-
Notifications
You must be signed in to change notification settings - Fork 5
/
ls.py
59 lines (50 loc) · 2.01 KB
/
ls.py
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
import numpy as np
import scipy.linalg
import sklearn.metrics as metrics
def train_ls_dual_model(K_train, y_train, reg):
K_train = K_train.astype('float64')
y = np.eye(np.max(y_train) + 1)[y_train]
idxs = np.diag_indices(K_train.shape[0])
K_train[idxs] += reg
model = scipy.linalg.solve(K_train, y, sym_pos=True)
K_train[idxs] -= reg
return model, 0.0
def train_ls_dual_model_loo_tilt(K_train, y_train, K_test, y_test, reg):
K_train = K_train.astype('float64')
y = np.eye(np.max(y_train) + 1)[y_train]
idxs = np.diag_indices(K_train.shape[0])
K_train[idxs] += reg
K_train_inv = np.linalg.inv(K_train)
K_train[idxs] -= reg
alpha_regular = K_train_inv.dot(y)
Q = np.diag(K_train_inv)[:, np.newaxis]
y_loo = y - Q/alpha
results = []
for tilt_fac in range(100):
tilt_fac /= 100
alpha_loo_tilt = K_train_inv.dot(y - tilt_fac*y_loo)
y_test_pred = np.argmax(K_test.dot(alpha_loo_tilt), axis=1)
acc_test = metrics.accuracy_score(y_test_pred, y_test)
results.append((acc_test, alpha_loo_tilt))
best_acc, model = max(results, key_func=lambda x: x[0])
return model, 0.0
def train_ls_dual_model_center(K_train, y_train, reg):
K_train = K_train.astype('float64')
K_row_sums = np.sum(K_train, axis=1)[:, np.newaxis]
K_train_c = K_train.copy()
K_train_c -= K_row_sums/K_train.shape[0]
K_train_c -= K_row_sums.T/K_train.shape[0]
K_train_c += np.sum(K_train)/(K_train.shape[0]*K_train.shape[0])
y = np.eye(np.max(y_train) + 1)[y_train]
y_c = y - np.mean(y, axis=0)
idxs = np.diag_indices(K_train.shape[0])
K_train_c[idxs] += reg
model = scipy.linalg.solve(K_train_c, y_c, sym_pos=True)
K_train_c[idxs] -= reg
bias = np.sum((y - K_train.dot(model)), axis=0)/K_train.shape[0]
return model, bias
def eval_ls_model(model, bias, K, y):
K = K.astype('float64')
logits = K.dot(model) + bias
preds = np.argmax(logits, axis=1)
return logits, preds, np.sum(preds == y)/y.shape[0]