acc_list = []
std_list = []
train_time_list = []
test_time_list = []
for L in landmarks:
mean_accuracies = []
for i in range(ITER):
accuracies = []
train_times = []
test_times = []
for train_inds, val_inds, test_inds in splits_generator(
dist_matrices, 3, sets):
lands = select_landmarks(train_inds, L)
train_x, train_y = twod_array(
get_kernels(dist_matrices[train_inds],
lands)), labels[train_inds]
val_x, val_y = twod_array(
get_kernels(dist_matrices[val_inds], lands)), labels[val_inds]
test_x, test_y = twod_array(
get_kernels(dist_matrices[test_inds],
lands)), labels[test_inds]
t1 = time.time()
# tuning
PATH = "results/{}/missing/lmvsvm/{}".format(DATASET, recons)
print("learning on {}, missing views completed by {}. results saved in {}".
format(DATASET, recons, PATH))
acc_list = []
std_list = []
times = []
for r in ratios_missing:
print(r, "\n")
accuracies = []
rec_times = []
for train_inds, val_inds, _ in splits_generator(X, ITER, None):
train_x, train_y = X[train_inds], Y[train_inds]
val_x, val_y = X[val_inds], Y[val_inds]
lands = select_landmarks(train_x, L)
k_train_x = get_kernels(train_x, lands, rbf_kernel)
k_val_x = get_kernels(val_x, lands, rbf_kernel)
# erase some views from data
k_train_x = set_random_views_to_value(k_train_x, r, recons)
k_val_x = set_random_views_to_value(k_val_x, r, recons)
t1 = time.time()
# erase some views from training
x = set_random_views_to_value(X, r, r_type="none")
test_x = set_random_views_to_value(test_X, r, r_type="none")
# kernelize and reconstruct views
t0 = time.time()
# k_x, mask, mask2 = laplacian_reconstruction(x, rbf_kernel, test_x)
mask, mask2 = np.ones(len(Y), dtype=bool), np.ones(len(test_Y), dtype=bool)
k_x = multiview_kernels(np.vstack((x, test_x)), np.vstack((x, test_x)), rbf_kernel)
y, test_y = Y[mask], test_Y[mask2]
t10 = time.time()
# cross-validation
for train_inds, val_inds, _ in splits_generator(y, CV, None):
train_y = y[train_inds]
val_y = y[val_inds]
k_train_x = get_view_dict(k_x[np.ix_(train_inds,train_inds)])
k_val_x = get_view_dict(k_x[np.ix_(val_inds,train_inds)])
t1 = time.time()
# tuning
tuning_acc = {}.fromkeys(c_range, 0.)
for c in c_range:
model = train(k_train_x, train_y, c)
pred = predict(k_val_x, val_y, model)
accuracies = []
times = []
# erase some views from training
x = set_random_views_to_value(X, r, r_type="none", sym=True)
t0 = time.time()
# kernelize and reconstruct views
# k_x, mask = laplacian_reconstruction(x)
k_x, mask = x, np.ones(len(x), dtype=bool)
t10 = time.time()
inds = np.arange(len(X))
# cross-validation
for train_inds, val_inds, test_inds in splits_generator(Y, CV, sets):
train_val_inds = np.hstack((train_inds,val_inds))
train_inds = np.isin(inds, train_inds)[mask]
val_inds = np.isin(inds, val_inds)[mask]
test_inds = np.isin(inds, test_inds)[mask]
train_val_inds = np.isin(inds, train_val_inds)[mask]
train_y = Y[mask][train_inds]
val_y = Y[mask][val_inds]
test_y = Y[mask][test_inds]
train_val_y = Y[mask][train_val_inds]
k_train_x = get_view_dict(k_x[np.ix_(train_inds,train_inds)])
k_val_x = get_view_dict(k_x[np.ix_(val_inds,train_inds)])