MNIST：手写数字数据集

1# required imports
2import pandas as pd
3import numpy as np
4import matplotlib.pyplot as plt
5import seaborn as sns
6sns.set(font_scale=1.3) # setting font size for the whole notebook
7sns.set_style("white") # want to set the style
8
9from sklearn import svm, linear_model, datasets
10from sklearn.model_selection import cross_val_score

1# loading handwritten digits dataset!
2mnist = datasets.load_digits()
3print(mnist.keys())

1df=pd.DataFrame(data=mnist.data,columns=mnist.feature_names)
2df['target']=mnist.target
3print("We have {} numbers/observations in the data.".format(len(df)))
4print("Unique target values are: \n{} \nand the respective class balance is: \n{}".format(
5    df.target.unique(), df.target.value_counts().sort_index().values))
6print(df.tail(2))

1index=-2 # 1, 5 or some other available index
2number_at_0_index=np.array(df.drop('target',axis=1).iloc[index]).reshape(8,8)
3print("target value is: {}".format(int(df.iloc[index][-1])))
4plt.imshow(number_at_0_index,cmap=plt.cm.viridis);

1# The same number as above
2plt.imshow(mnist.images[index], cmap=plt.cm.gray_r)#, interpolation='nearest')
3plt.show()

1# features and the target
2digits_X, digits_y = df.drop('target',axis=1), df.target
3
4# Logistic regression without doing anything
5print("Cross-validated mean scores:")
6LogR = linear_model.LogisticRegression(multi_class='multinomial')
7print("\tUsing logistic regression is\t\t: {}".format(
8    cross_val_score(estimator=LogR,X=digits_X,
9                    y=digits_y,cv=5,scoring='accuracy').mean()))
10
11# SVM with linear kernel
12svc_linear = svm.SVC(kernel='linear')#,decision_function_shape='ovr')
13svc_linear.fit(digits_X,digits_y)
14print("\n\tUsing svc with linear kernel is\t\t: {}".format(
15    cross_val_score(estimator=svc_linear,X=digits_X,
16                    y=digits_y,cv=5,scoring='accuracy').mean()))
17
18# SVM with polynomial kernel
19svc_poly = svm.SVC(kernel='poly')
20svc_poly.fit(digits_X,digits_y)
21print("\tUsing svc with polynomial kernel is\t: {}".format(
22    cross_val_score(estimator=svc_poly,X=digits_X,
23                    y=digits_y,cv=5,scoring='accuracy').mean()))
24
25# SVM with rbf (aka gaussian) kernel - default
26svc_rbf = svm.SVC(kernel='rbf')
27svc_rbf.fit(digits_X,digits_y)
28print("\tUsing svc with rbf kernel is\t\t: {}".format(
29    cross_val_score(estimator=svc_rbf,X=digits_X,
30                    y=digits_y,cv=5,scoring='accuracy').mean()))

1# This will take some time, let's note the time using the time module
2import time
3start=time.time()
4from sklearn.model_selection import GridSearchCV
5
6# The model
7model = svm.SVC()
8
9#parameters grid
10gamma_values = np.logspace(-5, 2, 10)
11C_values = np.logspace(-3, 2, 10)
12kernels = ['rbf', 'sigmoid', 'linear', 'poly']
13param_grid = dict(gamma=gamma_values, C=C_values, kernel=kernels)
14
15#grid-search
16grid = GridSearchCV(estimator=model,param_grid=param_grid,
17                    cv=5,scoring='accuracy',verbose=1)
18grid.fit(digits_X, digits_y)
19
20print("GridSearch took {} sec".format(time.time()-start))

1print("The best set of parameters is: ",grid.best_params_)
2print("The best score is: ",grid.best_score_)
3best_svm = grid.best_estimator_
4print("The best model is saved in 'best_svm'")

1grid_results = pd.DataFrame(grid.cv_results_)
2# we can separate the values of parameters in 
3# individual column from params dictionary as well.
4grid_results['C'] = [x['C'] for x in grid_results.params.values]
5grid_results['gamma'] = [x['gamma'] for x in grid_results.params.values]
6grid_results['kernel'] = [x['kernel'] for x in grid_results.params.values]
7grid_results['mean_score'] = grid_results['mean_test_score']
8#There are several columns, we are just grabbing a few!
9print(grid_results[['C','gamma','kernel','mean_score']]\
10        .sort_values('mean_score',ascending=False).head())