Projects

### Overview: 本项目旨在对OpenAI的Whisper语音识别模型进行微调，支持多种训练方式（无时间戳数据、有时间戳数据、无语音数据）和部署平台（Windows桌面、Android、Web）。通过使用Lora微调，加速推理并支持多种部署环境，旨在提升模型的性能和应用范围。 ### Key features: - 支持无时间戳数据训练 - 支持有时间戳数据训练 - 支持无语音数据训练 - 提供多种部署环境支持，包括Windows桌面应用、Android应用和Web部署 ### Tech stack: 使用Python 3.8、PyTorch 2.4.0、Ubuntu 18.04、GPU A100-PCIE-40GB*1。 ### Why it matters: 通过微调Whisper模型并支持多种训练方式和部署环境，可以显著提高模型的泛化能力和应用范围，特别是在需要实时处理和跨平台部署的场景中。

### Overview: Perceiver is an innovative model designed for general perception tasks, built on the Pytorch framework. This implementation is based on the research paper "Perceiver: General Perception with Iterative Attention" by Yannic Kilcher et al., published in 2021. It offers a flexible backbone that can handle sequences of varying lengths, making it highly versatile for a wide range of applications. ### Key Features: - **Perceiver Model**: Implements the Perceiver architecture, capable of processing sequences of varying lengths. - **PerceiverIO Model**: Extends Perceiver with a flexible backbone that can handle sequences of any length. - **PerceiverLM Model**: A language model variant that can be used for text generation tasks. ### Tech Stack: - **Python**: The primary programming language used for implementing the model. - **Pytorch**: The deep learning framework utilized for building and training the models. ### Why It Matters: Perceiver and its variants are designed to address the challenges in processing sequences of varying lengths, making them particularly suitable for applications in computer vision, natural language processing, and other domains where the input data can vary widely in terms of length and complexity. Their ability to handle sequences of any length makes them highly versatile and adaptable to a wide range of tasks. ### Repository: The repository is named `lucidrains/perceiver-pytorch` and has been starred by 1,221 developers, indicating its popularity and utility in the community. ### Install: To use the model, you can install it via pip: ```bash $ pip install

Overview: Ongrid is an AI-powered ops agent designed to understand your infrastructure, identify root causes, and fix issues directly from Slack, Telegram, Lark, or DingTalk. It automates the ops workflow, reducing manual intervention and improving efficiency. Key features: - 🤖 Coordinator + Specialist agents - 🚨 Auto-investigate on alert - 🔍 Root-cause RCA - 🔒 Zero inbound ports - 💻 Browser SSH - 🐳 Self-host in one command - 📊 Built-in observability - 🧠 Bring your own model - 💬 Two-way IM channels - 🛠️ Read-only host tools Tech stack: Go Why it matters: Ongrid simplifies ops management by automating tasks, reducing manual effort, and providing real-time insights directly from your favorite communication channels.

### Overview: GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond is an official implementation of the paper "GCNet: Non-local Networks Meet Squeeze-Excitation Networks and Beyond" on COCO object detection. This project leverages the strengths of Non-Local Networks (NLNet) and Squeeze-Excitation Networks (SENet) to provide a simple, fast, and effective approach for global context modeling. ### Key Features: - **Context Block**: The core operator GC block is available [here](https://github.com/xvjiarui/GCNet/blob/master/mmdet/ops/gcb/context_block.py). - **Performance**: The project achieves state-of-the-art results on various recognition tasks, outperforming NLNet and SENet on major benchmarks. - **Training Stability**: The code is refactored, and more results are provided, including all configurations in `configs/gcnet`. ### Tech Stack: - Python ### Why It Matters: - **Improvement on Recognition Tasks**: GCNet provides a simple, fast, and effective approach for global context modeling, which generally outperforms both NLNet and SENet on major benchmarks for various recognition tasks. - **Stability and Performance**: The project is supported by the official mmdetection repo and has been accepted by TPAMI, demonstrating its stability and effectiveness. ### GitHub Stars: 1220 ### Repository: - Repository: xvjiarui/GCNet - Primary Language: Python - GitHub Description: GCNet: Non-local Networks Meet Squeeze-

