Projects

### Overview: 本教程提供了一系列自然语言处理（NLP）任务的范例，涵盖了文本词向量、词法分析、预训练语言模型、文本分类、文本语义匹配、文本生成、信息抽取、机器翻译和对话系统。通过这些范例，读者可以学习如何使用Python和PyTorch实现这些任务的基础知识和实际应用。 ### Key Features: - **文本词向量**：提供了一个利用领域数据集提升词向量效果的例子。 - **词法分析**：实现了一个基于LSTM和CRF的词法分析任务。 - **语言模型**：展示了多个语言模型，如BERT、ELECTRA、ELMo、GPT等，并提供了支持语言模型在垂直领域数据上继续训练的工具包。 - **文本分类**：使用机器学习和深度模型完成文本分类任务。 - **文本匹配**：实现了一个基于SBERT的文本匹配算法，适用于搜索、推荐系统排序和召回等场景。 - **文本生成**：包含一个基于BERT面向生成任务的预训练+微调模型框架，以及一个GPT模型的应用。 - **信息抽取**：提供了一个实体识别的基线实现，包括人名、地址名称的命名实体识别。 - **机器翻译**：实现了一个带Attention机制的基于LSTM的多层RNN Seq2Seq翻译模型。 - **对话系统**：提供了一个基于BERT的问答模型。 ### Tech Stack: Jupyter Notebook 本教程使用Python和Jupyter Notebook进行

Overview: SMFS is a filesystem designed for agents, offering SOTA retrieval, automatic memory profiles, and a sync engine. It supports various file types and can be mounted as a local directory or integrated into agent toolsets. Key features: - SOTA retrieval - Automatic memory profiles - Sync engine - Supports multiple file types (pdf, images, videos) Tech stack: Rust Why it matters: SMFS simplifies the process of managing and accessing agent memory, making it easier for developers to work with and analyze data in real-time. Repository: supermemoryai/smfs GitHub stars: 490 GitHub description: A filesystem designed for agents, with SOTA retrieval, automatic memory profiles, sync engine. Drop any file type (pdf, images, videos), and grep through them.

### Overview: VisualWebArena is a benchmark designed to evaluate the capabilities of multimodal autonomous agents. It comprises a set of realistic and diverse web-based visual tasks that assess various aspects of these agents. ### Key Features: - **Realistic Tasks**: Tasks simulate real-world scenarios, ensuring the evaluation is as close to practical use as possible. - **Multi-Modal Agents**: Evaluates the performance of agents across multiple modalities, including visual, textual, and audio. - **Scalability**: Supports a wide range of environments and websites, making it suitable for both research and production use. - **Comprehensive**: Covers a broad spectrum of tasks, providing a comprehensive evaluation of agents. ### Tech Stack: - Python 3.10 - Pre-commit: Code style and linting tool - Black: Code formatting - MyPy: Static type checking - Beartype: Code beautification ### Why It Matters: VisualWebArena provides a robust and realistic benchmark for evaluating the performance of multimodal autonomous agents in real-world scenarios. It helps researchers and developers understand how agents perform across different modalities and environments, ensuring that their work is both practical and effective. ### Repository: Repository: **web-arena-x/visualwebarena** Primary Language: **Python** GitHub Stars: **490** GitHub Description: **VisualWebArena is a benchmark for multimodal agents.**

### Overview: vs-mlrt is a comprehensive VapourSynth ML runtime project that supports multiple platforms, including x86 CPUs, Intel GPUs, NVidia GPUs, AMD GPUs, and Apple SoCs. It provides optimized pure CPU and Intel GPU runtimes for popular AI filters, including waifu2x, DPIR, RealESRGANv2/v3, Real-CUGAN, RIFE, SCUNet, ArtCNN, and more. The project also offers a Python wrapper for easy usage and a unified interface to select different backends. ### Key features: - **vsov: OpenVINO-based Pure CPU & Intel GPU Runtime** - **vsort: ONNX Runtime-based CPU/GPU Runtime** - **vstrt: TensorRT-based GPU Runtime** - **vsmigx: MIGraphX-based GPU Runtime** - **vsort: ONNX Runtime-based CPU/GPU Runtime** - **vsncnn: NCNN-based GPU (Vulkan) Runtime** ### Tech stack: C++ ### Why it matters: vs-mlrt simplifies the usage of AI filters by providing optimized runtimes for different hardware platforms, making it easier and faster to integrate AI-powered effects into VapourSynth projects.

