Skip to content

Quick Start

Use this page for the shortest path from setup to first train/inference run.

Environment & Dependencies

  • Python: 3.12.3
  • Install runtime dependencies:
/work/envs/depth/bin/python -m pip install -r requirements.txt

The root requirements.txt installs this repository from the curated dependencies in pyproject.toml.

Quick Training

Pixel-space GeoTIFF training with the default super-config:

/work/envs/depth/bin/python train.py \
  --scenario temperature

Scenario choices:

/work/envs/depth/bin/python train.py --scenario temperature
/work/envs/depth/bin/python train.py --scenario salinity
/work/envs/depth/bin/python train.py --scenario joint

The default training config is src/depth_recon/configs/px_space/training_super_config.yaml. --scenario derives the field list, salinity data flag, generated channels, and condition channels. Use --set data.*, --set model.*, or --set training.* for one-off overrides after scenario resolution.

Latent diffusion uses the separate src/depth_recon/configs/lat_space/ configs. See Autoencoder + Latent Diffusion for architecture, goals, limitations, and workflow details.

Quick Inference

For public inference from PyPI, install the package and run one ISO week:

python -m pip install depth-recon

from depth_recon import run_week_inference

run_dir = run_week_inference(
    year=2015,
    iso_week=25,
    rectangle=(-20.0, 30.0, 10.0, 50.0),
    device="cuda",
)
print(run_dir)

The package downloads the public model artifacts from Hugging Face, downloads EN4/ARGO and OSTIA source files when needed, and writes prediction GeoTIFFs under inference/outputs/depthdif_argo_<YYYYMMDD>/. See Public Inference Package for the full API and CLI.

For repository-local smoke checks, set config/checkpoint constants at the top of src/depth_recon/inference/run_single.py, then run:

/work/envs/depth/bin/python -m depth_recon.inference.run_single

For EO multiband runs, use: - CONFIG_PATH = "src/depth_recon/configs/px_space/inference_super_config.yaml" - SCENARIO = "temperature", "salinity", or "joint"

The inference super-config has top-level data, model, training, and inference sections. Scenario resolution derives the same model/data channel contract used by training.

To export one stitched world raster and prepare the hosted Cesium assets afterward, use:

/work/envs/depth/bin/python -m depth_recon.inference.export_global --year 2010 --iso-week 1
/work/envs/depth/bin/python -m depth_recon.inference.export_cesium_globe_assets \
  --run-dir inference/outputs/global_top_band_<YYYYMMDD> \
  --public-base-url https://<bucket-or-site>/inference_production/globe/ \
  --rclone-remote r2:<bucket>/inference_production/globe \
  --rclone-sync-scope globe