Projects

### Overview: This repository is a curated list of resources and papers related to multimodal recommender systems. It aims to provide a comprehensive overview of the field, focusing on the latest research and methodologies. ### Key features: - **Code First**: The repository is primarily composed of code, showcasing the latest advancements in the field. - **Taxonomy, Evaluation, and Future Directions**: A comprehensive survey paper by Hongyu Zhou et al. provides a taxonomy, evaluation metrics, and insights into future research directions. - **General CF Models**: Integrates 10+ multimodal models and various general collaborative filtering (CF) models. ### Tech stack: The repository uses Python, TensorFlow, and other relevant technologies for data processing and model implementation. ### Why it matters: This repository is crucial for developers and researchers interested in the field of multimodal recommender systems. It serves as a valuable resource for staying updated with the latest research, understanding the current state of the art, and identifying potential areas for further exploration. ### Repository: The primary language is unknown, but the repository has been starred by 473 developers, indicating its popularity and utility in the community. ### Primary language: Unknown ### GitHub stars: 473 ### GitHub description: A curated list of awesome resources about multimodal recommender systems. Please contact us if your work is not in the list.

Overview: This repository contains sample applications, model recipes, agent skills, and utilities for LiteRT and LiteRT-LM, designed to help developers quickly integrate and deploy AI models on-device. Key features: - LiteRT and LiteRT-LM sample applications - Model recipes for model conversion and quantization - Agent skills for automating AI tasks - End-to-end model conversion and benchmarking recipes Tech stack: Python Why it matters: This repository provides a comprehensive set of tools and resources for developers to quickly integrate and deploy AI models on-device, making it easier to create and deploy AI applications.

Overview: StyleShot is a model designed to generate high-quality, personalized style transformations on any image, allowing users to achieve a variety of styles such as 3D, flat, abstract, and fine-grained without the need for test-time fine-tuning. Key features: - Text-driven style transfer - Image-driven style transfer - Support for various styles including 3D, flat, abstract, and fine-grained Tech stack: Python Why it matters: StyleShot simplifies the process of applying various styles to images by leveraging a style-aware encoder and a well-organized style dataset, enabling users to achieve high-quality style transformations without the need for extensive fine-tuning, making it highly versatile and efficient.

Overview: This project focuses on the development of a system that leverages large language models (LLMs) for knowledge graph construction and reasoning. It includes three main components: basic evaluation of large models, virtual knowledge extraction, and automatic knowledge graph construction. Key features: - Basic Evaluation: Assessing large models (text-davinci-003, ChatGPT, and GPT-4) using performance data from fully supervised state-of-the-art models. - Virtual Knowledge Extraction: Examining the virtual knowledge capabilities of large models on the constructed VINE dataset. - Automatic KG: Proposing methods to utilize multiple agents to facilitate the construction and reasoning of knowledge graphs. Tech stack: Python Why it matters: This project aims to explore the capabilities and opportunities for large language models in constructing and reasoning knowledge graphs, contributing to advancements in AI and natural language processing. Repository: zjunlp/AutoKG Primary language: Python GitHub stars: 473 GitHub description: [WWWJ 2024] LLMs for Knowledge Graph Construction and Reasoning: Recent Capabilities and Future Opportunities

Overview: JetStream is a throughput and memory optimized engine designed for efficient LLM inference on XLA devices, including TPUs and GPUs. Key features: - Support for both Jax and Pytorch models - Optimized for throughput and memory efficiency - Primarily implemented using Python Tech stack: Python Why it matters: JetStream provides a high-performance solution for LLM inference, making it ideal for applications requiring fast and resource-efficient inference on XLA devices.

Overview: This repository contains an official implementation of the paper "Deep ViT Features as Dense Visual Descriptors". The code is developed in Python and requires specific modules for efficient performance. Key features: - ViT Extractor: A wrapper class for a ViT model to extract dense visual descriptors. - Part Co-segmentation: A notebook for co-segmentation tasks using deep ViT features. - Co-segmentation: A notebook for co-segmentation tasks using deep ViT features. Tech stack: Python, PyTorch, Faiss, Timm, Matplotlib, OpenCV, Scikit-learn. Why it matters: This implementation demonstrates the effectiveness of deep features extracted from a self-supervised, pre-trained ViT model as dense patch descriptors, showcasing its potential for real-world vision tasks such as co-segmentation and point correspondence.