Overview: This repository contains the implementation of various state-of-the-art text classification models using PyTorch. Key features: - FastText model - TextCNN model - TextRNN model - RCNN model - CharCNN model - Seq2seq with attention model - Transformer model Tech stack: Python 3.5.0, Pandas 0.23.4, NumPy 1.15.2, Spacy 2.0.13, PyTorch 0.4.1.post2, Torchtext 0.3.1. Why it matters: This repository provides a comprehensive implementation of various text classification models, showcasing the latest advancements in the field. It is designed to be easily adaptable for various datasets and can be used for both academic research and practical applications. Repository: AnubhavGupta3377/Text-Classification-Models-Pytorch Primary language: Python GitHub stars: 489

### Overview: This repository provides a Tensorflow-based implementation of the FLAME 3D head model, demonstrating how to sample 3D head meshes, fit the model to 2D or 3D keypoints, and generate textured head meshes from images. The code is based on a scientific publication and includes additional functionalities for speech-driven facial animation. ### Key features: - Sample 3D face meshes - Fit the 3D model to 2D or 3D keypoints - Fit the 3D model to registered 3D meshes - Sample the texture space - Generate templates for speech-driven facial animation ### Tech stack: Python ### Why it matters: This repository is essential for researchers and developers interested in 3D face modeling, facial expression synthesis, and applications in computer graphics and virtual reality. It provides a practical toolset for creating realistic 3D face models and animations, making it valuable for both academic research and industrial applications.

Overview: MiVOS is a modular interactive video object segmentation framework that achieves state-of-the-art performance in both DAVIS and YouTube semi-supervised evaluations. It includes features for dense correspondences, propagation, and fusion, making it highly versatile and effective for interactive video object segmentation tasks. Key features: - Dense Correspondences - Propagation - Fusion Tech stack: Python Why it matters: MiVOS provides a comprehensive solution for interactive video object segmentation, offering advanced features that significantly improve segmentation accuracy and efficiency, making it a valuable tool for researchers and developers working in this field. Repository: hkchengrex/MiVOS Primary language: Python GitHub stars: 489 GitHub description: [CVPR 2021] Modular Interactive Video Object Segmentation: Interaction-to-Mask, Propagation and Difference-Aware Fusion. Semi-supervised VOS as well!

Overview: Rectified Flow is an implementation of a flow-based model that uses rectified flows for image synthesis and transformation tasks. This model is built on PyTorch and aims to improve upon existing flow-based models by incorporating rectified flows and their follow-up research. Key features: - Rectified flow implementation - Unet-based architecture - Support for training and inference using PyTorch - Examples provided for quick testing on datasets like Oxford Flowers Tech stack: Python Why it matters: Rectified Flow is significant because it provides a more efficient and effective way to generate and manipulate images, making it a valuable tool in various applications such as image synthesis, style transfer, and more. Repository: lucidrains/rectified-flow-pytorch GitHub stars: 489 GitHub description: Implementation of rectified flow and some of its followup research / improvements in Pytorch README excerpt: <img src="./rf.png" width="400px"></img>

Overview: Agency is a Python library designed to simplify the creation of agent-based applications, enabling developers to experiment and build upon minimal foundations. Key features: - Easy to use API - Performance and scalability - Observability and control - Multiple agent examples for experimentation Tech stack: Python Why it matters: Agency provides a minimal foundation for creating custom agent-based applications, making it a valuable tool for developers looking to experiment and build upon minimal frameworks. Repository: operand/agency Primary language: Python GitHub stars: 489 GitHub description: A fast and minimal framework for building agentic systems

