refactor: 将子模块转为普通目录，移除外部 git 依赖

- 移除 JEPA/lejepa-identifiability 子模块 gitlink
- 移除 research/multiply/MultiPLY 子模块 gitlink
- 删除 .gitmodules（不再有外部 URL 依赖）
- 两个目录内容作为普通文件纳入主仓库追踪
- 删除各自内部 .git 目录，消除嵌套 git 仓库

JEPA/lejepa-identifiability/experiments/slurm/launch_2d.sh

@@ -0,0 +1,27 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_2d
+#SBATCH --output=logs/2d_%A_%a.out
+#SBATCH --error=logs/2d_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=16G
+#SBATCH --time=02:00:00
+#SBATCH --array=0-7   # 8 runs; seeds loop inside
+
+RUNS=(spiral_lejepa spiral_whiten banana_lejepa banana_whiten \
+      sinusoid_lejepa sinusoid_whiten nvp_lejepa nvp_whiten)
+SEEDS=(1337 1338 1339)
+
+RUN=${RUNS[$SLURM_ARRAY_TASK_ID]}
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+
+mkdir -p logs
+for SEED in "${SEEDS[@]}"; do
+    echo "${RUN} seed=${SEED}"
+    python run.py --config configs/2d.yaml \
+        --run "${RUN}" --seed "${SEED}"
+done

JEPA/lejepa-identifiability/experiments/slurm/launch_gennorm.sh

@@ -0,0 +1,32 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_gennorm
+#SBATCH --output=logs/gennorm_%A_%a.out
+#SBATCH --error=logs/gennorm_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=16G
+#SBATCH --time=03:00:00
+#SBATCH --array=0-71   # 8 runs x 9 alphas; seeds loop inside
+
+RUNS=(spiral_lejepa spiral_whiten banana_lejepa banana_whiten \
+      sinusoid_lejepa sinusoid_whiten nvp_lejepa nvp_whiten)
+ALPHAS=(0.125 0.25 0.5 1.0 2.0 4.0 8.0 16.0 32.0)
+SEEDS=(1337 1338 1339)
+
+N_ALPHAS=${#ALPHAS[@]}
+RUN_IDX=$(( SLURM_ARRAY_TASK_ID / N_ALPHAS ))
+ALPHA_IDX=$(( SLURM_ARRAY_TASK_ID % N_ALPHAS ))
+RUN=${RUNS[$RUN_IDX]}
+ALPHA=${ALPHAS[$ALPHA_IDX]}
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+
+mkdir -p logs
+for SEED in "${SEEDS[@]}"; do
+    echo "${RUN} alpha=${ALPHA} seed=${SEED}"
+    python run.py --config configs/gennorm.yaml \
+        --run "${RUN}" --alpha "${ALPHA}" --seed "${SEED}"
+done

JEPA/lejepa-identifiability/experiments/slurm/launch_grid.sh

@@ -0,0 +1,29 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_grid
+#SBATCH --output=logs/grid_%A_%a.out
+#SBATCH --error=logs/grid_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=16G
+#SBATCH --time=12:00:00
+#SBATCH --array=0-8   # 9 lambda values; rhos and seeds loop inside
+
+LAMBS=(1e-6 1e-5 1e-4 1e-3 5e-3 1e-2 5e-2 1e-1 5e-1)
+RHOS=(0.3 0.5 0.7 0.8 0.9 0.95 0.99)
+SEEDS=(0 1 2)
+
+LAMB=${LAMBS[$SLURM_ARRAY_TASK_ID]}
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+
+mkdir -p logs
+for RHO in "${RHOS[@]}"; do
+    for SEED in "${SEEDS[@]}"; do
+        echo "lamb=${LAMB} rho=${RHO} seed=${SEED}"
+        python run.py --config configs/grid.yaml \
+            --lamb "${LAMB}" --rho "${RHO}" --seed "${SEED}"
+    done
+done

JEPA/lejepa-identifiability/experiments/slurm/launch_reacher_ou.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_ou
+#SBATCH --output=logs/ou_%A_%a.out
+#SBATCH --error=logs/ou_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=32G
+#SBATCH --time=24:00:00
+#SBATCH --array=0-6
+
+# Each task: prerender eval (skipped if exists) + 200k images for one rho,
+# then train 4 lambdas × 3 seeds × 3 inits = 36 training runs
+
+RHOS=(0.3 0.5 0.7 0.8 0.9 0.95 0.99)
+RHO_RAW=${RHOS[$SLURM_ARRAY_TASK_ID]}
+RHO=$(printf "%.2f" $RHO_RAW)
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+export MUJOCO_GL=egl
+mkdir -p logs
+
+echo "Node: $(hostname) | rho=${RHO} | Start: $(date)"
+
+# Step 1: Prerender (eval + this rho)
+python prerender.py eval
+python prerender.py ou --rho "${RHO}"
+
+# Step 2: Train
+python run_reacher.py --config configs/reacher.yaml \
+    --data_dir "data/reacher/ou/rho=${RHO}"
+
+echo "Done: $(date)"

JEPA/lejepa-identifiability/experiments/slurm/launch_reacher_traj.sh

@@ -0,0 +1,36 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_traj
+#SBATCH --output=logs/traj_%A_%a.out
+#SBATCH --error=logs/traj_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=32G
+#SBATCH --time=24:00:00
+#SBATCH --array=0-6
+
+# Each task: prerender eval (skipped if exists) + 200k images for one delta,
+# then train 4 lambdas × 3 seeds × 3 inits = 36 training runs
+
+DELTAS=(1 2 4 8 16 32 64)
+DELTA=${DELTAS[$SLURM_ARRAY_TASK_ID]}
+
+H5_PATH="data/reacher.h5"
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+export MUJOCO_GL=egl
+mkdir -p logs
+
+echo "Node: $(hostname) | delta=${DELTA} | Start: $(date)"
+
+# Step 1: Prerender (eval + this delta)
+python prerender.py eval
+python prerender.py traj --delta "${DELTA}" --h5_path "${H5_PATH}"
+
+# Step 2: Train
+python run_reacher.py --config configs/reacher.yaml \
+    --data_dir "data/reacher/traj/delta=${DELTA}"
+
+echo "Done: $(date)"

JEPA/lejepa-identifiability/experiments/slurm/launch_scaling.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+#SBATCH --job-name=lejepa_scale
+#SBATCH --output=logs/scale_%A_%a.out
+#SBATCH --error=logs/scale_%A_%a.err
+#SBATCH --partition=gpuq
+#SBATCH --qos=slow_nice
+#SBATCH --gres=gpu:v100:1
+#SBATCH --cpus-per-task=4
+#SBATCH --mem=16G
+#SBATCH --time=12:00:00     # ← bumped from 6h: 3 modes × 5 seeds = 15 runs per N
+#SBATCH --array=0-9   # 10 dims
+
+DIMS=(2 4 8 16 32 64 128 256 512 1024)
+SEEDS=(0 1 2 3 4)
+MODES=(lejepa whiten infonce)
+
+N=${DIMS[$SLURM_ARRAY_TASK_ID]}
+
+eval "$(conda shell.bash hook)"
+conda activate pytorch
+
+mkdir -p logs
+
+for MODE in "${MODES[@]}"; do
+    for SEED in "${SEEDS[@]}"; do
+        echo "N=${N} seed=${SEED} mode=${MODE}"
+        python -u run.py --config configs/scaling.yaml \
+            --N "${N}" --seed "${SEED}" --mode "${MODE}"
+    done
+done