Automating SAR-ATR Workflows on DSRC and Cloud HPC Clusters

In this abstract, we present the automation and parallel execution of two Synthetic Aperture Radar Automatic Target Recognition (SAR-ATR) object detection workflows using the Parsl Python library and the Parallel Works platform.

The on-premise workflow runs on the Navy DSRC cluster Koehr. The cluster connects to a user workspace within the Parallel Works platform via SSH and Kerberos for secure authentication. A first step uses Xpatch to apply the shooting-and-bouncing ray (SBR) method to predict realistic radar signatures for 3D target models, processing each target in parallel. The generated data undergoes processing to apply random scaling factors, phase shifts, and noise to some image properties. Following this, several convolutional neural networks are trained using both synthetic and measured data to evaluate accuracy on measured SAR data. The final step creates an ensemble model by averaging the probability vectors of the individual models, calculating performance metrics, and saving the ensemble confusion matrix for further analysis. The resulting model is portable and can be used for subsequent inference on the edge.

The cloud workflow operates on a SLURM elastic cluster hosted on AWS. Configured via the Parallel Works platform, users can specify the region, instance types, and additional configurations like EFA and shared filesystems (NFS or Lustre). When the cluster starts, a controller node is provisioned, and compute nodes are added on demand based on job submissions to the SLURM scheduler, and deleted when idle. This workflow aims to train SAR-ATR classifiers using the Distribution A SAMPLE dataset, containing both measured and synthetic SAR data. Initially, noise is added to the images in parallel using Python and MPI with the mpi4py library. These images are then processed using either Matlab or Python, demonstrating the workflow’s flexibility. Several ResNet-18 models are trained in parallel on the synthetic data, with performance evaluated on the measured data. Finally, results for each model are merged and summarized.

Both workflows use the Parsl HighThroughputExecutor with the SlurmProvider and PBSProProvider to manage the job. Parsl workers connect back to the user workspace through SSH tunnels, enabling secure, hybrid workflows across different clusters and sites. This setup highlights the efficiency and adaptability of the Parallel Works platform in handling complex SAR-ATR workflows in diverse computing environments.

PRESENTER

Shaxted, Matthew
shaxted@parallelworks.com
847-254-0230

Parallel Works

CO-AUTHOR

Vidal Torreira, Alvaro
alvaro@parallelworks.com

CATEGORY

Artificial Intelligence / Machine Learning usage for HPC Applications

SYSTEMS USED

Koehr, AWS Cloud

SECRET

No