### Overview: DNA-Diffusion is a diffusion-based model designed for the generation of 200bp cell type-specific synthetic regulatory elements. It utilizes a small subset of the DHS Index dataset for training and can be used to recreate the dataset and train the model. ### Key Features: - **Model Type**: Diffusion-based model - **Dataset**: Small subset of the DHS Index dataset for training - **Sequence Generation**: Generates synthetic regulatory elements - **Training**: Predefined config file for training the diffusion model on the DHS Index dataset ### Tech Stack: Python ### Why it Matters: DNA-Diffusion offers a novel approach to generating synthetic regulatory elements, potentially aiding in the development of personalized and cell type-specific therapeutic strategies.

Overview: FlashVideo is a project that aims to generate high-resolution videos from low-resolution inputs in a fast and efficient manner. It uses a combination of 3D VAE and stage models to enhance the fidelity and detail of the generated videos. Key features: - Stage-I weight for 270P video generation - Stage-II for enhancing 270P videos to 1080P - Inference code for both stages - Training code and related augmentation - Implementation with diffusers - Gradio interface Tech stack: Python Why it matters: FlashVideo offers a practical solution for generating high-quality videos from low-resolution inputs, making it useful for applications requiring real-time video generation, such as in virtual reality, augmented reality, or content creation tools. Repository: FoundationVision/FlashVideo Primary language: Python GitHub stars: 488 GitHub description: [AAAI-2026]FlashVideo: Flowing Fidelity to Detail for Efficient High-Resolution Video Generation

### Overview: A Computer Vision based Traffic Signal Violation Detection System is developed using Python and Tkinter. The system automates the process of detecting traffic violations in real-time, allowing traffic police departments to efficiently monitor and enforce traffic regulations. ### Key features: - Vehicle Classification - Violation Detection - Real-time Traffic Signal Violation Detection ### Tech stack: Python ### Why it matters: This system helps traffic police departments to monitor and enforce traffic regulations more efficiently, reducing traffic violations and accidents, and ensuring safer roads for all.

Overview: This repository contains a collection of prompts designed to challenge the reasoning abilities of large language models when presented with misguiding information, drawing parallels to human behavior. The prompts are variations of common thought experiments and paradoxes, and the goal is to observe how LLMs respond to them. Key features: - Prompts based on logic puzzles - Evaluation of LLMs' responses to misguiding information - Tags for tracking progress over time Tech stack: Python Why it matters: This project aims to highlight the potential pitfalls of relying on pre-existing knowledge and patterns, demonstrating the importance of careful reasoning and avoiding the Einstellung effect.

### Overview: PCN is a learning-based method that completes point clouds directly without voxelization, making it a significant advancement in 3D vision research. ### Key features: - Directly maps partial point clouds to dense, complete point clouds - Based on our 3DV 2018 publication - Supports multiple datasets and models - Compatible with various programming languages and frameworks ### Tech stack: Python, Tensorflow 1.12, CUDA 9.0 ### Why it matters: PCN offers a novel approach to shape completion, improving the accuracy and efficiency of 3D reconstruction and analysis tasks.

### Overview: LLM-grounded Diffusion is a project that enhances the understanding of text-to-image diffusion models by integrating large language models (LLMs). This approach leverages LLMs to parse prompts and generate intermediate representations, which are then processed by Stable Diffusion to produce high-quality images. The project aims to improve the interpretability and performance of text-to-image diffusion models. ### Key features: - **LLM-Prompt Parsing**: Converts text prompts into a structured format that LLMs can understand. - **Intermediate Representation**: Uses LLMs to generate intermediate representations that guide the image generation process. - **Stable Diffusion Integration**: Combines the structured representations with Stable Diffusion for high-quality image generation. ### Tech stack: Python ### Why it matters: LLM-grounded Diffusion provides a novel way to enhance the interpretability and performance of text-to-image diffusion models by leveraging the capabilities of large language models. This approach can lead to more efficient and effective image generation, making it a valuable contribution to the field of computer vision and AI.

