Description
In this programming assignment, you will implement techniques to learn a multi-class logistic regressor
for image digit classification on the MNIST dataset (https://en.wikipedia.org/wiki/MNIST_database).
The classifier will take an image of a hand-written numerical digit between 1 and 5 as input and classify
it into one of 5 classes corresponding to digit 1 to 5 respectively. Given the training data, you will follow
the equations in the lecture notes to train a multi-class logistic regressor using the gradient descent
method and then evaluate its performance on the given testing data. For training, you will learn the
discriminant function regression parameters Θ=(Wk,Wk,0)
t
, where k=1,2,3, 4, 5, Wk is a vector for kth
discriminant function, and Wk,0 is its bias.
Specifically, given the training data D={x[m], y[m]}, m=1,2, …, M, where x[m] is a grayscale image of
28×28 (a 784×1 vector) and y[m] is output vector that follows the 1 of K encoding, with kth element of
being 1 and the rest being 0. You will implement the following methods to learn the parameters Θ
1. For students taking this class at 4000 level, implement in Tensorflow the gradient descent method
to solve for Θ iteratively using all data in D. Initialize Θ to small values and literately update
Θ with appropriate learning rate until convergence. Save Θ and plot the (Wk) for each class
using matplotlib.pyplot functions (see below for details).
2. For students taking this class at 6000 level, implement in Tensorflow the Stochastic Gradient
Descent method with L2 regularization. Initialize Θ to small values and literately update Θ with
appropriate learning rate and regularization parameter λ until convergence. Save Θ and plot the
(Wk) for each class using matplotlib.pyplot functions (see below for details).
3. Using the given testing dataset T, evaluate the performance of your trained classifier by
computing the classification error for each digit as well as the average classification errors
for all five digits. The classification error for each digit is computed as the ratio of incorrect
classification to the total number images for that digit
Do not use Tensorflow’s existing gradient descent and stochastic gradient descent functions. Submit your
Tensorflow code via LMS, along with the required classification errors, weight plots, and the saved
weights W .
Data format
The training data D contains 25112 training images (00001.jpg – 25112.jpg) in train_data folder, and
their labels are stored in train_label.txt with each raw being the label of the corresponding image.
The testing data T contains 4982 testing images (00000.jpg – 04982.jpg) in test_data folder, and their
labels are stored in test_label.txt.
Load images and normalize
Load the images in the order
Convert it to vector and normalize it to [0,1] by dividing the intensity of each pixel by 255
The following link teaches how to load images:
http://learningtensorflow.com/lesson3/
Output
Use the code below to save the learned parameters matrix W in the following format
⎥
⎦
⎤ ⎢
⎣
⎡ = 1,0 2,0 3,0 4,0 5,0
1 2 3 4 5
W W W W W
W W W W W
W
Wk is the learnt weights for the kth digit and Wk,0 is the corresponding bias.
——————–
import pickle
filehandler = open(“multiclass_parameters.txt”,”wb”)
pickle.dump(W, filehandler)
filehandler.close()
—————–
Plot
Use the following function to plot Wk as an image for each digit:
—————–
import matplotlib.pyplot as plt
Img = W_k.reshape(28,28)
plt.imshow(img)
plt.colorbar()
plt.show()
—————–