Projects

### Overview: This repository contains the source code for my personal column "Machine Learning Without Tears" on Zhihu. It includes projects in Natural Language Processing (NLP) and Computer Vision, such as text generation, machine translation, deep convolutional GANs, and other practical code examples. ### Key features: - Text generation using LSTM models - Word2Vec model for skip-gram algorithm - Lyrics generation using RNN - Seq2Seq model for sorting words - Image denoising using autoencoder - MNIST handwritten digit generation using GANs - CIFAR image classification using CNNs - MNIST digit generation using DCNNs - Batch normalization implementation - Machine translation model using Seq2Seq - Attention mechanism and BiRNN in machine translation - Sentiment analysis model using DNN, LSTM, and CNN - Image style transfer model using TensorFlow ### Tech stack: Python 3.6 ### Why it matters: This repository provides a comprehensive collection of practical projects and code examples in Natural Language Processing and Computer Vision, making it useful for developers looking to enhance their skills in these areas.

Overview: ManimML is a project focused on creating animations and visualizations of common machine learning concepts using the Manim Community Library. Key features: - Animations of neural networks - Visualizations of convolutional neural networks - Demonstration of max pooling - Explanation of activation functions - Introduction to neural network dropout Tech stack: Python Why it matters: ManimML aims to provide a comprehensive set of visualizations for machine learning concepts, making it easier for users to understand complex ideas in a more intuitive way.

### Overview: PyTracking is a comprehensive Python framework for visual object tracking and video object segmentation, built on PyTorch. ### Key features: - **TaMOs (WACV 2024)**: A generico object tracker for multiple generic objects. - **RTS (ECCV 2022)**: Robust end-to-end trainable segmentation-centric pipeline. - **ToMP (CVPR 2022)**: Segmentation-centric pipeline with instance localization. - **KeepTrack (ICCV 2021)**: Efficient visual tracking with exemplar transformers. - **LWL (ECCV 2020)**: Lightweight, efficient tracker. - **KYS (ECCV 2020)**: Key-Value instance tracker. - **PrDiMP (CVPR 2020)**: Pruning-based instance tracker. - **DiMP (ICCV 2019)**: Deep instance-based tracker. - **ATOM (CVPR 2019)**: Advanced tracking model. ### Tech stack: PyTorch. ### Why it matters: PyTracking provides a robust and efficient toolkit for visual object tracking and segmentation, making it a valuable resource for researchers and developers in the field. Repository: visionml/pytracking Primary language: Python GitHub stars: 3520 GitHub description: Visual tracking library based on PyTorch.

Overview: AliceVision is a Photogrammetric Computer Vision Framework that provides robust software basis for 3D reconstruction and camera tracking algorithms. Key features: - 3D Reconstruction - Camera Tracking - Open-source framework Tech stack: C++ Why it matters: AliceVision offers cutting-edge computer vision algorithms that can be tested, analyzed, and reused for production usage. Repository: alicevision/AliceVision GitHub stars: 3517 GitHub description: 3D Computer Vision Framework

### Overview: This repository contains a comprehensive cheat sheet designed specifically for AI engineering interviews. It covers a wide range of topics from foundational knowledge to advanced concepts, ensuring that candidates are well-prepared for various roles in the field. ### Key Features: - **Must Know**: Essential terms like LLM, RAG, MCP, Agent, Fine-tuning, and Quantization are explained in detail. - **LLM Fundamentals**: Understanding of LLMs, their working, and how they've transformed AI engineering. - **Prompt Engineering**: Techniques for creating effective prompts. - **Retrieval-Augmented Generation (RAG)**: How RAG models work and their applications. - **AI Agents and Agentic Systems**: Concepts and best practices for designing AI-driven systems. - **Fine-Tuning and Model Adaptation**: Strategies for improving and adapting models. - **Vector Databases and Embeddings**: How vector databases and embeddings enhance AI systems. - **AI System Design**: Best practices for designing AI systems from scratch. - **LLMOps and Production AI**: Practical considerations for deploying AI in production environments. - **Evaluation and Testing**: Methods for evaluating AI models and ensuring their reliability. - **AI Safety, Ethics, and Responsible AI**: Key considerations for building responsible AI systems. - **Multimodal AI**: Integrating multiple types of data into AI models. - **AI Infrastructure and Scalability**: Best practices for building scalable AI infrastructure. - **Coding and Practical Implementation**: Practical examples and code snippets for implementing AI solutions. - **Behavioral and Scenario-Based Questions**: Common interview questions

Overview: OptaPlanner is an AI constraint solver designed to optimize planning problems such as vehicle routing, employee rostering, task assignment, and maintenance scheduling. Key features: - Vehicle routing optimization - Employee rostering - Task assignment - Maintenance scheduling - Conference scheduling Tech stack: Java Why it matters: OptaPlanner is a powerful tool for solving complex planning problems efficiently, making it essential for businesses looking to streamline their operations and improve efficiency.

