def test_blocks_Fisher_Q():
model, filters_info = build_model()
f_u = build_f_u(model, filters_info)
X_U = np.random.ranf(
(1, 3, 32, 32)) # test only with 1 for memory consumption
query = blocks_Fisher_Q(X_U, f_u)
import pdb
pdb.set_trace()
def test_blocks_Fisher_P():
model, filters_info = build_model()
f_u = build_f_u(model, filters_info)
f_l = build_f_l(model, filters_info)
X_U = np.random.ranf((50, 1, 28, 28))
X_L = np.random.ranf((100, 1, 28, 28))
Y_L = np.asarray([np.random.randint(10) for i in range(100)]).reshape(
(100, 1))
blocks_Fisher_P(X_L, Y_L, X_U, f_l, f_u, 50)
def test_kfac_unlabelled():
X = np.random.ranf((12, 3, 32, 32))
model, filters_info = build_model()
f = build_f_u(model, filters_info)
kfac_unlabelled(X, f)
def test_build_f_u():
model, filters_info = build_model()
f = build_f_u(model, filters_info)
x_value = np.random.ranf((100, 3, 32, 32))
dico = f(x_value)
def test_kfac_query():
model, filters_info = build_model()
f_u = build_f_u(model, filters_info)
X = np.random.ranf((10, 3, 32, 32))
batch_size = len(X)
kfac_query(X, f_u, batch_size)