Congratulations to the top 3 winners on this amazing competition organized by datafest.africa & sunbird.ai

So as to drive learning, can the winners please share their approach to the competion @danmaestro @ASSAZZIN @FADHLOUN. This is also open for all, what worked, what didn't work.

My approach

I didn't really have time on this competition so I couldn't run many experiments. My approach was most basic. Resize signals --> extract mel/mfcc spectrograms --> fit resnet18(bs: 8, epochs:15, One cycle policy schedule).

Extracted spectrograms only had one channel(depth), so I changed the pretrained models input parameters to accept the spec images. Also discovered that this didn't work well for some models like effnetb5, b4, seresnet, probably I had to tune it better or run more experiments.

Things I'd have loved to try

1. EDA - I din't do much of EDA
2. Try more varieties of single models
3. Try to extract spectral features like centroids, bandwidths etc from the spectrograms to include in my solution
4. Increase image channel from 1 to 3 with 2 extra spec information/transforms
5. Augmentations
6. Ensembles

I couldn't run more experiments so please share your approaches, thoughts, what worked for you and what didn't so we can learn. Cheers