I first learned about deep learning from the kaggle galaxy classification competition from 5 years ago. I always wanted to try this competition myself, and now with fastai I can!
Here the write-up of my solution on my blog: http://jimypbr.github.io/2019/09/galaxy-zoo-kaggle-competition-redux-with-fast-ai
I managed to get to the top 3 of the leader board scores. I’m a bit miffed I couldn’t beat the top score though. I struggled a lot with underfitting actually - my val score was always lower than the training score. I got over this by just training more and more cycles until the training score was lower than the val score.
I also had to write a custom layer to normalize the output of the network so that the output is a valid probability. I show how to do this in the post.
I’d love any questions and comments on this. Especially regarding my ‘underfitting’ problem! I hope I’m missing something here.