COMP338 – Computer Vision – Assignment 2 
 
o This assignment is worth 15% of the total mark for COMP338 
 
o Students will do the assignment individually. 
 
Submission Instructions 
 
o Send all solutions as a single PDF document containing your answers, results, and 
discussion of the results. Attach the source code for the programming problems as 
separate files (e.g., python or *.ipynb ), and compress the codes into one zip file. 
 
o Each student will make a single submission to the Canvas system. 
 
o The deadline for this assignment 12/12/2024, 5:00pm 
 
o Penalties for late submission apply in accordance with departmental policy as set 
out in the student handbook, which can be found at 
http://intranet.csc.liv.ac.uk/student/msc-handbook.pdf 
 and the University Code of Practice on Assessment, found at 
https://www.liverpool.ac.uk/media/livacuk/tqsd/code-of-practice-onassessment/code_of_practice_on_assessment.pdf
Image Classification with Deep Models 
 
In this project, we will do image classification using the Cifar100 dataset. The lab 
“COMP338_Lab_08_Fashion_MNIST_Classification.ipynb” on Canvas shows the example 
source code, you can refer it. 
 
Tasks: 
1. (30 marks) Design a deep neural network for image classification. 
2. (30 marks) Train and test your network on Cifar100 dataset. 
3. (40 marks) Write a report to clearly explain your network, the intuition behind your design, 
and discussion of your results. 
 
Rules: 
- You can refer to any papers and reuse any source code. However, you should clearly 
cite the references in your report. 
- Use free Google Colab account (https://colab.research.google.com/) for training. The 
maximum training time on a free Google Colab account is 12 hours. 
- You can choose one existing model as your baseline model, then you can add some 
novel modules or training strategies to improve the performance step by step. 
 
Our solution will be evaluated by: 
- Why do you choose this model as baseline, please explain your reasons? (15%) 
- The overall design concept or novelty of your network design (20%). 
- How you can improve the model step by step, and how to verify your concept? (some 
comparisons) (40%). 
- The completeness of your report (25%). 
 
Notes: 
1. We recommend using simpler baseline models to avoid excessively long training times. 
2. Designing and verifying experiments are crucial aspects of this assignment. 
3. Incorporating concepts discussed in our lectures is encouraged. 
4. Please add all your references into the reference section. 
請加QQ：99515681  郵箱：99515681@qq.com   WX：codinghelp