Optimizing inference of segmentation on high-resolution images in MLExchange

Area of Science:

• Scientific imaging
• Machine learning operations
• Data analysis at synchrotron facilities

Background:

• Synchrotron facilities generate large, high-resolution scientific images requiring efficient analysis.
• Current machine learning (ML) workflows for image segmentation can be computationally intensive, demanding significant manual annotation.
• MLExchange provides web-based user interfaces (UIs) for ML-driven data analysis pipelines.

Purpose of the Study:

• To optimize the segmentation inference workflow for large datasets within the MLExchange platform.
• To significantly reduce the computational time and manual annotation effort for ML-based image segmentation.
• To enhance the accessibility and speed of segmentation results for synchrotron users.

Main Methods:

• Optimized the sequence of data transfers between CPU and GPU.
• Introduced CPU parallelization for critical operations within the segmentation workflow.
• Implemented multi-GPU strategies for scalable inference.

Main Results:

• Achieved a nearly 3x speedup in segmentation inference runtime on a single GPU compared to the original workflow.
• Demonstrated over 40x speedup with 100 GPUs relative to the optimized single-GPU workflow.
• Significantly improved per-device, per-image frame computational efficiency.

Conclusions:

• The developed code optimizations substantially accelerate ML-based image segmentation for large scientific datasets.
• Faster segmentation inference workflows enhance user experience by reducing wait times for critical analysis results.
• These improvements make advanced ML tools more practical and accessible for synchrotron research.