NNModelling

Design neural networks as visual graphs, inspect tensor shapes and dtypes while editing, and train through an isolated backend. Export a Python wheel that runs the trained prediction model without this checkout.

Project directory → editor + Lua tensor inference → authenticated uploads
                  → container training worker → portable prediction wheel

Run the editor

pnpm install --frozen-lockfile
pnpm --dir front-end dev --host 127.0.0.1 --port 5174

Open the URL printed by Vite. Choose New project to create a writable project directory, or Open project to select one containing model.json. Graph changes save automatically. The web editor needs writable directory access through the browser's File System Access API.

You can also open the web distribution or build the Linux desktop application. The desktop host uses the same editor and project format. Neither distribution includes a training backend.

To try an existing model, copy an entire directory from examples/diagrams/package/models/, including its custom packages and datasets, then open the copy. Start with the VAE or ResNet guide. Prepare their data files before training; cloning the example does not download MNIST.

Train and use a model

The Training sidebar pairs with an operator-managed backend. Select a project dataset, match graph Input bindings to its named tensor slots, and configure the objective and job settings. A successful job provides a downloadable wheel with trained weights and the public Model prediction API.

• Training workflow
• Backend installation, pairing and operations
• Using an exported model in Python

The supported backend accepts package graphs and project datasets. Historical NNTree fixtures are not editable projects or inputs to this workflow.

Documentation

Build the public Sphinx guides and generated TypeScript reference:

pnpm run docs

Outputs are docs2/build/html/ and docs2/build/typedoc/. Documentation build instructions cover prerequisites, previewing both sites and strict checks.

• Getting started
• Projects and graph editing
• Project datasets
• Custom stereotypes
• Tensor type system
• Architecture
• Troubleshooting

Internal architecture and contributor contracts are indexed in docs/README.md. Repository instructions live in AGENTS.md, with package-specific guidance below it.

Development checks

Run the checks for the package you change:

pnpm --dir front-end check
pnpm --dir front-end test
pnpm --dir mcp-server test
cd converted && uv run pytest src/tests/ -m fast -q

Frontend and MCP builds use pnpm --dir <package> build. Slow backend and integration tiers have separate prerequisites; consult the relevant package's AGENTS.md before running them. Browser-backed MCP needs a running editor and selected tab; see the MCP architecture.