Skip to content

README.md

Latest commit

 

History

History
153 lines (118 loc) · 5.78 KB

README.md

File metadata and controls

153 lines (118 loc) · 5.78 KB

PINNacle Qing

This project is based on PINNacle.

Experiment Scripts

First, you need to replace all occurrences of /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source with your own pinnsagent path.

To run a single experiment (for training a single PDE)

For simplicity, we only use config file to run experiments, if --yaml_path is specified, the arguments in the config file will be used, otherwise the default arguments defined in benchmark.py will be used.

To run a simple experiment, you can use the following command: python benchmark.py [--yaml_path yaml_path]

cd /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source/pinnacle
source /gpfs/0607-cluster/miniconda3/etc/profile.d/conda.sh
conda activate pinnsagent
python benchmark.py \
 --yaml_path "./config/1_1-test_single_pde.yaml"
  • The results of the experiment will be saved in the ./{args.output_dir}/{args.name} directory
  • where args.output_dir and args.name are defined in the .yaml config file.

Database Preparation for PINNsAgent

(1) Generate Random Search .yaml files

To create database for PINNsAgent, we first generate a large number of PDEs with different hyperparameters. This is done by running a random search over the hyperparameter space.

Generate random search configuration files

cd /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source/pinnacle
bash ./scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/1_generate_random_search_yaml.sh

This script generates 10 batches (30 configs per PDE per batch) and saves them to scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/1_ICML_2025_configs/ directory, structured as follows:

scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/1_ICML_2025_configs/
├── Batch_1/
│   ├── 1d/
│   │   ├── Burgers1D-30/
│   │   │   ├── train_1.yaml
│   │   │   ├── train_2.yaml
│   │   │   └── ...
│   │   └── Wave1D-30/
│   ├── 2d/
│   │   ├── Burgers2D-30/
│   │   ├── Wave2D_Heterogeneous-30/
│   │   └── ...
│   ├── 3d/
│   │   └── Poisson3D_ComplexGeometry-30/
│   └── nd/
│       ├── PoissonND-30/
│       └── HeatND-30/
├── Batch_2/
│   └── ...
└── Batch_10/
    └── ...

(2) Generate Batch Training .sh files

Then you can generate training shell .sh scripts for all batches

cd /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source/pinnacle
python scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/2_generate_batches_shell.py
  • Do remember to update config_base, script_base, work_dir, and log_base in the script to your paths before running it, as it will generate the .sh files in the same directory.

This creates a complete training script system:

scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/2_run_experiments_per_batchs/
├── Batch_1/
│   ├── 1_train_1d.sh        # Train 1D PDEs (GPU 0)
│   ├── 2_train_2d.sh        # Train 2D PDEs (GPU 0-7, parallel)
│   ├── 3_train_3d.sh        # Train 3D PDEs (GPU 1)
│   ├── 4_train_nd.sh        # Train ND PDEs (GPU 7)
│   └── run_batch_1.sh       # Run all 4 scripts in parallel
├── Batch_2/ ... Batch_10/
├── generate_summary.py      # Generate training summary
└── run_all_batches.sh       # Master control script

Features:

  • Automatic GPU allocation (1D: GPU0, 2D: GPU0-7, 3D: GPU1, ND: GPU7)
  • Resume from checkpoint (skips *_ok.yaml files)
  • Detailed logging in log/2025_05_15/RandomSearch_Datasets_high_quality/

(3) Run the Batch Training .sh files

After generating the .sh files, you can run them to start the training.

(3).1 Run multiple batches experiments

You can run the following command to run multiple batches experiments:

cd /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source/pinnacle
cd scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/2_run_experiments_per_batchs
source /gpfs/0607-cluster/miniconda3/etc/profile.d/conda.sh
conda activate pinnsagent

# Serial execution (one batch after another)
./run_all_batches.sh 1-5     # Run batches 1 to 5
./run_all_batches.sh 6-10    # Run batches 6 to 10

(3).2 Run single batch experiments

Or you can run the following command to run single batch experiments:

cd /gpfs/0607-cluster/qingpowuwu/Project_4_PINNsAgent/1_Ours/PINNsAgent_Unified/pinnsagent_progress-open_source/pinnacle
cd scripts/2025_05_15/1_RandomSearch_Datasets_high_quality/2_run_experiments_per_batchs
source /gpfs/0607-cluster/miniconda3/etc/profile.d/conda.sh
conda activate pinnsagent
# run single batch experiments
./run_all_batches.sh 1
./run_all_batches.sh 2
./run_all_batches.sh 3
./run_all_batches.sh 4
./run_all_batches.sh 5
./run_all_batches.sh 6
./run_all_batches.sh 7
./run_all_batches.sh 8
./run_all_batches.sh 9
./run_all_batches.sh 10

Monitor progress:**

# View master log
tail -f log/2025_05_15/RandomSearch_Datasets_high_quality/master.log

# View specific batch logs
tail -f log/2025_05_15/RandomSearch_Datasets_high_quality/Batch_1/1d.log
tail -f log/2025_05_15/RandomSearch_Datasets_high_quality/Batch_1/2d.log

or you can also track them by running 3_1_get_statistics_for-1-batch.py or 3_1_get_statistics_for-10-batch.py

Results:

  • Training outputs: ./output/
  • Completed configs are renamed to *_ok.yaml
  • Summary reports: log/2025_05_15/RandomSearch_Datasets_high_quality/Batch_X/summary_report.txt