Projects

### Overview: This repository contains detailed lecture notes for the Stanford CS231n course on Deep Learning for Computer Vision, focusing on the Spring 2025 semester. The notes cover essential topics such as image classification, regularization, neural networks, convolutional networks, and more. ### Key Features: - **Lecture Notes**: Comprehensive coverage of CS231n topics. - **Image Classification**: Discusses the challenges of image classification and introduces key concepts like k-nearest neighbors and linear classifiers. - **Regularization and Optimization**: Explains regularization techniques like L1 and L2, gradient descent, and optimization methods like Adam and AdamW. - **Neural Networks and Backpropagation**: Covers the basics of neural networks, activation functions, and the chain rule. - **Convolutional Networks**: Delivers a deep dive into convolutional networks, including padding, pooling, and equivariance. - **CNN Architectures**: Explores CNN architectures like batch normalization, dropout, and residual connections. ### Tech Stack: - **Python** - The primary programming language used for the notes. ### Why It Matters: These notes are designed to be read in depth, providing a thorough understanding of deep learning concepts and techniques in computer vision. They are intended for developers and researchers looking to enhance their knowledge in this field. ### Repository: The repository is hosted on GitHub and has been starred by 42 users, indicating its popularity and usefulness among the developer community. ### Primary Language: The primary language used in this repository is Python. ### GitHub Stars: The repository has been starred

Overview: This project combines human expertise with Large Language Models (LLMs) like OpenAI's GPT-3.5 to simplify data annotation and model enhancement. It employs Iterative Active Learning for continuous improvement and integrates CleanLab (Confident Learning) to ensure high-quality datasets and better model performance. Key features: 1. Dataset Uploading and Annotation 2. Manual Annotation Corrections 3. CleanLab: Confident Learning Approach 4. Data Versioning and Saving 5. Model Training Tech stack: Python Why it matters: This framework saves time by leveraging the efficiency of LLMs and ensures high-quality datasets and better model performance through confident learning, making it valuable for data annotation and model fine-tuning.

Overview: This repository contains the code for "Post-pre-training for Modality Alignment in Vision-Language Foundation Models" (CVPR 2025), focusing on the application of post-pre-training techniques to enhance modality alignment in vision-language foundation models. Key features: - Implementation of post-pre-training for modality alignment in vision-language foundation models. - Utilization of the COCO Caption dataset for post-pre-training and ImageNet for evaluation. Tech stack: Python Why it matters: This project offers a practical approach to improving modality alignment in vision-language models through post-pre-training techniques, potentially leading to better performance in downstream tasks. Repository: yshinya6/clip-refine Primary language: Python GitHub stars: 42

Overview: EasyDetect is an easy-to-use framework for detecting hallucinations in multimodal large language models (MLLMs) like GPT-4V, Gemini, and LLaVA. It provides a unified approach to categorize and detect hallucinations across different modalities. Key features: - Unified Multimodal Hallucination Detection - MHalluBench dataset for benchmarking - Easy installation and quick start guide Tech stack: Python Why it matters: EasyDetect simplifies the process of detecting hallucinations in MLLMs, making it accessible for researchers and developers. Repository: zjunlp/EasyDetect GitHub stars: 42 GitHub description: [ACL 2024] An Easy-to-use Hallucination Detection Framework for LLMs.

Overview: 本项目提供了一个从零开始学习大模型Transformer、GPT2和BERT预训练和微调的全面教程。通过详细的代码示例和图文解析，帮助读者理解模型构建、预训练和微调的全过程。 Key features: - Transformer模型构建 - GPT2模型构建 - BERT模型构建 - Transformer预训练 - GPT2预训练 - GPT2微调 - BERT微调 - Transformer微调 Tech stack: Jupyter Notebook Why it matters: 本项目为学习和实践大模型的构建、预训练和微调提供了全面的指导，适合希望从零开始掌握这些技术的开发者。

