Scott Votaw

feat: Remove CNTK functionality and replace with ONNX

49

## What changes are proposed in this pull request? Currently the ImageFeaturizer uses CNTK models. This PR replaces this underlying dependency with ONNX models, effectively removing usage of CNTK from...

feat: Add LightGBM streaming execution mode

40

# Summary Add the streaming execution mode to LightGBM wrapper. This mode uses almost no memory on top of what LightGBM needs to execute. # Tests Tests will be modified...

PushRows APIs are not thread safe for sparse data

1

## Description The API `LGBM_DatasetPushRows `(and CSR version) are not thread safe for sparse data. Internally, there is some thread safety for the OpenMP parallelism, since each thread gets its...

## Summary Improve memory performance of Dataset loading by supporting a more _streaming_-based approach to loading Datasets, as opposed to _bulk_ loading from memory. ## Motivation The LightGBM `Dataset `format...

feature: Add serialization of reference dataset

1

## Summary This is in reference to feature request: https://github.com/microsoft/LightGBM/issues/5426 This PR adds APIs for serializing/deserializing Datasets without their data to a byte array, effectively creating a "schema" or "reference"...

## Summary Create APIs for serialization and deserialization of Dataset definition/schema to a byte array, without the data. It should just be the feature group definitions and overall Dataset properties....