agi-app-pytorch-playground

agi-app-pytorch-playground publishes the pytorch_playground_project AGILAB app as a self-contained PyPI payload. It turns an interactive neural-network playground into an executable AGILAB app with persisted arguments, worker execution, and deterministic evidence artifacts.

Purpose

Use this package to train a small PyTorch classifier on generated visual datasets, watch the decision boundary move with play/pause controls, inspect the resulting boundary/layers/loss terrain, and keep the configuration and artifacts replayable.

Classic browser neural-network playgrounds are still the shortest beginner teaching route. This package is the engineering-oriented step after that: PyTorch-native execution, replay tokens, deterministic manifests, evidence ZIPs, and reusable PyTorch/Lightning code from the same visual lesson.

Installed Project

The distribution name is agi-app-pytorch-playground; the AGILAB project name is pytorch_playground_project. The package exposes both pytorch_playground and pytorch_playground_project through the agilab.apps entry point group, so AgiEnv(app="pytorch_playground_project") resolves the project without a monorepo checkout.

Install

pip install agi-app-pytorch-playground

The app project itself installs PyTorch when AGILAB prepares its project environment. The payload package stays lightweight and only exposes the project root.

Direct Launch

When AGILAB is installed, launch the app-managed Streamlit surface directly:

agilab pytorch-playground

To open the hosted backend instead of a local app virtual environment:

agilab pytorch-playground --backend hf

Use AGILAB_PYTORCH_PLAYGROUND_HF_URL or --hf-space owner/agilab for a custom Hugging Face Space.

Run In AGILAB

Select pytorch_playground_project, then open ANALYSIS for the app-owned PyTorch Playground surface. That surface places persisted ORCHESTRATE arguments next to the decision boundary, training curves, neuron/loss views, evidence download, a one-click Run instant demo path, a deterministic Train / refresh path, and a live play/pause mode for watching bounded epoch ticks without leaving ANALYSIS. The boundary-first panel shows the final surface through a WebGL-first agi-web island with Canvas2D fallback, local epoch scrubbing, play/pause replay, clickable timeline, keyboard controls, confidence HUD, glowing uncertainty contour, and hover probability readouts, with Plotly detail tabs kept for evidence inspection and lesson cards for replayable circles, XOR feature engineering, spiral capacity, and gaussian sanity-check variants.

Open ORCHESTRATE when you want the reproducible AGILAB execution path: tune the sidebar fields, then run Deploy scheduler & workers and RUN. Enable loss-landscape computation only when you want the heavier 3D projection in the evidence bundle.

Expected Inputs

The default run generates a synthetic dataset. No external dataset, API key, notebook, cloud service, or private model is required.

Expected Outputs

The run writes the playground config, samples, training history, decision grid, network-layer summary, activation maps, optional loss landscape, a manifest, and a deterministic evidence ZIP.

Change One Thing

Switch the dataset from circles to XOR or spiral, then rerun the app. The manifest and training-history artifacts should make the changed behavior auditable.

Scope

This is an educational reproducibility app. It is not a production trainer, model registry, serving stack, or generic app-agnostic analysis page. The PyTorch-specific UI stays inside the app project; reusable apps-pages remain optional artifact readers for shared contracts.