K = min(train_count, kernel_count)
centroids, _ = kmeans(train_x, K)
groups, _ = vq(train_x, centroids)
mu = np.zeros((K, train_x.shape[1]))
sigma = np.zeros((K, train_x.shape[1], train_x.shape[1]))
t3 = time()
print '완료 (%f seconds)' % (t3 - t2)
#
# Training
#
print '평균, 표준편차, Weight 행렬 계산',
for j in range(K):
selected = train_x[groups[j]]
mu[j] = sum(selected) / selected.shape[0]
sigma[j] = np.cov(selected.transpose()) + 0.1 * np.eye(train_x.shape[1])
H = tool.get_H(train_x, mu, sigma)
Ht = H.transpose()
Y = tool.get_Y(train_y)
W = np.linalg.lstsq(Ht.dot(H), Ht.dot(Y))[0]
tEnd = time()
print '완료 (%f seconds)' % (tEnd - t3)
print ''
print '학습에 쓰인 총 시간: %f seconds' % (tEnd - t0)
print ''
centroids, _ = kmeans(train_x, K)
groups, _ = vq(train_x, centroids)
mu = np.zeros((K, train_x.shape[1]))
sigma = np.zeros((K, train_x.shape[1], train_x.shape[1]))
t3 = time()
print '완료 (%f seconds)' % (t3 - t2)
#
# Training
#
print '평균, 표준편차, Weight 행렬 계산',
for j in range(K):
selected = train_x[groups[j]]
mu[j] = sum(selected)/selected.shape[0]
sigma[j] = np.cov(selected.transpose()) + 0.1 * np.eye(train_x.shape[1])
H = tool.get_H(train_x, mu, sigma)
Ht = H.transpose()
Y = tool.get_Y(train_y)
W = np.linalg.lstsq(Ht.dot(H), Ht.dot(Y))[0]
tEnd = time()
print '완료 (%f seconds)' % (tEnd - t3)
print ''
print '학습에 쓰인 총 시간: %f seconds' % (tEnd - t0)
print ''
test_x, test_y = (test_set[0][:test_count], test_set[1][:test_count])
# PCA
test_x = learn.pca.dot(test_x.transpose()).transpose()
test_count = learn.dimension
t1 = time()
print '완료 (%f seconds)' % (t1 - t0)
#
# Test
#
print '퍼포먼스',
H = tool.get_H(test_x, learn.mu, learn.sigma)
Y = tool.get_Y(test_y)
f = H.dot(learn.W)
max_val = f.max(1)
max_idx = f.argmax(1)
t2 = time()
print '완료 (%f seconds)' % (t2 - t1)
#
# Grade
#
print '채점',
confusion_matrix = np.zeros((10, 10))
for i in range(test_y.shape[0]):
confusion_matrix[test_y[i], max_idx[i]] += 1
test_x, test_y = (test_set[0][:test_count], test_set[1][:test_count])
# PCA
test_x = learn.pca.dot(test_x.transpose()).transpose()
test_count = learn.dimension
t1 = time()
print '완료 (%f seconds)' % (t1 - t0)
#
# Test
#
print '퍼포먼스',
H = tool.get_H(test_x, learn.mu, learn.sigma)
Y = tool.get_Y(test_y)
f = H.dot(learn.W)
max_val = f.max(1)
max_idx = f.argmax(1)
t2 = time()
print '완료 (%f seconds)' % (t2 - t1)
#
# Grade
#
print '채점',
confusion_matrix = np.zeros((10,10))
for i in range(test_y.shape[0]):
confusion_matrix[test_y[i], max_idx[i]] += 1