Making better mistakes

This repository contains the code for the paper:

Making Better Mistakes: Leveraging Class Hierarchies with Deep Networks
Luca Bertinetto*, Romain Mueller*, Konstantinos Tertikas, Sina Samangooei, Nicholas A. Lord*.
IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2020

Resources:

• Latest version on arXiv
• CVPR 2020 open-access
• Short video presentation

If you find our work useful/interesting, please consider citing it as:

@InProceedings{bertinetto2020making,
author = {Bertinetto, Luca and Mueller, Romain and Tertikas, Konstantinos and Samangooei, Sina and Lord, Nicholas A.},
title = {Making Better Mistakes: Leveraging Class Hierarchies With Deep Networks},
booktitle = {The IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2020}
} 

Data preparation

• Download train+val sets of ImageNet-ILSVRC12 and/or train+val sets of iNaturalist'19
• Create the dataset by using the splits described in dataset_splits/splits_tiered.zip and dataset_splits/splits_inat19.zip.
• For our experiments, we resized the images (stretching them) to 224x224. This script is convenient for this purpose.
• Rename data_paths.yml.example and edit it to reflect the paths on your system.

Installation

The training/testing environment can be initialized using conda as:

conda env update -n better-mistakes -f environment.yml
source activate better-mistakes
pip install -e .

Alternatively, we provide a Dockerfile that can be built using:

docker build -t better-mistakes .

Hierarchies

The hierarchies are defined in ./data for the datasets tieredImageNet-H and iNaturalist19. ImageNet-H is also avaialble for future convenience. For each of these datasets we provide the following files:

• dataset_isa.txt, a text file containing all parent -> child relationships in the hierarchy.
• dataset_tree.pkl a pickled nltk.Tree containing the hierarchy.
• dataset_ilsvrc_distances.txt.xz, a compressed text file with every node pair and their ilsvrc distance. Note that the pairs are sorted lexicographically.
• dataset_ilsvrc_distances.pkl.xz, pickled dictionary of the above distances.

Running the code

The experiments of the papers are contained in the experiments/ directory. Inside of your environment (or docker) run for example:

cd experiments
bash crossentropy_inaturalist19.sh

The entry points for the code are all inside of scripts/:

• start_training.py runs training and validation for all the methods (note: the code has been tested on single-gpu mode only)
• plot_tradeoffs.py produces the main plots of the paper given the json files produced by start_training.py
• start_testing.py runs the trained model on the test set for the epochs output by plot_tradeoffs.py (as in experiment_to_best_epoch.json).

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Commercial licenses available upon request.