Speakers
Description
From the Bluesky project, Tiled [1] is a data service that removes several barriers by providing secure, authenticated remote access to data. Tiled abstracts variations in file formats and other data storage details across different beamline instruments, making data analysis and visualization code portable. It enables fast, targeted access to specific data regions and offers search and filtering capabilities over metadata. Designed for a wide range of use cases—from analysis notebooks to web-based visualization applications—Tiled is ideally suited for AI/ML workflows at facilities. It provides a unified interface across datasets with varied formats and access control requirements.
The MLExchange project [2], co-developed by LBNL, ORNL, ANL, BNL, and SLAC, builds machine learning tools that facilitate ML analysis at user facilities, focusing on software that supports beamtime experiments. This framework includes browser-based user interfaces along with ML training and inference workflows.
MLExchange applications were recently enhanced to use Tiled for centralized data management at several workflow stages. This updated framework has been installed locally at the DIAD beamline at Diamond Light Source and tested during a tomography beamtime. As data was acquired, tomographic reconstructions were performed and automatically integrated into MLExchange to be used in the Segmentation Application [3]. Once introduced, Tiled fed the reconstructed frames to a browser-based user application, where users can create segmentation classes using several convenient drawing tools. Classes created in the Segmentation Application were then stored in Tiled and made available to the DLSIA[4] framework, which provides access to various segmentation and denoising neural networks. DLSIA conducted training and segmentation, drawing datasets from Tiled and writing results back to Tiled. These results were then available in the Segmentation Application for users to browse. The MLExchange Segmentation application was well received and will soon be installed and used at the Advanced Light Source during a tomography beamtime, where we will scan the same samples and use the models trained at Diamond to segment data reconstructions at ALS.
[1] Bluesky Collaboration. Tiled https://github.com//bluesky/tiled.
[2] Z. Zhao et al., "MLExchange: A web-based platform enabling exchangeable machine learning workflows for scientific studies," 2022 4th Annual Workshop on Extreme-scale Experiment-in-the-Loop Computing (XLOOP), Dallas, TX, USA, 2022, pp. 10-15, doi: 10.1109/XLOOP56614.2022.00007.
[3] Hao G, Roberts EJ, Chavez T, et al. Deploying Machine Learning Based Segmentation for Scientific Imaging Analysis at Synchrotron Facilities. IS&T Int Symp Electron Imaging. 2023;35:IPAS-290. doi:10.2352/ei.2023.35.9.ipas-290
[4] Roberts, E. J., Chavez, T., Hexemer, A., & Zwart, P. H. (2024). Dlsia: Deep learning for scientific image analysis. Journal of Applied Crystallography, 57(2).
| Abstract publication | I agree that the abstract will be published on the web site |
|---|
