본문으로 건너뛰기
버전: docs v25.02

GCR Release Note

v3.1.0

May. 10, 2024

New Features

LIME support Graph XAI

- LIME(Local Interpretable Model-agnostic Explanation) algorithm is supported for local Graph XAI - Hyper Parameter Optimization implementation with Optuna

Improvements

- Learning stability secured by wide variety of dnn model options and HPO

Compatibility: ALO v2.3   Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11  


v3.1.0

Scheduled for release on Apr. 25, 2024

New Features

Lightweight Graph XAI

- A novel lightweight and fast Graph XAI algorithm that shows feature importance values in the original feature names

Bug Fixes

- The issues with respect to GCR 3.1.0 early release will be fixed, if any  

Compatibility: ALO v2.3   Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11  


v3.0.0

Apr. 11, 2024

New Features

Lightweight classification/regression DNN model  

- Further light-weighted XAI by using an attention-inspired technique - Cross-validation and early-stopping is supported to enhance learning optimization - Feature Importance is extracted during Inference

Improvements

- Numerical columns with high selectivity are automatically selected for the use of original data in the DNN.

Bug Fixes

Clean code and Refactor

- Erase unnecessary functions - Refactor Graph Embedding pipeline

Compatibility: ALO v2.3 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11


v2.1.0

Mar. 14, 2024

New Features

Global/local Graph XAI

- Supports the graph XAI, i.e., telling which original data columns have how much contribution on the prediction of a downstream task model even if the model is working with the graph embedded vectors. - Both the global XAI (explanation over the entire samples) and local XAI (explanation for a single specific sample) are supported

Improvements - Hyper parameter optimization algorithm has been upgraded with Optuna, one of the fast HPO solutions

Bug Fixes

The GCR installation issue on Mellerikat docker has been fixed

- The root cause was the confliction in the installation order between Mellerikat docker and AI Content; Docker requires AI content installation before pip libraries while AI contents requires pip libaries first. - It was fixed by installing the pip libraries first by the AI Content

The operational bug in the case of CPU-only environment (i.e., no GPU physically) has been fixed

Compatibility: ALO v2.3 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11


v2.0.0

Feb. 16, 2024

New Features

Inductive graph embedding algorithm

- Supports Inductive graph embedding (inference set can be graph-embedded independently from train set without losing consistency in vector space)

Bug Fixes

- The exessive disc usage issue has been fixed - The mismatch bug in data columns in between train and inference pipelines has been fixed

Compatibility: ALO v2.2 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11


v1.3.0

Dec. 29, 2023

Improvements

Supports for GPU

Compatibility: ALO v2.1 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11


v1.1.0

Nov. 20, 2023

Improvements

Execution speed

- Execution speed has been improved by adopting light weight data types

Bug Fixes

- Graph partition bug has been fixed - Output dimension setting bug has been fixed - Column type recognition bug has been fixed  

Compatibility: ALO v2.0 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11


v1.0.0

Oct. 16, 2023

New Features

- Full support for any tabular inputs - Preprocessing is unnecessary, including cases with missing values. Graphs are inherently robust to missing values, ensuring reliable performance even in the presence of incomplete data. - Using graph partitioning, the model can process large datasets without fully loading into memory - Easy of use - Users are accessible to some of the hyper-parameters in the training process, allowing them to flexibly optimize resources and to enhance embedding performance - Relational graph constructor enables formation of graph with hierarchical structure, enhancing performance depending on the contents of the data.

Compatibility: ALO v1.0 Tested on c6i.2xlarge - CPU: 6 / MEM: 14G, p3.2xlarge - CPU: 6 / MEM: 55G / GPU: 1(16G) / CUDA 11