def test_initial_value_shape():
shapedAns = answer.reshape(-1, 1)
[x, status, hist] = l1ls(A, y, lmbda, x0=shapedAns, tar_gap=rel_tol)
assert_allclose(x, shapedAns, atol=1e-5)
expect(x.shape).to_equal(shapedAns.shape)
expect(hist.shape[0]).to_equal(1)
expect(status).to_equal('Solved')
def RSC_identif(train_set, test_image, mean, reductor, dico_norm, nb_classes):
""" Perform the identification of a test_image thanks to the adapted RSC algorithm """
e = np.array((test_image - mean).astype(float))
norm_y = norm_column(test_image)
test_normalized = normalize_column(test_image)
for j in range(NB_ITER):
delta = f_delta(e)
mu = PARAM_C / delta
before_exp = mu * (e ** 2 - delta)
todiag = to_diag(before_exp)
w_train = normalize_matrix(train_set * todiag[:, np.newaxis])
w_test = normalize_column(todiag * test_image)
w_train_red = dim_reduct(w_train, reductor, DIM_REDUCTION)
w_test_red = dim_reduct(w_test, reductor, DIM_REDUCTION)
D = normalize_matrix(w_train_red)
y = normalize_column(w_test_red)
if REG_METHOD == 'l1':
[x, status, hist] = L.l1ls(D, y, LAMBDA, quiet=True)
else:
x = l2_ls(D, y, LAMBDA)
e = norm_y * (test_normalized - dico_norm.dot(x))
return classif(D, y, x, TRAINING_FACES, nb_classes)
def RSC_identif(train_set, test_image, mean, reductor, dico_norm, nb_classes):
""" Perform the identification of a test_image thanks to the adapted RSC algorithm """
e = np.array((test_image - mean).astype(float))
norm_y = norm_column(test_image)
test_normalized = normalize_column(test_image)
for j in range(NB_ITER):
delta = f_delta(e)
mu = PARAM_C / delta
before_exp = mu * (e**2 - delta)
todiag = to_diag(before_exp)
w_train = normalize_matrix(train_set * todiag[:, np.newaxis])
w_test = normalize_column(todiag * test_image)
w_train_red = dim_reduct(w_train, reductor, DIM_REDUCTION)
w_test_red = dim_reduct(w_test, reductor, DIM_REDUCTION)
D = normalize_matrix(w_train_red)
y = normalize_column(w_test_red)
if REG_METHOD == 'l1':
[x, status, hist] = L.l1ls(D, y, LAMBDA, quiet=True)
else:
x = l2_ls(D, y, LAMBDA)
e = norm_y * (test_normalized - dico_norm.dot(x))
return classif(D, y, x, TRAINING_FACES, nb_classes)
def test_initial_value_shape():
shapedAns = answer.reshape(-1, 1)
[x, status, hist] = l1ls(A, y, lmbda, x0=shapedAns, tar_gap=rel_tol)
assert_allclose(x, shapedAns, atol=1e-5)
expect(x.shape).to_equal(shapedAns.shape)
expect(hist.shape[0]).to_equal(1)
expect(status).to_equal("Solved")
def sift_identif(sift, dico):
print "sift"
e = np.array((sift - mean).astype(float))
norm_y = normColumn(sift)
NTest = normalizeColumn(sift)
for j in range(nbIter):
delta = fdelta(e)
mu = param_c / delta
before_exp = mu * (e**2 - delta)
todiag = toDiag(before_exp)
WTrain = normalizeMatrix(dico * todiag[:, np.newaxis])
WTest = normalizeColumn(todiag * sift)
WTrainRed = dimReduct(WTrain, reductor)
WTestRed = dimReduct(WTest, reductor)
D = normalizeMatrix(WTrainRed)
y = normalizeColumn(WTestRed)
[x, status, hist] = L.l1ls(D, y, lmbda, quiet=True)
