Examples¶
Runnable, self-contained scripts in examples/. Each writes its figures to examples/output/ as standalone HTML (and PNG, where a Chrome install is available for Plotly's kaleido/plotly_get_chrome export). Run them from the repo root after installing parx[dev] (see Home § Installation).
01 — Identity network¶
python examples/01_identity_network.py
A one-hidden-layer network with identity weights (2 → 2 → 1) partitions R² into exactly its four quadrants — a partition simple enough to verify by hand. Demonstrates both sparse_julia and exact_julia modes on the same network, the verification functions (check_no_overlaps, check_covers_space), and all three 2D visualizations (plot_partition_2d, plot_region_counts, plot_halfspaces).
02 — Random MLP¶
python examples/02_random_mlp.py
A deeper 2 → 5 → 5 → 5 → 1 network with random weights. Demonstrates the state_dict workflow (model.state_dict() → compute_partition, rather than passing the nn.Module directly), the method registry (parx.list_methods(), comparing sparse vs. exact and the Julia vs. Python backends), and coloring regions by a metric derived from the local affine map (affine_frobenius, affine_spectral).
03 — Partition evolution over training¶
python examples/03_epochs.py
Trains a small network to convergence, capturing a state_dict snapshot every epoch, then computes the exact partition for each snapshot. Produces a static before/after comparison, an interactive Plotly animation with a play button and epoch slider (animate_epochs), and an exported GIF (animate_epochs_video, requires pip install "parx[animate]"). This is the closest example to the primary research use case described in Usage § Training-time tracking.
04 — Feature embedding¶
python examples/04_feature_embedding.py
Trains a classifier on the two-moons dataset, computes the exact polyhedral partition, extracts penultimate-layer features, and plots a t-SNE embedding colored by region membership — connecting the partition back to the network's learned representation, not just its input-space geometry.
For a guided, narrated walkthrough of the same two-moons workflow (sparse and exact partitions before/during/after training, layer-by-layer views, save/reload), see the Notebooks page instead — the scripts here are meant to be read as source code and run standalone, while the notebooks are meant to be read top-to-bottom as a tutorial.