LECO: Learning with a Ever-Changing Ontology

This repository contains all the image classification code and experiments that appear in our paper for reproducibility.

We provide an environment yml file for conda user at environment.yml. Or else, you may install torch(==1.6.0) from official site.

We provide pretrain.py to save the model initialization file to ensure reproducibility. You may refer to pretrain.sh for examples of how to save an checkpoint from random initialization (used in our paper).

For CIFAR-LECO and iNat-LECO with two TPs, please refer to train.py.

For iNat-LECO with four TPs, please refer to train_for_more_tps.py.

You may visualize the long-tailed distribution of Semi-iNat at SemiInatStats.ipynb.