Bring your model to Seahorse

Every model in Seahorse has the same shape. Express your freshly built STPP model that way once, and it plugs into the entire harness — training, multi-seed benchmarks, every metric profile, and sampling — with no other wiring.

UnifiedSTPP

StateModel event history → hidden state encodeevolve

→

EventModel hidden state → log-probability log_probsample_nextintensity_grid

▼

you get, for free fit bench · multi-seed evaluate · all profiles predict_next CLI + Python API

Your model is comparable to every baseline by construction: the same data contract, the same normalization, the same metric definitions apply the moment it is registered.

Four steps from nn.Module to benchmarked

1. 1
   Wrap your model as two parts

   Split it into a StateModel (encodes history into a hidden state) and an EventModel (turns that state into a log-likelihood). Both are plain nn.Modules.

   class MyEventModel(nn.Module): def log_prob(self, times, locs, state, mask): ... Wrap an existing model →
2. 2
   Wire them in a config

   A ModelFamilyConfig owns the build-time parameters and returns a wired UnifiedSTPP from build_model().

   def build_model(self): return UnifiedSTPP(state, event, hidden_dim=self.hidden_dim) Register a preset →
3. 3
   Register a preset name

   One decorator makes your model resolve through fit, bench, evaluate, and the Python API — no import-path changes anywhere else.

   @ConfigRegistry.register("my_preset") Full checklist →
4. 4
   Declare what it can do

   The methods you implement decide which metrics run. An unimplemented capability is reported as a clean skip — never a silent wrong number.

   def sample_next(self, state, t_last, n_samples=1): ... # unlocks predictive metrics Declare capabilities →

Then test and ship

Run a tiny one-epoch fit, an evaluate, and a save/load round-trip before adding the preset to benchmark examples. The Testing Checklist is the short list to clear.

Pick your path

I already have a PyTorch model Wrap an existing model The StateModel / EventModel adapter pattern, with minimal working stubs. → I'm building a new family from scratch Add a model The end-to-end checklist, from components to a bundled preset. → My components are ready to expose Register a preset Config registry, YAML defaults, and CLI + Python API wiring. → I'm deciding which metrics apply Declare capabilities Map implemented methods to evaluation profiles and exact-vs-approximate NLL. →

See also

• Architecture — the full model-layer documentation.
• Model Capability Matrix — what each shipped preset declares.

Was this page helpful? Yes No Glad it helped — a ⭐ star on GitHub helps other researchers find Seahorse, and here is how to cite it. Sorry that missed — tell us what's missing.

Suggest edits Raise issue