Assignment 4
Due 04/13/2021 midnight EST. No exceptions.
Goal
The idea is to build on the ideas you proposed in your mid term report, while incorporating the midterm feedback. If your project feedback mentioned that the exact problem you are working is not clear in the report, this issue must be resolved for sure by now. We strongly advise you stop by the office hours of one of the teaching staff if this is the case. If your report said you need to start coding ASAP, we would like to see significant progress on your code in comparison to the midterm report.Deliverables
There are three deliverables for this milestone:
(1) Go through the feedback given to you in the mid-term and adapt your project according to the given feedback,
(2) Show significant progress in coding beyond what was submitted in the midterm. One complete experiment needs to be submitted. More details below.
(3) Give a clear account of experiments you will be giving in your final project report and a timeline for how you plan on completing these. While working on this, please try to be cognizant of the feedback given to you in your mid-term report so as to avoid running into those errors. A diagram or flow chart of the timeline would be the best way to approach this deliverable.
More details on expectations from the code:-
In terms of expected coding: you should aim to have one completed experiment: 1. hypothesis,
2. model trained, 3. results reported.
The idea is to have the pipeline setup, so that you can generate more results/test more hypothesis easily as you progress.
As a concrete example: if your project focuses on understanding the role of color in recognition, a good submission would contain the following:-
1. Load a model pre-trained on ImageNet.
2. Convert ImageNet images to gray-scale.
3. Load gray-scale images and run prediction on them given the loaded pre-trained model.
4. Report Results: what is the accuracy on gray-scale images as compared to normal RGB images? What is the performance for different categories? can you think of some interesting things to plot here?
5. Share your code base with us. It's ok if it's not clean, it doesn't need to be documented or in a form that we can run easily. But, we would like to see the code. Sharing as a google colab or as a github repository - both are ok.
For any clarifications, feel free to reach out to the teaching staff or stop by the office hours!
