Prerequisites
You need Elixir and Livebook installed. Use your preferred Elixir installation method for your operating system, then install or run Livebook using the official Livebook instructions.
The demo notebook uses Mix.install/2 to load this local package and notebook-only visualization dependencies. Kino and Vega-Lite are not required by the core library.
Clone the repo
git clone <repo-url>
cd micrograd_ex
If this repository is already on your machine, start from the repository root.
Start Livebook
livebook server
Open the URL printed by Livebook.
Open the demo notebook
Open:
notebooks/micrograd_demo.livemdThe setup cell expects the notebook to live under notebooks/ because it installs the local package with:
Mix.install([
{:micrograd_ex, path: ".."}
])If you move the notebook, update that path.
Run all cells
Use Livebook's run-all command from a clean runtime. Running top-to-bottom matters because later cells depend on values created earlier, such as dataset, model, and run.
Expected outputs
You should see:
- a scalar autodiff value and gradient;
- graph node and edge tables;
- a two-moons dataset table;
- a dataset scatter chart;
- parameter count
337; - initial loss and accuracy;
- a 100-step training history table;
- a loss chart;
- an accuracy chart;
- a decision-boundary chart.
The deterministic demo should show loss decreasing and accuracy improving.
Common setup problems
If Livebook cannot find micrograd_ex, confirm you opened notebooks/micrograd_demo.livemd from this repository and did not copy the file elsewhere.
If dependency installation fails, restart the Livebook runtime and run the setup cell again. Update notebook dependency versions only after testing them.
If training feels slow, remember that MicrogradEx is scalar and educational. Reduce sample count, hidden width, training steps, or use a coarser decision-boundary grid.
Next steps
Read Micrograd demo walkthrough for the math and workflow, then Elixir design notes for the immutable implementation choices.