def test02_sgd_sanity_with_epochs(self):
C, W0, X, _, _ = create_C_W_X_d()
optimizer = SGD(batch_size=256, m=X.shape[1])
W = W0.copy()
for epoch in range(15):
W = optimizer.optimize(W,
X,
C,
objective_soft_max,
objective_soft_max_gradient_W,
lr=1)
self.assertTrue(True)
def test01_linear_convergence_gradient_test_X(self):
max_iter = 17
C, W, X, _, d = create_C_W_X_d()
eps0 = 0.5
losses = []
loss_convergence = []
for i in range(max_iter):
eps = eps0 * (0.5**i)
objective1 = round(objective_soft_max(X + eps * d, W, C), 10)
objective2 = round(objective_soft_max(X, W, C), 10)
losses.append(abs(objective1 - objective2))
for i in range(1, len(losses)):
loss_convergence.append(losses[i] / losses[i - 1])
avg_val = average(round(loss_convergence[-5:], 4))
self.assertTrue(0.4 <= avg_val <= 0.6, msg=f'avg value = {avg_val}')
def test00_linear_convergence_gradient_test_W(self):
max_iter = 25
eps0 = 0.5
losses = []
loss_convergence = []
C, W, X, d, _ = create_C_W_X_d()
for i in range(max_iter):
eps = eps0 * (0.5**i)
losses.append(
abs(
objective_soft_max(X, W + eps * d, C) -
objective_soft_max(X, W, C)))
for i in range(1, len(losses)):
loss_convergence.append(losses[i] / losses[i - 1])
avg_val = average(round(loss_convergence[-5:], 4))
self.assertTrue(0.40 <= avg_val <= 0.6, msg=f'avg value = {avg_val}')
def test03_quadratic_convergence_gradient_X(self):
max_iter = 17
eps0 = 0.5
losses = []
loss_convergence = []
C, W, X, _, d = create_C_W_X_d()
for i in range(max_iter):
eps = eps0 * (0.5**i)
objective1 = objective_soft_max(X + eps * d, W, C)
objective2 = objective_soft_max(X, W, C)
objective3 = eps * trace(
d.T @ objective_soft_max_gradient_X(X, W, C))
losses.append(abs(objective1 - objective2 - objective3))
for j in range(1, len(losses)):
loss_convergence.append(losses[j] / losses[j - 1])
avg_val = average(round(loss_convergence[-5:], 4))
self.assertTrue(0.2 <= avg_val <= 0.3, msg=f'ans = {avg_val}')
def test02_objective_soft_max_sanity_test(self):
C, W, X, _, _ = create_C_W_X_d()
c1 = np.round(objective_soft_max(X, W, C), 10)
c2 = np.round(objective_soft_max_old(X, W, C), 10)
self.assertEqual(c1, c2)
def test01_sgd_sanity(self):
C, W0, X, _, _ = create_C_W_X_d()
optimizer = SGD(batch_size=256, m=X.shape[1])
W = optimizer.optimize(W0, X, C, objective_soft_max,
objective_soft_max_gradient_W)
self.assertTrue(True)