Presets and Configs¶

A preset is a named model configuration that works immediately with fit, bench, evaluate, and the Python API. Seahorse ships presets for all registered model families.

Available Presets¶

Family	CLI preset	Python class	NLL type
Poisson + GMM	`poisson_gmm`	`PoissonGMM`	Exact (factorized)
Hawkes + GMM	`hawkes_gmm`	`HawkesGMM`	Exact (factorized)
Self-correcting + GMM	`selfcorrecting_gmm`	`SelfCorrectingGMM`	Exact (factorized)
Poisson + CNF	`poisson_cnf`	`PoissonCNF`	Exact (factorized)
Hawkes + CNF	`hawkes_cnf`	`HawkesCNF`	Exact (factorized)
Self-correcting + CNF	`selfcorrecting_cnf`	`SelfCorrectingCNF`	Exact (factorized)
Poisson + TVCNF	`poisson_tvcnf`	`PoissonTVCNF`	Exact (factorized)
Hawkes + TVCNF	`hawkes_tvcnf`	`HawkesTVCNF`	Exact (factorized)
Self-correcting + TVCNF	`selfcorrecting_tvcnf`	`SelfCorrectingTVCNF`	Exact (factorized)
RMTPP + GMM	`rmtpp_gmm`	`RMTPPGMM`	Exact (factorized)
THP + GMM	`thp_gmm`	`THPGMM`	Exact (factorized)
AutoSTPP	`auto_stpp`	`AutoSTPP`	Exact
DeepSTPP	`deep_stpp`	`DeepSTPP`	Exact
NSMPP DeepBasis	`nsmpp`	`NSMPP`	Exact
NJSDE	`njsde`	`NJSDE`	Exact
Neural JumpCNF	`neural_jumpcnf`	`NeuralJumpCNF`	Exact
Neural AttnCNF	`neural_attncnf`	`NeuralAttnCNF`	Exact
SMASH	`smash`	`SMASH`	Approximate
Diffusion STPP	`diffusion_stpp`	`DiffusionSTPP`	Approx (ELBO)

Discover all registered presets at runtime:

from seahorse import list_available_models
print(list_available_models())

Config Sources¶

Every run needs a config source. In order of precedence when combined:

Preset (--preset <name>): loads defaults from the bundled YAML under seahorse/configs/<preset>.yaml.
Config file (--config path/to/config.yaml): loads a complete or partial YAML config.
Overrides (--override key=value): dotted-path overrides applied last.

Override Keys¶

Overrides use dotted paths into STPPConfig. Common groups:

# Training
--override training.n_epochs=20 training.lr=5e-4 training.batch_size=32

# Data
--override data.normalize=false data.num_workers=2

# Model-specific build params
--override model.build_overrides.hidden_dim=128

Run python -m seahorse fit --help for the full config schema.

Using a Preset¶

Python APICLI fitCLI bench

from seahorse import AutoSTPP

model = AutoSTPP(device="cpu", seed=42)
model.fit(train, val, test, epochs=10, batch_size=64)

python -m seahorse fit \
  --preset auto_stpp \
  --train data/my_dataset/train.jsonl \
  --val data/my_dataset/val.jsonl \
  --test data/my_dataset/test.jsonl \
  --out runs/fit

python -m seahorse bench \
  --presets poisson_gmm hawkes_gmm auto_stpp \
  --splits_dir splits \
  --seeds 1 2 3 \
  --out runs/bench

Saving and Loading a Config¶

After fitting, the resolved config is written to {run_dir}/config.yaml. Load it for a reproducible re-run:

python -m seahorse fit \
  --config runs/fit/auto_stpp/<run_id>/config.yaml \
  --train data/my_dataset/train.jsonl \
  --val data/my_dataset/val.jsonl \
  --out runs/rerun

See Add a Model to register your own preset.