$24
ASSIGNMENT 1 SOLUTION
What should I submit, where should I submit and by when?
Your submission for this assignment will be one PDF (.pdf) le and one python (.py) le. In-structions on how to prepare and submit these les is given below.
Assignment Package:
http://web.cse.iitk.ac.in/users/purushot/courses/ml/2019-20-a/material/assn1.zip
Deadline for all submissions: 07 September, 9:59PM IST
There is no provision for \late submission" for this assignment
1.1 How to submit the PDF le
The PDF le must be submitted using Gradescope in the group submission mode. Note that this means that auditors may not make submissions to this assignment.
Make only one submission per assignment group on Gradescope. Gradescope allows you to submit in groups - please use this feature to make a group submission.
To clarify the above, there are currently 49 assignment groups. We wish to have only 49 submissions on Gradescope, i.e. one submission per assignment group, not one submission per student.
Link all group members in your group submission. If you miss out on a group member while submitting, that member may end up getting a zero since Gradescope will think that person never submitted anything.
You may overwrite your group’s submission (submitting again on Gradescope simply overwrites the old submission) as many times as you want before the deadline.
Do not submit Microsoft Word or text les. Prepare your PDF le using the style le we have provided (instructions on formatting given later).
1.2 How to submit the Python le
Upload the Python le on your IITK (CC or CSE) website (for CC, log on to webhome. cc.iitk.ac.in, for CSE, log on to turing.cse.iitk.ac.in).
Fill in the following Google form to tell us the exact path to the le https://forms.gle/5VTTP5ty3qQ2YMph8
Do not submit Jupyter notebooks or else les in other languages such as C/Matlab/Java. We will use automated scripts to evaluate your code and will not bother to run code in other formats or other languages (we will simply give such submissions a zero).
1
The name of your le should preferably be \submit.py". Submit only one le. We will not download multiple les from your website even if you ask us to. We will download only one le which we expect to be a Python (.py) le (only one .py le will be accepted and not .ipynb or .m or .c or .cc or .java or .zip or .gz etc).
Do not host your Python code submission le on le-sharing services like Dropbox or Google drive. Host it on IITK servers only. We will be using a script to autodownload your submissions and if not careful, Dropbox or Google Drive URLs may send us an HTML page (instead of your submission) when we try to autodownload your le. Thus, it is best to host your code submission le locally within IITK servers.
While lling the form, give the complete URL to the le, not just to the directory that contains that le. The URL should contain the lename as well.
Example of a proper URL: https://web.cse.iitk.ac.in/users/purushot/mlassn1/submit.py
Example of an improper URL ( le name missing): https://web.cse.iitk.ac.in/users/purushot/mlassn1/
Example of an improper URL (incomplete path): https://web.cse.iitk.ac.in/users/purushot/
We will use an automated script to download all your les. If your URL is malformed or incomplete, or if you have hosted the le outside IITK and it is di cult for us to download automatically, then we will not bother to make any e orts to manually locate your le or else contact you for clari cations. We will simply give your group a zero in these cases.
Make sure you ll in the Google form with your le link before the deadline. We will close the form at the deadline.
Make sure that your le is actually available at that link at the time of the deadline. We will run a script to automatically download these les after the deadline is over. If your le is missing, we will simply assign your group zero marks.
We will entertain no submissions or mercy appeals over email, Piazza etc. All submissions must take place before the stipulated deadline over the Gradescope and the Google form. The PDF le must be submitted on Gradescope at or before the deadline and the Python le must be available on the link speci ed on the Google form at or before the deadline.
2
Problem 1.1 (The Squared SVM Solver). Consider the following problem formulation for n binary labelled data points (xi; yi)i=1;:::;n where xi 2 Rd and yi 2 f 1; +1g that uses the squared hinge loss instead of the hinge loss
n
1
2
w2Rd
2 k
k2
i=1
yi
w; x
i
+
w
+ C
(P 1)
arg min
1
2
X
The above optimization problem (P1) can be rewritten as a new optimization problem (P2)
1
n
Xi
2
2
n 2 kwk22
argd
min
+ C
i2
w R
; R
w; xi 1
=1
s.t. yi
i; for all i 2 [n]
(P 2)
We have also provided you with data on a 20 dimensional binary classi cation problem with 20000 data points in the assignment package (in a le called data). You may create (held out/k-fold) validation sets out of this data in whichever way you like.
