Coding & Development

PassGAN is an open-source deep learning tool for password guessing, implementing the approach described in the paper "PassGAN: A Deep Learning Approach for Password Guessing." This repository provides a modified TensorFlow implementation of Improved Training of Wasserstein GANs, making it easy to train and sample from the model. It includes a command-line interface for generating password samples and training custom models. A pretrained PassGAN model, trained on the RockYou dataset, is also provided. Users can train their own models using various password leaks and datasets, with instructions for downloading common datasets like the LinkedIn leak. The tool is released under an MIT License, acknowledging the original authors of the PassGAN paper and the underlying WGAN training code.

pnnx (PyTorch Neural Network eXchange) is an open standard designed for PyTorch model interoperability. It offers an open model format for PyTorch, meticulously defining computation graphs and high-level operators to strictly match PyTorch's architecture. The tool enables users to optimize their PyTorch models, reduce dependencies on extension packages, and convert models between various formats like TorchScript, ONNX, and NCNN. pnnx also facilitates the export of models to a portable pnnx format, ONNX-zero, or NCNN, making them suitable for deployment on different platforms. It supports both Python pip installation and portable binary packages, offering flexibility for developers.

bondingAI offers an AI Operating System designed for enterprises, integrating data, AI intelligence, and business workflows into a unified platform. Powered by xLLM, a proprietary enterprise language model, it focuses on delivering private, deterministic, and explainable AI solutions. The platform aims to address common enterprise AI challenges such as security, cost, and hallucinations, enabling businesses to run operations through a single intelligent interface. It supports information query, data analytics, actions, agentic rules, and explainable AI, allowing for custom AI model building tailored to specific business data, processes, and culture. bondingAI emphasizes privacy, accountability, and fairness, with features like human-in-the-loop governance and private deployment options.

Soft-NMS is an open-source algorithm designed to enhance the accuracy of object detection models. It works by intelligently re-scoring bounding box predictions, providing a more robust alternative to traditional Non-Maximum Suppression (NMS). The tool is integrated with popular object detectors such as R-FCN and Faster-RCNN, allowing users to easily incorporate Soft-NMS into their existing pipelines. It supports both linear and Gaussian weighting schemes, with configurable parameters for fine-tuning. Soft-NMS has demonstrated significant performance improvements in challenges like COCO 2017, where it was adopted by many top-performing submissions. The repository provides code for testing models and includes updated ROI Pooling layers for improved interpolation.

sgmse is an open-source repository offering official PyTorch implementations of Score-based Generative Models, also known as Diffusion Models, specifically tailored for speech enhancement and dereverberation tasks. It includes code for various research papers, allowing users to reproduce results and build upon existing models. The repository provides pretrained checkpoints for different datasets and tasks, such as speech enhancement on VoiceBank-DEMAND and WSJ0-CHiME3, and dereverberation on WSJ0-REVERB. It supports training and evaluation with options for various SDEs and backbone networks, catering to both 16 kHz and 48 kHz models. Detailed installation instructions and logging options (W&B or local CSV) are also provided, making it a valuable resource for researchers and practitioners in audio processing.

SparseR-CNN is an advanced end-to-end object detection model that leverages learnable proposals, eliminating the need for hand-crafted proposals common in traditional object detection systems. This approach allows for more efficient and potentially higher-performing detection across various computer vision applications. The tool provides different configurations with varying backbone models like ResNet and PVT, demonstrating competitive inference and training times. It is built upon established frameworks such as Detectron2 and DETR, ensuring a robust and scalable architecture. SparseR-CNN is suitable for researchers and developers working on object detection, offering detailed installation and usage instructions for training, evaluation, and visualization.

iml is an R package designed to provide interpretable machine learning capabilities. It allows users to understand the behavior and predictions of various machine learning models, regardless of their underlying architecture. The package implements several model-agnostic interpretability methods, including feature importance, partial dependence plots, individual conditional expectation plots (ICE), accumulated local effects, tree surrogate models, and Local Interpretable Model-agnostic Explanations (LIME). It also supports Shapley values for explaining single predictions. This makes iml a valuable tool for data scientists and developers who need to gain insights into complex AI models and ensure transparency in their decision-making processes.

SpectralCluster is a Python-based open-source library that re-implements advanced spectral clustering algorithms, particularly those used in Google's speaker diarization research. It provides functionalities for speaker diarization, including refined Laplacian matrix calculations, constrained spectral clustering, and multi-stage clustering. The tool allows users to customize various parameters such as minimum and maximum clusters, Laplacian type, refinement operations, and distance metrics for K-Means. It also supports auto-tuning for optimal performance and offers fallback clusterers for smaller datasets or specific conditions. SpectralCluster is designed for researchers and developers working on speech recognition and audio analysis, offering both standard and streaming prediction capabilities.

