Projects

Overview: XMem is a cutting-edge method for long-term video object segmentation that leverages an Atkinson-Shiffrin memory model to handle very long videos efficiently while maintaining high performance. Key features: - Handles very long videos with limited GPU memory usage. - Provides quite fast performance, achieving ~20 FPS even with long videos. - Comes with a user-friendly GUI based on MiVOS for easy integration into existing projects. Tech stack: Python Why it matters: XMem offers a novel approach to video object segmentation by integrating a memory model inspired by human memory, making it robust to long-term occlusions and capable of processing large video datasets efficiently. This method addresses the limitations of existing techniques by balancing short-term and long-term memory, ensuring high performance and ease of use.

### Overview: TimeMixer is an innovative model designed for time series forecasting, presented at ICLR 2024. It offers a novel approach to decomposing multiscale mixing for improved performance across various tasks. ### Key Features: - **Decomposable Multiscale Mixing**: Breaks down time series into smaller, manageable components for better analysis and prediction. - **State-of-the-Art Performance**: Achieves state-of-the-art results across 8 diverse analytical tasks. - **Versatile Applications**: Supports long-term forecasting, short-term forecasting, anomaly detection, imputation, and classification. ### Tech Stack: - **Python**: The primary programming language used for implementation. ### Why It Matters: TimeMixer provides a powerful tool for time series forecasting, offering a unique method for decomposing and mixing time series data to enhance predictive accuracy and efficiency. ### Repository: - **Repository**: kwuking/TimeMixer - **GitHub Stars**: 1993 - **GitHub Description**: Official implementation of "TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting" ### Updates: - **2024.10**: TimeMixer series included in PyPOTS. - **2024.10**: TimeMixer++ version released, featuring a time series pattern machine. - **2024.07**: TimeMixer included in NeuralForecast. - **2024.07**: TimeMixer upgraded to support comprehensive time series analysis. - **2024.06**:

### Overview: Featureform is a virtual feature store that integrates seamlessly with your existing data infrastructure, enabling data scientists to manage and serve machine learning model features. ### Key Features: - **Enhanced Collaboration**: Standardizes feature definitions for easy sharing and reusability across the team. - **Centralized Experimentation**: Features, transformations, and training sets can be easily pushed from notebooks to a centralized repository. - **Automated Deployment**: Orchestrates data infrastructure for feature deployment, reducing the need for custom solutions. - **Reliable Feature Management**: Ensures immutability and automatic retry logic for feature management. - **Compliance and Security**: Built-in role-based access control, audit logs, and dynamic serving rules for compliance. ### Tech Stack: Go ### Why It Matters: Featureform simplifies the process of defining, managing, and deploying machine learning model features, making it easier for data scientists to collaborate, experiment, and deploy models without the complexities of traditional feature stores.

Overview: ARTEX is an AI-driven autonomous penetration-testing system for global environments, featuring multi-agent capabilities and language-switching options. Key features: - Autonomous multi-agent penetration testing - Language switching for agent-generated reports and chat replies - Support for English, Korean, Chinese, Spanish, and Japanese Tech stack: Go backend + Next.js frontend Why it matters: ARTEX provides a valuable tool for defenders to understand how autonomous AI attacks work and build detection capabilities against them. Repository: jiwoochris/artex-global Primary language: Go GitHub stars: 1991

### Overview: Text Generator is a versatile plugin for Obsidian that allows you to generate text content using various AI providers, including OpenAI, Anthropic, Google, and local models. It offers free and open-source capabilities, flexibility in prompts, and a community of users sharing use cases and templates. ### Key features: - **AI Integration**: Seamless integration with OpenAI, Anthropic, Google, and local models. - **Prompt Flexibility**: Supports various prompts and configurations. - **Template Engine**: Enables the creation of templates for repetitive tasks. - **Community Templates**: Offers shared templates and easy sharing of use cases. - **Configuration Options**: Highly flexible configuration using frontmatter. ### Tech stack: TypeScript ### Why it matters: - **Powerful AI Integration**: Access to advanced AI tools for generating text. - **Community Support**: Engage with a community of users sharing use cases and templates. - **Ease of Use**: Simple installation and configuration options for both Obsidian and Discord integration. ### GitHub stars: 1991 ### Repository: nhaouari/obsidian-textgenerator-plugin

