XNAT ML 1.0.0 Release Notes
The datatype schema definitions have changed in the core ML plugin as of version 1.0.0, from using the "clara:" root to using the "ml:" root. This is a breaking change that will not support usage of data generated on XNAT ML Beta setups. If you wish to migrate data from an ML Beta instance, see Migrating XNAT ML data from Beta to Production
The bulk of the functionality of the XNAT ML Beta release has been retained in the 1.0.0 release. For the most part, this release consists of feature hardening, UI cleanup, and minor compatibility updates to coincide with finalized versions of XNAT 1.8 and dependent plugins. The 1.0.0 release also introduces semantic changes to the naming of core elements, including the datatype XSD namespace, Javascript namespaces and API endpoints.
Changelog
Semantic Changes
- "xnatx-clara-plugin" repository renamed as ml-plugin
- "xnatx-collections-plugin" repository renamed as datasets-plugin
- xsitypes refactored:
- clara:model → ml:model
- clara:trainConfig → ml:trainConfig
- clara:trainSession → ml:trainSession
- dataset resources converted to references
- "/clara" API paths refactored to "/ml" paths
- "Clara" renamed to "ML" or "Machine Learning" in several parts of the plugin UI
- JavaScript namespace "XNAT.plugin.clara" refactored to "XNAT.plugin.ml"
Demo Changes
- The "xnat/clara-train-v3" container image version has been updated to 1.0, and now includes a command for development testing that will launch training without requiring a local GPU setup
Access Changes
- Access to ML Dashboard and Datasets Dashboard restricted to Project Owners and Site Admins
- "View In TensorBoard" functionality disabled by default. This can be enabled in the Plugin Settings UI
Functional Improvements and Feature Hardening
- Improve scalability of datasets
- Improve communication of training container status back to XNAT
- Handle empty values in config for training/validation/test without failing
- Improve XML upload handling and rationalize XML structure of ML datatypes
- Handle dataset creation failures if a selected resource collection is empty
UI Cleanup
- Fix and rationalize links to the Training Dashboard and Datasets Dashboard throughout
- Include in-application help text and links to documentation
- Improve form validation in several places to keep users from falling into blind alleys or frozen screens
- Improve display of edit and report pages for models, configurations, training datasets, and training runs
- Enable "Delete" actions and fix Manage Files display on ML datatype report pages
- Improve error reporting
View JIRA Log
Login Required: https://issues.xnat.org/issues/?filter=17111#
Known Issues with XNAT ML 1.0.0 / XNAT 1.8.0
There are several known issues with the release version of the XNAT ML plugins. These issues are the highest priority.
| Issue | Description | Workaround |
|---|---|---|
| XNAT-6589 | Datasets are not tied to canonical copies of data files; changes to XNAT session metadata can invalidate previously saved datasets | Manual practices in data control can reduce this risk but can't eliminate it |
| XNAT-6622 | Dataset creation can deliver results different from what is predicted by the validation process, when more than one "match" is found | Refine your dataset filter parameters or your data structure to deliver a single matching file |
| XNAT-6657 | Datasets cannot be created using shared data in a project | Assemble all data to be included in a training dataset into a single project, without using sharing |
| XNAT-6680 | Downloaded model zip structure does not match the structure expected by XNAT when uploading models | Unzip and repackage downloaded models before uploading. See Installing a Model |
| There is no warning in the UI when empty dataset results are created | Delete empty datasets and revisit your filter parameters and/or data structure | |
| XNAT-6710 | The XNAT Event Service does not support XNAT ML datatypes (models, configs, training runs), or other "project assets" | None, yet |