Transformer-time-series-prediction is an open-source project offering a proof of concept for transformer-based time series prediction models. It features two distinct PyTorch models: one for single-step predictions and another for multi-step predictions. While designed as a demonstration, the repository highlights the models' ability to learn long-term trends from training data, as shown with the daily minimum temperature dataset. Users interested in serious applications are directed to the flow-forecast package, indicating this tool is primarily for research, experimentation, or understanding the underlying concepts of transformer models in time series forecasting rather than production-ready deployment. The project is available on GitHub under an MIT license.

Corgea is an AI-native application security platform designed to autonomously detect, triage, and fix insecure code, packages, infrastructure, and containers within a single workflow. It aims to replace fragmented security scanners with one control plane, enabling teams to maintain security without slowing down development. Key features include AI SAST for higher-signal static analysis with accurate fixes, Dependency Scanning to prioritize exploitable packages, IaC Scanning to prevent cloud misconfigurations, Container Scanning for image-level risk prioritization, and Secrets Scanning to detect leaked credentials. Corgea also offers Code Quality Scanning and Attack Surface Mapping, providing a comprehensive solution for modern application stacks across backend, frontend, and package managers. It integrates seamlessly with developer workflows, offering IDE and SCM integrations like GitHub, GitLab, Azure DevOps, and Bitbucket, and provides developer-friendly feedback and AI-generated fixes directly in pull requests.

VideoPipe is a cross-platform video structuring and analysis framework developed in C++. It is designed with minimal dependencies and an easy-to-use pipeline architecture, where independent nodes can be combined to create diverse video analysis applications. The framework supports various tasks including object detection, image classification, feature extraction, and behavior analysis, similar to NVIDIA's DeepStream and Huawei's mxVision but with greater portability and ease of use. It integrates with different inference backends like OpenCV::DNN, TensorRT, PaddleInference, and ONNXRuntime, and now supports Multimodal Large Language Model (mLLM) integration. VideoPipe is ideal for scenarios such as video structuring, image search, face recognition, and traffic/security behavior analysis.

EdgeCortix is an AI platform company focused on delivering energy-efficient AI processors and acceleration solutions. Founded in 2019, the company applies a software-first approach to designing AI-specific processor architectures from the ground up. Their flagship product, SAKURA-II, is highlighted as a flexible and energy-efficient AI accelerator capable of handling multi-billion parameter models for generative AI at the edge. EdgeCortix also offers the MERA compiler and framework, an industry-first software platform for AI inference across heterogeneous systems, and Dynamic Neural Accelerator technology for flexible, run-time reconfigurable architectures. The company aims to provide near cloud-level performance at the edge with significantly better energy efficiency and processing speed, targeting industries such as embedded systems, enterprise, industrial, robotics, smart city, smart home, space, aerospace & defense.

Oha AI is a domain name currently listed for sale on Spaceship.com. The domain is available for purchase at a price of $1,270,000 USD. Spaceship.com facilitates secure transactions and provides guided transfer support to ensure a smooth process for buyers. The platform offers various payment methods and a buyer protection program. While the domain's original intent was to be a marketplace for pre-trained AI models, accelerating the development of new AI applications, its current status is purely as a domain for sale. Interested parties can make an offer or proceed with a direct purchase through the secure checkout system.

MLX-VLM is a comprehensive open-source package designed for developers and researchers working with Vision Language Models (VLMs) and Omni Models on Apple Silicon Macs. Leveraging the MLX framework, it facilitates both inference and fine-tuning of these advanced AI models directly on macOS. The tool offers a flexible command-line interface (CLI) for quick generation tasks, a Gradio-based chat UI for interactive use, and a FastAPI server for robust API integration and continuous batching. Key features include support for multi-modal inputs (image, audio, text), speculative decoding for faster generation, and KV cache quantization to optimize memory usage. It also provides detailed documentation for various supported models, making it a powerful solution for local VLM development.

MMCV is a foundational library developed by OpenMMLab for computer vision research, offering a comprehensive suite of functionalities. It supports various tasks including image and video processing, image and annotation visualization, and image transformation. The library also provides high-quality implementations of common CPU and CUDA operations, making it suitable for developing and deploying deep learning models. MMCV is compatible with Linux, Windows, and macOS, and requires Python 3.7+. It offers two versions: `mmcv` for full features with CUDA ops, and `mmcv-lite` for a lighter version without CUDA ops, catering to different development needs.

MixtralKit is an open-source toolkit designed for the inference and evaluation of Mistral AI's 'mixtral-8x7b-32kseqlen' model. It offers an experimental implementation of inference code, allowing users to perform text completion and other tasks with the Mixtral model. The toolkit also provides detailed performance comparisons with other models like Mistral-7B-v0.1, Llama2-70B, and Qwen-72B across various datasets and metrics. It includes resources such as model architecture details, instructions for downloading model weights from Hugging Face or via magnet link, and a comprehensive installation guide. MixtralKit is intended for researchers and engineers working with large language models, offering tools for both running and evaluating the Mixtral model.

