Explore Projects

This proposal aims to provide Tensorflow with more object detection models and tutorial examples. The plan consists of adding transfer-learning-friendly models to the model garden and creating notebooks and tutorials for fine-tuning them. By adding these resources to the Model Garden, Tensorflow can make it more popular among ML beginners. Models like Faster R-CNN, Mask R-CNN and YOLO can be the starting point for learning advanced ones. Beginners can study the history and design ideas of object detection neural networks. The result of this project consists of the creation of several new fine-tunable transfer-learning-friendly object detection models, a detailed tutorial for building and fine-tuning the model, and a brief guide on the impacts of hyperparameters and how to configure them. Check out this Colab tutorial for keras faster-rcnn transfer learning: https://colab.research.google.com/drive/1-X9xcXYe2bYxichgomdY3FVN1Xap0Wui#scrollTo=Z_epOwYCf_4R Check out this Colab for YoloV5 Transfer Learning: https://colab.research.google.com/drive/1XkHcer0xh_LvQPkqALJFv-5kPOcXnlF8

Calorimeter is one of the most important components of the Large Hadron Collider (LHC) experiments, detecting energy loss of particles after collision. However, current Geant4 simulation of showering process in the calorimeter is inherently slow for large amounts of events, especially after High Luminosity Upgrading in the future. Therefore, generative models and other machine learning techniques can be used to accelerate the Geant4 simulation. This project aims to improve the performances of the current VQ-VAE Transformer for particle showering and explore architectures beyond (VQ-)VAE and/or Transformer, including new position embedding, hierarchical attention, an-isotropic attention, Fourier Transform linear mixer, and etc. We expect to provide deliverables including notes of extended numeric experiment and integrating new models into the project code.

The ticket-32159 (https://code.djangoproject.com/ticket/32519) originally requested a new feature to make Django provide generic database functions for manipulating JSON data on the database. Currently, in order to make partial updates to a JSONField, developers need tol oad model instance to the memory, modify JSONField data as objects, then call save() on the model instance, or write custom mapping for JSON functions as Func subclasses, and use them in update(). It would be better if Django has built-in support for doing partial JSONField updates more efficiently and more idiomatic to how developers can currently query JSONFields (i.e. with key transforms). In order to achieve that, this proposal intends to create an implementation for all databases supported that developers can use in order to update partial values of JSONField. This project aims to accomplish the problem by dividing it into two big parts: first, implement JSONSet and JSONRemove functions to allow partial updates to JSONField. Second, add support for key transforms in update() for JSONField to make it easier to use for developers. The project will be executed on creating JSONSet for SQLite, MySQL, PostgreSQL, Oracle first, then, create JSONRemove on same databases. After that, support for key transform on update() will be implemented.

Diffusion models have experienced rapid growth in usability, availability, and research. The classical algorithm consists of two main parts: a forward diffusion process and a learned denoising step. The forward diffusion process discretely applies noise to an image until it is unrecognizable. The denoising process reverses the noise, predicting the original image. Quantum techniques, such as parametric quantum circuits, could be applied to both parts to create a fully quantum model that could run on a quantum computer, speeding up training time. This model would be trained on multiple high energy physics datasets, such as the quark and gluon dataset for generating quark and gluon jets, alongside a classical diffusion model, and the results could be evaluated taking into account resources, time, and quality. Furthermore, a mathematically-driven approach would provide the theoretical foundations of quantum diffusion models in addition to experimental results.

<p>Flexible Suppression model</p>

<p>PyMC3 is a probabilistic programming language that uses Bayesian statistics to specify models and estimate unknown quantities. PyMC3 has about seven time-series distributions, an important class of models in Probabilistic Programming. This project works on extending this class in a few ways: Firstly, the addition of a new model - namely Auto-Regressive Integrated Moving Average(ARIMA). Expansion of the existing distributions(and their respective documentation), including Time-Series analysis functions. Thirdly, state-space implementation of possibly a time-varying linear and Gaussian time series model, based on Aesara. Additionally, I believe that Programming can be an efficient way of understanding applied mathematics, motivating me to focus on more pedagogical or explanatory notebooks about Probabilistic Programming (Time-Series in this case). Explanatory Notebooks for ARIMA and state-space model usage, exploring their Econometric applications will be a small step towards fostering an interdisciplinary bond between academia and the open-source community, mutually benefiting both.</p>

This project aims to create AI-powered 3D models of marine species for educational use. By converting text descriptions into 2D images and then transforming them into 3D structures, we will generate realistic models of marine life such as fish, octopuses, and corals. These models will be optimized and rendered in Blender or Unity, making them interactive and adaptable to different environmental conditions. The final output will include high-quality, dynamic 3D assets that can be integrated into AR applications, helping students explore marine biology in an engaging and immersive way.

