This is a stripped down repo of my 8th place solution in the 2018 Kaggle competition: Quick, Draw! Doodle Recognition. Blog post/paper pending. Rushed summary drafted upon competition closure here.
This repo does not contain training code. Managing the raw 50 million example training set was non-trivial. I assume that anyone getting value from this code will implement something of their own.
Weights for my best single model: sh download_weights.sh
Download kaggle data (requires cli setup): sh download_test_raw.sh
Requirements:
conda install --file requirements-conda.txt
pip install -r requirements-pip.txt
python test.py