Back to Education
Coursework

EuroSAT transfer-learning benchmark

A 7-paradigm comparison on satellite imagery: LoRA matches full fine-tuning at 0.36% of the trainable parameters and 3.7× faster.

Coursework (EuroSAT_DL_Assignment); numbers read straight from committed notebook outputs.

Results

Every value links to the committed artifact that produces it.

MetricValueEvidence
Best overall — EfficientNet-B0, full fine-tune98.37%eurosat_benchmark.ipynb (cell 100)
LoRA — ViT-B/16, r = 898.22%eurosat_benchmark.ipynb (cell 94)
Trainable params under LoRA0.36% (310,292 / 86,108,948)eurosat_benchmark.ipynb (cell 91)
LoRA vs full fine-tune, wall-clock3.70× faster (3.70 vs 13.69 min)eurosat_benchmark.ipynb (cell 94)

What didn’t work

SimCLR burned 490 minutes of A100 time to reach 89.0% — roughly 133× the compute for a worse result than a short transfer-learning run.

Reported as a null, not hidden — the same honesty policy applies to every number on this site.

Stack

  • PyTorch
  • LoRA / PEFT
  • ViT-B/16
  • CLIP
  • SimCLR
  • Transfer learning