### Overview: OpenFlowLM is an open-source project that provides optimized NPU kernels for AMD Ryzen™ AI chips, replacing the closed NPU kernels with open ones. It supports all Ryzen™ AI Series chips with XDNA2 NPUs (Strix, Strix Halo, Kraken, and Gorgon Point) and enables running LLMs, embedding models, and MoE models on AMD Ryzen™ AI NPUs without the need for a GPU. ### Key features: - **Open kernels**: The AIE designs the engine dispatches, sourced from the open source repository. - **Second embedding backend**: Six encoder models beyond the one upstream ships. - **GGUF and Q4_K containers**: Models are not confined to one weight format. - **Built from source**: Use CMake presets for building. ### Tech stack: - **C++**: The primary language used for development. ### Why it matters: - **Runs on the NPU**: Not the GPU, and not as a CPU fallback. - **Open kernel path**: The designs are here, built via CMake presets. - **Long context**: Up to 256k tokens on models that support it. - **Familiar CLI**: `run`, `serve`, `list`, `bench`. ### GitHub stars: 40 ### Acknowledgements: - Forked from [FastFlowLM](https://github.com/ROCm/FastFlowLM) - Powered by the **AMD Ryzen™ AI NPU** architecture - Inspired by [llama.cpp](https://github.com/gg

### Overview: SAR-SSL is a Python implementation of "Self-Supervised Learning of Spatial Acoustic Representation with Cross-Channel Signal Reconstruction and Multi-Channel Conformer" [TASLP 2024]. This project focuses on developing a novel method for self-supervised learning of spatial acoustic representation, utilizing cross-channel signal reconstruction and multi-channel conformer architectures. The repository includes contributions to the field of spatial acoustic representation learning and multi-channel audio signal processing. ### Key Features: - **Self-supervised learning of spatial acoustic representation (SSL-SAR)**: The project introduces the first self-supervised learning method specifically designed for spatial acoustic representation learning and multi-channel audio signal processing. - **Multi-channel audio Conformer (MC-Conformer)**: This architecture is unified for both the pretext and downstream tasks, learning the local and global properties of spatial acoustics present in the time-frequency domain, thereby boosting the performance of both pretext and downstream tasks. ### Tech Stack: Python ### Why it Matters: This project is significant for researchers and practitioners in the field of spatial acoustic representation and multi-channel audio signal processing, offering a novel approach to self-supervised learning that can be applied across various acoustic applications.

### Overview: ConsistencyTTA is an advanced code implementation for accelerating the diffusion-based text-to-audio (TTA) generation process. This project was developed by Microsoft Applied Science Group and UC Berkeley, focusing on a novel framework called Consistency Distillation. The model reduces the computational complexity of diffusion-based TTA models by a factor of 400, making it more efficient and faster to train. ### Key features: - **Consistency Distillation Framework**: A method to train TTA models using a single neural network query, significantly reducing the computational cost. - **Classifier-Free Guidance**: Incorporates this framework into the distillation process, preserving the quality and diversity of diffusion models. - **End-to-End Fine-Tuning**: Utilizes novel loss functions such as the CLAP score for improved performance, allowing for seamless fine-tuning with minimal additional complexity. ### Tech stack: The project is implemented using PyTorch as the central tool, with extensive use of HuggingFace's Accelerator package. ### Why it matters: This work introduces a significant improvement in the efficiency and performance of diffusion-based TTA models, making them more practical and accessible for various applications in audio synthesis and generation.

