Research & Education

Text2Human is an official PyTorch implementation for text-driven controllable human image generation, as presented in the SIGGRAPH 2022 paper. This open-source tool enables users to create human images by providing text descriptions that specify clothing shapes and textures. It includes a comprehensive framework for training and sampling, utilizing a large-scale, high-quality DeepFashion-MultiModal Dataset with rich multi-modal annotations. Researchers and developers can leverage its capabilities for tasks like generating images from parsing maps or human poses, and it offers a user interface for interactive text-to-human image generation. The project also provides pretrained models and detailed installation instructions, making it a valuable resource for AI research in computer graphics.

Solving Inverse Problems with FLAIR is an AI tool available on Hugging Face that allows users to tackle common inverse problems in image processing. It provides functionalities for both inpainting and super-resolution. For inpainting, users can upload a photo and draw a mask over the areas they wish to replace. For super-resolution, the tool takes a low-resolution picture and enhances its detail. The platform also allows users to write a short description of their desired outcome, guiding the AI in its processing. This tool is suitable for anyone needing to restore or enhance images through AI-driven solutions.

Colorization is an open-source project that leverages deep neural networks for automatic image colorization. Developed by Richard Zhang, Phillip Isola, and Alexei A. Efros, it was first presented at ECCV in 2016. The tool also incorporates functionality from "Real-Time User-Guided Image Colorization with Learned Deep Priors" from SIGGRAPH 2017, allowing for interactive colorization. Users can clone the GitHub repository, install dependencies, and then use Python scripts to colorize images. It provides pre-trained colorizers for both ECCV 2016 and SIGGRAPH 2017 models, with clear instructions for integration into Python projects, including necessary pre and post-processing steps like Lab space conversion and resizing.

BuildingMachineLearningSystemsWithPython is an open-source repository containing the complete source code for the book "Building Machine Learning Systems with Python" by Luis Pedro Coelho and Willi Richert. This resource is invaluable for students, teachers, and professionals looking to understand and implement machine learning systems using Python. The code corresponds to the second edition of the book, published in 2015, and provides practical, hands-on examples for various machine learning concepts. It serves as a direct companion to the book, allowing users to explore, run, and modify the code to deepen their understanding of the topics covered. The repository is hosted on GitHub, making it easily accessible for anyone interested in learning or teaching machine learning with Python.

BinaryNet.pytorch offers a PyTorch implementation of Binarized Neural Networks (BNN), specifically designed for VGG and ResNet models. This open-source tool allows researchers and developers to delve into the world of binarized neural networks, which are known for their efficiency in terms of memory and computational resources. The project is hosted on GitHub and provides the necessary code to run models like resnet18 for datasets such as cifar10. It serves as a valuable resource for those looking to understand, implement, or experiment with BNNs within the PyTorch framework, building upon existing work in the field.

ConvNetDraw is a small, open-source tool designed for creating multi-layer neural network diagrams within a web browser. Users can visualize complex neural network architectures by simply entering a script, making it accessible for quick diagram generation. The project is hosted on GitHub and encourages contributions, indicating an active development community and potential for future enhancements. While straightforward in its current functionality, it provides a valuable resource for researchers, students, and developers looking to illustrate their network designs without needing specialized software.

d2l-tvm is an open-source project dedicated to deep learning compilers, offering comprehensive resources for those looking to understand and optimize deep learning models. Hosted on GitHub, it provides a platform for learning about the TVM deep learning compiler stack. The project includes detailed documentation, practical examples, and guides on how to contribute, making it a valuable resource for developers and researchers. It covers various aspects of deep learning compilation, from common operators and CPU/GPU schedules to deployment strategies, enabling users to dive deep into the technical intricacies of optimizing AI models.

garage is a comprehensive, open-source toolkit designed for developing and evaluating reinforcement learning (RL) algorithms, emphasizing reproducibility in research. It offers a wide array of modular tools, including composable neural network models, high-performance samplers, replay buffers, and an expressive experiment definition interface. The toolkit supports logging to various outputs like TensorBoard, ensures reliable experiment checkpointing and resuming, and provides environment interfaces for popular benchmark suites. garage is compatible with Python 3.6+ and supports both PyTorch and TensorFlow for neural network implementations, with algorithms not requiring neural networks found in the `garage.np` package. Its robust testing strategy, including continuous integration and comprehensive benchmarks, ensures state-of-the-art performance and reliability.

dmol-book is an open-source project offering a comprehensive book on deep learning for molecules and materials. Hosted on GitHub, this resource allows users to access and build the book locally using Jupyter Book, providing a flexible and customizable learning experience. The repository includes all necessary files and instructions for local setup, making it ideal for researchers and students who want to delve into the intersection of deep learning and scientific applications. It covers various topics relevant to chemistry and materials informatics, serving as a valuable educational tool for those interested in the field.

