This is an another one approach to solve the competition from kaggle iMet Collection 2019 - FGVC6.
47th place out of 446 (silver medal) with 0.636 F2 score (top 1 -- 0.672).
pip install -r requirements.txt
First download the train and test data from the competition link and put them into ./data folder.
To train the model run
python ./src/train.py
As this competition is kernel only, this command generates only trained models. It also needed to do inference on best chechpoints and averege the results via simple mean.
Detailed solution see in this presentation.