Overview: PDDLStream is a planning framework that integrates symbolic planners and blackbox samplers, extending PDDL to include streams for declarative specifications of sampling procedures. It aims to solve PDDL problems with only a blackbox description of each sampler. Key features: - Integrates symbolic planners and blackbox samplers - Extends PDDL with streams for declarative specifications of sampling procedures - Solves PDDLStream problems with domain-independent algorithms Tech stack: Python Why it matters: PDDLStream provides a flexible and powerful framework for AI planning, particularly useful in robotics applications where domain-specific knowledge is often limited. Repository: caelan/pddlstream Primary language: Python GitHub stars: 487

Overview: This repository contains links to machine learning exams, homework assignments, and exercises from various Carnegie Mellon University (CMU) courses, including 10-601, 10-701, and 10-702. These materials are designed to help students test their understanding of machine learning concepts and algorithms. Key features: Links to midterm exams, final exams, and practice problems for courses such as 10-601, 10-701, and 10-702. The repository also includes solutions for midterm and final exams. Tech stack: Python, NumPy, Pandas, Matplotlib, Scikit-learn, TensorFlow, PyTorch, and other relevant machine learning libraries. Why it matters: These resources are invaluable for students and professionals looking to test their knowledge of machine learning. They provide a practical way to assess understanding and prepare for exams, as well as to practice solving problems.

### Overview: Whatlies is a toolkit designed to help developers understand what lies within word embeddings. It offers a range of tools for creating interactive visualizations and supports multiple language backends, including spaCy, fasttext, tfhub, Hugging Face, and BPemb. The library is maintained by koaning and is useful for benchmarking and understanding the nuances of word embeddings. ### Key Features: - **Interactive Visualizations**: Tools to create and visualize word embeddings. - **Support for Multiple Language Backends**: spaCy, fasttext, tfhub, Hugging Face, and BPemb. - **Lightweight Featurization**: Featurizes embeddings using scikit-learn, making it easy to integrate with other machine learning pipelines. - **Scikit-Learn Featurization**: Enables seamless integration with scikit-learn models for further analysis. ### Tech Stack: Python ### Why It Matters: Understanding word embeddings is crucial in natural language processing and machine learning. Whatlies simplifies the process of visualizing and analyzing these embeddings, making it easier for developers to interpret their results and make informed decisions.

Overview: Matchbox is a tool designed to simplify the writing of PyTorch code by allowing researchers to write examples at the individual level and then run them efficiently on minibatches. Key features: - Support for `MaskedBatch` type to track padding and masking for variable-size data. - `@matchbox.batch` decorator to rewrite control flow into a SIMT-like form with execution masking and synchronization primitives. - Convenience methods like `batch_ones`, `split_dim`, and `causal_mask` for common use cases in dynamic neural network code. - Compatibility with `torchtext` and a test wrapper for ensuring numerical equivalence with loop-based examples. Tech stack: Python, PyTorch 0.4, `gast`, `astor`, and `six`. Why it matters: Matchbox streamlines the process of writing and testing deep learning models by providing a more intuitive way to handle variable-sized data and dynamic neural network operations, making it easier to experiment and iterate on research code.

Overview: ChatTS is a Time Series Multimodal Language Model (TS-MLLM) designed to understand and reason about time series data, similar to vision and audio models. It retains raw numerical values, enabling precise answers to value-related questions. Key features: - Question-answering capabilities - Native support for multivariate time series of different lengths and dimensions - Retains raw numerical values for accurate answers - Easy integration into standard LLM pipelines Tech stack: Python Why it matters: ChatTS enhances the understanding and reasoning capabilities of time series data, making it invaluable for applications requiring precise analysis and interpretation of time series information. Repository: NetManAIOps/ChatTS GitHub stars: 486 GitHub description: [VLDB' 25] ChatTS: LLM for Time Series Understanding and Reasoning