PMW proxy pretraining¶

geo2wf-export geo-pmw creates a larger GEO-to-PMW paired dataset with the same one-channel target shape as GEO-to-SAR. The intent is initialization on a related satellite image reconstruction task before SAR fine-tuning.

Supported PMW channels¶

Each target remains one channel, even in 10-band GEO experiments. That preserves model output dimensionality between proxy pretraining and SAR fine-tuning.

Export¶

uv run geo2wf-export geo-pmw \
  --config configs/config_pretrain_geo_pmw.yaml

uv run geo2wf-export geo-pmw \
  --config configs/config_pretrain_geo_pmw_10bands.yaml

uv run geo2wf-export geo-pmw \
  --config configs/config_pretrain_geo_pmw_10bands_era5.yaml

The exporter is PMW-anchored: it finds a closest GEO occurrence for each acceptable PMW observation, builds the same geographic crop contract as the SAR exporter, and writes generic condition_path and target_path columns. This is why PairedImageDataset can load either task without a separate class.

Moving from pretraining to SAR¶

Transfer is explicit: use --weights-only-path so model weights load strictly while optimizer, scheduler, epoch, and step state start fresh.

uv run geo2wf-train \
  model=conditional_diffusion \
  --weights-only-path /path/to/pmw-pretraining.ckpt

The model architecture and condition/target widths must match exactly.