Training guide

Every command runs from the repository root. uv sync --all-packages first.

MAMBA_SSM_AVAILABLE=0 is recommended unless you have the CUDA Mamba extension installed: importing mamba_ssm can hang for minutes while it builds kernels.

Classical machine learning

uv run inmotion train classification \
    --data data/processed/dataset.csv

Runs EDA, cross-validates a set of scikit-learn classifiers, optionally tunes them with Optuna, computes feature importance, and writes plots plus a results CSV. Add --optimize for the Optuna stage and --dashboard to launch optuna-dashboard.

Deep learning baselines

MAMBA_SSM_AVAILABLE=0 uv run inmotion train dl \
    --data data/processed/dataset.csv --seed 42 --trials 50

Phases, in order, each skippable by flag:

phase	flag to skip	what it does
1	--no-optuna	trains each single model on its own GPU (round-robin)
2	--no-optuna	Optuna HPO per model
3	--no-optuna	NAS
4	--no-moe	mixture-of-experts variants
5	--no-deepstack	DeepStackEnsemble, a 10-base stack with a learned meta-model
—	--no-meta	metadata-fusion variants

Checkpoints land in --models-dir, results in --results-dir. An existing checkpoint is reused unless --force-retrain is given.

World models (SSL)

# find a configuration first
MAMBA_SSM_AVAILABLE=0 uv run inmotion train exotic \
    --model ts_jepa --data data/processed/dataset_augmented3.csv \
    --hpo --hpo-trials 50

# pretrain plus fine-tune
MAMBA_SSM_AVAILABLE=0 uv run inmotion train exotic \
    --model ts_jepa --data data/processed/dataset_augmented3.csv --seed 42 \
    --pretrain-epochs 500 --finetune-epochs 80 --batch-size 512

Two-stage by nature: self-supervised pretraining, then supervised fine-tuning. The JEPA models (t_jepa, ts_jepa, lejepa, cf_jepa) use both stages; sigreg and the mamba3_* hybrids are directly supervised and take --epochs.

Stage-1 checkpoint selection is driven by a linear-probe MCC, not by the masked-prediction loss, because the loss does not track representation quality. --probe-cadence and --probe-patience control that probe.

Resume from a pretrained encoder:

inmotion train exotic --model ts_jepa \
    --checkpoint 20-aug/models/exotic/normal-new-ds/ts_jepa_pretrain_best.pt \
    --finetune-only

The paper’s HPO-best configs

The Optuna databases were deleted, so the winning configurations are re-declared from the documented values and retrained:

MAMBA_SSM_AVAILABLE=0 uv run inmotion train hpo-paper \
    --data data/processed/dataset.csv --seed 42

Values come from docs/icaisf/hpo_supplementary.md.

Ensembles

# greedy selection over existing checkpoints (Caruana forward selection)
uv run inmotion evaluate ensemble \
    --data data/processed/dataset_augmented3.csv --seed 42

# OOF stacking over every family, with calibration
MAMBA_SSM_AVAILABLE=0 uv run inmotion evaluate mega-ensemble \
    --data data/processed/dataset_augmented3.csv --seed 42

mega-ensemble combines DL checkpoints, the JEPA family, classical ML and TabPFN. It needs the member checkpoints present; inmotion model list shows which are.

On a cluster

scripts/slurm/*.sh carry #SBATCH headers:

sbatch scripts/slurm/job_full_train.sh

scripts/README.md says which script does what and which are superseded. Credentials come from the environment (env.example), never from the scripts.

Reproducing a run

Record these four things and the run is repeatable: the command, --seed, the dataset (inmotion datasets show <id> gives its checksum), and the git commit. inmotion doctor prints the environment. See docs/reproducibility.md.