Overview: SwiftInfer is an efficient implementation of Streaming-LLM using TensorRT, designed to improve the performance and production readiness of LLM inference. Key features: - TensorRT-based implementation of Streaming-LLM - Production-grade inference with stable APIs - Supports LLM models like Llama-2-7b-chat-hf and Lmsys/Vicuna-7b-v1.3 Tech stack: Python, TensorRT, PyTorch Why it matters: SwiftInfer simplifies the process of deploying LLM models by providing a stable and efficient solution, making it easier to integrate LLM inference into production environments.

Overview: This repository contains the slides and Jupyter notebooks for Kat's ["Python from Space"](https://us.pycon.org/2017/schedule/presentation/364/) talk at Pycon 2017, focusing on Python tools for processing and analyzing open satellite imagery for Earth imaging satellites. Key features: - Python tools for using Open Street Maps data - Geospatial Data Abstraction Library (GDAL) - OpenCV and NumPy for image processing Tech stack: Jupyter Notebook Why it matters: This repository provides a practical guide for Python developers interested in using satellite imagery for data science and social good, covering topics such as satellite data ecosystem, Python tools for image processing, and real-world examples of analyzing satellite imagery for environmental and social research.

Overview: NimTorch is a PyTorch frontend written in Nim, aiming to provide a native C++ performance with ease of use in Python. Key features: - Automatically generated, from `Declarations.yaml`, the full ATen API - Cuda support (add `-d:cuda` when compiling with Nim) - WASM support (add `-d:wasm` when compiling with Nim) - Automatically generated, from `derivatives.yaml`, gradient procs - Autograd - More high-level PyTorch API (Module, Models, etc) Tech stack: Nim Why it matters: Nim compiles to C++, making it a bare-metal native C++ performance language, while still providing the ease of use of Python. GitHub stars: 472 Repository: sinkingsugar/nimtorch

Overview: Master-skill is a Buddhist AI persona framework powered by FoJin, designed for those interested in Buddhist studies, including practitioners, researchers, and students seeking specific insights into Buddhist teachings. It provides a source-grounded, boundary-aware, and fidelity-tested framework for understanding and engaging with Buddhist teachings across various traditions. Key features: - Source-grounded: The framework uses authentic and reliable sources for its teachings. - Boundary-aware: It respects the traditional boundaries and contexts of different Buddhist traditions. - Fidelity-tested: The framework ensures high fidelity in its translations and interpretations. - Runtime-ready: It is designed to be easily integrated into various applications and environments. Tech stack: Python Why it matters: Master-skill offers a practical and reliable way to explore and engage with Buddhist teachings, making it accessible to a wide range of users, including practitioners, researchers, and students. Its source-grounded and boundary-aware approach ensures that users can gain a deeper understanding of Buddhist teachings across various traditions. Repository: xr843/Master-skill GitHub stars: 472 GitHub description: FoJin-powered Buddhist AI persona framework — source-grounded, boundary-aware, fidelity-tested, runtime-ready.

Overview: This repository contains my solutions for labs and programming assignments from Coursera courses, focusing on deep learning techniques. Key features: - Custom Models, Layers, and Loss Functions with TensorFlow - Functional APIs - Multi-output models using Keras Functional API - Creating a custom loss function - Implementing a quadratic layer Tech stack: Jupyter Notebook Why it matters: This repository serves as a practical guide for understanding and implementing advanced deep learning techniques, providing insights into custom model creation, loss function design, and layer customization.

