Testing influence of dataset size on C

test_c.py

import numpy as np

from sklearn import datasets
from sklearn.cross_validation import ShuffleSplit
from sklearn.grid_search import GridSearchCV
from sklearn.svm import SVC
from sklearn.preprocessing import Scaler

#data = datasets.load_digits()
data = datasets.fetch_mldata("usps")
X, y = data.data, data.target
X = Scaler().fit_transform(X)

n_samples, n_features = X.shape

C_grid = dict(C=2. ** np.arange(0, 20))

cv = ShuffleSplit(n=n_samples, train_fraction=.7, test_fraction=.2, n_iterations=10)
grid_search = GridSearchCV(SVC(kernel='rbf'), param_grid=C_grid, cv=cv, n_jobs=12)

grid_search.fit(X, y)
print(grid_search.grid_scores_)