### Overview: PennyLane is an open-source quantum software platform designed for quantum computing, quantum machine learning, and quantum chemistry. It provides a seamless transition from inspiration to implementation, leveraging a comprehensive library of research demos, interactive tutorials, and cutting-edge components to build meaningful quantum algorithms. ### Key Features: - **Inspiration to Implementation, Quickly:** PennyLane simplifies the process of translating theoretical ideas into practical quantum algorithms, thanks to its intuitive interface and extensive library of demonstrations and tutorials. - **Fast where it matters. Scalable where it counts:** PennyLane excels in performance, offering fast execution, efficient compilation, and scalable workflows. It supports industrial-grade performance with resource estimation and the Catalyst compiler, and can run on a variety of hardware devices from different quantum computing platforms. - **Hardware agnostic, hardware ready:** PennyLane seamlessly integrates with a broad range of quantum hardware, ensuring compatibility with a wide array of devices, including superconducting qubits, trapped ion systems, neutral atoms, and photonics. - **Participate, collaborate, innovate:** PennyLane fosters a vibrant community by providing resources for hardware estimation, circuit compilation, and hardware device compatibility, encouraging collaboration and innovation in the quantum computing field. ### Tech Stack: Python ### Why it Matters: PennyLane is the go-to platform for anyone looking to explore and develop quantum algorithms, making the transition from theory to practice both accessible and efficient. Its comprehensive library and cutting-edge components ensure that users can quickly and effectively translate their ideas into practical quantum applications, making it a valuable tool for researchers, engineers, and developers

Overview: This project introduces Adversarial Latent Autoencoders (ALAE), a novel autoencoder architecture designed to address the limitations of traditional autoencoders and generative adversarial networks (GANs). The ALAE architecture combines the generative power of GANs with the representational properties of autoencoders, enabling it to produce high-quality image reconstructions and manipulations. Key features: - Adversarial Latent Autoencoder (ALAE) - StyleGAN-based StyleALAE - High-resolution face image generation - Real-image-based face reconstructions and manipulations - First autoencoder capable of surpassing GANs in quality and resolution Tech stack: Python Why it matters: ALAE represents a significant advancement in the field of image generation by combining the strengths of both autoencoders and GANs, offering superior performance and capabilities compared to traditional methods. Repository: podgorskiy/ALAE GitHub stars: 3508 GitHub description: [CVPR2020] Adversarial Latent Autoencoders

### Overview: KuiperInfer 是一个由 zjhellofss 开发的高性能深度学习推理库项目，旨在通过 C++ 20 标准实现大模型的推理，包括 LLaMA2、Unet、YOLOv5、ResNet 等模型。该项目通过 CMake+Git 管理项目，支持 Cuda 加速和 Int8 量化，为校招、秋招、实习提供高质量项目。 ### Key features: - 支持多种深度学习模型的推理，包括 LLaMA2、Unet、YOLOv5、ResNet 等。 - 使用 C++ 20 标准编写代码，采用统一、美观的代码风格和良好的错误处理。 - 采用 CMake+Git 管理项目，支持 CPU 和 CUDA 两种后端实现，对大模型有良好支持。 - 实现常见的算子，如卷积算子、池化算子、全连接算子等。 - 在课程中提供详细的计算图设计和实现，以及优化算子执行的手段。 - 提供从零开始的深度学习推理框架项目，涵盖从设计到实现的全过程，包括项目管理、单元测试和基准测试。 ### Tech stack: C++ ### Why it matters: KuiperInfer 项目为开发者提供了一个从零开始构建高性能深度学习推理库的框架，适合校招、秋招、实习等场景，同时也为技术进阶和项目实战提供了丰富的资源和经验。

Overview: This repository contains a collection of machine learning tutorials primarily in Python 3, focusing on various topics such as deep learning, model deployment, operation research, reinforcement learning, and more. The content aims to provide a balanced introduction to machine learning using both mathematical notations and practical implementation. Key features: - Softmax Regression from scratch - Softmax Regression - Tensorflow hello world - Multi-layer Neural Network - Tensorflow - Convolutional Neural Network (CNN) - Tensorflow - Recurrent Neural Network (RNN) - Vanilla RNN - Tensorflow - Long Short-Term Memory (LSTM) - Tensorflow - RNN, LSTM - PyTorch hello world - Word2vec (skipgram + negative sampling) using Gensim - Model selection - Data science is software - Unbalanced data - Clustering old - Linear regression - Regularization - Networkx - Association rule - Big data - Data science is software - Genetic Algorithm (GA) - Unbalanced data - Clustering old - Linear regression - Regularization - Networkx - Association rule - Big data - Data science is software Tech stack: Python 3, Jupyter Notebook, NumPy, Pandas, Matplotlib, SciPy, TensorFlow, PyTorch, Keras, Scikit-learn, FastText, HuggingFace, ONNX, XGBoost, LightGBM, PyTorch, Gensim, H2O, OR-Tools

