Introduced a single API call for getting features and labels (get_features_and_labels())) for PyTorch modeling. This call uses user model training parameters to greatly simplify the process of retrieving and splitting data for training, validation, and testing purposes.
Custom data transformation operations can now accept parameters via the No Code client for increased flexibility and reuse of transformation operations.
No-code ML modeling
Train/create API call now supports deploying no-code models. Currently, there is support for two types of clustering: K-Means and Spectral.
Support users’ parameters from the No Code client for these two no-code clustering models, e.g., number of clusters, elbow detection, etc.
Train/create API also supports deploying an LSTM text classification model in a no-code manner while using distributed Horovod.
All configurable parameters related to LSTM text classification are supported via the No Code client.
DaskHub notebooks
Jupyter notebooks and directories are implemented as a virtual filesystem on top of Elasticsearch.
Users may utilize Jupyter notebooks and automatically store their model cells as boss model objects by attaching relevant metadata to the cell containing model code.
Ability to connect models from Jupyter notebooks to the GUI client as well as open relevant notebooks from the GUI.
Jupyter enabled models are notated in the GUI client.
OAuth integration for DaskHub.
Refactored BOSS AI python library
Package updated to use new CAS mechanism.
Package structure (modules, naming conventions, etc.) has been updated for increased convenience and reflects changes in the BOSS AI backend.
Expanded data visualization capabilities
Various regression analysis plots are added to the No Code client. New plots support regression models implemented in PyTorch, TensorFlow, XGBoost, and Scikit-learn.
Filtering on correlation matrix (e.g., by correlation coefficient and/or features) has been added to help data analytics.
2D correlation matrix added, in addition to the current 3D matrix.
Histogram configuration has been added, enabling users to select the number of intervals/buckets to use in the chart.
Expanded federation view in No Code client
New view in client to see the whole federation across the globe.
Ability to view relative information like Virtual Datasets, Sources, and Trainings across federates.
User Experience Improvements in the No Code Client
New Main Top Menu for section and window flow
New Log Panel at bottom of the screen to see logs and open No Code Debug Console
Modelling section panels improved with more reliability.
Added Asset Training functionality to Modelling Section.
Data Pooling added to Data Tables and Modeling lists.
Added helping tooltips for Query Builder save and exit.
Improved information is shown to users on selected nodes in Data Transformation.
Added ability to copy the Model ID from the GUI.
Added ability to preview Model code from the GUI.
Back-end capabilities
Machine learning capabilities
Auto-activation of “eval mode” for PyTorch models, simplifying usage of trained models for inference purposes.
Horovod training failure on limited or insufficient resources are handled dynamically after each failed run where the next run adds 50% more number of Kubernetes pods for training.
Major updates to plotting functionality and introduction of new Plot & Plotter classes within the boss core libraries.
Major updates to confusion matrix capability and introduction of new ConfusionMatrix class within the boss core libraries.
Data loading performance improvement
Data loading and No Code client responsiveness improved due to the use of Elasticsearch for storage and querying.
Oauth2
User Credentials and Registration is now handled outside of the GUI.
ElasticSearch Integration
Support users’ authentication to ElasticSearch via Oauth2
Index-level security is provided by implementing security groups using Keycloak
Feature-level security is also supported via customized query DSL on top of ElasticSearch
Known Issues
Remote-federate plots not currently being collected
PyTorch PredictExplain testing not working properly
Search update not recording the update to queries
BOSS Client dynamic panels break on logout
Bug Fixes
Various No Code code cleanup
Various UI cleanup
Fixed missing tooltips in No Code client
Fixed missing error on Model upload
Fixed endpoint read formats for federates, virtual datasets, sources, plots, and training
Fixed parameters for Tabular, Image, and NLP operations
Fixed user flow for global application logout, close, and exit transitions
Fixed Model download
Minor format update to embeddings for better usage of ElasticSearch.