Nets4Learning

Design, train and explain deep learning models in the browser, for learning and teaching.

Open the app · Manual · Paper · Changelog · MIT License

---

Nets4Learning is a web platform for getting started with deep learning. It poses classic machine learning problems with well-known datasets. You can then build a neural network layer by layer, choose its training hyperparameters, train it and compare the results. You can also try pre-trained models and see why they make each prediction.

Everything runs in the browser with TensorFlow.js: there is no server to install and your data never leaves your computer.

The platform is described in the paper Nets4Learning: A Web Platform for Designing and Testing ANN/DNN Models (Electronics, 2024). If you use it in your work, please cite it.

Features

• Four tasks: tabular classification, regression, object detection and image classification.
• Training in the browser: layer editor, optimizers, loss functions, metrics and live training charts (tfjs-vis). You can train several models and compare them.
• Your own data: upload a CSV, then clean, encode and scale its columns before training.
• Pre-trained models for each task, ready to try with examples, your own images, a drawing canvas or the webcam.
• Explainability: SHAP for tabular data, regression and images, and LRP (layer-wise relevance propagation) for the handwriting models.
• Learning material: a step-by-step manual, a glossary (activation functions, optimizers, losses, metrics), dataset descriptions and an exploratory data analysis page.
• Three languages: English, Spanish and Japanese. The default language comes from the browser and is remembered once you change it.

Tasks, datasets and models

Task	Train with a dataset	Pre-trained models	Explainability
Tabular classification	Your CSV, Car evaluation, Iris, Lymphography	Car evaluation, Iris, Lymphography	SHAP
Regression	Your CSV, Salary, Auto MPG, Housing prices, Student performance, Wine quality	Auto MPG, Student performance, Wine quality	SHAP
Object detection	-	Face detector, Face mesh, MoveNet (pose), COCO-SSD, face-api (age, gender, expression), Hand signs	SHAP
Image classification	MNIST (digits), KMNIST (Japanese characters)	MNIST, KMNIST (98.3 % test accuracy), MobileNet V2	SHAP, LRP

Tech stack

React 19, TypeScript 5.9 and Vite 7, with:

• TensorFlow.js 4.22 and tfjs-vis for the models;
• danfo.js for dataframes;
• React-Bootstrap for the interface;
• Chart.js and Plotly for charts;
• i18next for translations;
• Vitest and Testing Library for tests.

Getting started

Requirements: Node.js ≥ 22 and pnpm 12. The exact pnpm version is pinned in package.json, and Corepack installs it for you.

corepack enable
pnpm install
pnpm dev

The app is served at http://localhost:5173/n4l/.

Scripts

Command	What it does
pnpm dev	Development server with hot reload
pnpm build:simidat	Type-check and build for simidat.ujaen.es/n4l (served under /n4l)
pnpm preview	Serve the last build locally
pnpm test	Run the tests in watch mode (pnpm test run runs them once)
pnpm test:ui / pnpm test:coverage	Tests in the Vitest UI / with coverage
pnpm lint	ESLint
pnpm version	Update CHANGELOG.md from the commit history

The build is memory-hungry. If it runs out of memory, run it with NODE_OPTIONS="--max-old-space-size=8192", as the Docker image does.

Environment variables

Vite reads .env and then .env.<mode>. The three files are versioned:

• .env for development;
• .env.simidat for SIMIDAT;
• .env.netlify for Netlify.

The app uses these variables:

Variable	Purpose
VITE_PATH	Path the app is served under: "/n4l" or "" for the domain root. It also sets Vite's base.
VITE_ENVIRONMENT	development enables i18next debug logs, shorter default training and development panels. Any value other than production also enables the development-only pages, /debug and the test pages.
VITE_GA_MEASUREMENT_ID	Google Analytics 4 measurement ID.
VITE_SHOW_NEW_FEATURE	"true" shows sections that are still in progress (glossary, datasets page).

Deployment

Docker

The Dockerfile builds the app with pnpm and serves it with an unprivileged nginx under /n4l/. It uses the simidat mode by default; change it with the BUILD_MODE build argument.

docker compose up --build -d

The app is then served at http://localhost:3000/n4l/. In production, Traefik routes requests to the container on port 80 (see the comments in docker-compose.yml and nginx.conf).

Project structure

public/
  datasets/ models/                 Datasets and pre-trained models
  docs/ locales/                    Manual and translations (en, es, ja)
src/
  pages/playground/<task>/models/   One class per dataset or model
  core/                             Training, dataframes, explainability
  components/                       Shared UI
  DATA_MODEL.ts                     What each menu lists
  MODEL_KEYS.ts                     Dataset and model identifiers
Scripts/                            Dataset and model generation
tests/                              Vitest tests

To add a dataset or a pre-trained model:

1. Create its class in src/pages/playground/<task>/models/ and register it in that folder's index.ts.
2. Add its key to MODEL_KEYS.ts and its menu option to DATA_MODEL.ts.
3. Add its texts to public/locales/*/translation.json and its files to public/datasets/ or public/models/.

Contributing

• Commits follow Conventional Commits; commitlint checks them.
• Before opening a pull request, run pnpm lint, pnpm test run and pnpm build:simidat. The build also type-checks the project.
• New interface texts go into the three translation files. Manual pages go into public/docs/{en,es,ja}.

Team

Developed by the SIMIDAT research group (Intelligent Systems and Data Mining) of the University of Jaén, with the support of the DaSCI institute.

• Direction: Antonio Jesús Rivera Rivas, María Dolores Pérez Godoy, María José del Jesus Díaz.
• Development: Antonio Mudarra Machuca, David Valdivia Vico, Carlos Requena.

Citation

Mudarra Machuca, A., Valdivia, D., Ducange, P., Germán Morales, M., Rivera Rivas, A. J. and Pérez Godoy, M. D. (2024). Nets4Learning: A Web Platform for Designing and Testing ANN/DNN Models. Electronics, 13(22), 4378. https://doi.org/10.3390/electronics13224378

@article{mudarra2024nets4learning,
  author  = {Antonio Mudarra Machuca and David Valdivia and Pietro Ducange and Manuel Germ{\'a}n Morales and Antonio Jes{\'u}s Rivera Rivas and Mar{\'\i}a Dolores P{\'e}rez Godoy},
  title   = {Nets4Learning: A Web Platform for Designing and Testing ANN/DNN Models},
  journal = {Electronics},
  year    = {2024},
  volume  = {13},
  number  = {22},
  pages   = {4378},
  issn    = {2079-9292},
  doi     = {10.3390/electronics13224378},
  url     = {https://www.mdpi.com/2079-9292/13/22/4378},
}

License

The code is released under the MIT License, © 2023 Universidad de Jaén | SIMIDAT.

Datasets and third-party pre-trained models keep their own licenses and references. The app shows them in each dataset and model description. For example, KMNIST (Kuzushiji-MNIST) is © ROIS-DS Center for Open Data in the Humanities, under CC BY-SA 4.0.