Overview: AnyLabeling is an AI-assisted data labeling tool that leverages YOLO and Segment Anything for efficient and accurate image annotation. Key features: - Auto-labeling with YOLOv8 for object detection - Auto-labeling with Segment Anything family, including SAM, SAM 2, SAM 2.1, and SAM 3 - Text detection, recognition, and KIE labeling - Hardware acceleration with CUDA, CoreML, DirectML, OpenVINO, and vendor NPU providers - Multiple languages support: English, Vietnamese, Chinese Tech stack: Python Why it matters: AnyLabeling simplifies the data labeling process by automating the labeling of images and text, making it an essential tool for developers and data scientists looking to speed up their workflows.

### Overview: OML 1.0: Fingerprinting is a tool designed to generate and embed secret fingerprints into Large Language Models (LLMs) through fine-tuning. This method allows AI model owners to protect their models by identifying them uniquely through specific queries and responses. The fingerprints act as distinct secret signatures, ensuring that models can only be verified by their owners. ### Key Features: - **Fingerprint Generation**: Generates secret fingerprints for LLMs through fine-tuning. - **Model Verification**: Allows model owners to verify the authenticity of their models by testing them with specific queries. - **Model Distribution**: Enables model users to verify the exact model they are interacting with by using their fingerprints. ### Tech Stack: - Python ### Why It Matters: Model owners can protect their LLMs by embedding them with fingerprints, making it impossible for unauthorized users to access or use the models. This ensures compliance and reduces the risk of misuse.

Overview: This is a comprehensive Code-First Introduction to Natural Language Processing (NLP) course, designed to provide a practical understanding of the field through hands-on projects and real-world applications. Key features: - Code-based learning approach - Focus on practical applications and real-world projects - Coverage of various NLP techniques and tools - Use of libraries such as sklearn, nltk, pytorch, and fastai - Topics include topic modeling, sentiment classification, language modeling, and translation with deep learning Tech stack: Python with Jupyter Notebooks Why it matters: This course is designed to help developers and data scientists understand and apply NLP techniques to real-world problems, providing a practical understanding of the field and its applications.

Overview: min(DALL·E) is a fast, minimal port of DALL·E Mini to PyTorch, designed for efficient inference and minimal resource usage. Key features: - Fast inference speed - Minimal dependencies - Supports seamless grid generation - Supports both mega and non-mega models Tech stack: Python Why it matters: This lightweight implementation is ideal for applications requiring high performance and low resource consumption, such as real-time generation or embedded systems. Repository: [kuprel/min-dalle](https://github.com/kuprel/min-dalle) GitHub stars: 3490 GitHub description: min(DALL·E) is a fast, minimal port of DALL·E Mini to PyTorch. It has been stripped down for inference and converted to PyTorch. The only third-party dependencies are numpy, requests, pillow, and torch. This project is a fast, minimal port of Boris Dayma's [DALL·E Mini](https://github.com/borisdayma/dalle-mini) (with mega weights). It has been stripped down for inference and converted to PyTorch. The only third-party dependencies are numpy, requests, pillow, and torch. To generate a 3x3 grid of DALL·E Mega images, it takes: - 55 sec with a T4 in Colab - 33 sec with a P100 in Colab - 15 sec with an A10G on Hugging Face Here's a more detailed breakdown of performance on an A

### Overview: This project showcases the use of Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs) for human activity recognition using smartphone sensor data. The dataset consists of six activities: walking, walking up stairs, walking down stairs, sitting, standing, and lying down. The approach leverages raw, preprocessed data without extensive feature engineering, making it a simpler alternative to classical signal processing methods. ### Key Features: - **Activity Recognition**: Classifies six different human activities. - **Dataset**: Uses a smartphone dataset with preprocessed sensor data. - **Model**: LSTM RNN architecture for time-series data processing. - **Preprocessing**: Filters out gravitational force from accelerometer signals. - **GitHub Stars**: 3,489 ### Tech Stack: - **Python**: Main programming language. - **TensorFlow**: Deep learning library for building and training the LSTM model. ### Why it Matters: The project demonstrates a straightforward yet effective approach to human activity recognition, making it a valuable example for researchers and developers interested in simplifying machine learning applications.