<p>MPSoCBench developed by the ArchC team currently provides four ISA simulators namely ARM, PowerPC, MIPS and SPARC processor models. The ultimate goal of this project is to make the existing MPSoCBench platform more powerful by adding a new processor model based on the open-source RISC-V architecture. By the end of the project one will have a fully functioning RV32IMAFD (also known as RV32G) processor model which will be directly executable on the ArchC simulator. Also, the model will contain files that will allow it to be used with the MPSoCBench platform.</p>

The existing Neural-lam probabilistic forecasting model, GraphEFM, currently exists in an isolated branch (prob_model_lam) with a monolithic ARModel class encapsulating both the training, rollout, and single-step prediction functionality in a single file. Meanwhile, PR #208 refactors the deterministic models into a clean three-layer hierarchy: ForecasterModule (Lightning training harness), Forecaster (rollout strategy), and StepPredictor (single-step prediction). The GraphEFM model cannot be merged into main until it is ported to this new hierarchy. This project ports the GraphEFM model into the new hierarchy by refactoring it into three new classes that mirror the existing separation of concerns: GraphEFMPredictor (extends StepPredictor : owns the encoder, prior, decoder, latent variable logic), EnsembleForecaster (extends Forecaster : owns the AR rollout with separate prior and encoder paths), and EnsembleForecasterModule (extends ForecasterModule : owns the ELBO training, ensemble metrics, and visualization). Key design decisions include using explicit forward() and forward_encoder_path() instead of gating on self.training (handled automatically by PyTorch Lightning), making the predictor self-contained with its own per_var_std buffer, and replacing all hardcoded constants with the datastore API introduced in PR #208. Deliverables: - All components of the encoder/decoder moved, GraphEFM predictor that fulfills the contract of StepPredictor, EnsembleForecaster with both paths tested, and end-to-end graph efM training of --model graph_efm on MEPS, along with ensemble metrics. - KL/CRPS warmup schedules, ensemble parameters defined via configuration, Zarr-based ensemble prediction saving, additional ensemble visualizations, spread maps, and spaghetti plot, checkpoint backward compatibility, and architecture documentation.

This project aims to develop a vision foundation model for strong gravitational lensing by comparing various self-supervised learning techniques, including contrastive learning, masked autoencoders, and the physics-informed I-JEPA model. The objective is to evaluate and benchmark these methods on diverse gravitational lensing datasets and to integrate physical priors into the I-JEPA framework for more meaningful representation learning. The resulting foundation model will be fine-tuned for multiple downstream tasks such as lens classification, super-resolution, parameter regression, and lens detection. Deliverables include a unified lensing dataset, comparative analysis of self-supervised methods, and a robust, physics-aware foundation model for gravitaional lensing images.

At the CMS detector, jet images are stored at reduced resolution, discarding ~75% of total deposited energy. Standard super-resolution fails here because it optimises for visual similarity, not physical validity — a model without energy constraints converges to a response of 0.52, recovering only half the missing energy. As part of the ML4Sci/CMS test task, I developed a complete physics-aware conditional GAN pipeline (github.com/rajveer43/cms-superres-reconstruction) that maps 3×64×64 LR jet images to 3×125×125 HR reconstructions using a residual CNN generator, PatchGAN discriminator with spectral normalisation, and a three-component hybrid loss (LSGAN + L1 + energy response). The best configuration achieves raw-space L1 of 0.00698 vs 0.00846 for bicubic (17.5% improvement), PSNR 14.32 vs 14.13 dB, and energy response 1.0098 — within 1% of simulation ground truth. This GSoC project extends that validated baseline by implementing and benchmarking SwinIR-style transformer SR adapted for η-φ calorimeter coordinates, and a conditional DDPM/flow-matching diffusion variant as a stretch goal. Every architecture is evaluated on the same physics-constrained suite: energy response must remain within [0.99, 1.01] or the model is rejected regardless of pixel quality. Deliverables: reproducible GAN baseline with full metric suite, trained transformer SR model with benchmark comparison table, mid-project GAN vs transformer report, final benchmark across all architectures, technical report, and organized codebase for CMS collaboration use.

<p>Tiny-dnn is a lightweight dependency free library. It is extremely easy to get your first boilerplate code running as it requires nearly no additional setup. Since it is specifically suited for Mobile and iOT devices, or broadly speaking, devices with low compute capability, it is implied that the focus of library is more on inference rather than training huge models.<br> That said, there is a need of shipping some standard models, while paying specific attention to ease of loading pre trained weights in them. This project aims to build a Model Zoo for tiny-dnn, providing pretty standard CNN architectures such as AlexNet, VGG, Inception etc. Moreover, the goal is to make transfer learning easily understandable and implementable with tiny-dnn, as one might need to fine tune a huge model on a low end hardware device.</p>

<p>This proposal consists of the implementation of a 3D model repository for OpenStreetMap, improving upon the basic system offered by the Simple 3D Buildings tags. By allowing 3D models to be used in OSM 3D visualizers, the quality and ease of creation of the rendering can be improved. To implement this repository by the proposed deadline, I have decided to use the Django/Python framework, with PostgreSQL as the database backend, and for the front end Bootstrap was chosen. The OpenLayers library will also be used to render slippy maps to the user. The repository includes an user interface allowing uploading models from an authenticated OSM account through the OAuth mechanism, and an API for 3D visualizers to use, so that they can query the repository for models. Finally, while there were a plethora of interesting features that could be included in the deadline checklists, I have opted for only having the essential features required for a 3D repository as deliverables, with the rest as extra features, to be added if time allows it.</p>

The goal of this project is to allowing moving models between apps.. The model movement will be auto detected by a new generate_moved_model method in auto-detector to decrease the manual work load. The model will be moved using all the existing operations.

