COMP3670/6670 Programming Assignment 3 - Linear
Regression
Enter Your Student ID:
Your Name:
Deadline:
Submit: Write your answers in this file, and submit a single Jupyter Notebook file (.ipynb) on
Wattle. Rename this file with your student number as 'uXXXXXXX.ipynb'. Note: you don't
need to submit the .png or .npy files.
Enter Discussion Partner IDs Below: You could add more IDs with the same markdown
format above. Please implement things by yourself. If you use any external resources, list
them here.
In [ ]:
The following section provides some helper functions.
import numpy as np
import matplotlib.pyplot as plt
In [ ]:
Task 0: Warming Up
The following code block visualises the difference between different methods of performing
linear regression.
## GENERAL FUNCTIONS - DO NOT MODIFY ##
def lr_mle(X, y):
# maximum likelihood (least squares) for linear regression
XtX = np.dot(X.T, X)
Xty = np.dot(X.T, y)
theta = np.linalg.solve(XtX, Xty)
return theta
def lr_map(X, y, alpha=0.1):
# maximum a-posteriori (regularised least squares) for linear regression
N, D = X.shape[0], X.shape[1]
XtX = np.dot(X.T, X) + np.diag(alpha*N*np.ones(D))
Xty = np.dot(X.T, y)
theta = np.linalg.solve(XtX, Xty)
return theta
def lr_bayes(X, y, alpha=0.1, noise_var=0.01):
# exact posterior for Bayesian linear regression
N, D = X.shape[0], X.shape[1]
XtX = np.dot(X.T, X) + np.diag(alpha*N*np.ones(D))
Xty = np.dot(X.T, y)
mean = np.linalg.solve(XtX, Xty)
# note: calling inv directly is not ideal
cov = np.linalg.inv(XtX) * noise_var
return mean, cov
def predict_point(X, theta):
# predict given parameter estimate
return np.dot(X, theta)
def predict_bayes(X, theta_mean, theta_cov):
# predict gien parameter posterior
mean = np.dot(X, theta_mean)
cov = np.dot(X, np.dot(theta_cov, X.T))
return mean, cov
def add_bias_col(x):
# add an all-one column
n = x.shape[0]
return np.hstack([x, np.ones([n, 1])])
## END GENERAL FUNCTIONS ##
第2页 共9页
In [ ]:
Task 1: What makes a good regression?
As can be seen from the visualisation above, the regressed line seems to be far from the
# load data
data = np.loadtxt("./data/ass3_data1_train.txt")
x_train, y_train = data[:, 0][:, None], data[:, 1][:, None]
data = np.loadtxt("./data/ass3_data1_valid.txt")
x_valid, y_valid = data[:, 0][:, None], data[:, 1][:, None]
# some data for visualisation
N_plot = 100
x_plot = np.linspace(-2.5, 2.5, N_plot).reshape([N_plot, 1])
# add one col to the inputs
x_train_with_bias = add_bias_col(x_train)
x_plot_with_bias = add_bias_col(x_plot)
# MLE = least squares
theta_mle = lr_mle(x_train_with_bias, y_train)
f_mle = predict_point(x_plot_with_bias, theta_mle)
# MAP = regularised least squares
alpha = 0.1
theta_map = lr_map(x_train_with_bias, y_train, alpha)
f_map = predict_point(x_plot_with_bias, theta_map)
# exact Bayesian
theta_mean, theta_cov = lr_bayes(x_train_with_bias, y_train, alpha)
f_bayes_mean, f_bayes_cov = predict_bayes(
x_plot_with_bias, theta_mean, theta_cov)
# plot utility
def plot(x, y, x_plot, f_mle, f_map, f_bayes_mean, f_bayes_cov):
# plot utility
plt.figure(figsize=(6, 4))
plt.plot(x, y, '+g', label='train data', ms=12)
if f_mle is not None:
plt.plot(x_plot, f_mle, '-k', label='mle')
if f_map is not None:
plt.plot(x_plot, f_map, '--k', label="map", zorder=10)
if f_bayes_mean is not None:
plt.plot(x_plot, f_bayes_mean, '-r', label="bayes", lw=3)
f_std = np.sqrt(np.diag(f_bayes_cov))
upper = f_bayes_mean[:, 0] + 2*f_std
lower = f_bayes_mean[:, 0] - 2*f_std
plt.fill_between(x_plot[:, 0], upper, lower, color='r', alpha=
plt.legend()
plt.xlabel('x')
plt.ylabel('y')
plt.ylim([-3, 3])
# plot the training data and predictions
plot(x_train, y_train, x_plot, f_mle, f_map, f_bayes_mean, f_bayes_cov
第3页 共9页
datapoints. Are there any ways we can improve the regression?
