def least_squares_SGD(y, tx, initial_w, max_iters, gamma):
"""
Linear regression using stochastic gradient descent
Args:
y: labels
tx: features
initial_w: initial weight vector
max_iters: number of steps to run
gamma: step-size
Returns:
w: optimized weight vector for the model
loss: optimized final loss based on mean squared error
"""
threshold = 1e-8
ws = [initial_w]
losses = []
w = initial_w
for _ in range(max_iters):
random_index = np.random.randint(len(y))
# sample a random data point from y vector
y_random = y[random_index]
# sample a random row vector from tx matrix
tx_random = tx[random_index]
error_vector = compute_error_vector(y_random, tx_random, w)
loss = compute_mse(error_vector)
gradient_vector = compute_gradient(tx_random, error_vector)
w = w - gamma * gradient_vector
ws.append(w)
losses.append(loss)
if len(losses) > 1 and np.abs(losses[-1] - losses[-2]) < threshold:
break # convergence criterion met
return ws[-1], losses[-1]
def least_squares_GD(y, tx, initial_w, max_iters, gamma):
"""
Linear regression using gradient descent
Args:
y: labels
tx: features
initial_w: initial weight vector
max_iters: number of steps to run
gamma: step-size
Returns:
w: optimized weight vector for the model
loss: optimized final loss based on mean squared error
"""
threshold = 1e-8
ws = [initial_w]
losses = []
w = initial_w
for _ in range(max_iters):
error_vector = compute_error_vector(y, tx, w)
loss = compute_mse(error_vector)
gradient_vector = compute_gradient(tx, error_vector)
w = w - gamma * gradient_vector
ws.append(w)
losses.append(loss)
if len(losses) > 1 and np.abs(losses[-1] - losses[-2]) < threshold:
break # convergence criterion met
return ws[-1], losses[-1]
def least_squares_SGD(y, tx, initial_w, max_iters, gamma):
"""
Linear regression using stochastic gradient descent
:param y: labels
:param tx: training data
:param initial_w: initial value of weights
:param max_iters: maximum iterations used in gradient descent process
:param gamma: learning rate
:return: optimized loss value based on MSE, optimized weight vectors for the model
"""
threshold = 1e-9
ws = [initial_w]
losses = []
w = initial_w
for _ in range(max_iters):
random_index = np.random.randint(len(y))
y_random = y[random_index]
tx_random = tx[random_index]
error_vector = compute_error(y_random, tx_random, w)
loss = compute_mse(error_vector)
gradient = compute_gradient(tx_random, error_vector)
w = w - gamma * gradient
ws.append(w)
losses.append(loss)
if len(losses) > 1 and np.abs(losses[-1] - losses[-2]) < threshold:
break
return ws[-1], losses[-1]
def least_squares(y, tx):
"""
Least squares regression using the normal equation
:param y: labels
:param tx: training data
:return: optimized loss value based on MSE, optimized weight vectors for the model
"""
coefficient_matrix = tx.T.dot(tx)
constant = tx.T.dot(y)
w = np.linalg.solve(coefficient_matrix, constant)
loss = compute_mse(compute_error(y, tx, w))
return w, loss
def ridge_regression(y, tx, lambda_):
"""
Ridge regression using the normal equation
:param y: labels
:param tx: training data
:param lambda_: regularization parameter
:return: optimized loss value based on MSE, optimized weight vectors for the model
"""
coefficient_matrix = tx.T.dot(tx) + 2 * len(y) * lambda_ * np.identity(
tx.shape[1])
constant_vector = tx.T.dot(y)
w = np.linalg.solve(coefficient_matrix, constant_vector)
error_vector = compute_error(y, tx, w)
loss = compute_mse(error_vector)
return w, loss
def least_squares(y, tx):
"""
Least squares regression using normal equations
Args:
y: labels
tx: features
Returns:
w: optimized weight vector for the model
loss: optimized final loss based on mean squared error
"""
coefficient_matrix = tx.T.dot(tx)
constant_vector = tx.T.dot(y)
w = np.linalg.solve(coefficient_matrix, constant_vector)
error_vector = compute_error_vector(y, tx, w)
loss = compute_mse(error_vector)
return w, loss
def ridge_regression(y, tx, lambda_):
"""
Ridge regression using normal equations
Args:
y: labels
tx: features
lambda_: regularization parameter
Returns:
w: optimized weight vector for the model
loss: optimized final loss based on mean squared error
"""
coefficient_matrix = tx.T.dot(tx) + 2 * len(y) * lambda_ * np.identity(tx.shape[1])
constant_vector = tx.T.dot(y)
w = np.linalg.solve(coefficient_matrix, constant_vector)
error_vector = compute_error_vector(y, tx, w)
loss = compute_mse(error_vector)
return w, loss
def least_squares_GD(y, tx, initial_w, max_iters, gamma):
"""
Linear regression using gradient descent
:param y: labels
:param tx: training data
:param initial_w: initial value of weights
:param max_iters: maximum iterations used in gradient descent process
:param gamma: learning rate
:return: optimized loss value based on MSE, optimized weight vectors for the model
"""
# Define parameters to store w and loss
threshold = 1e-9
ws = [initial_w]
losses = []
w = initial_w
for _ in range(max_iters):
loss = compute_mse(compute_error(y, tx, w))
gradient = compute_gradient(y, tx, w)
w = w - gamma * gradient
ws.append(w)
losses.append(loss)
if len(losses) > 1 and np.abs(losses[-1] - losses[-2]) < threshold:
break
return ws[-1], losses[-1]