The following enumerates 7 parts to the question. Parts 1-5 need to be answered in the PDF le. Part 6 needs to be answered in the Python le. Part 7 is a bonus part and needs to be answered in the PDF le itself.
Introduce a dual variable i for each of the n constraints in (P2) and write the expression for the Lagrangian. Remember, the Lagrangian has no max; min terms and is simply a
function of the form L(w; ; ) where = [ 1; : : : ; n] 2 Rn and = [ 1; : : : ; n] 2 Rn. Also take care that in this case, the primal problem (P2) has not only a w variable that
represents the model, but also a variable that encodes the slacks. (5 marks)
Find out the dual problem by eliminating the primal variables w; using the rst order optimality trick i.e. by setting @@wL = 0 and @@L = 0. Write down the dual optimization problem that you get as a result { let us call this problem (D2). Note that (P1), (P2),
(D2) are all the same problem written di erently. (10 marks)
Implement at least 3 methods to solve the problem of learning with the squared hinge loss. Coordinate maximization on D2 (similar to the method we saw in class to solve the dual of the CSVM problem) must be one of the 3 methods you implement. The other 2 can be any one of (a)-(e) methods given below
SGD on P1 (or mini-batch variant)
Coordinate descent on P1
Coordinate minimization on P1
Projected stochastic gradient ascent on D2 (or mini-batch variant)
Coordinate ascent on D2
Coordinate maximization on D2 (compulsory)
Please do not implement gradient descent on P1 or projected gradient ascent on D2. We wish to implement only stochastic or else coordinate methods in this assignment. For each of the methods you implement, tell us the name of the method and write down all details about how you implemented the method. For example, if you used mini-batch SGD, write down the gradient expression and tell us what batch size and step lengths did you choose. For another example, if you used coordinate minimization, write down the formula for the minimizer along a chosen coordinate as well as which method you used to select the next
coordinate along which to minimize. (3 x 3 = 9 marks)
3
4. Also tell us, for each of these 3 methods you implemented, how did you arrive at the best
values for the hyperparameters mentioned above { for example, you might say \We used p
step length at time t to be = t where we checked for = 0:1; 0:2; 0:5; 1; 2; 5 using held out validation and found = 2 to work the best". For another example, you might say, \We tried random and cyclic coordinate selection choices and found cyclic to work best using 5-fold cross validation". Thus, you must tell us among which hyperparameter choices did
you search for the best and how. (3 x 2 = 6 marks)
For your chosen 3 methods, plot the convergence curves o ered by those 3 methods in a manner as we do in lecture notebooks. The x axis in the graph should be time taken and the y axis should be the value of the objective function in (P1) (warning: do not use the objective function in (P2) or (D2) but rather the objective function in (P1) for all 3 methods). Plot all 3 curves on the same graph (not on 3 di erent graphs) so that they can be compared. Include this graph in your PDF le submission as an image. Write down which method you think performs the best.
Warning: this is not a toy 2D dataset like we have in class. Any method (GD/SDCA etc) may take several seconds/minutes to run before giving decent results. You may have to
run these methods for several tens of thousands of iterations as well. (10 marks)
Of the methods you experimented with above, choose the one you think is the best method (i.e. which achieves the smallest objective value of (P1) in the fastest time in the most reliable manner) and submit code for that method in submit.py (we do not need code for the other 2 methods, just the one you thought was best). We will evaluate your method
on a slightly di erent dataset than the one we have given you and check how good is the
method you submitted. (40 marks)
Bonus: Notice that (P2) does not have the positivity constraints i 0 that the CSVM formulation had. Can you show that the inserting these additional constraints into (P2) does not change the optimization problem at all (i.e. the solution remains the same)?
