Structural Uncertainty

This repository contains the implementation of our work "Learning Probabilistic Topological Representations Using Discrete Morse Theory", accepted to ICLR 2023 (Spotlight).

Getting started - compile dipha

You only need to run cmake & make once

• (dipha-graph-recon folder)

run the following commands in this folder to build dipha

rm -rf build/ (this removes my build directory)

mkdir build

cd build

cmake ..

make

now that dipha is built, you are ready

Training from scratch:

CREMI

python3 train_model.py  --params params/CREMI_train.json
                        --train_batch 24

ISBI2013

python3 train_model.py  --dataset ISBI2013
                        --params params/ISBI2013_train.json
                        --train_batch 24

DRIVE

python3 train_model.py  --dataset DRIVE 
                        --params params/DRIVE_train.json
                        --train_batch 8

Finetune from baseline:

CREMI

python3 train_model.py  --params params/CREMI_train.json 
                        --train_batch 24
                        --pretrain False
                        --resume baseline

ISBI2013

python3 train_model.py  --dataset ISBI2013 
                        --params params/ISBI2013_train.json 
                        --train_batch 24 
                        --pretrain False 
                        --resume baseline

DRIVE

python3 train_model.py  --dataset DRIVE 
                        --params params/DRIVE_train.json 
                        --train_batch 8 
                        --pretrain False 
                        --resume baseline

Finetune from best:

CREMI

python3 train_model.py  --params params/CREMI_train.json 
                        --train_batch 24 
                        --pretrain False 
                        --resume best

ISBI2013

python3 train_model.py  --dataset ISBI2013 
                        --params params/ISBI2013_train.json 
                        --train_batch 24 
                        --pretrain False 
                        --resume best

DRIVE

python3 train_model.py  --dataset DRIVE 
                        --params params/DRIVE_train.json 
                        --train_batch 8 
                        --pretrain False 
                        --resume best

Validation:

CREMI

python3 infer.py  --params params/CREMI_validation.json

ISBI2013

python3 infer.py --dataset ISBI2013 --params params/ISBI2013_validation.json

DRIVE

python3 infer.py --dataset DRIVE --params params/DRIVE_validation.json

Citation

Please consider citing our paper if you find it useful.

@inproceedings{hu2021topology,
  title={Topology-Aware Segmentation Using Discrete Morse Theory},
  author={Hu, Xiaoling and Wang, Yusu and Fuxin, Li and Samaras, Dimitris and Chen, Chao},
  booktitle={International Conference on Learning Representations},
  year={2021}
}

@inproceedings{hu2023learning,
  title={Learning Probabilistic Topological Representations Using Discrete Morse Theory},
  author={Hu, Xiaoling and Samaras, Dimitris and Chen, Chao},
  booktitle={International Conference on Learning Representations},
  year={2023}
}