Overview: LynseDB is an embedded-first vector database designed for AI applications moving from local prototypes to self-hosted services and lightweight sharded clusters. It supports a variety of vector types and provides a unified collection API for dense and sparse vectors, metadata filtering, and more. Key features: - Support for multiple vector types: embeddings, distance paths, binary fingerprints, aligned numeric profiles, and probability distributions. - Built-in clustering: sharded group sharding, stable hash bucket routing, and replica write mirroring for a self-hosted cluster. - Python client for easy integration: embedded storage, HTTP server, and coordinator-backed cluster. - Rust core: vector storage, search, indexes, filters, and server execution. - AI-native retrieval: supports Haversine, Tanimoto, Dice, and more for various data types. - Small operational footprint: single Python process for development, HTTP service for production. - Cluster mode: shard groups, health checks, primary promotion for multi-node setups. - Cost-conscious: local mode for development, network and cluster overhead for larger datasets. Tech stack: Rust. Why it matters: LynseDB simplifies the transition from local prototypes to self-hosted services and lightweight sharded clusters, making it ideal for applications requiring strong retrieval primitives without a heavyweight database stack. Repository: BirchKwok/lynsedb GitHub stars: 41 GitHub description: Embedded-first vector database that grows into a self-hosted service and lightweight cluster.

Overview: DengBoCong/prompt-tuning is a pipeline designed for Prompt-tuning, integrating various Prompt-tuning methods and search template strategies. It provides a complete execution pipeline for Prompt-tuning. Key features: - Integrates multiple Prompt-tuning methods and search template strategies. - Offers a complete execution pipeline for Prompt-tuning. Tech stack: Python Why it matters: This repository is valuable for developers looking to implement Prompt-tuning pipelines efficiently, leveraging the latest methods and strategies to enhance text classification and natural language inference tasks.

Overview: Dive into the latest research from ICML 2025 with this comprehensive collection of papers, including cutting-edge deep learning, generative AI, optimization techniques, reinforcement learning, and more. All code implementations are included for practical application. Key features: - Comprehensive coverage of ICML 2025 papers - Latest research in deep learning, generative AI, optimization, and reinforcement learning - Code implementations for practical application Tech stack: Python, Jupyter Notebook, Markdown, and GitHub Actions Why it matters: Stay updated with the cutting-edge of machine learning research, ensuring your projects are at the forefront of innovation. Repository: https://github.com/DmitryRyumin/ICML-2025-Papers GitHub stars: 41

### Overview: This project is an Android version assistant designed to help users with visual impairments manage their devices, even when the screen is broken. It is a continuation of the original V.I.S.O.R. project, aiming to provide a solution for blind users to interact with their devices without the need for a working screen. ### Key Features: - **Voice Command Recognition:** Allows users to control the phone using voice commands. - **Device Management:** Manages the phone's settings, notifications, and other features. - **Screen Reader Integration:** Integrates with screen readers to provide alternative input methods. ### Tech Stack: Java ### Why it Matters: This project is essential for users with visual impairments who need to manage their devices without the use of a screen. It provides a practical solution for those who cannot rely on traditional screen readers or other assistive technologies. ### Repository: Edw590/VISOR---Android-Version-Assistant GitHub Stars: 41 ### Primary Language: Java ### Download: The most recent release can be downloaded from the [Releases](https://github.com/DADi590/VISOR---A-better-Android-assistant/releases). ### Roadmap: - Enhance user interface for better accessibility. - Expand feature set to include more advanced functionalities. - Improve performance and stability. ### Project Status: The project is in active development and improvement. ### License: The project is licensed under the [MIT License](https://github.com/DADi590/VISOR

### Overview: This repository contains a minimal multi-gpu implementation of EDM2, a paper focused on analyzing and improving the training dynamics of diffusion models. The implementation includes configs and experiments on small or medium-scale datasets such as CIFAR-10/100 and Tiny-ImageNet. ### Key Features: - Config C: Unconditional EDM2C model - Config E: Unconditional EDM2E model - Config G: Final EDM2G model ### Tech Stack: PyTorch ### Why it Matters: This repo provides insights into the improvements proposed by the authors in the paper "Analyzing and Improving the Training Dynamics of Diffusion Models" by Karras et al. by comparing the performance of different configurations. The configurations include Config C, Config E, and Config G, which differ in their network sizes and training dynamics. The repo also includes sample scripts for generating images and training models on CIFAR-10/100 and Tiny-ImageNet datasets. ### Results: - Config C shows consistent improvements over the original EDM. - Config E could perform better but has significantly slower convergence. - Config G favors latent-space modeling and performs inferior to Config E on pixel-space generation. - Post-hoc EMA (and power function EMA) tends to favor longer training durations, which is not beneficial for small datasets like CIFAR. ### Usage: - Use 4 GPUs to train Config C, E, and G models on CIFAR-100. - Generate 50,000 images with different checkpoints

