experiments_v2/scripts/run_ablate_lambda.sh

#!/usr/bin/env bash
# Lambda sweep on K=2 (default). Used for the rate vs sensing trade-off ablation.
set -e
cd "$(dirname "$0")/../.."
K=2

for LAM in 0.0 0.1 0.3 0.5 0.7 0.85 1.0 2.0; do
  TAG=$(echo $LAM | tr '.' 'p')
  echo "=== Training K=${K} lambda=${LAM} ==="
  CUDA_VISIBLE_DEVICES=0 python train_joint_dual.py \
    --cache experiments_v2/caches/v2_K${K}_15k.pt \
    --layout LLNLNLN --share-mask 1100000 \
    --epochs 100 --batch-size 96 \
    --lambda-sense ${LAM} --warmup-epochs 30 \
    --ckpt experiments_v2/checkpoints/v2_K${K}_ablate_lambda${TAG}.pt \
    2>&1 | tee experiments_v2/logs/train_K${K}_lambda${TAG}.log
done