We design A4NN, a highly efficient, composable workflow that leverages existing NAS and in situ parametric predictions to classify protein conformations from protein diffraction image datasets. We rigorously record neural architecture histories, model states, and metadata to reproduce the search for near-optimal NNs and publish the output results, rather than the software, as a reproducibility artifact in this GitHub repository and our Harvard Dataverse repository (see Quick Links). We demonstrate A4NN’s ability to reduce training time and resource consumption on a real dataset generated by an X-ray Free Electron Laser (XFEL) experiment. We discuss how the workflow can deploy across a broader spectrum of NAS and datasets in our ICPP 2023 paper.
This artifact contains metadata and results from our workflow executions on several protein diffraction datasets for different GPU distributions. Additionally, the artifact contains build information for the Conda environments and Python scripts that reproduce the figures in our paper and allow the user to produce their predictions from the created models. All build information, including Python scripts, XFEL datasets, model architectures, and metadata, are included to reinforce the reproducibility of our results.
What's Changed
- Fixed bug in plots by @patelria007 in #1
- Fixed labeling in script by @patelria007 in #2
- Refactoring figures notebook & added matching script by @patelria007 in #3
- Adjusted colors in graphs by @patelria007 in #4
- Fixed other small details for plots by @patelria007 in #5
- Cleaned up repo by @patelria007 in #6
New Contributors
- @patelria007 made their first contribution in #1
Full Changelog: https://github.com/TauferLab/Reproducibility_A4NN_ICPP23/commits/v1.0
