Results and Artifacts

Every experiment run produces a RunResult. When a local result backend is configured, the result record is written to disk and model checkpoints are registered as downloadable artifacts.

Enable Local Storage

Pass save_dir directly:

from torch_timeseries import Experiment

results = Experiment(
    model="DLinear",
    task="Forecast",
    dataset="ETTh1",
    windows=96,
    pred_len=96,
    save_dir="./results",
).run(seeds=[1])

or attach a backend explicitly:

Experiment(
    model="DLinear",
    task="Forecast",
    dataset="ETTh1",
    windows=96,
    pred_len=96,
).with_local("./results").run(seeds=[1])

RunResult Fields

Important fields include:

metrics: Final test metrics, such as mse and mae for forecasting.
history: Per-epoch training loss and validation metrics when the engine records them.
hparams: The full hyperparameter snapshot saved with the run.
run_config: The normalized result-affecting configuration used to identify equivalent experiment setups.
config_hash: A stable fingerprint of run_config. It excludes seed and storage metadata.
run_id: A seeded run identity, currently shaped as seed{N}-{config_hash}.
artifacts: Downloadable files produced by the run, such as the best model checkpoint.

Storage Layout

Result records are grouped by configuration:

results/
  records/
    DLinear/
      Forecast/
        ETTh1/
          {config_hash}/
            config.json
            seed1.json
            seed2.json

Model artifacts are stored separately:

results/
  artifacts/
    DLinear/
      Forecast/
        ETTh1/
          {config_hash}/
            seed1/
              best_model.pth

The training engine may also keep its own run directory:

results/
  runs/
    DLinear/
      Forecast/
        ETTh1/
          {config_hash}/
            seed1-{config_hash}/
              best_model.pth

Why Hash Configurations?

Older result filenames were based on model, task, dataset, and seed. That meant two different configurations with the same seed could overwrite each other.

The current layout uses config_hash so different settings are stored independently and the leaderboard can compare arbitrary configurations instead of only a few hard-coded columns.

What Goes Into the Hash?

The hash is built from result-affecting configuration:

model, task, and dataset;
task settings such as windows, pred_len, horizon, masks, and splits;
model settings such as architecture dimensions;
training settings such as lr, batch_size, epochs, loss, and scaler.

It excludes storage and infrastructure settings such as save_dir, device, num_worker, and pin_memory.

Artifact Backends

The local artifact backend is implemented today. Database, object storage, or model-hub backends can implement the same artifact backend interface and attach downloadable model locations to RunResult.artifacts without changing the experiment workflow.