Understanding outputs¶

RAITAP writes its outputs to the Hydra run directory. By default, this is a timestamped folder under ./outputs relative to the directory where you launched RAITAP.

If needed, you can override the output location via Global configuration.

Here is the output of an illustrative example run:

outputs/                                      # Hydra's default output directory
└── 2026-02-28/                               # Run date
    └── 14-30-45/                             # Run time
        ├── __main__.log                      # Run log
        ├── REPRODUCIBILITY.md                # Only when the run used a stochastic method
        ├── .hydra/                           # Hydra configuration logs
        │   ├── config.yaml
        │   ├── hydra.yaml
        │   └── overrides.yaml
        ├── metrics/                          # Metric module outputs (see below for details)
        │   ├── artifacts.json
        │   ├── metadata.json
        │   └── metrics.json
        ├── reports/                          # HTML report and report-only assets
        │   ├── report.html
        │   ├── report.zip
        │   ├── report_manifest.json
        │   └── _assets/
        ├── transparency/                     # Transparency module outputs (see below for details)
        │   ├── explainerA/
        │   │   ├── attributions.pt
        │   │   ├── visualisation1.png
        │   │   ├── visualisation2.png
        │   │   └── metadata.json
        │   └── explainerB/
        │       ├── attributions.pt
        │       ├── visualisation1.png
        │       └── metadata.json
        └── robustness/                       # Robustness module outputs (see below for details)
            ├── pgd/
            │   ├── robustness_data.pt
            │   ├── ImagePairVisualiser_0.png
            │   └── metadata.json
            └── fgsm/
                ├── robustness_data.pt
                ├── ImagePairVisualiser_0.png
                └── metadata.json

The transparency/ and robustness/ directories keep the full assessor artifacts for debugging and tracking. The reports/ directory contains the compact report, a ZIP package with the HTML, manifest, and curated report-only figure assets.

REPRODUCIBILITY.md is written only when the run used a stochastic method (for example SHAP GradientExplainer, a PGD attack, or an image-corruption assessor). It lists which artefacts are not bit-reproducible. The same caveat appears as a banner in the report and as a warning after the run. Deterministic runs write no such file. Each stochastic method's metadata.json also carries a semantics.stochastic: true field.

You may want to look at each module's output directory for more details:

Transparency

Robustness

Metrics

Reporting