Spaces:

competitions

/

aiornot

like 137

Runtime error

Introduction

Congrats to all participants and thanks to Hugging Face and especially to @abhishek for the initiative and for such a fun and well-prepared competition, I can’t imagine a topic more in vogue than this!

Motivation

One of the goals when I entered the competition was to sharpen my CV skills. I was recently in Kaggle Days Paris edition and 2 people from Hydrogen Torch (Philip Singer and Yauhen Babakhin) shared a lot of stuff about their mental process during CV competitions which I wanted to put in practice and interiorize. In fact, I think that Pascal (also from Hydrogen Torch) is our 3rd place. I was afraid when I saw him in the LB but noticing that he was taking part in 2 hard Kaggle competitions at the same time (and probably he couldn’t spend much time on this one) brought me some relief 😅. Anyway, you cand find a lot of these CV tricks on this video: https://www.youtube.com/watch?v=_mzrfMA8Qx4. My solution was inspired on it (and on Hydrogen Torch documentation 😄) so shoutout to them as well.

The solution

The final solution is quite simple, an ensemble (plain average) of 3 models:

All modes were trained with all the available data and each one of them was in turn an average of several random seeds. The public/private split was 50/50, this means that public LB was calculated with 21,721 examples which is higher than the train set (18618), so at some point I started to validate new ideas only on public LB.
Not much diversity in the approach but I didn’t have time to experiment with other backbone families. The augmentation that worked the best in my case was Cutmix by far.

Link to the full code pipeline: https://github.com/delai50/aiornot (EDIT)

PD: I was using for the first time the Sandesh package created by Abhishek to send the results of the experiments to my mobile phone (https://www.youtube.com/watch?v=jDqjSd42024), it is a wonderful tool! I recall going out for dinner with my girlfriend and checking the phone from time to time and she was like "what are you looking at so eagerly?" 😄 .

Feedback

Everything went smooth and clean but maybe displaying a countdown to the end of the competition somewhere could be an improvement. Also, a couple of times I experienced errors when submitting but I think this is something expected in all competition platforms.
I am happy that Hugging Face is taking part in the data science competitions world, and I hope we build a good community around that. See you in the next one!