### Overview: 本次调研旨在对信息抽取领域的学术界和工业界进行总结，涵盖实体识别、关系抽取、属性抽取、属性值抽取和事件抽取五个子任务。本次调研将为自然语言处理信息抽取领域的相关学者和研究人员提供帮助，保持定期更新和跟踪前沿技术。 ### Key features: - 实体抽取与链指 - 关系抽取 - 属性抽取 - 属性值抽取 - 事件抽取 ### Tech stack: Python, NLP libraries, and tools. ### Why it matters: 信息抽取是将海量内容自动分类、提取和重构的重要技术，广泛应用于知识图谱、信息检索、问答系统、情感分析和文本挖掘等领域。通过本次调研，可以了解学术界和工业界对信息抽取领域的最新研究进展和实际应用情况，为相关研究提供参考和借鉴。

Overview: TokenFlow is an innovative image tokenizer designed for multimodal understanding and generation, surpassing state-of-the-art models like LLaVA-1.5 and EMU3. It introduces a dual-codebook architecture that decouples semantic and pixel-level feature learning while maintaining alignment through shared mapping. Key features: - Unified image tokenizer for multimodal understanding and generation - Innovative dual-codebook architecture - Significantly outperforms LLaVA-1.5 and EMU3 in multimodal understanding - Comparable performance to SDXL in 256×256 resolution for text-to-image generation Tech stack: Python Why it matters: TokenFlow offers a novel approach to bridging the gap between multimodal understanding and generation, making it a valuable tool for researchers and developers working in the field of AI and computer vision. Repository: ByteVisionLab/TokenFlow GitHub stars: 471 GitHub description: Official impl. of "TokenFlow: Unified Image Tokenizer for Multimodal Understanding and Generation".

### Overview: Amadeus is a real-time multimodal desktop agent designed to integrate voice, character presence, and agent work into a single, seamless interface. It enables users to communicate naturally, have their character present, delegate tasks to specialized providers, and stay in control of their projects and progress. ### Key Features: - **Real-time Communication:** Seamless integration of voice and text communication. - **Character Presence:** Voice, subtitles, lip sync, and scene behavior all play on the same timeline. - **Task Delegation:** Specialized providers handle tasks, while the host manages identity, state, and persistence. - **Control and Recovery:** Projects, progress, and permissions remain visible and recoverable. ### Tech Stack: Python ### Why It Matters: Amadeus addresses the limitations of traditional voice assistants, desktop characters, and execution agents by providing a unified interface for communication, presence, and task delegation. This allows for a more efficient and intuitive way of managing tasks and maintaining control over projects, ensuring that users can stay engaged and informed throughout their work. Repository: Code-Amadeus/Amadeus

Overview: EgoLife is an ambitious project that captures daily activities of six participants over a week using Meta Aria glasses, synchronized third-person cameras, and mmWave sensors. The goal is to provide a rich dataset for long-term video understanding and to develop AI assistants that can support memory, habit tracking, event recall, and task management, advancing real-world egocentric AI applications. Key features: - **EgoGPT**: An omni-modal vision-language model fine-tuned on egocentric datasets for continuous video captioning and extracting key events, actions, and context. - **EgoRAG**: A retrieval-augmented generation module that retrieves relevant past events and synthesizes contextualized answers to user queries. Tech stack: Python Why it matters: EgoLife aims to create a comprehensive dataset and AI solutions that enhance the understanding and management of daily activities, particularly for individuals with limited mobility or visual impairments, by leveraging the power of egocentric AI. Repository: EvolvingLMMs-Lab/EgoLife GitHub stars: 471 GitHub description: [CVPR 2025] EgoLife: Towards Egocentric Life Assistant

Overview: Sunsimiao中文医疗大模型是一个基于Qwen2-7B微调的7B量级模型，旨在提供安全、可靠、普惠的中文医疗大模型。 Key features: - 提供安全、可靠、普惠的中文医疗大模型 - 由Qwen2-7B微调而来 - 在医学问答和医学考试中表现出色 Tech stack: Python Why it matters: Sunsimiao中文医疗大模型致力于传承孙思邈的医疗经验，通过高质量医疗数据微调，提供安全、可靠、普惠的中文医疗大模型，助力医疗领域的发展。