Morph-Net is an open-source method for learning deep network structure during training, focusing on continuous relaxation of the network-structure learning problem. It induces activation sparsity by adding regularizers that target resource consumption like FLOPs or model size. The tool allows users to shrink existing models to satisfy constraints such as memory or latency by adjusting the number of output channels in each convolution layer. Morph-Net supports various regularizer types, including LogisticSigmoid for probabilistic channel regularization, GroupLasso for models without BatchNorm, and Gamma for models with BatchNorm. It provides functionality to export learned model structures in JSON format and offers examples for adding FLOPs regularizers.

The Local AI Playground is a free and open-source native application designed to simplify experimenting with AI models locally, requiring zero technical setup and no GPU. It features a Rust backend for memory efficiency and compactness, offering CPU inferencing that adapts to available threads and supports GGML quantization. The tool provides robust model management capabilities, including a resumable, concurrent downloader, usage-based sorting, and directory-agnostic storage. It also ensures model integrity through BLAKE3 and SHA256 digest verification. Users can easily start a local streaming server for AI inferencing in just two clicks, supporting streaming, quick inference UI, and remote vocabulary.

CLIKA Inc. offers an Automatic Compression Engine (ACE) SDK that functions as a universal compiler, optimizer, and translator for AI models. This proprietary engine intelligently preserves model performance while maximizing efficiency, reducing memory footprint by up to 90% and enhancing speed by up to 18x. It supports all types of AI models, including custom and fine-tuned ones, with current limitations under 15 billion parameters. CLIKA's ACE works in on-premise or air-gapped environments, ensuring data privacy. It supports various hardware platforms like Nvidia, Intel, AMD GPUs and CPUs, and Qualcomm (coming soon), by converting unsupported elements into optimized, supported alternatives. The compression pipeline outputs automatically compressed and compiled models, resulting in faster inference speeds with minimal accuracy loss.

TrainLoop is a post-training research and product lab dedicated to advancing AI capabilities. They specialize in developing algorithms, methods, and tooling to reliably train, steer, and deploy specialized AI systems. Their work focuses on training reasoning models that are aligned with specific organizational goals, offering expertise in areas like Life Sciences for biological reasoning, Continual Training to prevent catastrophic forgetting, Information Theory for stable reasoning, and Evaluation & Interpretability tools. TrainLoop collaborates with organizations possessing unique data to train state-of-the-art reasoning models, frequently achieving state-of-the-art or pareto-optimal performance. They offer a structured research-to-production workflow to co-define objectives and sustain performance.

pytorch-segmentation-detection is an open-source library designed for image segmentation and object detection tasks using PyTorch. It offers a comprehensive set of tools, including pre-trained models and scripts, to facilitate the reproduction of reported results on various image segmentation and object detection datasets. The library supports models tested on datasets like PASCAL VOC 2012, Endovis 2017, and Cityscapes, with detailed performance metrics such as Mean IOU, pixel accuracy, and inference time. Developers and researchers can leverage this library to implement and evaluate deep learning models for computer vision applications, with provided installation instructions and a citation for academic use.

SOLIDER is a Semantic Controllable Self-Supervised Learning Framework designed to learn general human representations from large datasets of unlabeled human images. Unlike existing self-supervised learning methods, SOLIDER incorporates prior knowledge from human images to build pseudo semantic labels, enriching the learned representation with more semantic information. It also features a conditional network with a semantic controller, allowing it to adapt to different downstream tasks that require varying ratios of semantic and appearance information. This flexibility ensures that the learned representations can benefit a wide range of human-centric visual tasks, including person re-identification, person search, pedestrian detection, attribute recognition, human parsing, and pose estimation. The framework is implemented with Python and PyTorch, utilizing Swin-Transformer as its backbone.

TDK SensEI is an Edge AI platform designed to transform industrial operations by offering a comprehensive suite of tools, services, expertise, and systems. It enables the development, deployment, and scaling of industrial-grade Edge AI applications across diverse industrial environments. The platform is backed by a team of experienced AI engineers, ensuring robust support and advanced capabilities for users. TDK SensEI focuses on integrating AI directly into edge devices, facilitating real-time data processing and decision-making closer to the source of data generation, which is crucial for efficiency and responsiveness in industrial settings.

SReC is a PyTorch implementation of "Lossless Image Compression through Super-Resolution." This open-source tool leverages neural networks to achieve state-of-the-art compression rates on large datasets with practical runtimes. It supports full training, compression, and decompression functionalities, allowing users to convert images into highly compressed .srec files and then back to PNG without loss. SReC provides pre-trained models for datasets like ImageNet64 and Open Images, making it accessible for immediate use. The project is hosted on GitHub, offering a comprehensive solution for efficient and high-quality image data management.