This repository contains my code for the assignments in the 'Machine Learning Techniques' course from National Taiwan University on Coursera.
In the libsvm folder, I put two files: svm.h and svm.c. The source of this library can be found here. I always put these two files along with my C++ code files and #include "svm.h" to use the library.
cvxopt: for Quadratic Programming
scipy: a stack of libraries containing numpy, matplotlib, sklearn, ...
/hw1q3/hw1q3.py
sklearn, numpy, cvxopt, matplotlib used
In this question, I implemented hard-margin linear SVM (primal) and kernel SVM (dual) using a QP solver.
The boundary is plotted.
/hw1q15_cpp/: A c++ version of question 15 in function hw1q15(), using LibSVM.
/hw1q15_py/hw1q15.py: A python version of question 15 in function hw1q15, using sklearn.
/hw1q15_cpp/: A c++ version of question 15 in function hw1q16(), using LibSVM.
/hw1q15_py/hw1q15.py: A python version of question 15 in function hw1q16, using sklearn.
Note: in /hw1q15_cpp/, use
make
./hw1q15
will run both question 15 and 16.
/hw1q15/hw1q15.py: these three questions in functions hw1q18, hw1q19, hw1q20.
Note: in /hw1q15_py/, use
python hw1q15.py
will run all questions 15, 16, 18, 19, 20.
/hw2q12/hw2q12.py: an AdaBoost with stumps implemented
numpy, matplotlib used
Note: in /hw2q12/, simply use
python hw2q12.py
will run the AdaBoost and print out the answers to all questions 12 - 18. The script will also draw the decision boundary.
/hw2q19/hw2q19.py: a kernel ridge regression implemented
numpy used
Note: in /hw2q19/, use
python hw2q19.py
will run the regression and print out the answers to questions 19, 20.
/hw3q13/hw3q13.py: solutions to all questions in homework 3 in this file.
numpy, matplotlib used
Note: in /hw3q13/, use
python hw3q13.py
will run all solutions to questions 13 - 20.
hw3q13_14_15(): trains a hand-written decision tree; dumps the branches; prints out E_in and E_out; plots the trained decision boundary.
hw3q16(): trains a lot of decision trees using bagging; computes the average E_in.
note that I only trained 1000 trees instead of 30000, since they're already enough to get a quite stable average E_in
hw3q17_18(): trains a lot of random forests with C&RT's; computes the average E_in and E_out; plots the first trained random forest as an example.
hw3q19_20: trains a lot of random forests with pruned trees; computes the average E_in and E_out; plots the first trained random forest as an example.
hw4q11/hw4q11_handwritten.py: a handwritten neural network (nnets are so complicated!).
Node: in /hw4q11/, use
python hw4q11_handwritten.py
will run all solutions to questions 11 - 14. Be careful, my neural networks are super slow. So, when choosing parameters, you might want to change nexperiments to some smaller value first to eliminate some clearly bad candidates.
hw4q15/hw4q15.py: a handwritten KNN
hw4q19/hw4q19_handwritten.py: a handwritten kmeans
hw4q19/hw4q19_sklearn.py: kmeans using sklearn
Both should be fine.