Projects

Overview: This repository contains implementations of various online continual learning algorithms and tricks for computer vision tasks, including state-of-the-art methods like ASER, SCR, and survey. It also includes datasets and requirements for running experiments. Key features: - ASER: Adversarial Shapley Value Experience Replay - EWC++: Efficient and online version of Elastic Weight Consolidation - iCaRL: Incremental Classifier and Representation Learning - LwF: Learning without forgetting - AGEM: Averaged Gradient Episodic Memory - ER: Experience Replay - MIR: Maximally Interfered Retrieval - GSS: Gradient-Based Sample Selection - GDumb: Greedy Sampler and Dumb Learner - CN-DPM: Continual Neural Dirichlet Process Mixture - SCR: Supervised Contrastive Replay Tech stack: Python, PyTorch, PyTorch Vision, Scikit-learn, and YAML. Why it matters: This repository provides a comprehensive collection of online continual learning algorithms and tricks, making it a valuable resource for researchers and practitioners in the field of computer vision.

Overview: 深度学习/人工智能/机器学习资料汇总，持续更新。涵盖深度学习、神经网络、Python开发、爬虫、数据科学等多个领域，提供最新课程、论文、代码和实例练习。 Key features: - 深度学习/机器学习入门指南 - 神经网络与深度学习速查表 - Python基础环境与开发语言教程 - 爬虫框架与实例练习 - 机器学习算法概述 - 在线课程推荐 Tech stack: Python, TensorFlow, PyTorch, Scrapy, Libra Why it matters: 提供全面的深度学习资源，帮助开发者快速入门和提升技能，解决人工智能和机器学习领域的高门槛问题。

### Overview: OpenDWM is an open-source initiative focused on autonomous driving video generation. The project provides a high-quality, controllable tool for generating autonomous driving videos using the latest technology. It aims to build a codebase that is both user-friendly and highly reusable, with the community's collective wisdom continuously improving the project. ### Key Features: - **Transparent and Reproducible Training:** Complete training codes and configurations are provided, allowing anyone to reproduce experiments, fine-tune on their own data, and customize development features as needed. - **Significant Improvement in Environmental Diversity:** Enhanced generalization ability through the use of multiple datasets, enabling generation of diverse scenarios such as snowy city streets or lakeside highways with distant snow mountains. - **Improved Generation Quality:** Support for popular model architectures (SD 2.1, 3.5) and various training techniques (multitasking, self-supervision) allow the model to utilize information effectively, improving generation quality. - **Convenient Evaluation:** Evaluation follows the popular framework `torchmetrics`, with public configurations (e.g., FID, FVD on the nuScenes validation set) provided to align with other research works. ### Tech Stack: Python ### Why It Matters: OpenDWM provides a high-quality, controllable tool for generating autonomous driving videos using the latest technology, aiming to build a codebase that is both user-friendly and highly reusable. The project's continuous improvement through the community's collective wisdom makes it an essential resource for researchers and developers in the field of autonomous driving.

Overview: The kalman filter library is a powerful tool for optimization problems, particularly in visual odometry, sensor fusion localization, and SLAM. It provides highly accurate results, operates both online and offline, is computationally efficient, and is easy to design with Python. Key features: - Symbolic computation of Jacobians for Extended Kalman Filters - Support for Error State Kalman Filters for 3D orientation estimation - Multi-State Constraint Kalman Filters for integrating feature-based visual odometry Tech stack: Python Why it matters: This library simplifies the implementation of complex Kalman filters by automating the computation of Jacobians and handling the complexities of non-linear systems, making it an essential tool for developers working on optimization problems in robotics and computer vision.

Overview: This project, "Selective Context for LLMs," compresses prompts and contexts to enhance the processing efficiency and performance of large language models like ChatGPT. Key features: - Efficient context management - Evaluates lexical units for informativeness - Extensive evaluations on arxiv, BBC news, and conversation transcripts Tech stack: Python Why it matters: This project optimizes the performance of LLMs by maximizing the utility of fixed context length, allowing them to handle longer documents and conversations more effectively. Repository: liyucheng09/Selective_Context GitHub stars: 426

### Overview: Ivy is an open-source video-based object counting software designed to tally any object in a video. It supports various detectors like YOLO, Tensorflow Object Detection API, Detectron2, and Haar cascades, making it versatile for different use cases. ### Key Features: - **Supports Multiple Detectors**: YOLO, Tensorflow Object Detection API, Detectron2, and Haar cascades. - **Python 3 Compatible**: Requires Python 3.7. - **Easy Setup**: Clone the repository, create a virtual environment, and install dependencies. - **Demo Data**: Included in the repository for easy setup and testing. - **Run**: Simple command to start the application. - **Debug Mode**: Provides a window to monitor and debug the object counting process. ### Tech Stack: Python ### Why it Matters: Ivy is essential for anyone looking to automate the counting of objects in videos, making it a valuable tool for traffic monitoring, surveillance, and more. ### Repository: nicholaskajoh/ivy ### Primary Language: Python ### GitHub Stars: 426 ### GitHub Description: Video-based object counting software. ### README Excerpt: Ivy is an open-source video-based object counting software for tallying pretty much anything (vehicles, people, animals — you name it).

