Changelog

All notable changes to AiDotNet are documented here.

[Unreleased]

Added

• Comprehensive documentation site with tutorials
• 30+ sample applications covering all feature categories
• End-to-end applications: ChatbotWithRAG, SpeechAssistant, ImageClassificationWebApp
• Expanded “Why AiDotNet” comparison against TorchSharp, TensorFlow.NET, ML.NET
• Navigation configuration for GitHub Pages
• Community documentation (contributing guide, roadmap)

Changed

• Migrated Vector/Matrix/Tensor data from T[] to Memory for better performance
• Updated GPU training infrastructure for LSTM/GRU, GNN, and activations

Fixed

• Diffusion conv GPU training with auto eigenbasis
• Applied dotnet format across codebase

[0.x.x] - Recent Releases

GPU and Performance

• SIMD and BLAS optimizations for CPU tensor operations
• GPU training infrastructure for LSTM/GRU, GNN, and activation functions
• Memory migration for zero-copy tensor operations
• GPU kernel optimizations for common operations

Neural Networks

• 100+ neural network architectures including:
  • Transformer variants (BERT, GPT, ViT, CLIP)
  • Diffusion models (Stable Diffusion components)
  • Graph Neural Networks (GCN, GAT, GraphSAGE)
  • 3D and NeRF models

Classical ML

• 106+ classical machine learning algorithms
• 28 classification algorithms
• 41 regression algorithms
• 20+ clustering algorithms

Computer Vision

• 50+ computer vision models
• YOLO v8-11 object detection
• DETR and Faster R-CNN
• Mask R-CNN instance segmentation
• SAM (Segment Anything Model)
• OCR models

Audio Processing

• 90+ audio processing models
• Whisper speech recognition
• Text-to-Speech synthesis
• Audio classification
• Music generation

Reinforcement Learning

• 80+ RL agents
• DQN, Double DQN, Dueling DQN, Rainbow
• PPO, A2C, TRPO
• SAC, DDPG, TD3
• Multi-agent systems (MADDPG, QMIX, MAPPO)

RAG Components

• 50+ RAG components
• Sentence transformer embeddings
• In-memory and FAISS vector stores
• Dense, sparse, and hybrid retrievers
• Cross-encoder rerankers

LoRA Fine-tuning

• 37+ LoRA adapters
• QLoRA (4-bit quantized)
• DoRA (Weight-Decomposed)
• AdaLoRA (Adaptive)
• VeRA, LoKr, LoHa variants

Distributed Training

• 10+ distributed strategies
• DDP (Distributed Data Parallel)
• FSDP (Fully Sharded Data Parallel)
• ZeRO optimization (Stage 1/2/3)
• Pipeline and Tensor parallelism

Meta-Learning

• 15+ meta-learning methods
• MAML, Reptile
• Prototypical Networks
• iMAML

Self-Supervised Learning

• 10+ SSL methods
• SimCLR, MoCo
• DINO, MAE
• BYOL, Barlow Twins

Version History

The detailed version history is available on GitHub: https://github.com/ooples/AiDotNet/releases

Versioning

AiDotNet follows Semantic Versioning:

• MAJOR: Breaking API changes
• MINOR: New features, backward-compatible
• PATCH: Bug fixes, backward-compatible

Upgrade Guide

When upgrading to a new major version, check the release notes for:

• Breaking changes and migration paths
• Deprecated features to update
• New features to take advantage of