Build LitData chunks

Training I/O is dominated by GDAL random reads. To saturate multiple A6000s, MarsRecon pre-builds LitData chunks once and streams them at training time.

Two builders

Script	Output format
scripts/training/build_litdata.py	Processed: includes sun-vector estimates, topographic residuals, fills.
scripts/training/build_litdata_raw.py	Raw: elevation + bounds + CRS only — for ablations / custom pipelines.

Both write to the directory specified by data.litdata_root in your config.

Run

PYTHONPATH=src uv run python scripts/training/build_litdata.py \
    --config configs/train_hirise.yaml \
    --workers 96

--workers should match your CPU count. The script:

1. Instantiates the same MarsHiRISEDTM + HiRISEGeoSampler the trainer will use.
2. Iterates all patches in the chosen split.
3. Runs the same adapter (DepthFMHiRISEAdapterCached) — sun-vector, fills, normalization.
4. Serializes each patch to a binary chunk.

Verifying chunks

PYTHONPATH=src uv run python -m dataset.stats.compute_stats_litdata \
    --litdata-root /scratch/litdata/hirise --split train

This runs a Welford accumulator over the pre-built chunks for a sanity-check of the per-channel statistics.

See: dataset.stats.compute_stats_litdata.