# MACS experiments

Outputs of the MACS v2 experiments (`experiments/macs_v2` in the MACS-2 repository), backed up per host.

```
macs_v2/
  results/<user>@<host>.results.csv        one row per finished run on that host (task, target, variant, budget, seed,
                                           test metrics, timing, host, commit); the same run may appear on two hosts
                                           when it was re-run - keep the latest row per (task, target, variant, budget, seed)
  runs/<task>/<target>/<variant>/budget-<percent>/seed-<seed>/
      result.json                          full record: config, source weights (prior and corrected), teacher diagnostics,
                                           per-round records, test metrics (patch- and image-level, calibration), resources
      rounds.csv                           per active-learning round
      log.txt                              the run's log
  models/<task>/<target>/<variant>/budget-<percent>_seed-<seed>.pth
                                           student checkpoints: seed 1 at 10% and 100% for the paper's variants
                                           (macs-v2-tta, macs-v7, macs-v10, neuroadda, single-labeled, scratch)
```

Tasks: `membrane` (9 domains, each target uses the other 8 as sources) and `mitochondria` (5 domains, 4 sources).
Budget = annotation budget as a percentage of the target's training patches. Budgets: 2, 5, 10 (seeds 1-3) and 20, 40, 60, 80, 100 (seed 1). Variants named `macs-*` are MACS recipes
(`macs-v2-tta`, `macs-v7`, `macs-v10` are the reported ones; the others are ablations and development history,
see `config` inside each result.json). Source (teacher) models live in `imAniksahA/blab` under
`c/MACS/source-models/v3/`.