#!/usr/bin/env bash
# SPDX-License-Identifier: MIT
# YS-EN-0003 · the full experiment, in the order it was run.
#   bash run_all.sh            # all phases (~7 h on a GTX 1060 6 GB)
#   bash run_all.sh 2          # one phase (5 needs ~/ys-work/ogle/ogle_slc.npz from build_ogle.py)
# Every run appends one line to results/image_runs.jsonl.
set -euo pipefail
cd "$(dirname "$0")"
PY=${PY:-python3}
T="$PY train_image.py ${SKIP_DONE:+--skip-done}"   # SKIP_DONE=1 resumes an interrupted grid
phase=${1:-all}
R="0 1 2"

if [[ $phase == all || $phase == 1 ]]; then   # which encoding? (ResNet-18, ImageNet weights)
  for r in $R; do for e in gasf gadf mtf rp line stack; do
    $T --encoding $e --backbone resnet18 --pretrained --resample $r
  done; done
fi

if [[ $phase == all || $phase == 2 ]]; then   # which backbone? does pretraining matter?
  E=${E:-stack}
  for r in $R; do
    $T --encoding $E --backbone resnet18 --resample $r                      # same net, random init
    for b in resnet50 convnext_t efficientnet_b0 vit_s16 dinov2_s; do
      $T --encoding $E --backbone $b --pretrained --resample $r
    done
    for b in resnet50 dinov2_s; do                                         # frozen features only
      $T --encoding $E --backbone $b --probe --resample $r
    done
  done
fi

if [[ $phase == all || $phase == 3 ]]; then   # re-creation of the 2023 notebook
  for r in $R; do $T --original --resample $r; done
fi

if [[ $phase == all || $phase == 4 ]]; then   # learning curve: is pretraining worth more with less data?
  E=${E:-stack}; B=${B:-resnet18}
  for r in $R; do for n in 25 50 100 250 500; do
    $T --encoding $E --backbone $B --pretrained --resample $r --ntrain $n
    $T --encoding $E --backbone $B --resample $r --ntrain $n
  done; done
fi

if [[ $phase == all || $phase == 5 ]]; then   # the same image model on OGLE-SLC (shape only)
  for r in $R; do $T --dataset ogle --encoding stack --backbone resnet18 --pretrained --resample $r; done
fi