I've used Deep Learning, and particularly fastai, for some time now. Normally I was reluctant to share anything thinking is not going to be of much use, but motivated by this competition, I decided to try and share a notebook in pure Pytorch to show how to implement a Multilabel Stratification and Cross Validation strategy https://www.kaggle.com/ronaldokun/multilabel-stratification-cv-and-ensemble because I thought it was very important to have a competitive edge. I am glad I did that because the notebook proved useful for some, it motivated me to also try to write more pure Pytorch.

I used this competition to learn the latest version of the fastai library, fastai2 (still in development but already ready to use).

fastai is awesome, with a lot of advanced stuff and neat tricks already build in the library, but it can be somewhat of a "black box", hard to inspect and adapt the code because is vastly different than familiar python codebases. That's why I share my entire workflow here in fastai2 so it can be useful to others to get to know the library, use it and adapt to other competitions.

Imports

Statistics of the Dataset.

Many top competitors in Kaggle don't just use the Imagenet Statistics but calculate the Statistics of the current whole dataset ( train + test ) and use those instead. I lack expertise to know which is best but nevertheless I calculated the statistics of the current dataset. The values are already calculated, uncomment the following cell if you want to make the calculation yourself.