Overview: This tutorial provides a comprehensive guide on how to utilize the kedro-mlflow plugin to synchronize training and inference, and serve a Kedro pipeline, focusing on best practices for code organization and reproducibility. Key features: - Synchronizing training and inference - Maintaining consistency between training and inference - Serving a Kedro pipeline - Best practices for code organization and reproducibility Tech stack: Python Why it matters: This tutorial offers valuable insights into using kedro-mlflow for advanced mlops practices, ensuring that training and inference processes are synchronized and that models are easily deployable and reproducible. Repository: Galileo-Galilei/kedro-mlflow-tutorial Primary language: Python GitHub stars: 41

### Overview: LabLLM is a free, native macOS app that streamlines the entire process of training and fine-tuning small language models. It allows users to build a Transformer from scratch, train models on their own datasets, leverage Apple Silicon and MLX for local training, and monitor live training metrics and loss curves. The app also supports checkpoint saving and resuming, fine-tuning with LoRA and DPO, inspecting tokens and embeddings, and chatting with the model. ### Key Features: - **Build Transformer from Scratch:** Start with a blank slate and build your own Transformer. - **Train on Custom Datasets:** Use your own datasets to train the model. - **Local Training on Apple Silicon:** Train models locally with MLX acceleration for faster performance. - **Live Metrics and Loss Curves:** Monitor training progress in real-time. - **Save and Resume Checkpoints:** Save checkpoints and resume training from any point. - **Fine-Tuning with LoRA and DPO:** Fine-tune the model using LoRA and DPO for better performance. - **Inspect Tokens and Embeddings:** Analyze the model's tokens and embeddings. - **Chat with the Model:** Engage with the trained model for interactive sessions. - **Model Management:** Manage models, datasets, and experiments seamlessly. ### Tech Stack: Swift ### Why It Matters: Training a small language model should be a hands-on experience rather than a complex task. LabLLM simplifies the process, making it accessible and straightforward. It allows users to explore and experiment with language models directly, without the

Overview: Kõnele service is an Android app that provides speech-to-text functionality, specifically for the Kõnele app, utilizing an open-source speech recognition server. Key features: - Speech-to-text service integration - Compatible with Android's SpeechRecognizer interface - Open-source speech recognition server software (Kaldi-GStreamer-Server) Tech stack: Java Why it matters: This app enhances communication between Kõnele and other apps by offering a speech-to-text feature, improving user interaction and accessibility. Repository: Kaljurand/K6nele-service Primary language: Java GitHub stars: 41

Overview: This project is a reverse-engineered API for Stable Diffusion XL 1.0, offering alternative text-to-image capabilities via Midjourney. It generates beautiful 1024x1024 images. Key features: - Reverse-engineered Stable Diffusion XL API - Generates 1024x1024 images - Supports text prompts and advanced parameters Tech stack: Python Why it matters: This project provides an alternative to Midjourney, offering similar image generation capabilities in Python, making it useful for developers and artists who prefer Python-based solutions. Repository: KoushikNavuluri/stable-diffusion-xl-api Primary language: Python GitHub stars: 41 GitHub description: Reverse engineered API of Stable Diffusion XL 1.0 (Midjourney Alternative), A text-to-image generative AI model that creates beautiful 1024x1024 images. README excerpt: # Stable Diffusion XL ( API ) Reverse engineered API of Stable Diffusion XL 1.0 ( Midjourney Alternative ) via https://replicate.com/ , A text-to-image generative AI model that creates beautiful 1024x1024 images. <img src="https://github.com/KoushikNavuluri/stable-diffusion-xl-api/assets/103725723/50a81ef7-4f4e-4b46-a881-a55726b8e21

