Chapter 1 - Exercise

[michaelshekasta]

Hey everyone,

I'm excited to introduce a new exercise for this unit that I believe will be both engaging and beneficial. My concept revolves around the idea of finding a picture within another picture, offering a practical application of the concepts we've been exploring.

I've outlined the details in the draft PR, which you can find here: Link to PR

Looking forward to your feedback and collaboration on this!

Michael

[michaelshekasta]

@bellabf

[bellabf]

Hi, Michael! I think this would be cool but maybe to keep things more cohesive we could do it with [SIFT and feature matching(https://huggingface.co/learn/computer-vision-course/unit1/feature-extraction/feature-matching)] instead of NN? It would be fit better content wise.

[michaelshekasta]

@bellabf
Thank you for your response. It seems there may have been a misunderstanding. To tackle the exercise without ANN, you should use the sift technique.

[bellabf]

@bellabf Thank you for your response. It seems there may have been a misunderstanding. To tackle the exercise without ANN, you should use the sift technique.

Yes, sorry. I read the file quickly and didn't realize you said without NN. I think it might just be easier/clearer to say using the SIFT technique (it leaves less room for guessing).

I would be interested to hear more about other people's perspective on adding exercises that are not graded, cc @mmhamdy @ATaylorAerospace @merveenoyan and also @johko (who might a take a bit longer to reply to this) considering there is also another discussion #295 on adding exercise to another unit 2, It might be worth it to discuss whether to take a broader effort and how they would be structured.

[michaelshekasta]

P.S. Here is the link for checking progress in the audio course:
https://huggingface.co/spaces/MariaK/Check-my-progress-Audio-Course

[bellabf]

This discussion rings a bell. The audio course uses the gradio to check the progress in order to get a certificate of completion. I believe the criteria for the certificate of completion was made by the organization team. In my opinion, it all comes down to what it is easier for the person that has to issue the certificate of completion.

One thing to keep in mind that is that exercises to check understanding, certificate of completion criteria, and leaderboard challenges serve different goals (which would learning, passing, and engagement, respectively). It might be a good idea to reflect about what it is that we want to accomplish by adding exercises to the unit.

[michaelshekasta]

@bellabf I totally agreed!
@mmhamdy @johko @merveenoyan What do you think?

[mmhamdy]

I agree. Leaderboards are more engaging and fun but this requires reevaluating our certification process (as @bellabf said), which I don't actually mind. But @johko and @merve would know better, so let's see what they think about that.

[johko]

To be fair - I don't know that much better 😄 But I do have an opinion on the whole thing.

I do like your proposal @michaelshekasta but it also sounds like a lot of work maintaining and creating an exercise like that for every unit and I'm not sure we can handle that in a good way

I also think it would be a good way to make the certification multi-step, similar to the audio course. Right now I feel like somebody could just take some model, do some fine-tuning, upload it and basically be done without having to look at the course too much.

I think we should come up with some key exercises that are taken into account for the overall certification. I don't know which ones that would be yet and we probably have to involve @merveenoyan in the whole planning, as she already mentioned that she wants to create a leaderboard.

For the other units it might be good enough to have a multiple choice quiz at the end (you know I like quizzes) and they are not that hard to create as you can see in this file: https://github.com/huggingface/audio-transformers-course/blob/main/chapters/en/chapter1/quiz.mdx?plain=1

tl;dr graded exercises yes, but not for every unit

[michaelshekasta]

@johko, thank you for your comment! I believe it's important to keep in mind that utilizing advanced units like VLAM models requires substantial GPU resources, which could pose challenges for students in the course.

I've outlined some assignment directions already:

Unit 1: Identify small images within larger images using SIFT.
Unit 2: Train a neural network on established datasets like MNIST.
As for labs, which serve as valuable exercises, for Unit 2, one idea is to compare a single CNN with numerous kernels against the approach of using two CNNs sequentially with fewer kernels each.