### Overview: RuoYi-AI User Frontend is the user-facing component of the RuoYi-AI platform, designed for AI conversations, agent interactions, and knowledge-base Q&A. It provides a seamless interface for users to interact with AI-powered applications. ### Key Features: - **Vue 3 + TypeScript**: The frontend is built using Vue 3 with TypeScript, leveraging TypeScript for type safety and Vue 3 for modern web development practices. - **Ant Design Vue**: Utilizes Ant Design Vue for a clean and intuitive user interface. - **Pinia**: State management is handled by Pinia for efficient and reactive data management. - **Docker Deployment**: Supports two deployment methods: one-command deployment for a quick start and separate deployment for custom builds. ### Tech Stack: - **Framework**: Vue 3 + TypeScript - **UI Components**: Ant Design Vue - **State Management**: Pinia - **Build Tool**: Vite ### Why it Matters: RuoYi-AI User Frontend is a crucial component of the RuoYi-AI platform, enabling users to interact with AI-powered applications seamlessly. The Vue-based frontend, coupled with Ant Design Vue for a clean interface, ensures a user-friendly experience. The integration with Pinia for state management and Docker deployment options make it a robust and scalable solution for enterprise-grade AI platforms. ### GitHub Stars: 425 ### Repository: ageerle/ruoyi-web

### Overview: `flame` is a minimal and efficient framework built on `torchtitan` for language models with blazing efficiency. It simplifies the training process, making it easy to integrate with `fla` and `transformers`, and supports zero-cost data preprocessing, including online tokenization, dataset shuffling, and multiple datasets support. ### Key Features: - 🚀 Minimal, easy-to-use, extensible training framework - 🤗 Seamless integration with `fla` and `transformers` - 🔄 Zero-cost data preprocessing: online tokenization, dataset shuffling, and multiple datasets support - 🔮 4D parallelism (coming soon) ### Tech Stack: Python ### Why it Matters: `flame` is designed to be a minimal and efficient framework for language models, making it easier to scale and integrate with other tools. Its zero-cost data preprocessing capabilities, including online tokenization, dataset shuffling, and support for multiple datasets, make it highly versatile and useful for various language model training tasks.

### Overview: AGI-Papers is a curated archive of breakthroughs in Agents, Architecture, Training, RAG, and On-Device AI, focusing on the journey towards Artificial General Intelligence (AGI) in 2026. ### Key Features: - **Agents**: Reviews of self-driving agents, planning models, and multi-agent architectures. - **Architecture**: Innovations in LLM architectures (Transformer, Mamba, MoE). - **Pre-Training**: Overview of training data, scaling laws, and foundational models. - **Post-Training**: Discussion on RLHF, DPO, GRPO, and alignment. - **Evaluation**: Benchmarks, evaluation methods, and comparisons. - **RAG & Knowledge**: Search-enhanced generation, knowledge graphs, and memory. - **On-Device AI**: Local deployment, edge computing, and optimization. - **Projects**: Directly implemented projects and experimental results. - **Trends & Industry**: Industry trends, insights, and major news in AI. ### Tech Stack: Unknown ### Why it Matters: This repository aims to provide a comprehensive overview of groundbreaking research in AGI, focusing on key areas such as agents, architecture, training, and on-device AI. It serves as a valuable resource for researchers, developers, and enthusiasts interested in advancing the field of artificial general intelligence.

Overview: The Link Grammar Parser is a powerful tool for natural language parsing, offering a detailed analysis of sentence structure and providing various parsing formats. Key features: - Supports English, Thai, Russian, Arabic, and limited subsets of Persian and other languages. - Offers a graph-based representation of sentence structure, providing fine-grained and detailed information. - Allows conversion between conventional HPSG and dependency style parses. - Dramatically improved parse coverage of English, and additional languages like Thai and Russian. Tech stack: C Why it matters: The Link Grammar Parser is a significant advancement in natural language parsing, offering a more detailed and nuanced analysis of sentence structure compared to traditional parsers. Repository: opencog/link-grammar GitHub stars: 425 GitHub description: The CMU Link Grammar natural language parser README excerpt: The Link Grammar Parser exhibits the linguistic structure of English, Thai, Russian, Arabic, Persian, and limited subsets of a half-dozen other languages. This structure is a graph of typed links between the words in a sentence. One may obtain the more conventional HPSG (constituent) and dependency style parses by applying a collection of rules to convert to these different formats. The theory of Link Grammar parsing was originally developed in 1991 by Davy Temperley, John Lafferty, and Daniel Sleator, at the time professors of linguistics and computer science at Carnegie Mellon University. The current package has evolved significantly since its initial version, with improvements in performance, multi-threading, UTF-8 support, and enhanced parsing coverage. It

