def test_ridge():
m_ridge = Var(np.ones(X.shape[1]))
b_ridge = Var(0)
earlyStop = 0 #1e-8
forward = False
epochs = 300
lr = 0.001
m_ridge, b_ridge, loss = regression.iterative_regression(
X, y, m_ridge, b_ridge, regression.ridge_loss, lr, epochs, earlyStop,
forward)
clf_l2 = Ridge().fit(X, y)
assert m_ridge.val == approx(clf_l2.coef_, abs=1e-3)
assert b_ridge.val == approx(clf_l2.intercept_, abs=1e-3)
def test_lasso():
m_lasso = Var(np.ones(X.shape[1]))
b_lasso = Var(0)
earlyStop = 0 #1e-8
forward = True
epochs = 300
lr = 0.1
m_lasso, b_lasso, loss = regression.iterative_regression(
X, y, m_lasso, b_lasso, regression.lasso_loss, lr, epochs, earlyStop,
forward)
clf_l1 = Lasso().fit(X, y)
assert m_lasso.val == approx(clf_l1.coef_, abs=1e-3)
assert b_lasso.val == approx(clf_l1.intercept_, abs=1e-3)
def test_polynomial():
X = np.linspace(-10, 10, 20).reshape(-1, 1)
y = X.reshape(-1)**2
X = PolynomialFeatures(2, include_bias=False).fit_transform(X)
m = Var(np.random.rand(X.shape[1]), seed=1.0)
b = Var(0)
earlyStop = 0
forward = False
m, b, loss = regression.iterative_regression(X, y, m, b, regression.MSE,
0.0001, 10000, earlyStop,
forward)
predict = np.sum(m.val * X, axis=1) + b.val
assert predict == approx(y, abs=1e-2)
def test_elastic():
m_el = Var(np.ones(X.shape[1]))
b_el = Var(0)
earlyStop = 0 #1e-8
forward = False
epochs = 300
lr = 0.1
m_el, b_el, loss = regression.iterative_regression(X, y, m_el, b_el,
regression.elastic_loss,
lr, epochs, earlyStop,
forward)
clf_el = ElasticNet().fit(X, y)
assert m_el.val == approx(clf_el.coef_, abs=1e-3)
assert b_el.val == approx(clf_el.intercept_, abs=1e-3)
def test_store():
m = Var(np.ones(X.shape[1]))
b = Var(0)
earlyStop = 1e-8
forward = True
history = {}
m, b, loss = regression.iterative_regression(X, y, m, b, regression.MSE,
0.1, 100, earlyStop, forward,
history)
clf = LinearRegression().fit(X, y)
assert m.val == approx(clf.coef_, abs=1e-3)
assert b.val == approx(clf.intercept_, abs=1e-3)
assert history['m'][-1] == m.val
assert history['b'][-1] == b.val
assert history['loss'][-1] == loss.val