### Overview: FastMOT is a high-performance multiple object tracker that leverages YOLO, Deep SORT, OSNet ReID, and KLT for real-time tracking. It significantly speeds up the system by only running the detector and feature extractor every N frames, while KLT fills in the gaps efficiently. ### Key Features: - **YOLO Detector**: Tracks objects using the YOLOv4 model. - **Deep SORT + OSNet ReID**: Provides object tracking and identity verification. - **KLT Tracker**: Fills in gaps between frames for moving camera scenarios. - **TensorRT Backend**: Optimized for faster inference. - **Numba Optimization**: Optimized algorithms for better performance. ### Tech Stack: - Python - CUDA 10 - cuDNN 7 - TensorRT 7 - OpenCV 3.3 - Numpy 1.17 - Scipy 1.5 - Numba 0.48 - CuPy 9.2 - TensorFlow < 2.0 (for SSD support) ### Why it Matters: FastMOT achieves state-of-the-art MOTA scores on the MOT20 dataset, with tracking speed ranging from 50 to 150 FPS depending on the number of objects, making it suitable for both constrained devices like Jetson Nano and desktops. ### Repository: - Repository: [GeekAlexis/FastMOT](https://github.com/GeekAlexis/FastMOT) - GitHub Stars: 121

### Overview: Startup Skill is a revolutionary AI tool designed to help startups validate their ideas, conduct competitive intelligence, and plan their strategies. It provides market research, competitive battle cards, positioning analysis, financial projections, and a 30-day action plan—all in one afternoon. ### Key Features: - **Market Research**: Comprehensive market research to validate your startup idea. - **Competitive Analysis**: Battle cards for competitors, pricing landscape, and strategic reports. - **Positioning Analysis**: Dunford's market positioning framework. - **Financial Projections**: Detailed financial projections and validation experiments. - **Action Plan**: A 30-day action plan to guide your startup's growth. ### Tech Stack: The tool leverages Claude Code and any agent that supports skills, ensuring a high level of accuracy and efficiency. ### Why It Matters: Startup Skill offers a cost-effective and efficient solution for startups, providing them with the necessary tools to validate their ideas, conduct competitive analysis, and plan their strategies. It streamlines the entire process from idea to validated strategy, making it a game-changer for entrepreneurs. ### Repository: Repository: [ferdinandobons/startup-skill](https://github.com/ferdinandobons/startup-skill) ### GitHub Stars: GitHub stars: 1216 ### Installation: - **Claude Code Plugin (Recommended)**: Install via `claude plugin marketplace add ferdinandobons/startup-skill` and `claude plugin install startup@startup-skill`. - **Claude.ai (Web App)**

### Overview: This repository contains implementations of various self-attention mechanisms focused on computer vision tasks using PyTorch. The project aims to provide a comprehensive set of tools for developers working on computer vision applications. ### Key features: - Multi-head self-attention - Axial attention - Vanilla Transformer Encoder - Vision Transformer (ViT) with ResNet50 backbone - Re-implementation of the Unet model - Bottleneck Attention block - Positional embeddings (1D and 2D) ### Tech stack: PyTorch ### Why it matters: This project is essential for developers working on computer vision applications, providing a robust set of tools for implementing self-attention mechanisms in PyTorch. The inclusion of various attention mechanisms and the ability to integrate with a pre-trained backbone like ResNet50 makes this project highly valuable for researchers and developers alike.

