Repository for Code for the Rove-Tree-11:The not-so-Wild Rover. A hierarchically structured image dataset for deep metric learning research paper/dataset
This code is associated with the following paper: https://openaccess.thecvf.com/content/ACCV2022/html/Hunt_Rove-Tree-11_The_not-so-Wild_Rover_A_hierarchically_structured_image_dataset_for_ACCV_2022_paper.html
The Rove-Tree-11 dataset and code is provided to further research into automatic generation of phylogenetic trees (trees that show how closely related species are) from images using deep learning. To do this, segmented images of specimens of 13,887 beetles from 215 different species of Rove Beetles (Family Staphylinidae) are provided taken from the Natural History Museum of Denmark's collections. A current gold standard phylogeny is provided with the dataset.
The latent features are all generated using the forked submodule. If you are installing locally, there is a requirements.txt file you can use with pip. I used a docker container inside vscode. The relevant docker files are provided in the submodule.
The dataset is available at https://erda.ku.dk/archives/118d9022feb67f8eb7f7bc8bce71187f/published-archive.html (DOI: http://doi.org/10.17894/ucph.39619bba-4569-4415-9f25-d6a0ff64f0e3)
Once you have a system with the python requirements installed, you can run the main code via the command line, an example is given below:
python main.py --dataset=rove --suffix=tripD0 --source_path=/home/ngw861/01_abbey_rove/data --seed=0 --bs=112 --samples_per_class=2 --loss=triplet --batch_mining=distance --arch=resnet50_frozen_normalize --augmentation=rove --use_tv_split
If your gpu cannot run such a large batch size (I had issues with this), you can use gradient accumulation, with the gradient_accumulation_steps argument (be mindful it will slightly alter the results even if you are using the same seed):
python main.py --dataset=rove --suffix=tripD0 --source_path=/home/ngw861/01_abbey_rove/data --seed=0 --bs=8 --gradient_accumulation_steps=14 --samples_per_class=2 --loss=triplet --batch_mining=distance --arch=resnet50_frozen_normalize --augmentation=rove --use_tv_split
The runs for the main results provided in the paper are listed in sbatch.sh.
The align scores are generated in two steps:
- Running the latent features through RevBayes, which is a program to use bayesian inference to generate phylogenetic trees from continuous features.
- Checking the resulting tree against the ground truth phylogeny using the align score
This requires revbayes to be installed. I did this on linux using the similarity image provided by revbayes. This requires Singularity to be installed. This can be done on ubuntu with:
sudo apt-get install -y singularity-container
The revbayes singularity image can then be downloaded from: https://revbayes.github.io/download
The alignment score is calculated in the jupyter notebook '00. Run RevBayes for Single Result'. You will need to change the paths to the output folders of your training runs
If you use this code/dataset, please cite:
@InProceedings{Hunt_2022_ACCV,
author = {Hunt, Roberta and Pedersen, Kim Steenstrup},
title = {Rove-Tree-11: The not-so-Wild Rover, A hierarchically structured image dataset for deep metric learning research},
booktitle = {Proceedings of the Asian Conference on Computer Vision (ACCV)},
month = {December},
year = {2022},
pages = {2967-2983}
}
This code is provided with an MIT license (see license file). The original code in the forked submodule (Revisiting_Deep_Metric_Learning_PyTorch) has the same MIT license.
Made with much effort and frustration in d3, based on this example: https://observablehq.com/@d3/tree-of-life
