Run these examples from a source checkout. The package's examples extra installs dependencies; the scripts live in this repository:

pip install -e ".[examples]"
python examples/01_quickstart_2d.py

Examples run on one CPU thread, except 07, which uses eight; set POLYSTEP_THREADS to change it. Examples 02, 04, 05, 06, 10 and 11 use CUDA when available, and 05 runs in BF16 there; pass --device cpu to run them on the CPU. Example 03 runs on the CPU, accepts --device cuda, and skips its GIF with --no-render. Example 04 accepts --small for a 2,000-variable task.

Example 10 trains in about 20 seconds on an RTX 5090. On one CPU thread its 40 default epochs take about an hour, so shorten a CPU run with --epochs and --train-size.

For a short CPU check, I use:

python examples/06_loihi_snn_polystep.py --device cpu --hidden 32 --num-steps 5 --max-train 1024 --max-test 256 --pretrain-epochs 1 --adapt-epochs 1
python examples/10_cnn_mnist.py --device cpu --epochs 2 --train-size 1000 --test-size 256 --batch-size 32
python examples/11_transformer_selective_copy.py --device cpu --epochs 5 --train-size 512 --test-size 256

These small budgets check execution; use the defaults to assess learning quality.

Examples 05, 06 and 10 download MNIST to mnist under the system temp directory, /tmp/mnist on Linux. Example 06 takes --data-dir to use another directory.

Examples 10 and 11 take --compare to score the same batches with fully built candidates and with the site-local evaluator. This checks that the losses agree and times both paths, including first-call setup. On the CPU, example 10's fully built candidates take about 15 GB of memory. See performance for subspace settings and the API reference for amortization.

Examples 05, 06, 10 and 11 select checkpoints on validation data and report test accuracy afterward. Example 06 runs a PyTorch SNN simulation on the selected device. It adapts the readout and neuron parameters under Gaussian input noise. Its one-state neuron uses subtractive reset, so it does not reproduce Lava LIF or Loihi hardware. --backend cpu_sim remains accepted; --backend loihi2 fails before loading data.

Examples 07 and 08 report the mean over five seeds. Their OpenAI-ES baselines use the noise scale and learning rate selected on 20 separate tuning seeds.

Examples 07 to 09 measure the objective on the data used for optimization; those scores demonstrate fitting, not generalization. Their gradient baselines have a separate training-step budget. Example 08 reports sample standard deviations across its five seeds.

For published comparisons, see experiments.

I ran the default budgets on an RTX 5090 with v1.1.1, Python 3.14 and PyTorch 2.13. Test accuracy reached 94.4% in example 05, 94.35% in example 10, and 100% on example 11's synthetic copy task. Example 06 recovered 13.3 percentage points on shifted test inputs, while clean accuracy fell from 73.2% to 72.0%. These are the scripts' default seeds, not multi-seed estimates. The CPU CartPole run reached its 200-step evaluation limit on all 32 episodes.