Source code for:
- MisConv: Convolutional Neural Networks for Missing Data (WACV 2022)
- Estimating conditional density of missing values using deep Gaussian mixture model (ICONIP 2020)
Python 3.8 or higher is required. Models have been implemented with PyTorch.
To install the requirements, running:
pip install -r requirements.txt
should suffice.
To train the DMFA model, see the script:
python scripts/train_inpainter.py --h
To run classifier / WAE experiments, see the scripts:
python scripts/train_classifier_v2.py --h
python scripts/train_wae_v2.py --h
respectively.
Moreover, in the scripts/
directory we provide the *.sh
scripts which run the model trainings
with the same parameters as used in the paper.
All experiments are runnable on a single Nvidia GPU.
In order to run a classifier / WAE with DMFA, one must train the DMFA model first with the above script.
For some of the inpainters we compare our approach to, additional repositories must be cloned or installed:
- Partial Convolutions: https://github.com/NVIDIA/partialconv
- ACFlow: https://github.com/lupalab/ACFlow
- k-NN: either sklearn (runs very slowly) or cuml + cudf
- classical MFA: https://github.com/mareksmieja/gmm_missing
We provide DMFA training results (among which are JSONs, weights and training arguments) here.
We provide results for following models, trained on complete and incomplete data:
- MNIST - linear heads
- SVHN - fully convolutional
- CIFAR-10 - fully convolutional
- CelebA - fully convolutional, trained on 64x64 images
There are several Jupyter Notebooks in the notebooks directory. They were used for initial experiments with the DMFA models, as well as analysis of the results and calculating metrics reported in the paper.
The notebooks are not guaranteed to run 100% correctly due to the subsequent code refactor.
If you find our work useful, please consider citing us!
@InProceedings{Przewiezlikowski_2022_WACV,
author = {Przewi\k{e}\'zlikowski, Marcin and \'Smieja, Marek and Struski, {\L}ukasz and Tabor, Jacek},
title = {MisConv: Convolutional Neural Networks for Missing Data},
booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
month = {January},
year = {2022},
pages = {2060-2069}
}
@article{Przewiezlikowski_2020,
title={Estimating Conditional Density of Missing Values Using Deep Gaussian Mixture Model},
ISBN={9783030638368},
ISSN={1611-3349},
url={http://dx.doi.org/10.1007/978-3-030-63836-8_19},
DOI={10.1007/978-3-030-63836-8_19},
journal={Lecture Notes in Computer Science},
publisher={Springer International Publishing},
author={Przewięźlikowski, Marcin and Śmieja, Marek and Struski, Łukasz},
year={2020},
pages={220–231}
}