Homework 4 Binary classification (happy/sad) on facial expressions

CSCI 631 Homework 4 
Guidelines
For this assignment, you will perform binary classification (happy/sad) on facial expressions
using a CNN on the same face images we used in the last homework. The main goal is for
you to use a deep learning library of your choice to design and run a 2-layer CNN on the
images. Although the original dataset contained 7 emotional classes, similar to Homework
3, we will only use two of them (happy and sad) for this assignment.
You can download the data from from myCourses Contents→Programming assignments→Homework3
if you did not save a copy from the last homework. Feel free to use any library of your choice
to train your 2-layer CNN on fer3and4train.csv. And as we did before, validate your
network using the dataset fer3and4test.csv.
Feel free to go on the internet to familiarize yourself with one of the popular deep learning
libraries. I would strongly recommend PyTorch as its learning curve is less steep than some
of the others, but it is not a requirement for this assignment. Below is a tutorial that can be
valuable for building your CNN if you are new to deep learning: https://www.analyticsvidhya.
com/blog/2019/10/building-image-classification-models-cnn-pytorch/
Feel free to check out other tutorials or implementations online but ensure you state them in your report.
Requirements
You should perform this assignment using Python along with any deep learning CUDA-based library of your
choice. It is due on Friday December 13th by 11:59pm. You are required to submit your code in a
Jupyter notebook along with a brief report containing short write-ups from the question(s) in the assignment.
Your PDF and IPYB files should be in a zipped file and uploaded to myCourses via Assignments (formerly
known as Dropbox) before the due date.
Your submitted zipped file for this assignment should be named LastnameFirstname hw4.zip and should
contain at least two files - LastnameFirstname hw4.pdf and LastnameFirstname hw4.ipynb. Feel free to
submit any other auxiliary files required to run your code. We should be able to execute your code for the
assignment from your submitted Jupyter notebbok. Include a Readme file if necessary.
Problem 1. Two-layer CNN implementation
Write your own 2-layer CNN to take in the input training data from the file provided in your last homework
3. Once training is completed, test your results on the validation set provided. In your PDF report, please
address out the following:
i State your network parameters, such as your batch size, number of epochs trained for and learning
rate
ii Provide the accuracy scores from (a) training and (b) testing (or really validation) . If they are
significantly different, explain briefly why this is the case.
iii Lastly, compare the numbers you are getting for this exercise with the results from the networks you
wrote previously in Homework 3 on the same exact datasets.
1
CSCI 631 Homework 4 November 22, 2019
Using Google Colab
Google Colab will allow you to use high-end resources including access to a GPU for free, for up to 12 hours,
at which time the system will refresh and you will lose any code or data not stored in another location.
Ensure you save constantly and download your notebook often to avoid loss of data.
To access Colab, follow this:
• Log on to Google with your personal gmail id.
• Go to https://colab.research.google.com/
• From the Menu, select File→Upload notebook and upload Homework4.ipynb
• Next, from the Menu, select Edit→Notebook settings and check that the Runtime type is Python
3 and select Hardware accelerator to GPU.
• Click on the arrow at the left of the window (see red arrow) as shown in the image below, to expand
the tab as shown on the right image.
• Select the sub-menu on the tab named Files (blue arrow on right image); note that this is different
from the high-level option.
• You can drag-and-drop helper.py and the data file to make them accessible to the notebook. Uploading the data file is VERY SLOW; You might want to save it in your google drive and simply
mount when needed.
2