Overview: MCP Local RAG is a local-first RAG server designed for developers, providing semantic and keyword search capabilities for code and technical documents without sending them to an embedding API. It works with MCP or CLI and is fully private and zero setup. Key features: - Semantic search - Keyword search - Indexes PDF, DOCX, Markdown, and text files - Combines semantic similarity with keyword matching for precise results - Supports MCP or CLI clients - Works offline after initial model download - No API key, Docker, Python, or external database required Tech stack: TypeScript Why it matters: MCP Local RAG simplifies access to private documents, making it easier for developers to find relevant information without exposing sensitive data to external APIs.

Overview: This repository contains a practical implementation of the Linformer model, a novel attention mechanism that achieves linear complexity in sequence lengths. It is designed to be used with PyTorch and is currently validated on the WMT2019 English dataset. Key features: - Linformer model implementation - Supports attention heads with linear complexity - Includes an encoder and decoder module - Visualization of attention heads possible - Validated on wikitext-2 - Supports both causal and non-causal configurations Tech stack: Python Why it matters: Linformer is a significant advancement in attention mechanisms, offering linear complexity for sequence lengths, making it highly efficient for processing long sequences. Its implementation in PyTorch makes it accessible for researchers and developers looking to experiment with attention models. GitHub stars: 425 Repository: tatp22/linformer-pytorch

### Overview: This repository contains the official implementation of our paper "se(3)-TrackNet: Data-driven 6D Pose Tracking by Calibrating Image Residuals in Synthetic Domains" presented at IROS 2020. The paper proposes a novel approach for long-term, 6D pose tracking, specifically designed for synthetic domains. The key contributions include a neural network architecture that disentangles feature encoding and an effective 3D orientation representation using Lie Algebra, making the model robust even when trained with synthetic data. ### Key Features: - Data-driven 6D Pose Tracking - Synthetic Domain Training - Neural Network Architecture - 3D Orientation Representation - Consistent Robust Estimates - High Tracking Frequency (90.9Hz) - Applications: Model-based RL, Manipulation, AR/VR, Human-Robot Interaction, Automatic 6D Pose Labeling ### Tech Stack: Python ### Why it Matters: The approach is computationally efficient and achieves consistently robust estimates, outperforming alternatives even with real images. It is particularly useful in scenarios where real images are unavailable, such as when dealing with novel unknown objects. The model's ability to handle significant occlusions and its effectiveness in real-world applications make it a valuable tool in various robotic manipulation tasks. ### GitHub Stars: 424 ### Primary Language: Python ### Data Preparation: A novel RGBD-based dataset for 6D pose tracking in robotic manipulation is developed in this work. The dataset includes 9 videos with 74

Overview: The Alan AI SDK for Power Apps allows developers to quickly create AI agents for their applications, enabling human-like conversations and actions through voice commands. Key features: - Embedding an intelligent layer into Android applications - Generating business logic and UI in real time - Integrating AI-driven interfaces into existing apps in days - Creating a safe and validated environment for accurate, context-aware code generation Tech stack: Java, Android Why it matters: This repository provides the Alan AI SDK for Android, enabling developers to integrate AI-driven features into their Android applications efficiently.

Overview: DeepSIM is an official PyTorch implementation of the paper "DeepSIM: Image Shape Manipulation from a Single Augmented Training Sample" presented at ICCV 2021 Oral. The project focuses on learning to manipulate images using a single augmented training sample, making it a valuable tool for image manipulation tasks. Key features: - Training on a single augmented training sample - Support for different primitive representations: segmentation, edges, and hybrid - Automatic, manual, and hybrid primitive representations - Training on both CPU and GPU - Training on various image datasets including "face", "car", and "face_video" Tech stack: Python Why it matters: DeepSIM offers a unique approach to image manipulation by leveraging a single augmented training sample, which significantly reduces the amount of data required for training compared to traditional methods. This makes it particularly useful for applications where data is limited or expensive to obtain.