Note that parts 1-5 (and the bonus part 7 if you are attempting it) need to be answered in the PDF le. Part 6 needs to be answered in the Python le. You may want to check the references section of the course website and refer to [DAU] S. 7.7 S. 11.5-11.6, [SSBD] S. 15.3-15.5 for
help in solving parts 1 and 2 of this problem. (80 + bonus marks)
How to Prepare the PDF File
Use the following style le to prepare your report. https://media.neurips.cc/Conferences/NeurIPS2019/Styles/neurips_2019.sty
For an example le and instructions, please refer to the following les https://media.neurips.cc/Conferences/NeurIPS2019/Styles/neurips_2019.tex https://media.neurips.cc/Conferences/NeurIPS2019/Styles/neurips_2019.pdf
You must use the following command in the preamble
\usepackage[preprint]{neurips_2019}
instead of \usepackage[final]{neurips_2019} as the example le currently uses. Use proper LATEX commands to neatly typeset your responses to the various parts of the problem. Use neat math expressions to typeset your derivations. Remember that parts 1-5 (and the bonus part
4
7 if you are attempting it) need to be answered in the PDF le. All plots must be generated electronically - no hand-drawn plots would be accepted. All plots must have axes titles and a legend indicating what the plotted quantities are. Insert the plot into the PDF le using proper LATEX \includegraphics commands.
How to Prepare the Python File
The assignment package contains a skeleton le submit.py which you should ll in with the code of the method you think works best among the methods you tried. You must use this skeleton le to prepare your Python le submission (i.e. do not start writing code from scratch). This is because we will autograde your submitted code and so your code must have its input output behavior in a xed format. Be careful not to change the way the skeleton le accepts input and returns output.
The skeleton code has comments placed to indicate non-editable regions. Do not remove those comments. We know they look ugly but we need them to remain in the code to keep demarcating non-editable regions.
We have provided you with data points in the le data in the assignment package that has 20000 data points, each being 20 dimensional. You may use this as training data in any way to tune your hyperparameters (e.g. step length) by splitting into validation sets in any fashion (e.g. held out, k-fold). You are also free to use any fraction of the training data for validation, etc. but your job is to do really well in terms of minimizing the objective function of (P1) and that too really fast.
The code le you submit should be self contained and should not rely on any other les (e.g. other .py les or else pickled les etc) to work properly. Remember, we will download only one Python (.py) le from your website. This means that you should store any hyperparameters that you learn (e.g. step length etc) inside that one Python le itself using variables/functions.
We will test your submitted code on a secret dataset which would be slightly di erent from the one we have provided you. Thus, your code should not assume it will be given 20000 data points (for example, our secret dataset may have 2000 or even 23000 data points) nor should it assume that each data point has 20 dimensions (for example, our secret dataset may have 15 or even 25 dimensions). However, we can assure you that our secret dataset looks similar to the dataset we have provided you so that hyperparameter choices that seem good to you during validation, should work decently on our secret dataset as well.
Certain portions of the skeleton code have been marked as non-editable. Please do not change these lines of code. Insert your own code within the designated areas only. If you tamper with non-editable code (for example, to make your code seem better or faster than it really is), we may simply refuse to run your code and give you a zero instead (we will inspect each code le manually).
You are allowed to freely de ne new functions, new variables, new classes in inside your submission Python le while not changing the non-editable code.
You are not allowed to use any machine learning libraries such as sklearn, scipy, pandas etc in your submission Python le. Do not use the sys library either. You are expected to implement your method yourself from scratch. You are allowed to use only basic Python
5
libraries such as numpy, time and random which have already been included for you. Do take care to use broadcasted and array operations as much as possible and not rely on loops to do simple things like take dot products, calculate norms etc.
The assignment package also contains a le called eval.py which is an example of the kind of le we will be using to evaluate the code that you submit. Before submitting your code, make sure to run eval.py and con rm that there are no errors etc.
You do not have to submit eval.py to us { we already have it with us. We have given you access to the le eval.py just to show you how we would be evaluating your code.
6