### Overview: GANimation: Anatomically-aware Facial Animation from a Single Image (ECCV'18 Oral) is a project that introduces a novel GAN conditioning scheme based on Action Units (AU) annotations to generate an anatomically consistent facial animation from a single image. This approach allows for precise control over the activation of each AU and combines multiple AU activations to create realistic expressions. ### Key features: - Anatomically-aware facial animation generation from a single image. - Novel GAN conditioning scheme based on Action Units (AU) annotations. - Enables controlling the magnitude of activation for each AU. - Combines multiple AU activations to create realistic expressions. ### Tech stack: Python ### Why it matters: This project is significant because it addresses the challenge of generating realistic facial animations from a single image, which is crucial in fields such as computer-generated imagery, virtual reality, and entertainment. The use of Action Units (AU) annotations provides a continuous and anatomically accurate representation of facial movements, making the generated animations more believable and realistic.

### Overview: Make-A-Video is a cutting-edge project that implements the advanced text-to-video generator model from Meta AI, known as Make-A-Video. This project leverages PyTorch to achieve state-of-the-art results in the field of text-to-video generation. The implementation focuses on pseudo-3D convolutions and temporal attention, showcasing significant improvements in temporal fusion compared to previous methods. ### Key Features: - **Pseudo-3D Convolutions**: Combines axial convolutions with temporal attention to enhance temporal fusion. - **Temporal Attention**: Implements attention across time to improve the model's ability to capture temporal dependencies. - **Frame Interpolation**: Ensures correct frame interpolation to produce high-quality videos. - **SOTA Performance**: Achieves state-of-the-art results in text-to-video generation. ### Tech Stack: - **Python**: The primary programming language used for the implementation. - **PyTorch**: The deep learning framework used to build and train the model. ### Why it Matters: - **State-of-the-Art Performance**: The implementation of Make-A-Video surpasses existing methods in text-to-video generation, demonstrating the effectiveness of the proposed techniques. - **Research Contribution**: This project contributes to the advancement of generative artificial intelligence, specifically in the domain of text-to-video generation, by improving upon existing models. ### Repository: - **Repository**: lucidrains/make-a-video-pytorch - **GitHub Stars**: 1989 - **Description**: Implementation of Make-A-Video, a new SOTA text-to-video generator

### Overview: 天机（Tianji）是一款专注于传统人情场景的大语言模型应用。它涵盖了Prompt工程、RAG知识库、Agent智能体应用、模型微调等多种实现，并提供LLM开发入门教程，帮助用户快速入门大语言模型的开发。 ### Key features: - **Prompt工程**：构建人情世故的提示词工程。 - **RAG知识库**：基于LangChain和LlamaIndex构建的知识库问答系统。 - **Agent智能体**：基于Metagpt等框架开发的智能体应用，支持工具调用。 - **模型微调**：使用Transformers、Xtuner等框架的Lora与全量微调。 - **数据制造**：提供常见社交领域的语料库和数据制造与清洗工具。 ### Tech stack: Python ### Why it matters: 通过天机，用户可以快速掌握大语言模型的开发流程，从基础的模型支持到高级的微调与应用，全面覆盖，帮助开发者轻松构建符合传统人情场景的应用。

### Overview: Perplexity AI is an unofficial Python module that provides a powerful API wrapper for Perplexity.ai, including account generation and a web interface for managing accounts. This tool bypasses query limits by creating new accounts, making it ideal for users who need to perform unlimited pro queries. ### Key Features: - **Account Generation**: Automatically generate Gmail accounts using Emailnator. - **Unlimited Pro Queries**: Bypass query limits by creating new accounts. - **Web Interface**: Automate account creation and usage via a browser. - **Sync & Async APIs**: Full support for both synchronous and asynchronous programming. - **Type Safety**: Complete type hints for better IDE support. - **Robust Error Handling**: Custom exceptions for better error management. - **Comprehensive Logging**: Structured logging for debugging. - **File Upload**: Support for document analysis and Q&A. - **Streaming Responses**: Real-time response streaming. - **Retry Logic**: Automatic retry with exponential backoff. - **Rate Limiting**: Built-in rate limiting to prevent abuse. ### Tech Stack: Python ### Why it Matters: Perplexity AI simplifies the process of generating new accounts for Perplexity.ai, making it easier to perform unlimited pro queries without hitting query limits. The web interface allows users to automate account creation and usage, making it an excellent tool for those who prefer a GUI-based approach. The comprehensive features and robust error handling make it a reliable solution for managing Perplexity.ai accounts. Repository: helallao/perplexity-ai GitHub Stars: 1987