Overview: ArcadeDB is a multi-model database that supports SQL, Cypher, Gremlin, HTTP/JSON, MongoDB, and Redis. It is a conceptual fork of OrientDB, designed for extreme performance and support for vector embeddings. Key features: - Supports multiple data models: SQL, Cypher, Gremlin, HTTP/JSON, MongoDB, and Redis. - Vector embeddings for advanced search capabilities. - Built-in support for distributed systems with Raft and embedded Raft replication. - High availability and resilience features for fault tolerance. Tech stack: Java Why it matters: ArcadeDB offers a comprehensive solution for developers needing a multi-model database with advanced features like vector embeddings, high performance, and robust fault tolerance, making it suitable for applications requiring both relational and non-relational data management. Repository: ArcadeData/arcadedb GitHub stars: 1207

### Overview: VoxFormer is an innovative deep learning framework designed to predict complete 3D semantic scene completions from 2D images. This project, presented at CVPR 2023, showcases a cutting-edge baseline that leverages Transformer-based techniques to achieve superior performance in semantic scene completion tasks. ### Key Features: - **Transformer-based Semantic Scene Completion**: VoxFormer utilizes Transformers to predict complete 3D semantic scene completions directly from 2D images. - **Sparse Voxel Queries**: It starts with sparse voxel queries from depth estimation, which are then refined using deformable cross-attention and self-attention. - **Mask Tokens for Deformation**: The framework introduces mask tokens to complete the voxel features, ensuring that only visible and occupied areas are processed. - **GPU Memory Efficiency**: By reducing GPU memory usage, VoxFormer achieves better performance and more efficient training compared to state-of-the-art methods. ### Tech Stack: - **Python**: The primary programming language used for the implementation. ### Why It Matters: VoxFormer's ability to predict complete 3D semantic scene completions from 2D images opens up new possibilities for AI systems in tasks such as scene understanding and recognition, making it a significant advancement in the field of computer vision and 3D semantic scene completion. ### Repository: The project is hosted on GitHub at [NVlabs/VoxFormer](https://github.com/NVlabs/VoxFormer). ### GitHub Stars: The repository has received 12,121 stars, indicating high interest and engagement from the developer community. ### Primary

Overview: PaperJury is a pre-submission AI review tool designed to help researchers identify and address issues in their research papers before they are submitted. It uses a model-driven approach to review, verdict, revise, and verify the quality of the submitted papers. Key features: - Real-time feedback on paper quality - Automated detection of issues and suggestions for improvement - Integration with LaTeX to ensure bounded revision - Support for multiple languages and formats Tech stack: JavaScript Why it matters: PaperJury is crucial for researchers who need to ensure the quality and validity of their work before submitting it. It helps identify and correct issues that could be overlooked, improving the overall credibility of the research. Repository: Spark-To-Paper-Skills/paperjury Primary language: JavaScript GitHub stars: 1210 GitHub description: Pre-submission AI review stress-test for research papers. A Claude Code skill: review, verdict, revise, verify. README excerpt: <p align="center"> <img src="docs/paperjury-mark.png" width="170" alt="PaperJury logo"> </p> <h1 align="center">PaperJury</h1> <h3 align="center">Due-Process Review for Bounded LaTeX Revision</h3> <p align="center"> <b>正式投稿前，先让 AI reviewer 指出论文中需要解决的问题。</b><br> <b>模型负责阅读、判断和起草；状态管理、投票规则、补丁应用与停止条件由确定性

### Overview: CleanVision is a Python package designed to automatically detect potential issues in image datasets. It identifies common problems such as blurry, under/over-exposed images, and duplicates, helping to ensure the quality of your dataset before applying machine learning techniques. ### Key Features: - Automatically detects issues in image datasets like blurry, under/over-exposed, and duplicate images. - Quick and easy to use, with just a couple lines of Python code to audit any image dataset. - Supports various image formats, including JPEG, PNG, and more. ### Tech Stack: Python ### Why It Matters: CleanVision is a valuable tool for anyone working on computer vision projects, ensuring that your dataset is of high quality before applying machine learning techniques. This helps in obtaining more accurate and reliable results. ### Repository: - Repository: cleanlab/cleanvision - GitHub Stars: 1212 - GitHub Description: Automatically find issues in image datasets and practice data-centric computer vision. ### Installation: ```shell pip install cleanvision ``` ### Quickstart: 1. **Download Example Dataset** (optional): Use `wget` to download an example dataset. 2. **Run CleanVision**: ```python from cleanvision import Imagelab # Specify path to folder containing the image files in your dataset imagelab = Imagelab(data_path="FOLDER_WITH_IMAGES/") # Automatically check for a predefined list of issues within your dataset imagelab.find_issues() # Produce a neat report of the issues found in

