CLI Reference¶

Run repository commands with /work/envs/depth/bin/python. Every parser listed below supports --help; that output is authoritative for optional flags and defaults.

Installed console scripts¶

Command	Purpose
`depth-recon-infer-week`	Run public legacy-checkpoint inference for one ISO week.
`depth-recon-download-argo`	Download the EN4 archives needed by one ISO week.
`depth-recon-download-ostia`	Download the OSTIA inputs needed by one ISO week.
`depth-recon-export-paper-week`	Export a configured paper-week model bundle.
`depth-recon-export-paper-metrics`	Compute metrics from paper-week artifacts.

depth-recon-infer-week --help

Training and diagnostics¶

/work/envs/depth/bin/python train.py --scenario temperature
/work/envs/depth/bin/python train_autoencoder.py
/work/envs/depth/bin/python src/depth_recon/scripts/benchmark_dataloader_settings.py --help

train.py is the maintained pixel/baseline training entry point. Autoencoder and latent components are experimental; see Autoencoder.

The complete held-out-2016 baseline workflow uses two independent GPU workers, stable W&B run identities, resumable checkpoints, and the paper/spectral export pipeline:

/work/envs/depth/bin/python src/depth_recon/scripts/run_baseline_2016_suite.py \
  --phase all --gpu-indices 0 1 \
  --output-root logs/baseline_2016_global \
  --validation-examples 100000 --max-epochs 8 --patience 2 \
  --checkpoint-every-n-train-steps 5000 --max-task-hours 6 \
  --resume-incomplete

Use --dry-run to inspect all model/scenario commands without requiring CUDA or W&B. Normal execution requires both requested GPUs and valid online W&B authentication; it never falls back to CPU.

Packaged data downloads¶

/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.data_download_packaged.download_aligned_argo_zarr --help
/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.data_download_packaged.download_exported_geotiff_dataset --help

Raw Copernicus/EN4 download shell workflows live below data_download_raw/. They are documented in Dataset downloads because their credentials, product candidates, and date ranges differ.

Dataset production¶

Check source coverage, export enriched profiles, export dense rasters, and package the result in that order:

/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.export_aligned_argo.a_check_export_sourcefiles --help
/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.export_aligned_argo.b_export_enriched_argo_profiles --help
/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.export_dataset_geotiff.export_dataset_geotiff --help
/work/envs/depth/bin/python -m depth_recon.data.dataset_creation.export_aligned_argo.c_package_huggingface_aligned_argo --help

The production default temperature source is potential, backed by EN4 POTM_CORRECTED. Select in-situ only to reproduce older TEMP exports.

Synthetic prior¶

/work/envs/depth/bin/python -m depth_recon.data.synthetic_dataset_creation.fit_vertical_offset_prior --help
/work/envs/depth/bin/python -m depth_recon.data.synthetic_dataset_creation.plot_vertical_offset_examples --help

Model inference and exports¶

/work/envs/depth/bin/python -m depth_recon.inference.run_single
/work/envs/depth/bin/python -m depth_recon.inference.export_global --help
/work/envs/depth/bin/python -m depth_recon.inference.export_global_variables --help
/work/envs/depth/bin/python -m depth_recon.inference.export_validation_error_summary --help

The single-variable exporter writes one stitched run. The paired wrapper owns the temperature/salinity production bundle and can trigger temporal and spectral products.

Paper and comparative evaluation¶

/work/envs/depth/bin/python -m depth_recon.inference.export_paper_week --help
/work/envs/depth/bin/python -m depth_recon.inference.export_paper_metrics --help
/work/envs/depth/bin/python -m depth_recon.inference.export_spectral_comparison_bundle --help

These commands require explicit models/checkpoints and preserve their run metadata. export_paper_metrics has both bundle and legacy per-method modes; prefer bundle mode for new results.

Analysis bundles¶

/work/envs/depth/bin/python -m depth_recon.inference.export_error_analysis_dashboard --help
/work/envs/depth/bin/python -m depth_recon.inference.export_cesium_globe_assets --help
/work/envs/depth/bin/python -m depth_recon.inference.export_wavenumber_spectra --help
/work/envs/depth/bin/python -m depth_recon.inference.export_temporal_global_variables --help
/work/envs/depth/bin/python -m depth_recon.inference.export_temporal_consistency_dashboard --help
/work/envs/depth/bin/python -m depth_recon.inference.export_temporal_cesium_globe_assets --help

export_wavenumber_spectra includes incomplete ocean patches by default and masks invalid pixels during processing. Pass --require-complete-patches for the stricter alternate policy.

Developer visualization utilities¶

Sampler-comparison and plotting CLIs are contributor diagnostics rather than production inference entry points:

/work/envs/depth/bin/python -m depth_recon.utils.compare_ddpm_ddim_sampling --help
/work/envs/depth/bin/python -m depth_recon.experiments.compare_ddpm_ddim_step_grid --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_argo_corrected_depth_distribution --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_argo_corrected_depth_histogram --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_argo_glorys_depth_mapping --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_glorys_target_alignment_shift --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_land_fraction_filter_examples --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.plot_loss_explanations --help
/work/envs/depth/bin/python -m depth_recon.utils.visualization.create_paper_header_image --help

Examples with all options are kept at the top of each task-specific script. The diagnostic samplers do not define quality thresholds or a recommended universal step count.