This directory contains Slurm job scripts for running experiments on the Klab HPC infrastructure.

Trains an RNN model with temporal stability loss as an alternative objective to energy efficiency.

Usage:

1. Update the following placeholders in the script:

   • [email protected] → your actual email address
   • your_env_name → your conda environment name
   • your_username → your username on the cluster

2. Make the script executable:

   chmod +x jobscripts/submit_temporal_stability.sh

3. Submit the job:

   sbatch jobscripts/submit_temporal_stability.sh

4. Monitor the job:

   squeue -u $USER

Parameters:

• Partition: klab-gpu (uses H100 80GB GPU)
• Resources: 8 CPUs, 64GB RAM, 1 GPU
• Time limit: 48 hours
• Model: Temporal stability loss with L2 objective
• Dataset: MS-COCO with DeepGaze fixations

Output:

• Logs: logs/temporal_stability_<job_id>.out/err
• Models: models/temporal_stability/
• Wandb logging enabled

• Single experiment: Use H100 GPU (klab-gpu partition)
• Multiple experiments: Use L40S GPUs (klab-l40s partition) to run parallel jobs
• CPU-only baseline: Use klab-cpu partition

Check GPU usage during training:

watch -n 1 nvidia-smi

View job logs in real-time:

tail -f logs/temporal_stability_<job_id>.out