def test_gradient_descent_MSE2():
X_data = np.array([1, 2, 3, 4, 5]) # Input x-data
Y_true = 3 * X_data # Actual y-values
# Create initial weights for the data
w = ad.autodiff('w', [1 for i in range(0, len(Y_true))])
f1 = w * X_data # Functional form
# Run MSE-loss gradient descent
g = ad.gradient_descent(f1,
Y_true,
loss='MSE',
beta=0.001,
max_iter=5000,
tol=0.05)
assert g['loss_array'][-1] <= 0.05
def test_gradient_descent_weightname():
x = np.array([[2, 0], [5, 1]]) #Data
w = ad.autodiff('t', [0.6, 0.4]) #Weights
# Set up parameters for gradient descent
max_iter = 40000
beta = 0.00001
f = 3 + w * x / 2.0
y_act = [3, 4]
tol = 0.2
loss = "RMSE"
with pytest.raises(ValueError):
# Run gradient descent
g = ad.gradient_descent(f,
y_act,
beta=beta,
loss=loss,
max_iter=max_iter,
tol=tol)
def test_gradient_descent_MAE():
x = np.array([[5, -2], [3, -4]]) #Data
w = ad.autodiff('w', [3, 0.5]) #Weights
# Set up parameters for gradient descent
max_iter = 10000
beta = 0.1
f = 1 + admath.exp(-1 * w * x)
y_act = [1.0, 1.05]
tol = 1E-4
loss = "MAE"
# Run gradient descent
g = ad.gradient_descent(f,
y_act,
beta=beta,
loss=loss,
max_iter=max_iter,
tol=tol)
# Assert correct values within tolerance
assert (np.absolute(g['f'].val[0] - y_act[0]) +
np.absolute(g['f'].val[1] - y_act[1])) / len(y_act) <= tol
def test_gradient_descent_RMSE():
x = np.array([[2, 0], [5, 1]]) #Data
w = ad.autodiff('w', [0.6, 0.4]) #Weights
# Set up parameters for gradient descent
max_iter = 40000
beta = 0.00001
f = 3 + w * x / 2.0
y_act = [3, 4]
tol = 0.2
loss = "RMSE"
# Run gradient descent
g = ad.gradient_descent(f,
y_act,
beta=beta,
loss=loss,
max_iter=max_iter,
tol=tol)
# Assert correct values within tolerance
assert np.sqrt(((g['f'].val[0] - y_act[0])**2 +
(g['f'].val[1] - y_act[1])**2) / len(y_act)) <= tol
def test_gradient_descent_MSE():
x = np.array([[1, -2, 1], [3, 0, 4]]) #Data
w = ad.autodiff('w', [3, -1, 0]) #Weights
# Set up parameters for gradient descent
max_iter = 5000
beta = 0.005
f = w * x
y_act = [5.5, 9.5]
tol = 1E-8
loss = "MSE"
# Run gradient descent
g = ad.gradient_descent(f,
y_act,
beta=beta,
loss=loss,
max_iter=max_iter,
tol=tol)
# Assert correct values within tolerance
assert ((g['f'].val[0] - y_act[0])**2 +
(g['f'].val[1] - y_act[1])**2) / len(y_act) <= tol