.net 8 and CUDA 12

Version 4 is binary compatible with 3.x but incorporates major API changes.

Updated to .net 6 and CUDA 11.5

3.0.3

bumped package version numbers

.net 5

updated to .net 5
added user notification interface and cancellation token handling
added index to buffer add method
renamed ColumnType to BrightDataType

SIMD, Self Attention, Memory Fixes

Added SIMD support for basic linear algebra operations
Self attention for Seq2Seq training
Small linear algebra refactoring
Memory usage fixes

3.0.0 Release

3.0 has been significantly refactored from the 2.0 branch and features a simpler architecture and improved performance.

3.0 is not backwards compatible with models from Bright Wire 2 - they will need to be exported into the new data format.

Along with .NET Core support, 3.0 also offers support for CUDA 11.

3.0.0 Beta Release

3.0 has been significantly refactored from the 2.0 branch and features a simpler architecture and improved performance

2.1.1

Enhancements:

• csv parser now parses numbers with commas
• exposed data table adaptor base classes
• exposed vectorisation model from default data table adaptor
• added suport for boolean columns as binary vectorisation columns
• weighted index list to index list conversion
• added naive bayes training from data tables
• added column constraints to data table analysis xml
• vector distances refactor and performance improvements
• added confusion matrix from data table
• added reverse normalise and reverse vectorise for output columns from normalisation model
• matrix columns/rows as vectors
• batch normalisation
• added output nodes and slots so that graphs can output multiple results
• split convolutional stride into x and y components

Bug Fixes:

• fixed filter initialisation when dropout percentage is 0
• bin overflow correction

Cuda Toolkit 10

Breaking change: data table binary format updated (use CreateLegacyDataTable to load data tables from old format)
Breaking change: in place conversion of tensors, matrices and vectors now explicitly prefixed with "Reshape..."
Added: data table frequency analysis
Added: data table normalisation considers columns of type vector
Added: training data error calculation while training
Fixed: kmeans++ initialisation performance improved

Updated: Updated to cuda toolkit 10
Fixed: CUDA performance improvements

2.0.6

performance improvements and bug fixes
.net standard version of brightwire.cuda
added data table summariser and column conversion

2.0.5

additional method to execute sequential neural networks
changed "score" to "error" if a percentage is not being maximised
several bug fixes