Deep-Learning-for-Tracking-and-Detection is a comprehensive open-source repository on GitHub, offering a curated collection of papers, datasets, code, and other resources specifically focused on object tracking and detection using deep learning. This tool is invaluable for AI researchers, engineers, and students who are actively engaged in computer vision projects. It covers a wide array of topics including static detection (RCNN, YOLO, SSD, RetinaNet, Anchor Free), video detection (Tubelet, FGFA, RNN), and multi-object tracking (Joint-Detection, Identity Embedding, Association, Deep Learning, RNN, Unsupervised Learning, Reinforcement Learning, Network Flow, Graph Optimization). The repository also provides resources for single object tracking, various deep learning techniques, and a multitude of datasets, making it a central hub for cutting-edge research and development in this field.

DANN provides a PyTorch implementation of the Domain-Adversarial Training of Neural Networks (DANN) paper, enabling unsupervised domain adaptation through backpropagation. This open-source tool is designed for researchers and developers working with neural networks who need to improve model performance across different data distributions or domains without extensive labeled data for the target domain. It includes the necessary network structure and training scripts, with specific instructions for setting up the environment using PyTorch 1.0 and Python 2.7. Users can download the required mnist_m dataset from provided links to begin training. The project also offers a separate version, DANN_py3, for Python 3 and Docker environments, indicating ongoing development and support for modern setups. Its primary utility lies in allowing models trained on one domain to generalize effectively to another, reducing the need for costly data annotation in new environments.

CityFlow is an open-source multi-agent reinforcement learning environment specifically designed for large-scale city traffic scenarios. It features a microscopic traffic simulator that models the behavior of individual vehicles, offering a high level of detail for traffic evolution. The tool supports flexible definitions for road networks and traffic flow, making it adaptable to various urban layouts. With its friendly Python interface, CityFlow is well-suited for reinforcement learning applications in traffic management. It boasts fast simulation capabilities due to elaborately designed data structures and multithreading, allowing it to simulate city-wide traffic efficiently. This makes it a valuable resource for researchers and engineers working on urban traffic management and planning, enabling them to test and develop advanced traffic control algorithms.

pyRiemann is an open-source Python machine learning package designed for processing and classifying real or complex-valued multivariate data. It leverages the Riemannian geometry of symmetric or Hermitian positive definite matrices, offering a high-level interface that mimics the scikit-learn API. While generic for multivariate data analysis, it's specifically tailored for biosignals like EEG, MEG, or EMG in brain-computer interface (BCI) applications, including motor imagery, event-related potentials, and steady-state visually evoked potentials. It also supports multisource transfer learning and remote sensing applications, such as processing radar images. The package provides functionalities for estimating covariance matrices and classifying them, making it a powerful tool for researchers and developers in these fields. It can be easily integrated into scikit-learn pipelines for comprehensive data analysis workflows.

smartcore is a comprehensive, fast, and ergonomic open-source library designed for machine learning and numerical computing in Rust. It enables developers to apply machine learning algorithms leveraging first principles, covering a broad range of methods including linear models, tree-based methods, ensembles, SVMs, neighbors, clustering, decomposition, and preprocessing. The library emphasizes production-friendly APIs, strong typing, and good defaults, while remaining flexible for research and experimentation. It features strong linear algebra traits with optional ndarray integration, WASM-first defaults for portability, and practical utilities for model selection, evaluation, and data access. smartcore is ideal for developers building AI applications in Rust who need robust and efficient ML capabilities.

Self-Driving-Cars is an open-source repository hosted on GitHub, offering a comprehensive collection of Coursera open courses from the University of Toronto. This resource is specifically designed for individuals interested in the field of self-driving car technology, providing access to videos, subtitles, and PDF materials. It's particularly beneficial for postgraduate students and researchers aiming to work on automotive motion planning, offering a structured and in-depth learning experience. The repository includes courses covering topics from an introduction to self-driving cars to state estimation, visual perception, and motion planning. Users can download and watch the content, and a rough notebook based on subtitles is provided for better review.

wilds is an open-source machine learning benchmark designed to evaluate models under real-world distribution shifts. It offers a comprehensive package including data loaders that automate downloading, processing, and splitting of datasets, along with standardized evaluators for consistent model assessment. The benchmark covers a wide range of data modalities and applications, from medical imaging (tumor identification) to environmental monitoring (wildlife monitoring) and socio-economic analysis (poverty mapping). It also provides example scripts with default models, optimizers, and training/evaluation code, making it easy for researchers to integrate new algorithms and run experiments across its 10 included datasets. The package is installable via pip and supports optional integration with Weights & Biases for experiment tracking.

theMLbook is an open-source GitHub repository offering Python code designed to replicate the illustrations found in 'The Hundred-Page Machine Learning Book'. This resource is invaluable for students and professionals seeking to deepen their understanding of machine learning concepts through practical, visual examples. By providing the exact code used for the book's figures, theMLbook allows users to interact directly with the algorithms and models discussed, facilitating a hands-on learning experience. It covers a range of machine learning topics, from fundamental algorithms like linear regression and K-means to more advanced concepts such as autoencoders and UMAP, making it a comprehensive companion for the book's readers.

