#!/bin/bash
#SBATCH --job-name=snn_hinge
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuA40x4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=8
#SBATCH --gpus-per-node=1
#SBATCH --mem=64G
#SBATCH --time=48:00:00
#SBATCH --output=runs/slurm_logs/%j_hinge.out
#SBATCH --error=runs/slurm_logs/%j_hinge.err

# ============================================================
# Hinge Loss Lyapunov Regularization Experiment
# ============================================================
# Hypothesis: Using hinge loss (only penalize chaos, not stability)
# will allow the network to learn while still preventing chaotic
# dynamics.
#
# Hinge loss: max(0, lambda)^2
# - Only penalizes positive Lyapunov (chaos)
# - Allows negative Lyapunov (stable dynamics) without penalty
# - Combined with warmup to let network start learning first
# ============================================================

set -e

PROJECT_DIR="/projects/bfqt/users/yurenh2/ml-projects/snn-training"
cd "$PROJECT_DIR"

mkdir -p runs/slurm_logs data

echo "============================================================"
echo "HINGE LOSS Lyapunov Regularization"
echo "Job ID: $SLURM_JOB_ID | Node: $SLURM_NODELIST"
echo "Start: $(date)"
echo "============================================================"
nvidia-smi --query-gpu=name,memory.total --format=csv,noheader
echo "============================================================"

# Test depths: 4, 8, 12, 16 conv layers
# Using hinge loss + 20 epoch warmup
python files/experiments/depth_scaling_benchmark.py \
    --dataset cifar100 \
    --depths 4 8 12 16 \
    --T 4 \
    --epochs 150 \
    --batch_size 128 \
    --lr 0.001 \
    --lambda_reg 0.3 \
    --lambda_target -0.1 \
    --reg_type hinge \
    --warmup_epochs 20 \
    --data_dir ./data \
    --out_dir runs/depth_scaling_hinge \
    --device cuda

echo "============================================================"
echo "Finished: $(date)"
echo "============================================================"