### Overview: Show-o is a repository for a series of models that unify multimodal understanding and generation. It aims to provide a single unified model that can handle various modalities such as text, image, and video, offering improved performance and versatility. ### Key Features: - **Unified Model**: Show-o offers a single model that can handle multimodal inputs, including text, images, and videos. - **Improved Performance**: The repository includes ablation studies and improved models, demonstrating enhanced performance in various tasks. - **Comprehensive Benchmarks**: The repository includes models that have been evaluated on leading benchmarks, showcasing strong performance metrics. ### Tech Stack: - Python ### Why It Matters: Show-o is a significant contribution to the field of multimodal understanding and generation, providing a unified model that can handle multiple modalities efficiently. This model has the potential to revolutionize applications in natural language processing, computer vision, and more, offering improved performance and versatility compared to existing models.

Overview: Amazon Q CLI is a powerful command-line interface designed for developers to interact with Amazon Q, a natural language-based chat experience. Key features: - Amazon Q CLI allows seamless integration of Amazon Q into your terminal. - It supports various operating systems, including macOS, Linux, and alternative Linux builds. - The project is open-source and available as the Kiro CLI, a closed-source product. Tech stack: Rust Why it matters: Amazon Q CLI simplifies the interaction with Amazon Q, making it easier for developers to utilize its features directly from the command line, enhancing productivity and efficiency. Repository: aws/amazon-q-developer-cli GitHub stars: 1983

Overview: This repository compiles a collection of papers focused on the application of deep learning in protein design, a rapidly evolving field in computational biology. Key features: - Generative artificial intelligence for enzyme design and biocatalysis - The Human Bindome: A Proteome-scale Atlas of Designed Binder Candidates - Sequence dataset/benchmarks - Structure datasets/benchmarks - Public database - Similar list - Guides Tech stack: Python, TensorFlow, PyTorch, and other deep learning frameworks. Why it matters: This repository provides a comprehensive resource for researchers and developers interested in the intersection of deep learning and protein design, offering insights and tools for advancing computational biology and drug discovery. Repository: Peldom/papers_for_protein_design_using_DL GitHub stars: 1982 Community Values, Guiding Principles, and Commitments for the Responsible Development of AI for Protein Design: [details](https://responsiblebiodesign.ai/)

Overview: EOS is a lightweight 3D Morphable Face Model fitting library written in modern C++11/14, providing basic functionality for face models, camera and shape fitting. Key features: - MorphableModel and PcaModel classes for 3DMMs - Surrey Face Model (SFM), 4D Face Model (4DFM), Basel Face Model (BFM) 2009 and 2017, and Liverpool-York Head Model (LYHM) support - Low-resolution Surrey Face Model (share/sfm_shape_3448.bin) - Fast, linear pose, shape, and expression fitting - Edge and contour fitting - Texture extraction - Python bindings - Experimental non-linear fitting cost functions using Ceres Tech stack: Modern C++11/14 Why it matters: EOS simplifies the process of using and fitting 3D Morphable Face Models, making it easier for developers to incorporate this technology into their projects.

### Overview: This repository provides a script for fine-tuning the Qwen-VL series models (Qwen2-VL, Qwen2.5-VL, Qwen3-VL, and Qwen3.5) using only Hugging Face and Liger-Kernel. It supports various training modes and datasets, making it a versatile tool for researchers and developers. ### Key features: - Fine-tuning Qwen-VL series models - Support for reasoning mode training - Compatibility with mixed-modality datasets - Use of Liger-Kernel for efficient training - Support for classification tasks ### Tech stack: Python ### Why it matters: This script simplifies the process of fine-tuning large language models, making it accessible for researchers and developers without requiring extensive knowledge of deep learning frameworks. It leverages the power of Liger-Kernel for efficient training, making it suitable for various datasets and training modes. ### Repository: 2U1/Qwen-VL-Series-Finetune GitHub stars: 1981 GitHub description: An open-source implementation for fine-tuning Qwen-VL series by Alibaba Cloud.