Task 1.1
Explain why the above linear regression fails.
-----Your answer here-----
Task 1.2
What kind of features would lead to a better result? Why?
-----Your answer here-----
Task1.3
Implement featurise function that takes raw datapoints as the input and output a
reasonable design matrix Φ according to the method you mentioned in Task 1.2.
In [ ]:
Task 2: Estimating noise variance through the marginal
likelihood
One commonly asked question in Bayesian linear regression is how can we define the noise
level of the target. In previous questions, we set the noise variance in lr_bayes to be 0.01
def featurise(x):
# TODO: Try to come up with proper features
features = add_bias_col(x) # change this!
return features
x_train_feat = featurise(x_train)
x_valid_feat = featurise(x_valid)
x_plot_feat = featurise(x_plot)
# repeat but now with features
# MLE
theta_mle = lr_mle(x_train_feat, y_train)
f_mle = predict_point(x_plot_feat, theta_mle)
# MAP
alpha = 0.1
theta_map = lr_map(x_train_feat, y_train, alpha)
f_map = predict_point(x_plot_feat, theta_map)
# exact Bayesian
theta_mean, theta_cov = lr_bayes(x_train_feat, y_train, alpha)
f_bayes_mean, f_bayes_cov = predict_bayes(
x_plot_feat, theta_mean, theta_cov)
plot(x_train, y_train, x_plot, f_mle, f_map, f_bayes_mean, f_bayes_cov
第4页 共9页
- a fixed constant. But intuitively, after we have observed some datapoints, the noise level
can actually be inferred or estimated. This tasks is designed for you to investigate the
marginal likelihood (a.k.a. model evidence) and how we can use this to pick the noise
variance.
Task 2.1
Implement the negative log marginal likelihood, given the noise level of the likelihood,
training inputs and outputs, and the prior variance. We can pick prior_var using the
same procedure, but assume prior_var = 0.5 for this exercise. The form of the
marginal likelihood is provided in Week 7's lecture slides.
In [ ]:
Task 2.2
Select the most appropriate noise level that minimises the negative log marginal likelihood. In
practice, we can do this mimimisation by gradient descent, but for this exercise, we assume
we have access to a predefined set of potential noise levels and just need to pick one.
In [ ]:
Task 2.3
We visualise the predictions using the estimated noise variance, and compare to those when
the noise is very large or very small. Based on these graphs and the negative log marginal
likelihood corresponding to these noise levels, explain why finding a proper noise level of the
likelihood is important.
# 2a
def negative_log_marginal_likelihood(noise_var, x, y, prior_var=0.5):
# TODO: implement this
return 0
# 2.2
# a predefined list
potential_noise_vars = np.logspace(-4, 1.5, 50)
## YOUR CODE HERE! ##
noise_var_estimated = potential_noise_vars[0] # change this!
#####################
第5页 共9页
In [ ]:
-----Your answer here-----
** Task 2.4 - Optional **
The naive implementation of the negative log marginal likelihood above would require the
inverse of an N by N matrix, which is of time complexity . This is computationally
intractable for a large dataset (large N). Can we speed this up?
Θ(?3)
In [ ]:
Task 3: Regularisation
In machine learning, regularisation is an important technique to reduce overfitting.
Regularisation also provides better generalisation in general. This task aims to show how
regularisation affects the parameter estimates.
Task 3.1
Implement L1 , L2 . Both functions take the weight as input, output the regularisation
value and the gradient of the regularisation term (NOT THE GRADIENT OF THE ENTIRE
?
