Files and links (1)
PDFOpen Access (License Unspecified), Open Access
Thesis
Composable machine learning pipeline for fast deployment with neural architecture search
Master of Science (M.S.), Drexel University
Jun 2025
DOI:
https://doi.org/10.17918/00011069
Abstract
Deploying neural networks in real-time, low-latency, and power-constrained environments, such as high-speed imaging systems, requires efficient hardware-aware machine learning workflows. This thesis presents a composable framework designed to automate and accelerate the design, evaluation, and deployment of optimized neural network models for FPGA-based inference. The architecture is implemented using a modular, containerized approach, where each stage of the machine learning lifecycle is encapsulated within dedicated Docker containers. These are orchestrated using Kubernetes, supporting horizontal scalability and efficient parallel experimentation across diverse environments. The pipeline integrates QKeras for model pruning and precision tuning, HLS4ML for high-level synthesis (HLS), and Vivado for hardware compilation and resource reporting. The system aims to minimize latency and power consumption while maintaining accuracy, making it ideal for real-time applications. A specific use-case container, referred to as the "Fire Over Many Overlaps (FOMO) Frame Grabber", shows the pipeline's practical application. This use case utilizes automated Neural Architecture Search (NAS) for hyperparameter optimization, along with quantization-aware training, pruning, and HLS compilation. To address the challenge of tracking, monitoring, and reusing experimental results, the pipeline integrates DataFed for structured metadata logging and experiment management. It captures and stores training metrics, model artifacts, Vivado reports, and environment configurations. A Globus-enabled container ensures seamless and secure data transfer to remote DataFed storage, promoting long-term reproducibility and enabling intelligent, data-driven NAS by building upon prior runs. This composable machine learning pipeline streamlines deployment, reduces manual intervention, and accelerates design spaces exploration for low-power, low-latency AI applications. Its scalable and reproducible design offers a solution for both academic research and industrial use cases requiring real-time, hardware-efficient inference.
Metrics
26 File views/ downloads
26 Record Views
Details
- Title
- Composable machine learning pipeline for fast deployment with neural architecture search
- Creators
- Veronica Obute
- Contributors
- Joshua Agar (Advisor)
- Awarding Institution
- Drexel University
- Degree Awarded
- Master of Science (M.S.)
- Publisher
- Drexel University; Philadelphia, Pennsylvania
- Number of pages
- xi, 38 pages
- Resource Type
- Thesis
- Language
- English
- Academic Unit
- College of Engineering (1970-2026); Electrical (and Computer) Engineering [Historical]; Drexel University
- Other Identifier
- 991022058732804721