I TAed ECE276a in WI21, and I’d like to record this journey.
For the office hour, initially I thought I could spend first few minutes to recap what is covered in the lecture, but later on I found that’s not the best use of time for these reasons:
the instructor already answers a lot of questions for the lecture during the lecture or office hours, I don’t have to repeat them
different students have different questions, since they are at different stages of their assignments. It is thus hard to cover something that will be useful for all the students at that moment.
Therefore, I switched to normal Q&A style. One thing I could improve is to have some sort of queue system so that I could answer the questions on a FIFO basis. However, it depends on how many students how up for one particular office hour, and it will not be useful if only few students are there. Sometimes I told the students to type questions in zoom chat but sometimes people just ask at random order.
Later on we started project 1. It uses python and gradescope’s autograder. The autograder is not easy to use, as the error message is very unintuitive. For instance, at first I got the no module error but it turned out that I need to install some dependencies.
During the first project, I helped to debug students’ issues, and often times it is path issues. Again, autograder cannot output meaningful traceback for students to figure this out, and TAs need to manually run their code one by one, which is very inefficient.
I was also in charge of designing a project for VI SLAM. In order to make the project different from the previous ones, I changed the dataset from KITTI to KITTI 360. It is a very new dataset and there’s very few available resources, so I started from scratch by playing around with the devkit. Here are several video demos:
Afterwards, it turned out that the KITTI 360 dataset suffers the same issue with KITTI, i.e. the groundtruth pose is measuremed from GPS and sometimes not very reliable. For instance, the car’s z position can change a lot in a short time. I should have investigated this more carefully to provide a much cleaner data to the students.
Another issue is how to evaluate the performance of SLAM, one easy way is to use a trajectory with loop closure. The trajectories in KITTI 360 dataset with loop closures are super long and not ideal for a course project, so I did not have the chance to implement this.
At the end of the quarter, we received the evaluation report from some students. Some positive feedbacks are
- One of the best TA’s I’ve had in a long time! Would definitely recommend for the hours he spent answering hundreds of questions on piazza.
- Expertise in the field but here are also constructive feedbacks, such as
- can’t really understand his explanations in OH, piazza. could be faster at grading.
- Professor is great, TA needs improvement. I would like to thank the students who provide the feedbacks. This is my first TA experience and I will strive to do better the next time!