File size: 826 Bytes
abbb696 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 | #!/usr/bin/env bash
# Linear-only ablation: same depth (7 layers all-L) vs nonlinear LLNLNLN.
# Validates the C1 motivation: linear SIM cannot perform nonlinear range-angle inference.
set -e
cd "$(dirname "$0")/../.."
K=2
# Linear-only: 7 L layers, share_mask same prefix
echo "=== Training linear-only LLLLLLL ==="
CUDA_VISIBLE_DEVICES=0 python train_joint_dual.py \
--cache experiments_v2/caches/v2_K${K}_15k.pt \
--layout LLLLLLL --share-mask 1100000 \
--epochs 100 --batch-size 96 \
--lambda-sense 0.7 --warmup-epochs 30 \
--ckpt experiments_v2/checkpoints/v2_K${K}_linear_only.pt \
2>&1 | tee experiments_v2/logs/train_K${K}_linear_only.log
# Reference: nonlinear baseline (already trained as main, re-link here for clarity)
echo "Reference: experiments_v2/checkpoints/v2_K${K}_main.pt (nonlinear LLNLNLN)"
|