#x = l2_ls(D, y, lmbda)
if j == 0:
alpha = x
else:
# eta=find_eta(y,D,alpha,x,mu,delta,nbDim)
alpha = alpha + eta * (x - alpha)
if not (silence):
#debug_alpha(alpha)
pass
e = norm_y * (NTest - dico_norm.dot(alpha))
cl = classif(D, y, alpha, nbLabels)
print cl
return cl
def test_initial_value():
[x, status, hist] = l1ls(A, y, lmbda, x0=answer, tar_gap=rel_tol)
assert_allclose(x, answer, atol=1e-5)
expect(x.shape).to_equal(answer.shape)
expect(hist.shape[0]).to_equal(1)
expect(status).to_equal('Solved')
def test_high_accuracy():
[x, status, hist] = l1ls(A, y, lmbda, tar_gap=rel_tol / 10)
assert_allclose(x, answer_high_accuracy, atol=1e-5)
expect(hist.shape).to_equal((16, 5))
expect(status).to_equal('Solved')
def test_small_example_sparse():
[x, status, hist] = l1ls(S.csr_matrix(A), y, lmbda, tar_gap=rel_tol)
assert_allclose(x, answer, atol=1e-5)
expect(hist.shape).to_equal((12, 5))
expect(status).to_equal('Solved')
def test_small_example_zero():
[x, status, hist] = l1ls(A, np.zeros_like(y), lmbda, tar_gap=rel_tol)
expect(status).to_equal('Failed')
def test_shaped_y():
# Shape of 'y' does not make a difference
[x, status, hist] = l1ls(A, y.reshape(-1, 1), lmbda, tar_gap=rel_tol)
expect(x.shape).to_equal((A.shape[1], ))
expect(status).to_equal('Solved')
def test_initial_value():
[x, status, hist] = l1ls(A, y, lmbda, x0=answer, tar_gap=rel_tol)
assert_allclose(x, answer, atol=1e-5)
expect(x.shape).to_equal(answer.shape)
expect(hist.shape[0]).to_equal(1)
expect(status).to_equal("Solved")
def test_high_accuracy():
[x, status, hist] = l1ls(A, y, lmbda, tar_gap=rel_tol / 10)
assert_allclose(x, answer_high_accuracy, atol=1e-5)
expect(hist.shape).to_equal((16, 5))
expect(status).to_equal("Solved")
def test_small_example_sparse():
[x, status, hist] = l1ls(S.csr_matrix(A), y, lmbda, tar_gap=rel_tol)
assert_allclose(x, answer, atol=1e-5)
expect(hist.shape).to_equal((12, 5))
expect(status).to_equal("Solved")
def test_small_example_zero():
[x, status, hist] = l1ls(A, np.zeros_like(y), lmbda, tar_gap=rel_tol)
expect(status).to_equal("Failed")
def test_shaped_y():
# Shape of 'y' does not make a difference
[x, status, hist] = l1ls(A, y.reshape(-1, 1), lmbda, tar_gap=rel_tol)
expect(x.shape).to_equal((A.shape[1],))
expect(status).to_equal("Solved")
# Reconstruction with L2 (Ridge) penalization
rgr_ridge = Ridge(alpha=0.2)
start1 = timeit.default_timer()
rgr_ridge.fit(proj_operator, proj.ravel())
rec_l2 = rgr_ridge.coef_.reshape(l, l)
stop1 = timeit.default_timer()
print(stop1 - start1)
print('rec', rec_l2.shape)
# Try l1ls
lmbda = 0.1
rel_tol = 0.05
start = timeit.default_timer()
[x, status, hist] = L.l1ls(proj_operator, proj.ravel(), lmbda, tar_gap=rel_tol)
stop = timeit.default_timer()
print(stop - start)
xzzz = x.reshape(l, l)
print(x.shape)
# Hyper Para Tuning
lmbda = 0.01
rel_tol = 0.05
start = timeit.default_timer()
[x2, status2, hist2] = L.l1ls(proj_operator,
proj.ravel(),
lmbda,
tar_gap=rel_tol)
stop = timeit.default_timer()
print(stop - start)