Overview: Memento MCP is a knowledge graph memory system designed for large language models, providing resilient, adaptive, and persistent long-term ontological memory. It leverages semantic retrieval, contextual recall, and temporal awareness to enhance the performance and reliability of LLM clients. Key features: - Scalable and high-performance knowledge graph storage - Semantic retrieval capabilities - Contextual recall for improved performance - Temporal awareness with version history - Resilient and persistent long-term ontological memory Tech stack: TypeScript, Neo4j as the storage backend, and Neo4j Desktop for easy setup. Why it matters: Memento MCP offers a unified solution for large language models, providing enhanced performance and reliability through semantic graph storage and vector search capabilities, making it ideal for applications requiring persistent and resilient memory.

Overview: This is an implementation of the Recurrent Memory Transformer (RMT) model, a NeurIPS 2022 paper, written in Pytorch. It has been shown to be capable of copying information across 1 million tokens, making it potentially superior to the Adaptive Agent (AdA) and Recurrent Action Transformer with Memory (RATE) models. Key features: - RMT model architecture - Implementation in Pytorch - Demonstrated capability to copy information across 1 million tokens Tech stack: Pytorch Why it matters: RMT is expected to significantly enhance the performance of reinforcement learning agents, particularly in tasks requiring long-term memory and information transfer. Repository: lucidrains/recurrent-memory-transformer-pytorch Primary language: Python GitHub stars: 425 GitHub description: Implementation of Recurrent Memory Transformer, Neurips 2022 paper, in Pytorch README excerpt: <img src="./rmt.png" width="450px"></img>

### Overview: Agentok Studio is a powerful tool designed to simplify the creation and management of complex agent-based workflows using AG2, a cutting-edge agent framework. It offers intuitive visual tools to streamline the process, making it easier to create and manage tools, and provides full visibility into the underlying code and data for better understanding and debugging. ### Key Features: 1. **Visual Tools**: Intuitive drag-and-drop interface for creating and managing tools. 2. **Conversation Relations**: Facilitates the configuration of tools between nodes, ensuring optimal operation. 3. **Tool Editor**: A tool editor for creating and managing tools, allowing users to configure variables. 4. **Code Generation**: Generates native Python code with minimal dependencies, making it easy to execute as a normal Python program. 5. **Full Visibility**: Provides full visibility into the underlying data representation for better understanding and debugging. ### Tech Stack: - TypeScript ### Why It Matters: Agentok Studio simplifies the creation and management of complex agent-based workflows, making it easier to create and manage tools, and provides full visibility into the underlying code and data, making it easier to diagnose and debug issues.

### Overview: PromptCache is a Go-based semantic caching solution designed to reduce LLM costs and latency by up to 80%. It optimizes performance and resource usage by caching similar prompts, making it an essential tool for GenAI applications. ### Key Features: - **Sub-millisecond Response Times:** Cache hits provide near-instantaneous responses, reducing latency significantly. - **Cost Reduction:** Up to 80% reduction in LLM provider usage, lowering costs. - **Self-Hosted and Provider-Agnostic:** Runs on any system with a Go environment, eliminating the need for external dependencies. - **Two-Threshold Semantic Matching:** High similarity → cache hit, low similarity → cache miss, gray zone → optional verification. ### Tech Stack: Go ### Why it Matters: PromptCache is crucial for reducing LLM costs and improving performance in GenAI applications. By caching similar prompts, it significantly reduces the number of model calls, leading to lower provider usage and faster response times. This makes it an indispensable tool for optimizing resource utilization and improving the overall performance of AI applications.

Overview: 将AI Studio Playground和Build反代成API，兼容OpenAI、Anthropic、Gemini等API。 Key features: - 双额度通道：每个账户拥有独立的Playground和Build额度。 - Build对话模型：提供特定模型的Build对话面板，用量计入Build对话额度。 - 多账户高并发：识别账户权益并优先复用模型。 - 两种WAA后端：默认由Camoufox持有，设置`WAA_BACKEND=go`后由纯Go生成官方证明。 - 四套API协议：支持OpenAI Chat Completions、OpenAI Responses、Anthropic Messages、Gemini GenerateContent。 - 主流agent客户端：支持Claude Code、Codex、OpenCode、pi、omp、OpenClaw、Hermes的文件读写工具调用。 - TTS语音生成：支持Gemini TTS模型的单/多说话人音频生成。 - 图片生成：支持Nano Banana图片生成。 - 视频生成：支持Veo视频生成和图片转视频；Gemini Omni通过四套生成接口接收文本、图片与视频输入，输出文本与MP4视频。 - YouTube输入：粘贴视频URL即可作为外部视频附件读取。 - 智能模型切换：从AI Studio实时发现模型并按`model`字段路由。 - Google工具：支持Search、Image Search、URL Context、Code Execution和Maps。 - 文件与转录：支持文件上传、查询、内容读取、删除