README.md

Examples

These notebooks are self-contained. Each one builds or simulates its own small dataset, so no dataset download is needed. Most run in under a minute, and the model comparison (06) takes a few minutes. Open them with Jupyter (the examples extra installs JupyterLab).

jupyter lab examples/00_tpp_introduction.ipynb

We recommend reading them in order, since each one builds on the previous one.

Notebook	What it teaches	Runtime
00_tpp_introduction.ipynb	What a TPP is: the intensity, cumulative intensity, survival, next-event densities, the log-likelihood and sampling, each computed with TorchPP and checked against its formula on a hand-set Hawkes process. Optional sliders (pip install ipywidgets) change the parameters.	~10 s
01_quickstart.ipynb	The data structures (MVEventData, BatchedMVEventData) and the interface every model shares: density_*, prob_*, time_*, likelihood, sample, and the intensity view, on a Poisson process.	~5 s
02_fit_evaluate.ipynb	Fit a Poisson process with fit, recover the true rates, and score held-out data with evaluate_tpp.	~10 s
03_intensity_and_sampling.ipynb	Forecasting with a Hawkes process: the intensity on a history, many sampled futures, and how an event-free observation window (T_obs) changes the forecast.	~5 s
04_add_a_model.ipynb	Write your own model as an IntensityBasedTPP subclass (an intensity hook plus a closed-form or numerical cumulative intensity), then train and evaluate it like a built-in model. Companion to docs/extending.md.	~10 s
05_neural_piecewise_constant.ipynb	Build a neural TPP from a shared transformer encoder and a small head, fit it to Hawkes data and compare it with a Poisson process. Uses the base class's internal hooks (_get_params, _get_params_at_queries, …), which may change between versions.	~25 s
06_compare_models.ipynb	Simulate from a known Hawkes process, fit a Poisson process, two Hawkes processes and a neural TPP, compare them on held-out data, and recover the excitation matrix and kernel shapes.	~2–4 min

Survival analysis

survival/ is a separate series on survival and competing-risks models built on TorchPP. It covers Cox models, a spline hazard, the density-based route and a neural covariate model. See survival/README.md for details.

For training on the benchmark datasets with train.py, or on your own data, see getting started. The introduction gives the background on the quantities.