How namanmakkar/ece5545-ml-hardware-systems works
Main Components
• Core Architectural Elements Within Each Module:
– Notebooks: Each assignment contains one or more Jupyter notebooks (e.g., a0.ipynb, various numbered notebooks in a2, a3, a4) that drive the experiments and tests.
– Source Code (src): Each assignment has a src folder containing scripts (for example, model.py, train_utils.py in a0; various Python files in a2, a3, and a4) that implement core algorithms, processing pipelines, or specialized operations.
– Tests: Each assignment includes a tests directory with unit tests ensuring correctness.
– Supplementary Resources: PDFs (e.g., a1_submission.pdf, a2_submission.pdf, etc.), images, and additional documentation add context to the work.
• Cross-Assignment or External Resources: The repository-level README provides the high-level context for the assignments. The term paper represents an independent document summarizing research on communication compression.
Relationships and Interactions
• Data and Artifacts: Notebooks may process data (inputs in assignments like audio pre-processing in a2 or MNIST files in a4) that is then passed to the src code for model training, processing, or optimization. The output (results, plots, logs) is analyzed and validated using the tests.
• External Dependencies: Make sure to denote libraries/frameworks used in each assignment (for example, TVM in a3, the Arduino toolchain in a2, and any deep learning frameworks).
Incorporate Key Architectural Patterns and Design Principles
• Separation of Concerns: Break down within each module the different responsibilities: interactive exploratory work (notebooks), core logic and processing (src), and quality verification (tests).
Essential Elements
• Each major module label (e.g., “Assignment 1: Roofline Benchmarking” for a0, “Assignment 2: Keyword Spotting and Compression” for a2, “Assignment 3: DNN Compiling with TVM” for a3, “Assignment 4: DNN Primitives Approximation” for a4, and “Term Paper: Gradient Compression”).
• Inside each module, every sub-folder (Notebooks, src, tests, extra docs) should be mapped out as nodes.
– Data flow from inputs (raw data, hardware benchmarks) to processing (src logic) to outputs (visualizations, PDF reports).
• Annotations or legends to explain architectural patterns (e.g., “Microcontroller Deployment Flow” for a2, “Compiler Optimizations” for a3).