summaryrefslogtreecommitdiff
path: root/research/flossing/run_srm_after_W.sh
diff options
context:
space:
mode:
authorYurenHao0426 <blackhao0426@gmail.com>2026-06-13 12:35:36 -0500
committerYurenHao0426 <blackhao0426@gmail.com>2026-06-13 12:35:36 -0500
commit66e0d8b9fd4d0f7a2231d689c055e26fdf1cf04a (patch)
treec29cba61124018755a19b02c9d33e3ad5f2e05cc /research/flossing/run_srm_after_W.sh
rrm workspace: TRM/HRM/SRM code, Maze dataset, dynamical-analysis pipelineHEADmain
Curated export for clone-and-run Maze training (2x A6000) + diagnostics. trm/hrm pretrain.py carry trajectory-augmentation code (backward-compatible). Heavy artifacts (checkpoints/wandb/npz) gitignored; see PROVENANCE.md. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Diffstat (limited to 'research/flossing/run_srm_after_W.sh')
-rwxr-xr-xresearch/flossing/run_srm_after_W.sh32
1 files changed, 32 insertions, 0 deletions
diff --git a/research/flossing/run_srm_after_W.sh b/research/flossing/run_srm_after_W.sh
new file mode 100755
index 0000000..5ecf4e6
--- /dev/null
+++ b/research/flossing/run_srm_after_W.sh
@@ -0,0 +1,32 @@
+#!/bin/bash
+# Wait for the W-runner (X→Y→Z→W pipeline) to finish, then launch SRM-AOL training.
+set -e
+cd /home/yurenh2/rrm/srm
+source /home/yurenh2/miniconda3/etc/profile.d/conda.sh
+conda activate rrm
+export CUDA_VISIBLE_DEVICES=2
+
+W_RUNNER_PID=$1
+LOG_DIR=/home/yurenh2/rrm/research/flossing
+echo "[$(date '+%H:%M:%S')] SRM-runner waiting for W-runner PID $W_RUNNER_PID..." >> $LOG_DIR/step4_runner.log
+while kill -0 "$W_RUNNER_PID" 2>/dev/null; do
+ sleep 30
+done
+echo "[$(date '+%H:%M:%S')] W done. Starting SRM-AOL from-scratch training (Sudoku 1k, hidden=512, 3000 steps)" >> $LOG_DIR/step4_runner.log
+
+# Make sure log/ckpt dirs exist
+mkdir -p runs ckpts
+
+# SRM-AOL training: same setup as XYZ baselines but with SRM model + Lyapunov diag every 25 steps
+python scripts/train_srm.py \
+ --n-steps 3000 --batch-size 8 \
+ --hidden-size 512 --n-iters 12 --n-aol-layers 2 \
+ --kappa 0.9 --eta 1.0 --alpha 1.0 \
+ --warmup-steps 200 \
+ --k-lyap 2 --lyap-iters 8 --lyap-every 50 \
+ --eval-every 300 --eval-n 512 --eval-batch-size 32 \
+ --out runs/srm_aol_v1_sudoku1k_3k.json \
+ --save-ckpt ckpts/srm_aol_v1_3k.pt \
+ > $LOG_DIR/srm_aol_run.log 2>&1
+
+echo "[$(date '+%H:%M:%S')] SRM-AOL training complete" >> $LOG_DIR/step4_runner.log