scripts/_retry_gpu3_depthmask.sh

#!/usr/bin/env bash
# One-off retry of the GPU3 chain for run_4scenes_gtpose_50k_depthmask.sh
# (original failed due to network_gui port collision with the GPU2 job).
set -euo pipefail

BASE=/home/szha0669/storage/blur_slam_exp
REPO_BAGS=/srv2/szha0669/blur_slam_exp/repos/BAGS
LOG="$BASE/outputs/logs"
ITERS=50000
DEPTH_ALPHA=0.1
SAVE="10000 20000 30000 40000 50000"

declare -A DATA_DIR=(
  [scene0031_00]="${BASE}/data/i2slam_trigsplat_gt/scene0031_00_abl2_gtdepth_evssmcolor"
  [tum_fr2_xyz]="${BASE}/data/i2slam_trigsplat_gt/tum_fr2_xyz_abl2_gtdepth"
)

run_scene() {
  local SCENE=$1 GPU=$2
  local DATA="${DATA_DIR[$SCENE]}"
  local OUT_BAGS="$BASE/outputs/bags_i2slam_gtall_50k_depthmask/${SCENE}"
  mkdir -p "$OUT_BAGS"

  echo "[BAGS][$SCENE] start $(date) gpu=$GPU"
  conda run -n bags --cwd "$REPO_BAGS" \
    env CUDA_VISIBLE_DEVICES=$GPU OMP_NUM_THREADS=4 \
    python train.py \
      -s "$DATA" -m "$OUT_BAGS" \
      --images images -r 2 \
      --iterations $ITERS \
      --ms_steps 6000 \
      --test_iterations $SAVE \
      --save_iterations $SAVE \
      --checkpoint_iterations 25000 50000 \
      --use_depth_loss --depth_loss_alpha $DEPTH_ALPHA \
      --use_mask_loss --mask_loss_alpha 0.001 \
      --port $((6020 + GPU)) \
      --quiet
  echo "[BAGS][$SCENE] training done $(date)"

  conda run -n bags --cwd "$REPO_BAGS" \
    env CUDA_VISIBLE_DEVICES=$GPU \
    python render.py -m "$OUT_BAGS" --iteration $ITERS --quiet
  echo "[BAGS][$SCENE] render done $(date)"
}

run_scene scene0031_00 3
run_scene tum_fr2_xyz 3
echo "=== GPU3 chain done [$(date)] ==="