def alternating_from_data(cluster_num,
x_train,
y_train,
alpha_v,
epochs=100,
index_init=None):
n, tmax = np.shape(x_train)
assert (np.shape(x_train) == np.shape(y_train))
D0 = np.matmul(x_train, x_train.T) / tmax
D1 = np.matmul(x_train, y_train.T) / tmax
def _func1(v, ind):
result = 0
_z = get_index_matrix(cluster_num, ind)
for j in range(cluster_num):
result += np.dot(v[j, :], np.dot(D1, _z[:, j]))
return result
def _func0(ind):
result = 0
_z = get_index_matrix(cluster_num, ind)
for j in range(cluster_num):
result += np.dot(_z[:, j], np.dot(D0, _z[:, j]))
return result
if index_init is None:
res = kmeans_greedy.kmeans_greedy(_func0,
cluster_num,
n,
iter_limit=100)
ind_start = res.index
else:
ind_start = index_init
for e in range(epochs):
if e == 0:
ind_est = ind_start
nochange = False
else:
#print("K={}: epochs {}/{}".format(cluster_num, e, epochs))
res = kmeans_greedy.kmeans_greedy(lambda ind: _func1(v_est, ind),
cluster_num,
n,
iter_limit=50)
ind_est = res.index
nochange = res.nochange
z_est = get_index_matrix(cluster_num, ind_est)
v_est, loss = v_step_from_data(x_train, y_train, alpha_v, z_est)
if nochange:
break
theta_est = np.matmul(z_est, v_est)
return _ResultInstance(theta_est, z_est, v_est, ind_est, loss)
def alternating(cluster_num, D0, D1, alpha_v, epochs=10, index_init=None):
n, _ = np.shape(D0)
assert (np.shape(D0) == (n, n))
assert (np.shape(D1) == (n, n))
if index_init is None:
mat = np.matmul(random_basis(n, cluster_num).T, D1)
res = kmeans_greedy.kmeans_greedy(
lambda ind: np.linalg.norm(
np.matmul(mat, get_index_matrix(cluster_num, ind)), ord='nuc'),
cluster_num, n)
ind_start = res.index
else:
ind_start = index_init
for e in range(epochs):
if e == 0:
ind_est = ind_start
z_est = get_index_matrix(cluster_num, ind_start)
else:
z_est, ind_est, _ = z_step(cluster_num, D1, v_est, ind_old=ind_est)
v_est, loss = v_step(D0, D1, alpha_v, z_est)
theta_est = np.matmul(z_est, v_est)
#print("loss : {}".format(loss))
return _ResultInstance(theta_est, z_est, v_est, ind_est, loss)
def z_step(cluster_num, D1, v, ind_old=None):
k, n = v.shape
assert (np.shape(D1) == (n, n))
assert (k <= n)
mat = np.matmul(v, D1)
#check this stupid function!!! <- apparently works...
res = kmeans_greedy.kmeans_greedy(lambda ind: np.matrix.trace(
np.matmul(mat, get_index_matrix(cluster_num, ind))),
cluster_num,
n,
init_index=ind_old)
ind_mat = get_index_matrix(cluster_num, res.index)
return ind_mat, res.index, .0
def direct_from_data(cluster_num, x_train, y_train, alpha_v, index_init=None):
n, tmax = np.shape(x_train)
assert (np.shape(x_train) == np.shape(y_train))
func = lambda ind: -v_step_from_data(x_train, y_train, alpha_v,
get_index_matrix(cluster_num, ind))[
1] # loss
res = kmeans_greedy.kmeans_greedy(func,
cluster_num,
n,
init_index=index_init)
z_est = get_index_matrix(cluster_num, res.index)
v_est, loss = v_step(x_train, y_train, alpha_v, z_est)
theta_est = np.matmul(z_est, v_est)
# print("loss : {}".format(loss))
return _ResultInstance(theta_est, z_est, v_est, res.index, loss)
def direct(cluster_num, D0, D1, alpha_v, index_init=None):
n, _ = np.shape(D0)
assert (np.shape(D0) == (n, n))
assert (np.shape(D1) == (n, n))
func = lambda ind: -v_step(D0, D1, alpha_v,
get_index_matrix(cluster_num, ind))[1] #loss
res = kmeans_greedy.kmeans_greedy(func,
cluster_num,
n,
init_index=index_init)
z_est = get_index_matrix(cluster_num, res.index)
v_est, loss = v_step(D0, D1, alpha_v, z_est)
theta_est = np.matmul(z_est, v_est)
#print("loss : {}".format(loss))
return _ResultInstance(theta_est, z_est, v_est, res.index, loss)