### Overview: MM-INF is an efficient multi-modal instruction-following data synthesis tool that leverages multiple Large Language Models (LLMs) and Model-Linking Language Models (MLLMs). It automatically configures a pipeline using a single YAML file, making it highly flexible and scalable. The tool is particularly useful for generating diverse and high-quality multimodal data, including instruction-response pairs and detailed image captions, without requiring extensive manual configuration. ### Key Features: - **Automated Pipeline Configuration**: Configured using a single YAML file, making it highly flexible and scalable. - **Efficient Data Generation**: Generates diverse and high-quality multimodal data, including instruction-response pairs and detailed image captions. - **Versatile Tasks**: Supports a wide range of tasks, such as generating detailed image captions, multimodal training data based on images and captions, and instruction-response pairs. - **Minimal Configuration**: Requires only a single YAML file to configure, making it easy to extend and customize the pipeline. ### Tech Stack: - Python ### Why It Matters: - **Efficiency**: Automatically configures a pipeline using a single YAML file, making it highly efficient and scalable. - **Versatility**: Supports a wide range of tasks, including generating detailed image captions, multimodal training data, and instruction-response pairs. - **Ease of Use**: Minimal configuration required, making it easy to extend and customize the pipeline. ### Repository: - **Repository**: Letian2003/MM_INF - **GitHub Stars**: 41 - **GitHub Description**: An efficient multi-modal instruction-following data synthesis

### Overview: This repository is an official PyTorch implementation of the VSPBFR project, which focuses on Visual Style Prompt Learning Using Diffusion Models for Blind Face Restoration. The project aims to address the challenges of blind face restoration by introducing a visual style prompt learning framework that utilizes diffusion probabilistic models to generate visual prompts for high-quality restoration. ### Key Features: - Training and testing codes - Pre-trained models ### Tech Stack: Python ### Why it Matters: This project provides a comprehensive solution for blind face restoration, leveraging advanced techniques in pattern recognition and diffusion models to achieve high-quality restoration. The use of visual style prompts enhances the restoration process, making it more effective in capturing fine details and extracting informative patterns. The project's implementation and extensive experiments demonstrate its superiority in achieving high-quality blind face restoration.

### Overview: mAIcro is an open-source AI infrastructure designed to centralize organizational knowledge and answer questions via RAG (Retrieval-Augmented Generation). It is built and maintained by the Dev Department of MicroClub, a computer science club at USTHB (University of Science and Technology Houari Boumediene, Algiers). ### Key Features: - **RAG-Powered Q&A**: Google Gemini with hybrid search (semantic vector + keyword BM25) and Reciprocal Rank Fusion. - **Real-Time Discord Sync**: Gateway WebSocket listener handles message CREATE, EDIT, and DELETE instantly. - **Temporal Query Intelligence**: Understands "what happened today?" and "what was the last message?" - **Question Normalization**: Rewrites slang ("whats", "wanna", "gonna"), augments time-aware queries. - **Startup Audit**: Reconciles offline Discord edits and deletes on every restart. - **Stateless Architecture**: All cursors and vectors live in Qdrant Cloud; no local database. - **Rate Limit Resilience**: Exponential backoff with jitter plus optional secondary LLM fallback. - **Production-Ready**: Multi-stage Docker, health checks, graceful reconnection. ### Tech Stack: Python ### Why It Matters: mAIcro is designed to centralize organizational knowledge and provide accurate answers to questions via RAG, making it a valuable tool for communities and organizations. Its stateless architecture and production-ready performance make it suitable for cloud deployment and integration with Discord.