Overview: FastFlowLM (FLM) is a fast, lightweight, and AMD Ryzen™ AI NPU-first runtime that enables the execution of large language models on AMD's XDNA2 NPUs without the need for GPUs. It supports Vision, Audio, Embedding, and Multi-Headed Attention (MoE) and optimizes for context lengths up to 256k tokens, making it ultra-lightweight at just 17 MB. Key features: - Runs large language models on AMD Ryzen™ AI NPUs - Supports Vision, Audio, Embedding, and MoE - Optimized for context lengths up to 256k tokens - Ultra-lightweight at 17 MB - Supports all Ryzen™ AI Series chips with XDNA2 NPUs (Strix, Strix Halo, Kraken, and Gorgon Point) - No GPU required, faster and more power-efficient - Installs within 20 seconds Tech stack: C++ Why it matters: FastFlowLM (FLM) provides a fast, lightweight, and efficient solution for running large language models on AMD's AI NPUs, making it ideal for applications requiring low power consumption and fast execution times.

Overview: Easy Machine Learning is a general-purpose dataflow-based system designed to simplify the process of applying machine learning algorithms to real-world tasks, particularly on distributed platforms like Hadoop and Spark. Key features: - Simplifies the definition and execution of machine learning tasks. - Enables sharing and reusing of algorithms, job DAGs, and experimental results. - Seamlessly integrates standalone and distributed algorithms within a single task. Tech stack: Java Why it matters: Easy Machine Learning streamlines the process of applying machine learning algorithms, making it accessible and efficient even on distributed systems, thus unlocking the full potential of machine learning in big data applications.

### Overview: This repository contains PyTorch implementations of various Bayesian neural network inference methods, including Bayes by Backprop (BBP), Local Reparametrization Trick, MC Dropout, SGLD, Kronecker-Factorised Laplace Approximation, Stochastic Gradient Hamiltonian Monte Carlo (SG-HMC), and more. It also includes code for Bootstrap MAP Ensemble. ### Key features: - PyTorch implementations - Support for PyTorch 1.0.1 - Support for both CPU and GPU - Includes code for Bayesian inference methods such as BBP, LR, MC dropout, SGLD, Kronecker-Factorised Laplace Approximation, and SG-HMC - Provides code for Bootstrap MAP Ensemble - Written in Python 2.7 and PyTorch 1.0.1 - Supports both CPU and GPU - Includes Google Colab notebooks for running on a GPU for free ### Tech stack: PyTorch ### Why it matters: This repository is essential for researchers and developers looking to implement and experiment with Bayesian neural network inference methods using PyTorch. It provides a comprehensive set of tools for training and evaluating models using various inference techniques, making it a valuable resource for anyone working in the field of Bayesian neural networks.

Overview: AgentShell is a powerful tool that integrates your favorite AI agents into Emacs, making them accessible and familiar to users who prefer a desktop environment. It simplifies the process of running agents without the need for terminal emulation, providing a seamless experience. Key features: - Seamless integration of AI agents - Familiar Emacs buffers and bindings - No terminal emulation required - Built on ACP (Agent Computing Platform) for enhanced performance Tech stack: Emacs Lisp Why it matters: AgentShell enhances productivity by allowing developers to leverage their favorite AI tools within the familiar environment of Emacs, making the workflow more efficient and enjoyable.

Overview: 本项目基于Meta最新发布的Llama-3模型开发，提供了中文Llama-3基座模型和指令精调大模型，旨在提升中文基础语义和指令理解能力。 Key features: - 开源Llama-3-Chinese基座模型和指令模型 - 提供预训练脚本和指令精调脚本 - 提供alpaca_zh_51k, stem_zh_instruction, ruozhiba_gpt4 (4o/4T) 指令精调数据 - 提供快速本地量化和部署教程 Tech stack: Python Why it matters: 本项目是Chinese-LLaMA-Alpaca系列项目的第三期，提供了中文Llama-3基座模型和指令模型，旨在提升中文基础语义和指令理解能力，是研究和应用中文语言模型的重要资源。

### Overview: This project provides a curated list of papers on dataset distillation and related applications, offering comprehensive information on the field. ### Key features: - **Dataset Distillation Papers**: A collection of papers on dataset distillation, including seminal works and recent advancements. - **Contributions Welcome**: Encourages contributions to improve the list of papers. - **Stars and Forks**: 1,978 stars and forks, indicating high community interest and active participation. - **Project Maintainers**: Managed by Guang Li, Bo Zhao, and Tongzhou Wang. ### Tech Stack: HTML ### Why it Matters: This project serves as a comprehensive resource for researchers and practitioners interested in dataset distillation, providing a wealth of information on the field's latest developments and applications.