Code for the paper: "A data-centric approach for improving ambiguous labels with combined semi-supervised classification and clustering" by Lars Schmarje, Monty Santarossa, Simon-Martin Schröder, Claudius Zelenka, Rainer Kiko, Jenny Stracke, Nina Volkmann and Reinhard Koch
Datasets available here
The code is extended / forked version of the official FixMatch repository. The full original README is given in deprecated_README.md:
All code execution is designed to be run within docker. A working version of the latest Docker with GPU Support is expected.
create an image with docker file
docker build -t dc3 .
All following commands should be executed within the created docker image. You can create an appropriate container with
docker run -it --rm --gpus all --name dc3 --shm-size 16G -v <DIRECTORY_SOURCECODE>:/src -w /src -v <DIRECTORY_DATASETS>:/data-ssd -v <DIRECTORY_LOGS>:/data1 dc3 bash
You need to update the paths to the three directories to appropriate locations on your device.
You need to download the datasets from Zenodo and extract them into <DIRECTORY_DATASETS>.
You should now have 4 subfolders within the specified directory.
Create a fifth folder fixmatch inside this directory.
You need to import and preprocess these datasets now with ./generate_datasets.sh. Remember to run it within the docker container you created above.
The generated files will be stored under the created folder fixmatch.
By default the GPU with the index 0 is used to run these elements. If you have no GPU or an unsupported device this might lead to errors.
You can run the experiments within the docker container with the commands below
# Before you can start anything you need to export some enviroment variables for the fixmatch repository source code
export ML_DATA=/data-ssd/fixmatch
export PYTHONPATH=$PYTHONPATH:.
# original source code SSL method e.g. mean teacher
# this uses additional weighted cross entropy for unbalanced datasets
python mean_teacher.py --filters 32 --dataset plankton.1@-1-10 --train_dir /data1/ --train_kimg 1024 --wou 0 --wol 0 --lr 0.01 --ceinv_labels 0 --arch resnet --augment d.d.d --combined_output 0 --wf 0 --ws 0 --prior_fuzzy_distribution 0 --prob_gradient 0 --use_loss_rescale 0 --use_soft_prob 0 --use_weighted_xe 1
# dc3 in combination with SSL e.g. mean teacher
# the parameters are explained in `libml/overclustering_training.py`
python mean_teacher.py --filters 32 --dataset plankton.1@-1-10 --train_dir /data1/ --train_kimg 1024 --wou 10 --wol 10 --lr 0.01 --ceinv_labels 1 --arch resnet --augment d.d.d --combined_output 4 --wf 0.1 --ws 0.1 --prob_gradient 0 --use_loss_rescale 0 --use_soft_prob 1 --prior_fuzzy_distribution 0.6 --use_weighted_xe 1
# useable datasets, the last part (.1@-1-10) is a left over of the original fixmatch repository and can be ignored
plankton.1@-1-10 miceBone.1@-1-10 turkey.1@-1-10 cifar10h.1@-1-10
# supported methods, the other models could be added but are currently not modified
ce (supervised baseline), fixmatch, mean_teacher, pseudo_label, pi_model
Additional notes for running the experiments:
- All experiments were run on a RTX 2080 Ti with 12 GB VRAM
- For MiceBone data were we used 4 GPUs and used the following different arguments
--train_kimg 256 --batch 8 --report_kimg 8 --filters 8 - For FixMatch we used for all datasets 6 GPUs in parallel.
- For the method pseudo_label the augmentation is d.d not d.d.d
@inproceedings{schmarje2022dc3,
title={A data-centric approach for improving ambiguous labels with combined semi-supervised classification and clustering},
author={Schmarje, Lars and Santarossa, Monty and Schr{\"o}der, Simon-Martin and Zelenka, Claudius and Kiko, Rainer and Stracke, Jenny and Volkmann, Nina and Koch, Reinhard},
booktitle={European Conference on Computer Vision},
pages={363--380},
year={2022},
organization={Springer}
}