Experiment Roadmap

Prioritized order

Use this order to maximize signal per GPU hour:

  1. Base architecture, decoder-only tuning.
  2. Base architecture, LoRA (rank=8, then rank=16).
  3. Backbone scale sweep with best method from steps 1-2.
  4. RGB plus height adapter with best current backbone.
  5. Progressive unfreezing with layer-wise LR decay.
  6. Pseudo-label self-training cycle.
  7. Optional domain-adaptive pretraining.

Stop and escalation rules

  • If Dice gain is less than 2% and AP50 gain is less than 1% across two consecutive experiments, escalate to the next architecture class.
  • If inference slowdown is greater than 25% for less than 2% quality gain, reject for production path.
  • Accept only changes that improve all held-out geographic blocks, not just global average.

Concrete first experiment set

  1. decoder_only: freeze encoder/prompt, train decoder, lr=1e-4, 20 epochs.
  2. lora_rank_8: LoRA on decoder and top encoder attention, lr=2e-4, 20 epochs.
  3. lora_rank_16: same setup, compare quality versus memory/time.
  4. rgb_plus_height_adapter: 4->3 adapter plus decoder LoRA, lr=1e-4, 25 epochs.
  5. bitfit_decoder_plus_top_encoder: bias-only tuning, lr=5e-4, 20 epochs.
  6. progressive_unfreeze_llrd: unfreeze top two encoder stages, continue 10-15 epochs.
  7. backbone_scale_sweep: repeat best recipe on small/base/large candidates.
  8. pseudo_label_round_1: confidence-filtered pseudo-label retraining.

Timeline template (4 weeks)

  • Week 1: baseline evaluation harness and zero-shot benchmark.
  • Week 2: decoder-only and LoRA experiments.
  • Week 3: RGB plus height adapter and variant comparisons.
  • Week 4: checkpoint integration, regression checks, and report.

See also:

PreviousImplementation Plan NextEvaluation and Decision Matrix