# fit with the estimated noise variance
N = x_train_feat.shape[0]
prior_var = 0.5
alpha = noise_var_estimated / prior_var / N
theta_mean, theta_cov = lr_bayes(x_train_feat, y_train, alpha, noise_var_estimated
f_bayes_mean, f_bayes_cov = predict_bayes(
x_plot_feat, theta_mean, theta_cov)
plot(x_train, y_train, x_plot, None, None, f_bayes_mean, f_bayes_cov)
# fit with a very large noise
noise_var = 5
alpha = noise_var / prior_var / N
theta_mean, theta_cov = lr_bayes(x_train_feat, y_train, alpha, noise_var
f_bayes_mean, f_bayes_cov = predict_bayes(
x_plot_feat, theta_mean, theta_cov)
plot(x_train, y_train, x_plot, None, None, f_bayes_mean, f_bayes_cov)
# fit with a very small noise
noise_var = 0.00001
alpha = noise_var / prior_var / N
theta_mean, theta_cov = lr_bayes(x_train_feat, y_train, alpha, noise_var
f_bayes_mean, f_bayes_cov = predict_bayes(
x_plot_feat, theta_mean, theta_cov)
plot(x_train, y_train, x_plot, None, None, f_bayes_mean, f_bayes_cov)
data = np.loadtxt("./data/ass3_data1_train_large.txt")
x_large, y_large = data[:, 0][:, None], data[:, 1][:, None]
x_large_feat = featurise(x_large)
def negative_log_marginal_likelihood_v2(noise_var, x, y, prior_var=0.5
# TODO: implement this
return 0
第6页 共9页
OBJECTIVE FUNCTION).
In [ ]:
Task 3.2
We now run gradient descent and plot the predictions. Comment on the results.
def L1(theta):
# TODO: implement this
return 0, np.zeros_like(theta) # change this
def L2(theta):
# TODO: implement this
return 0, np.zeros_like(theta) # change this
def data_fit(theta, x, y):
diff = y - np.dot(x, theta) # N x 1
f = np.mean(diff**2) # 1 x 1
df = - 2 * np.dot(diff.T, x).T / x.shape[0]
return f, df
def objective(theta, x, y, alpha, l2=True):
reg_func = L2 if l2 else L1
reg, dreg = reg_func(theta)
fit, dfit = data_fit(theta, x, y)
obj = fit + alpha * reg
dobj = dfit + alpha * dreg
return obj, dobj
第7页 共9页
In [ ]:
-----Your answer here-----
D = x_train_feat.shape[1]
theta_l2_sgd_init = np.random.randn(D, 1)
theta_l2_sgd = theta_l2_sgd_init
no_iters = 2000
learning_rate = 0.1
alpha = 0.1
l2 = True
for i in range(no_iters):
obj, dobj = objective(theta_l2_sgd, x_train_feat, y_train, alpha,
theta_l2_sgd -= learning_rate * dobj
if i % 100 == 0:
print(i, obj)
f_l2_sgd = predict_point(x_plot_feat, theta_l2_sgd)
theta_l1_sgd = theta_l2_sgd_init
l2 = False
for i in range(no_iters):
obj, dobj = objective(theta_l1_sgd, x_train_feat, y_train, alpha,
theta_l1_sgd -= learning_rate * dobj
if i % 100 == 0:
print(i, obj)
f_l1_sgd = predict_point(x_plot_feat, theta_l1_sgd)
# Without any regularisation
theta_noreg_sgd = theta_l2_sgd_init
for i in range(no_iters):
obj, dobj = objective(theta_noreg_sgd, x_train_feat, y_train, 0, l2
theta_noreg_sgd -= learning_rate * dobj
if i % 100 == 0:
print(i, obj)
f_noreg_sgd = predict_point(x_plot_feat, theta_noreg_sgd)
theta_map = lr_map(x_train_feat, y_train, alpha)
f_map = predict_point(x_plot_feat, theta_map)
# plot utility
plt.figure(figsize=(6, 4))
plt.plot(x_train, y_train, '+g', label='train data', ms=12)
plt.plot(x_plot, f_l2_sgd, '-k', lw=3, label='sgd l2')
plt.plot(x_plot, f_l1_sgd, '-r', lw=2, label='sgd l1')
plt.plot(x_plot, f_map, '--o', label="map", zorder=10, lw=1)
plt.plot(x_plot, f_noreg_sgd, '-x', label="no reg", lw=2)
plt.legend()
plt.xlabel('x')
plt.ylabel('y')
plt.ylim([-3, 3])
plt.show()