In the scheduling progress, the karmada-scheduler makes decisions as per a bunch of factors, one of the factors is the resource details of the cluster. We don't want to collect and store each node's resources in detail(That's a burden for Karmada to maintain the information), but we want to build a resource model for each cluster.

Creation of Mobility and Orbit models for LEO Satellites along with a NTN Helper for GEO and LEO Satellites, and an example using the helper integrating LEO Satellites with 5G NR.

<p>Is to implement MobileNetV1 which will also include implementing depthwise separable convolutions and a resnet model builder that can be used to create resnet18, resnet34, resnet50, resnet101, and resnet152. This project would fall under the idea: Ready to use models in mlpack, the resnet builder and MobileNetV1 will fall under the models repository and depthwise separable convolutions will fall under the mlpack repository as a layer inside the artificial neural network codebase.</p>

<p>Gaussian Mixture Model (GMM) is widely used in computer vision as a state-of-the-art clustering algorithm. This project proposes Quantum Gaussian Mixture Model (QGMM) for Quantum Clustering. According to the paper, QGMM outperforms classical GMM in every aspect of the estimations. Therefore, in this project, we'll implement QGMM and conduct some experiments to see if how fast it trains, how better it models the data, and what edge cases there are, compared with classical GMM.</p>

<p>Model Visualization Plugin for App4MC - Eclipse IDE</p> <p>The world in which we live in is evolving faster than anyone can expect. New Technologies and Systems assisting us in our general life, from the invention of the electric power, over smartphones, up to self-driving cars. This was a big step in human evolution and gives us a small insight on what is to come in the future. In this future the technical requirements are going to be higher than ever. This will increase the importance of multi core real time systems more and more.</p> <p>The goal of this project will be to increase the important aspects of the App4MC platform with a user friendly and customizable representation of this data in the form of a plugin for the Eclipse IDE.</p>

This proposal outlines the development of an AI-powered Image Analysis App for Rocket.Chat. The project aims to introduce a feature that allows users to simply upload images, while the app automatically analyzes them using a multi-modal Large Language Model (LLM). The app will process images to identify locations, landmarks, points of interest, or event venues and retrieve up-to-date event and happening information from the internet. By integrating advanced image reasoning with real-time data retrieval, this solution enhances situational awareness and provides users with valuable insights. This proposal seeks to streamline information discovery, improve engagement, boost productivity, and elevate the overall user experience within the Rocket.chat ecosystem.

The main aim of the project is to improve the current model stack of deep learning models, in terms of inference speed with minimum loss of precision, for autonomous driving applications. JdeRobot organization has created Behavior Metrics a tool for comparing deep learning architectures for autonomous driving on different circuits with the support of Gazebo and Ros Noetic. The organization also provides another tool called DeepLearningStudio, which has datasets and some model implementations for training deep learning models. We will use available tools and techniques such as TensorRT, Quantization, Pruning, and architectural changes to optimize the correct model stack available in both PyTorch and Tensorflow.

My GSoC 2025 proposal focuses on enhancing the "Transformed Cubic Grids" framework (Issue #264) within the QC-Devs organization by implementing adaptive quadrature methods. Current Becke-style molecular grids suffer from overlapping atomic grids, leading to redundant computations and reduced accuracy in quantum chemistry simulations. Building on prior work (Issues #7, #15, #96), I will develop algorithms to dynamically adjust grid points and weights based on integrand behavior, such as electron density, ensuring high accuracy with fewer points. My deliverables include refreshing and merging Pull Request #136, implementing adaptive methods (e.g., trapezoidal or Simpson’s rule with refinement), writing comprehensive tests and documentation, and creating Jupyter notebook tutorials for practical use in molecular property calculations.

3DMR is a community-driven repository for 3D models and their metadata, designed to enhance 3D map rendering. It was originally developed during Google Summer of Code 2017 by Pedro Amaro, under the mentorship of the OpenStreetMap community, with Jan Marsch and Tobias Knerr serving as mentors. Since then, the project has been left largely abandoned and now runs on an outdated Django 2.0.5 codebase. This proposal aims to revive 3DMR by modernizing its foundation—upgrading the framework, adopting glTF as the standard model format, and introducing a range of improvements including robust model validation, enhanced preview functionality, and automated quality analysis feedback.