WebGPU Depth Anything is an AI-powered tool hosted on Hugging Face Spaces that enables users to generate depth maps from uploaded images. Utilizing WebGPU technology, it processes images to estimate the distance of objects, providing a visual representation of depth. This tool is particularly useful for researchers and developers in computer vision, offering a straightforward way to analyze spatial relationships within images. Its web-based nature makes it easily accessible for quick demonstrations and experiments without requiring complex local setups.

Contextualized Topic Models (CTM) is a powerful Python package designed for advanced topic modeling. It integrates pre-trained language representations, such as BERT embeddings, with traditional topic models to produce highly coherent topics. The package offers two main models: CombinedTM, which merges contextual embeddings with bag-of-words for enhanced topic coherence, and ZeroShotTM, ideal for tasks with missing words in test data and cross-lingual topic modeling when trained with multilingual embeddings. CTM supports various languages through HuggingFace models and allows for the use of different embedding methods, ensuring adaptability to new advancements. It also includes 'Kitty,' a submodule for human-in-the-loop classification to quickly categorize documents and create named clusters. The tool is particularly effective when the bag-of-words size is restricted to around 2000 elements, and it provides a preprocessing pipeline to manage this. CTM uses SBERT for embedding creation, offering flexibility in choosing embedding models and handling multilingual data.

Introduction to Machine Learning with Python is a comprehensive open-source repository designed to accompany the book of the same name by Andreas Mueller and Sarah Guido. It provides all the notebooks and code examples used in the book, making it an invaluable resource for students and practitioners looking to learn machine learning with Python. The repository includes helper functions from the `mglearn` library for creating figures and datasets, and all necessary datasets are included, with the exception of `aclImdb`. Users can set up their environment using `conda` or `pip` to install required packages like `numpy`, `scipy`, `scikit-learn`, `matplotlib`, `pandas`, `pillow`, and `graphviz`. It also supports `nltk` and `spacy` for text processing chapters.

Machine-Learning-in-Action is an open-source GitHub repository offering practical code implementations for various machine learning algorithms, all based on the popular book "Machine Learning in Action." Developed in Python 3, this resource is designed to help users understand and apply machine learning concepts through hands-on examples. The repository includes code for algorithms such as K-Nearest Neighbors, Decision Trees, Naive Bayes, Logistic Regression, Support Vector Machines, AdaBoost, and different regression techniques. It also provides datasets to accompany the code, making it a comprehensive learning resource for students and developers looking to deepen their understanding of machine learning.

makeyourownneuralnetwork is an open-source code repository hosted on GitHub, designed to accompany the 'Make Your Own Neural Network' book. It offers practical examples and implementations of neural network concepts, making it an invaluable resource for individuals looking to learn and understand the fundamentals of neural networks through hands-on coding. The repository includes various Jupyter Notebooks covering topics such as MNIST dataset handling, neural network implementation, loading custom images, and backquerying. This resource is ideal for students and self-learners who want to dive deep into the mechanics of neural networks and build their own models from scratch.

Tilde Research is a moonshot AI lab dedicated to advancing the frontier of intelligence through fundamental research. The lab focuses on three core areas: mechanistic interpretability, new AI architectures, and pretraining science. By building a foundational understanding of AI models, Tilde Research aims to contribute significantly to the broader AI research community. Their work is geared towards pushing the boundaries of what AI can achieve, emphasizing ambitious and innovative research goals rather than immediate commercial applications.

Chinese-number-gestures-recognition is an open-source Android application designed to recognize Chinese number gestures from 0 to 10 using a convolutional neural network (CNN). The project includes both the Android app code for real-time gesture recognition via a mobile camera and PC-side code for data processing and model training. It supports development environments like Python 3.6 with TensorFlow-gpu and Android Studio with TensorFlow Lite and OpenCV. The project also provides datasets, including raw images, data-augmented images, and compressed H5 datasets, along with pre-trained models. While the PC-trained models show high accuracy, the app's real-world performance can vary in complex environments.