A serial crystallography dataset for developing machine learning models
DiffraNet: A Dataset of Serial Crystallography Diffraction Patterns
DiffraNet is a dataset with over 25,000 labeled serial crystallography diffraction images.
Dataset
DiffraNet is comprised of 512x512 grayscale images, divided into:
- 25,000 synthetic images, labeled as blank, no-crystal, weak, good, or strong.
- 457 real raw images, labeled as diffraction or no-diffraction.
- 457 real preprocessed images, labeled as diffraction or no-diffraction.
Baseline models
Alongside DiffraNet, we provide a suite of baseline models in our GitHub repo:
- GLCM, LBP, and SIFT+BoVW feature extractors.
- RF and SVM classifiers.
- Our novel CNN topology dubbed DeepFreak.
- BOHB and Hyperopt AutoML optimization scripts.
For details on DiffraNet and our baselines models, see our DeepFreak paper.
References
If you use DiffraNet in scientific publications, we would appreciate citations to the following paper:
[link] Artur Souza, Leonardo B. Oliveira, Sabine Hollatz, Matt Feldman, Kunle Olukotun, James M. Holton, Aina E. Cohen, Luigi Nardi. DeepFreak: Learning Crystallography Diffraction Patterns with Automated Machine Learning. In arXiv preprint arXiv:1904.11834, April 2019.
Bibtex entry:
@article{Souza2019deepfreak,
title = {{DeepFreak: Learning Crystallography Diffraction Patterns with Automated Machine Learning}},
author = {Artur Souza and Leonardo B. Oliveira and Sabine Hollatz and Matt Feldman and Kunle Olukotun and James M. Holton and Aina E. Cohen and Luigi Nardi},
journal = {arXiv preprint arXiv:1904.11834},
year = {2019}
}