This is a Pytorch implementation of DELF that is based on official tensorflow implementation. This repository provides clean training codes on The Quick Draw Dataset.
According to the paper, training process is divided into two steps: 1. Train resnet backbone; 2. Train attention layers. You may choose this in settings.py
Dataset should be generated from Simplified Drawing files (.ndjson) of Quick Draw Dataset.
Simply starting training by:
python main.py -a train