinking=True, tol=0.001, verbose=False)
>>> df2.predict(svm)
0 0
1 0
2 0
...
147 2
148 2
149 2
Length: 150, dtype: int64
Cross Validation¶
scikit-learn has some classes for cross validation. model_selection.train_test_split splits data to training and test set. You can access to the function via model_selection accessor.
>>> train_df, test_df = df.model_selection.train_test_split()
>>> train_df
.target sepal length sepal width petal length petal width
124 2 6.7 3.3 5.7 2.1
117 2 7.7 3.8 6.7 2.2
123 2 6.3 2.7 4.9 1.8
65 1 6.7 3.1 4.4 1.4
133 2 6.3 2.8 5.1 1.5
.. ... ... ... ... ...
93 1 5.0 2.3 3.3 1.0
46 0 5.1 3.8 1.6 0.2
121 2 5.6 2.8 4.9 2.0
91 1 6.1 3.0 4.6 1.4
147 2 6.5 3.0 5.2 2.0
[112 rows x 5 columns]
>>> test_df
.target sepal length sepal width petal length petal width
146 2 6.3 2.5 5.0 1.9
75 1 6.6 3.0 4.4 1.4
138 2 6.0 3.0 4.8 1.8
77 1 6.7 3.0 5.0 1.7
36 0 5.5 3.5 1.3 0.2
.. ... ... ... ... ...
14 0 5.8 4.0 1.2 0.2
141 2 6.9 3.1 5.1 2.3
100 2 6.3 3.3 6.0 2.5
83 1 6.0 2.7 5.1 1.6
114 2 5.8 2.8 5.1 2.4
[38 rows x 5 columns]
You can iterate over Splitter classes via ModelFrame.model_selection.split which returns ModelFrame corresponding to training and test data.
>>> kf = df.model_selection.KFold(n_splits=3)
>>> for train_df, test_df in df.model_selection.iterate(kf):
... print('training set shape: ', train_df.shape,
... 'test set shape: ', test_df.shape)
training set shape: (112, 5) test set shape: (38, 5)
training set shape: (112, 5) test set shape: (38, 5)
training set shape: (112, 5) test set shape: (38, 5)
Grid Search¶
You can perform grid search using ModelFrame.fit.
>>> tuned_parameters = [{'kernel': ['rbf'], 'gamma': [1e-3, 1e-4],
... 'C': [1, 10, 100]},
... {'kernel': ['linear'], 'C': [1, 10, 100]}]
>>> df = pdml.ModelFrame(datasets.load_digits())
>>> cv = df.model_selection.GridSearchCV(df.svm.SVC(C=1), tuned_parameters,
... cv=5)
>>> df.fit(cv)
>>> cv.best_estimator_
SVC(C=10, cache_size=200, class_weight=None, coef0=0.0, degree=3, gamma=0.001,
kernel='rbf', max_iter=-1, probability=False, random_state=None,
shrinking=True, tol=0.001, verbose=False)
In addition, ModelFrame.model_selection has a describe function to organize each grid search result as ModelFrame accepting estimator.
>>> df.model_selection.describe(cv)
mean std C gamma kernel
0 0.974108 0.013139 1 0.0010 rbf
1 0.951416 0.020010 1 0.0001 rbf
2 0.975372 0.011280 10 0.0010 rbf
3 0.962534 0.020218 10 0.0001 rbf
4 0.975372 0.011280 100 0.0010 rbf
5 0.964695 0.016686 100 0.0001 rbf
6 0.951811 0.018410 1 NaN linear
7 0.951811 0.018410 10 NaN linear
8 0.951811 0.018410 100 NaN linear