Kaggle-Challenge-Kernels-Methods-For-Machine-Learning/start.py at master · Ulrich777/Kaggle-Challenge-Kernels-Methods-For-Machine-Learning · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
from kernels import load_sequence, save_spectrum_kernels, save_exponential_kernels, get_true_mismatch
import numpy as np
import pandas as pd
from classifiers import MultiKerOpt, CvSearch, get_best

sequences_train = load_sequence(number=3)
sequences_test = load_sequence(number=3, root='te')
labels_train = load_sequence(kind='Y' ,root='tr', number=3)


all_labels = labels_train.Bound.values.astype(int)

def _normalize(K, normalize=True):
    if normalize:
        diag = np.sqrt(np.diag(K))[:, np.newaxis]
        K_ = (1/diag) * K * (1/diag.T)
        k_xx, k_yx = K_[:2000, :2000], K_[2000:,:2000]
    else:
        k_xx, k_yx = K[:2000, :2000], K[2000:,:2000]
    return k_xx, k_yx

def build_training(k, sizes=[4,5], gamma=0.4, spectrum=False, m=1, normalize=True):

    K_xx, K_yx = [], []

    for size in sizes:
        if spectrum:
            print('loading spectrum kernel for size %d '%size)
            K = np.loadtxt('./mykernels/spectrum/K_%d_%d.txt'%(k, size))
            k_xx, k_yx = _normalize(K, normalize=normalize)

            K_xx.append(k_xx)
            K_yx.append(k_yx)

        if gamma is not None:
            print('loading exponential kernel for size %d '%size)
            K = np.loadtxt('./mykernels/exponential/K_%d_%d_%d.txt'%(k, size, 10*gamma))
            k_xx, k_yx = _normalize(K, normalize=normalize)

            K_xx.append(k_xx)
            K_yx.append(k_yx)

        if m is not None and size in [4,5]:
            print('loading mismatch kernels for size %d '%size)
            K = get_true_mismatch(sequences_train, sequences_test, ind=k, k=size, m=1, build=False)
            k_xx, k_yx = _normalize(K, normalize=normalize)

            K_xx.append(k_xx)
            K_yx.append(k_yx)


    y_train = all_labels[2000*k:2000*(k+1)]

    return np.array(K_xx), np.array(K_yx), y_train

#==========Building Kernels =================#
sizes =  [3, 4, 5, 6, 7]
YET_TO_SAVE_SPECTRUM = True
if YET_TO_SAVE_SPECTRUM:
    save_spectrum_kernels(sequences_train, sequences_test, k=0, sizes=sizes, rang=(4,4))
    save_spectrum_kernels(sequences_train, sequences_test, k=1, sizes=sizes, rang=(4,4))
    save_spectrum_kernels(sequences_train, sequences_test, k=2, sizes=sizes, rang=(4,4))

YET_TO_SAVE_EXPONENTIAL = True

if YET_TO_SAVE_EXPONENTIAL:

    for gamma in [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.9]:
        print(gamma)
        save_exponential_kernels(sequences_train, sequences_test, k=0, sizes=sizes, gamma=gamma)
        save_exponential_kernels(sequences_train, sequences_test, k=1, sizes=sizes, gamma=gamma)
        save_exponential_kernels(sequences_train, sequences_test, k=2, sizes=sizes, gamma=gamma)


YET_TO_SAVE_MISMATCH = True
if YET_TO_SAVE_MISMATCH:
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=0, k=3, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=1, k=3, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=2, k=3, m=1, build=False)


    Kk = get_true_mismatch(sequences_train, sequences_test, ind=0, k=4, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=1, k=4, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=2, k=4, m=1, build=False)


    Kk = get_true_mismatch(sequences_train, sequences_test, ind=0, k=5, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=1, k=5, m=1, build=False)
    Kk = get_true_mismatch(sequences_train, sequences_test, ind=2, k=5, m=1, build=False)
#====================================================================###

REG_PARAMS_SPAN = [10**i for i in range(-3, 2)]
GAMMA = 0.1

LAUNCH = True

if LAUNCH:
    print(GAMMA)

    K_xx_0, K_yx_0, y_train_0 = build_training(0, sizes=[3,4,5,6,7],  gamma=GAMMA, spectrum=True, m=1, normalize=True)
    df  = CvSearch(K_xx_0, K_yx_0, y_train_0, method='svm', degrees=[1, 2, 3],
               alphas=REG_PARAMS_SPAN, cv=5)
    df.to_csv('./summary/global/X0_normed_%d.csv'%(10*GAMMA), index=False)
    best_degree_0, best_alpha_0, best_0 = get_best(df)


    K_xx_1, K_yx_1, y_train_1 = build_training(1, sizes=[3,4,5,6,7],  gamma=GAMMA, spectrum=True, m=1, normalize=True)
    df  = CvSearch(K_xx_1, K_yx_1, y_train_1, method='svm', degrees=[1, 2, 3],
               alphas=REG_PARAMS_SPAN, cv=5)
    df.to_csv('./summary/global/X1_normed_%d.csv'%(10*GAMMA), index=False)
    best_degree_1, best_alpha_1, best_1 = get_best(df)

    K_xx_2, K_yx_2, y_train_2 = build_training(2, sizes=[3,4,5,6,7],  gamma=GAMMA, spectrum=True, m=1, normalize=True)
    df  = CvSearch(K_xx_2, K_yx_2, y_train_2, method='svm', degrees=[1, 2, 3],
               alphas=REG_PARAMS_SPAN, cv=5)
    df.to_csv('./summary/global/X2_normed_%d.csv'%(10*GAMMA), index=False)
    best_degree_2, best_alpha_2, best_2 = get_best(df)

    clf0 = MultiKerOpt(alpha=best_alpha_0, tol=1e-07, degree=best_degree_0, method='svm', hide=False)
    clf0.fit(K_xx_0, y_train_0, n_iter=5)
    y_pred_0 = clf0.predict(K_yx_0)

    clf1 = MultiKerOpt(alpha=best_alpha_1, tol=1e-07, degree=best_degree_1, method='svm')
    clf1.fit(K_xx_1, y_train_1, n_iter=5)
    y_pred_1 = clf1.predict(K_yx_1)

    clf2 = MultiKerOpt(alpha=best_alpha_2, tol=1e-07, degree=best_degree_2, method='svm')
    clf2.fit(K_xx_2, y_train_2, n_iter=5)
    y_pred_2 = clf2.predict(K_yx_2)

    y_pred = np.hstack((y_pred_0, y_pred_1, y_pred_2)).astype(int)
    Ids = np.array(range(3000))
    predictions = pd.DataFrame({'Id':Ids, 'Bound':y_pred.flatten()})

    predictions.to_csv('./predictions/Yte.csv', index=False)