### Overview: StudioGAN is a Pytorch library that provides implementations of representative Generative Adversarial Networks (GANs) for conditional/unconditional image generation. It aims to offer an identical playground for modern GANs, enabling machine learning researchers to easily compare and analyze new ideas. ### Key features: - **7 GAN Architectures**: StudioGAN includes 7 GAN architectures as representatives. - **9 Conditioning Methods**: It supports 9 conditioning methods. - **4 Adversarial Losses**: The library implements 4 adversarial losses. - **13 Regularization Modules**: 13 regularization modules are included. - **6 Augmentation Modules**: 6 augmentation modules are provided. - **8 Evaluation Metrics**: StudioGAN offers 8 evaluation metrics. - **5 Evaluation Backbones**: 5 evaluation backbones are supported. ### Tech stack: Pytorch ### Why it matters: StudioGAN provides an unprecedented-scale benchmark for generative models, including results from GANs (BigGAN-Deep, StyleGAN-XL), auto-regressive models (MaskGIT, RQ-Transformer), and Diffusion models (LSGM++, CLD-SGM, ADM-G-U). It offers a comprehensive benchmark with standard datasets (CIFAR10, ImageNet, AFHQv2, FFHQ) and supports various acceleration methods for training, including single GPU, data-parallel, distributed data-parallel, multi-node distributed data-parallel, and mixed-precision training. ### Repository: POSTECH-CVLab/PyTorch-StudioGAN - GitHub

### Overview: Maestro is a cross-platform desktop app designed for power users who live on the keyboard and rarely touch the mouse. It orchestrates AI agents and projects, allowing for long-running unattended sessions with clean context. ### Key Features: - 🌳 **Git Worktrees**: Run AI agents in parallel on isolated branches, creating worktree sub-agents from the git branch menu. - 🤖 **Auto Run & Playbooks**: File-system-based task runner that processes markdown checklists or pursues free-text objectives through AI agents. - 💬 **Group Chat**: Delegate moderation, appointing one agent as moderator to route questions and follow up with answers. - 📣 **Cross-Agent Mentions**: Inline consultation with other agents in chat, forwarding relevant slices of conversation. ### Tech Stack: TypeScript ### Why It Matters: Maestro is a high-velocity solution for hackers who are juggling multiple projects in parallel, designed for power users who live on the keyboard and rarely touch the mouse. It allows for long-running unattended sessions with clean context, making it ideal for collaborative workflows and detailed specification documents.

Overview: AnyCrawl is a powerful Node.js/TypeScript crawler designed for turning websites into LLM-ready data and extracting structured SERP results from multiple engines. Key features: - SERP crawling - Web scraping - Site crawling - Multi-threading - Batch tasks - LLM extraction Tech stack: TypeScript Why it matters: AnyCrawl is essential for developers looking to automate web scraping and crawling tasks efficiently, extracting structured data from various search engines, and preparing data for large language models. Repository: any4ai/AnyCrawl Primary language: TypeScript GitHub stars: 3482 GitHub description: AnyCrawl 🚀: A Node.js/TypeScript crawler that turns websites into LLM-ready data and extracts structured SERP results from Google/Bing/Baidu/etc. Native multi-threading for bulk processing.

Overview: The OpenMMLab YOLO series toolbox and benchmark includes implementations of various YOLO variants such as RTMDet, RTMDet-Rotated, YOLOv5, YOLOv6, YOLOv7, YOLOv8, YOLOX, PPYOLOE, and more. Key features: - RTMDet - RTMDet-Rotated - YOLOv5 - YOLOv6 - YOLOv7 - YOLOv8 - YOLOX - PPYOLOE Tech stack: Python Why it matters: This repository is part of the OpenMMLab project, which is a collection of open-source computer vision tools and benchmarks. It provides a comprehensive suite of YOLO-based object detection models and benchmarks, making it a valuable resource for researchers and developers working in the field of computer vision.

### Overview: PicoGPT is an unnecessarily tiny implementation of GPT-2 in plain NumPy, featuring minimal features and a fast forward pass. The entire forward pass code is 40 lines of code, making it the smallest implementation of GPT-2 available. ### Key features: - Fast? ❌ Nah, picoGPT is megaSLOW 🐌 - Training code? ❌ Error, 4️⃣0️⃣4️⃣ not found - Batch inference? ❌ picoGPT is civilized, single file line, one at a time only - top-p sampling? ❌ top-k? ❌ temperature? ❌ categorical sampling?! ❌ greedy? ✅ - Readable? `gpt2.py` ✅ `gpt2_pico.py` ❌ - Smol??? ✅✅✅✅✅✅ YESS!!! TEENIE TINY in fact 🤏 ### Tech stack: Python ### Why it matters: A quick breakdown of each of the files: - `encoder.py` contains the code for OpenAI's BPE Tokenizer, taken straight from their gpt-2 repo. - `utils.py` contains the code to download and load the GPT-2 model weights, tokenizer, and hyper-parameters. - `gpt2.py` contains the actual GPT model and generation code which we can run as a python script. - `gpt2_pico.py` is the same as `g