### Overview: ECON is a cutting-edge project designed to digitize human clothing from color images, combining the strengths of implicit and explicit representations to achieve high-fidelity 3D clothed humans. It supports multi-person reconstruction and utilizes SMPL-X for realistic animations, making it suitable for challenging poses and loose clothing. ### Key Features: - Multi-person reconstruction - SMPL-X based animation - Supports loose clothing and challenging poses ### Tech Stack: - Python ### Why it Matters: ECON is particularly valuable for its ability to accurately reconstruct and animate clothed humans from in-the-wild images, even in challenging poses and with loose clothing, making it a significant advancement in the field of human digitization.

Overview: QuiLLMan is a complete voice chat app powered by a speech-to-speech language model and bidirectional streaming, utilizing Kyutai Lab's Moshi model and Mimi streaming encoder/decoder. It features near-instantaneous response times and seamless audio streaming across the network. Key features: - Bidirectional streaming for continuous audio transmission - Speech-to-speech language model for natural conversation - Moshi model for efficient audio compression and real-time response - WebSocket-based connection for seamless communication Tech stack: Python Why it matters: QuiLLMan provides a practical application of language models in voice chat, showcasing the potential of integrating AI in real-time communication applications.

Overview: Neural Network Compression Framework (NNCF) is a comprehensive tool designed to optimize the inference performance of neural networks in OpenVINO, with minimal accuracy loss. Key features: - Post-Training Quantization - Weights Compression - Activation Sparsity - Post-Training Quantization - Quantization Aware Training - Weight-Only Quantization Aware Training with LoRA and NLS - Pruning Tech stack: Python Why it matters: NNCF is essential for developers and researchers working with neural networks, providing a robust solution for efficient inference with minimal performance degradation.

### Overview: Chorus is a powerful agent harness designed to streamline the collaboration between coding agents and human developers, enabling seamless proposal, verification, and software delivery. ### Key Features: - **Multi-Agent Collaboration**: Seamlessly integrates multiple coding agents with human developers. - **Proposal and Verification**: Agents propose code ideas, and humans verify them before proceeding. - **Pipeline Management**: Facilitates a complete pipeline from idea generation to software delivery. - **Fine-Grained Permissions**: Agents and humans have fine-grained permissions to control access and workflow. ### Tech Stack: TypeScript ### Why It Matters: Chorus is inspired by the AI-DLC methodology, focusing on the reversed conversation approach where AI proposes, and humans verify. This ensures a more efficient and effective collaboration process, improving the quality and speed of software development. ### Repository: Chorus-AIDLC/Chorus ### GitHub Stars: 1208 ### Primary Language: TypeScript ### GitHub Description: The Agent Harness for AI-Human Collaboration, inspired by the AI-DLC (AI-Driven Development Lifecycle) ### Documentation: 📖 Documentation ### What's New: - **[v0.23.0]**: Added workspace-local account management and personalization, proposal dependency validation, optional pre-submit checks, and shared Redis auth rate limits. - **[v0.22.2]**: Hosted on Linux or macOS with Docker Compose, PostgreSQL 17, and improved deployment, LAN HTTPS, and Cloudflare Tunnel guidance. - **[v0.22.0]**

Overview: MeanFlow and iMF are PyTorch implementations of Mean Flows and Improved Mean Flows, respectively. These models are designed for one-step generative modeling, focusing on stability and efficiency. Key features: - MeanFlow and iMF training modes - Dual-head DiT for MeanFlow and flow matching for iMF - CFG scale as model input with CFG distillation - JVP under `no_grad` for `dudt` only; separate grad-enabled forward pass for optimization - Config-based training for MNIST, CIFAR-10, and ImageNet latent Tech stack: Python Why it matters: These models are useful for applications requiring one-step generative modeling, such as image synthesis, where they offer stability and efficiency in training and inference. Repository: haidog-yaqub/MeanFlow Primary language: Python GitHub stars: 1208 GitHub description: PyTorch implementation of MeanFlows and iMF (one-step generative modeling).

