This repository contains implementations of SimSiam and BYOL, which were tested with the quickdraw dataset. To make the data loading easy, we use TFDS. TFDS allows us to load popular datasets as ImageNet or build our own datasets. In the next lines, we provide simple steps to run this code.
We have prepared a subset of QuickDraw with 1100 instances per class. You can download it from here. You should unpack it into $HOME/tensorflow_datasets (by default, python will look the datasets there). However, you can put the data into another folder, but you'll need to specify the data_dir in the tfds.load function in the train_ssl.py file.
In addition, we add configuration files that may facilitate experimentation. In our example we use qd.ini.
python train_ssl.py -config config/qd.ini -model SIMSIAM -gpu 0
python test_by_search.py -config config/qd.ini -model SIMSIAM -gpu 0
For testing, you should indicate the chekpoint in the configuration file (see CKP_FILE). Note that the testing process will create images with the results.