### Overview: Graph-Cut RANSAC is a novel algorithm designed for robust estimation of homography, fundamental matrix, essential matrix, and 6D pose estimation. This algorithm was introduced in the paper "Graph-Cut RANSAC" by Daniel Barath and Jiri Matas, which was presented at the Conference on Computer Vision and Pattern Recognition (CVPR) in 2018. The method is based on the Graph-Cut RANSAC framework and offers improved performance and robustness compared to traditional RANSAC techniques. ### Key features: - **Robust Estimation**: Graph-Cut RANSAC provides accurate and reliable estimates for homography, fundamental matrix, essential matrix, and 6D pose estimation. - **Enhanced Performance**: The algorithm is designed to handle noisy data more effectively, improving overall accuracy. - **Versatile Applications**: Suitable for various computer vision tasks such as structure from motion, correspondence-based pose estimation, and more. ### Tech stack: C++ The implementation is written in C++ and leverages modern C++ features like C++17 support. ### Why it matters: Graph-Cut RANSAC offers significant improvements in accuracy and robustness compared to traditional RANSAC methods, making it a valuable tool for computer vision applications requiring high precision and reliability in estimating geometric transformations. ### Repository: danini/graph-cut-ransac The repository contains the source code for the Graph-Cut RANSAC algorithm, including the paper's implementation and additional experiments. It is available at [this link](http://openaccess.thecvf

### Overview: Label Maker is a Python library designed to prepare satellite imagery for machine learning training by downloading and saving OpenStreetMap QA tiles and satellite imagery tiles as `.npz` files. ### Key features: - Downloads OpenStreetMap QA tiles and satellite imagery tiles. - Saves them as `.npz` files for machine learning training. - Supports Python 3.6 and requires `tippecanoe` as a peer dependency. ### Tech stack: Python ### Why it matters: Label Maker simplifies the process of preparing satellite imagery for machine learning, making it easier to train models on satellite data.

### Overview: CortexDB is a lightweight, pure Go implementation of AI memory and a knowledge graph. It stores all data in a single SQLite file, making it highly efficient and scalable. Unlike traditional models, it doesn't require an embedding model, making it ideal for applications that need to interact with existing databases or systems. ### Key Features: - **Pure Go Implementation:** CortexDB is written entirely in Go, ensuring high performance and low resource usage. - **Knowledge Graph and Memory:** It supports both knowledge graph and memory functionalities, allowing for the storage and retrieval of complex data structures. - **Vector-based Retrieval:** Uses vectors for efficient querying, making it suitable for large datasets. - **RAG (Reinforcement for Agents) Integration:** Seamlessly integrates with Claude Code and Codex for advanced AI capabilities. - **RDF/SPARQL Support:** Provides RDF and SPARQL support, enabling more advanced querying and data integration. - **Cypher Query Language:** Supports the Cypher query language for powerful graph operations. - **80+ MCP Tools:** Compatible with a wide range of machine learning and knowledge graph tools. - **No Embedding Model Required:** Works without the need for an embedding model, making it versatile for various use cases. ### Tech Stack: - **Go:** Purely written in Go, ensuring high performance and minimal dependencies. ### Why It Matters: - **Efficiency and Scalability:** CortexDB is designed to be highly efficient and scalable, making it suitable for large-scale applications. - **Versatility:** Its compatibility with a wide range of

### Overview: The project "Conditional Image-to-Video Generation with Latent Flow Diffusion Models" (LFDM) is a PyTorch implementation of a paper presented at CVPR 2023. This project focuses on generating video sequences from images using latent flow diffusion models, with a particular emphasis on text-conditioned image-to-video generation. The LFDM model is designed to handle various datasets, including MUG, MHAD, and NATOPS, and includes pretrained models for these datasets. ### Key Features: - **Latent Flow Diffusion Models (LFDM)**: A novel approach for image-to-video generation that leverages latent flow diffusion models. - **Text-Conditioned Generation**: The project supports text-conditioned generation, allowing for the creation of videos with specific textual prompts. - **Multi-Dataset Support**: The LFDM model is designed to work with multiple datasets, including MUG, MHAD, and NATOPS. - **Pretrained Models**: The project provides pretrained models for MUG, MHAD, and NATOPS datasets, which can be used as starting points for further training. ### Tech Stack: - **Programming Languages**: Python - **Frameworks/Library**: PyTorch ### Why It Matters: The LFDM project is significant because it offers a practical solution for generating high-quality video sequences from images, especially with the ability to incorporate textual prompts. This technology has applications in various fields, including video editing, content creation, and even in the creation of realistic videos for training AI models. ### Repository: The project is hosted on GitHub under the repository