Overview: RAGLite is a Python toolkit designed for Retrieval-Augmented Generation (RAG) using DuckDB or PostgreSQL as a keyword and vector search database. It supports various LLM providers, lightweight dependencies, and advanced features like PDF to Markdown conversion and optimal sentence splitting. Key features: - 🧠 Configurable LLM providers - 📖 PDF to Markdown conversion - 🧬 Multi-vector chunk embedding - ✏️ Optimal sentence splitting - 🔍 Hybrid search with native keyword & vector search - 💰 Improved cost and latency - 🌀 Optimal query adapter - 💬 Optional frontend for ChatGPT-like integration Tech stack: Python Why it matters: RAGLite simplifies RAG implementation with a lightweight and permissive approach, making it suitable for various LLM providers and database types, while offering advanced features like PDF conversion and optimal sentence splitting for better output quality and cost efficiency.

Overview: AGenUI is a high-performance A2UI SDK designed for building native generative UI experiences on iOS, Android, and HarmonyOS. It leverages a shared C++ core engine to render interactive UI components in real time, ensuring smooth performance across all platforms. Key features: - Real-time streaming rendering - High performance on native rendering - 22 built-in components - Custom component API - Function Call integration - Design Token & theming - Light / Dark mode Tech stack: C++ Why it matters: AGenUI provides a powerful and efficient solution for creating interactive UIs on multiple platforms, ensuring a smooth and engaging user experience.

### Overview: Autoprompt is a cutting-edge coding-agent skill designed to significantly reduce failure rates on agentic coding tasks. By automating the review, fixing, and rechecking of code, it achieves a 45% reduction in errors, making it an invaluable tool for developers. ### Key Features: - **Code Review and Fixing**: Autoprompt automatically reviews and fixes code, ensuring high-quality output. - **Rechecking**: It rechecks the code to catch any overlooked issues, further enhancing reliability. - **Customization**: Supports multiple coding agents, including Claude Code, Codex, OpenCode, Kilo Code, VS Code, Prime Agent, Oh My Pi, DeepSeek Harness, Reasonix, and Hermes Agent. ### Tech Stack: - JavaScript ### Why It Matters: Autoprompt significantly improves the quality and reliability of code, reducing the need for manual interventions and lowering the risk of errors. This makes it an essential tool for developers looking to streamline their workflow and maintain high standards of code quality. ### Repository: The repository is maintained by [Spielewoy](https://github.com/Spielewoy), and it has been starred by 1,217 developers, indicating its popularity and usefulness in the developer community. ### Primary Language: The primary language used in the project is JavaScript. ### GitHub Stars: The repository has received 1,217 stars from the developer community, reflecting its widespread adoption and usefulness in the field of coding. ### Installation: - **CLI Installation**: Use the provided CLI to install Aut

Overview: This project contains a minimal and readable coding agent implemented in Python, designed to explain the core components of coding agents. Key features: - Workspace snapshot collection - Stable prompt with turn state - Structured tools - Approval handling for risky actions - Transcript and memory persistence - Bounded delegation Tech stack: Python Why it matters: This project provides a concise and practical guide to understanding the key components of a coding agent, making it accessible for developers looking to improve their coding workflows. Repository: rasbt/mini-coding-agent GitHub stars: 1209 GitHub description: Minimal and readable coding agent harness implementation in Python to explain the core components of coding agents.