Code Editor

The workbench

The Explorer is built on Monaco, the editor behind VS Code. Its header names the project and where it is kept (in a folder on your computer, or in this browser) and holds what a run needs: the dataset, the runtime and Run. Files are on the left, the editor in the middle (it splits in two from the end of the tab strip), and the dock underneath.

Running code

Run (Ctrl Shift R, ⌘R on a Mac) runs the open file:

• JavaScript and TypeScript that use TensorFlow.js run in a TF.js worker, on WebGPU when the browser offers it. Other scripts run in a sandboxed worker.
• Python files and Jupyter notebooks run in Pyodide, a CPython build compiled to WebAssembly, or on a Jupyter kernel you attach.

The runtime pill in the header says which runtime the open file goes to and whether it is ready, and for TF.js which GPU the browser gave it. The dataset pill picks the Data Hub dataset a run receives as hubData (one object per row), with dataHeaders and dataName. It follows the project's active dataset, as the Model Builder does, and is read when a run starts. With a dataset chosen, hubData in the editor's path bar inserts a snippet that reads it, in JavaScript or Python.

The dock

Under the editor, one line says what the runner is doing: the epoch and loss while a model trains, how the last run ended, or what Run will do next. Its tabs hold:

• Terminal: a shell for the project. node and python run a file, python alone starts a REPL, pip install installs a package, and ls, cat and npm work as you'd expect. help lists everything.
• Output: the last run's output, with the figures it draws (matplotlib). When a Python run can't import a package, Output offers to install it.
• Metrics: loss and accuracy by epoch, from what the script prints (Epoch 3/20 … loss: 0.41) or reports with dlway.reportMetrics().
• Problems: what the editor finds in the open files, plus the browser's Python syntax check. Click one to go to it.
• Resources: the page's memory, TF.js tensors and GPU memory, the GPU adapter, and what the open file's model needs.
• Packages: Python packages installed with pip, into Pyodide or the attached kernel. npm packages are recorded but not downloaded: the browser sandbox gives require() TensorFlow.js and the project's own files.

Working with files

The Files panel shows the project's tree. A dot marks unsaved changes and a letter marks what git sees since the last commit. You can filter the tree, make files and folders, rename with F2, and delete (it asks first). Open a folder from your computer and saving writes back to it. A project kept in the browser can move to a folder with Keep in a folder…, and from then on saving writes there and trained models go to its models/ folder. DLWAY mirrors the project to a .dlway/project.json manifest so its state survives a refresh.

Search, source control and Copilot

While the Explorer is open, the sidebar shows its own tools: search across files, source control, and the AI Copilot.

More

See Keyboard shortcuts for every key binding, and Remote compute to run Python on a Jupyter or Kaggle kernel instead of in the browser.