6.8.0
Release date: 11-15-2024
Automatic Tests for Classification Models and Results Visualization in Generative AI Lab - 6.9
We are excited to introduce Generative AI Lab version 6.9.0, a feature-packed update that enhances your AI-powered project workflows with innovative capabilities and improved functionality. This release introduces new project types, advanced detection features, and model versioning, empowering users with greater flexibility and control over their projects.
With this update, you can streamline tasks such as checkbox identification, handwritten text detection, and model management, ensuring an optimized and user-friendly experience. Let’s dive into the highlights of this release.
Identify and Validate Checkboxes with Precision
Version 6.9.0 introduces a feature, a new project type called Checkbox Detection. With the new update, users are now able to use the model offered by Generative AI Lab to identify checkboxes in the tasks, including the checked and unchecked status in the respective tasks.
This project type can be selected from the Content Type page under Image tab during project configuration. The default model associated with Checkbox Detection is automatically downloaded from the Models Hub page and added to the project configuration.
After the project is configured, users can add relevant tasks and leverage the model to detect checkboxes and their respective checked and unchecked statuses.
This new update integrates seamlessly with the existing workflow, ensuring no changes or disruptions to the current application processes.
Detect and Validate Handwritten Text and Signatures
Version 6.9.0 expands the capabilities of the platform with the addition of Handwritten Text and Signature Detection. This new feature enables the automatic identification and annotation of handwritten text and signatures within documents, powered by a specialized pre-trained model. The new project type can be selected from the Content Type page under Image tab during project configuration. Upon selection, the default model for Handwritten Text and Signature Detection is automatically downloaded from the Models Hub and integrated into the project configuration.
Users can then add relevant tasks to the project and use the model to identify and annotate handwritten content and signatures in documents efficiently.
This feature doesn’t change the existing application workflow, maintaining consistency while enhancing functionality.
Model Versioning with Training
Generative AI Lab 6.9 introduces model versioning for the following project types: Named Entity Recognition (NER), Classification, Assertion, Relation, and Visual NER. In the TRAINING SETTINGS section of the Train page, a toggle labeled Enable Versioning is now available. By default, model versioning is disabled. To enable it, toggle Enable Versioning to on.
When enabled, models are saved with versioned names following the format projecttype_projectname_v1, projecttype_projectname_v2, and so on. If model deployment is enabled after training is complete, the most recently trained model is automatically applied to the project configuration. If model deployment after training is not enabled, the project configuration remains unchanged. All versions of trained models are accessible on the Reuse Resource page, allowing users to browse and select specific model versions for reuse in other projects.
Model versioning is also supported for previously created projects. If versioning is disabled, subsequent training overwrites the most recent model without creating a new version. When re-enabled, versioning resumes from the latest version rather than starting over from v1. This feature simplifies model management by enabling version tracking and reusability, offering seamless integration for new and existing projects.
Note: The Enable Versioning toggle is disabled during training.
Improvements
Enhanced label readability by using high-contrast default colors.
To improve readability and accessibility, randomly assigned colors will now meet a minimum contrast standard by default, ensuring they are easy to read. This eliminates the need for users to adjust colors manually, saving time and enhancing usability.
Improvements to the updater script
- A retry mechanism has been implemented to pull images during updates and installations. Each image will now be attempted up to three times to ensure successful retrieval.
When training or pre-annotation servers are running during an upgrade, a prompt appears: “Do you want to proceed with deleting these pods? (y/N):”
- If “No” is selected, the upgrade process is exited.
- If “Yes” is selected, the running training/pre-annotation pods are deleted, and the upgrade continues.
- If no matching training or pre-annotation pods are running, the message “No preannotator/training pods matching criteria found.” is displayed, and the upgrade proceeds.
- A new “silent” flag can now be passed while executing the upgrade script.
./annotationlab-updater.sh --silent
When “silent” flag is used:
- It automatically deletes running training/pre-annotation servers without prompting.
- It uses the default admin password for the upgrade procedure.
Bug Fixes [@dipendra]
- Prediction results are missing in Section Based Classification Projects when exported tasks are re-imported
Previously, in Section-Based Classification projects, prediction results were missing when exported tasks were re-imported back to Generative AI Lab project. This issue has been resolved. Prediction results now remain intact when tasks are re-imported for SBA-enabled projects, ensuring data consistency.
- Multiple API calls are made during pre-annotation on the task page
In earlier versions, multiple API calls were made during pre-annotation on the Task page for de-identification projects, causing unnecessary random page refreshes. This issue has been resolved. The Task page now operates efficiently without making multiple API calls or unexpected refreshes.
- De-identify dialog box is not updated after the server is ready with the de-identification pipeline
Previously, the de-identification dialog box was not displayed after the de-identification pipeline was deployed and ready for de-identification. This issue has been resolved. The pop-up now auto-refreshes and appears as expected once the de-identification pipeline is ready for deployment.
- Rule for section classifier is set to default value when the user navigates back to Relevant Page
Previously, when users wanted to edit the section configuration and navigated to the Relevant Sections page, the classifier rule for the section classifier would reset to its default value. This issue has been resolved. For SBA-enabled projects, the selected classifier model is now correctly displayed in the dropdown when users return to the page for reconfiguration or confirmation.
- Training status still showing as ‘Running’ when the user checks the training history after aborting the training
Previously, if a user aborted the deletion of a training pod from the cluster page and checked the training history, the status was incorrectly displayed as “Running” and the “Failed” status was shown in green. This issue has been resolved. The training history now accurately reflects the correct training status with the appropriate status color.
- When de-identification pipeline is deployed glove_100 is shown as deployed embeddings
In earlier versions, the default embedding glove_100 was incorrectly displayed as deployed when the de-identification pipeline was deployed, even if no embeddings were available. This issue has been resolved. The correct embedding is now displayed when an embedding is used. An empty embedding is shown when no embedding is used. This ensures an accurate representation of the deployed embeddings in the pipeline.
- “Model type not available “ error is seen frequently while navigating to the Train page
Previously, users frequently encountered the “Model type not available” error when navigating to the Train page. This issue has been resolved, and the error no longer appears while accessing the Train page from different locations within the project.
- Backup fails for Azure Blob Storage
An issue affecting backups to Azure Blob Storage has been resolved. Backups now function correctly, ensuring reliable data storage and recovery.
- Labels with instructions cannot be deleted and added from the project configuration
Previously, labels with instructions could not be added to or deleted from the project configuration. This issue has now been resolved, allowing users to seamlessly add labels with instructions or delete labels that contain instructions as needed.
- Active learning is not triggered for the second trigger If there is only one floating license, and it is being used by another server, the system previously failed to trigger training or wait for the license to become available when the active learning condition was met. This issue has now been resolved. Active learning will now wait for the license to be free before triggering training, ensuring a smoother and more reliable process. Additionally, when the active learning condition is met and the system is waiting for the license, this status will be displayed on the training page.