How minahilraza/brevitas_fixed_point works
Project Type and Purpose
• This is an open-source training toolkit for fixed-point quantized deep learning classification models (Lenet, Alexnet, VGG) built on PyTorch and Brevitas.
• Its primary function is to launch training/evaluation runs from the command line, manage checkpoints, and leverage Brevitas quantization modules.
File Structure Analysis
• Top-level: README.md, training_scripts/
• training_scripts/
• run.py – CLI entrypoint and orchestration script
• models/ – network definitions (Lenet, Alexnet, VGG) using Brevitas
• checkpoint/ – stored .pth files for pretrained weights
Key Insights from README
• External dependencies: PyTorch (≥1.1.0), Brevitas
• Usage: python run.py --network <ModelName> --dataset <Dataset> [--resume] [--evaluate]
• Supports training & evaluation on MNIST and CIFAR10, with weight and activation quantization
a. Main Components
• User/CLI Layer
• run.py (Orchestrator)
• Data Loader (PyTorch datasets: MNIST, CIFAR10)
• Model Definitions (Brevitas‐wrapped network modules)
• Training Engine (Loss, Optimizer, Backpropagation loop)
• Checkpoint Manager (save/load .pth files)
• External Libraries (PyTorch, Brevitas)
b. Relationships & Interactions
• The user issues a CLI command → run.py parses arguments →
• Data Loader loads batches →
• Model Definition is instantiated (with fixed-point quantization layers) →
• Training Engine consumes batches + model → forward/backward →
• Checkpoint Manager writes weights →
• Optional resume/evaluate hooks feed back into the engine
• All core operations call into PyTorch and Brevitas APIs
c. Architectural Patterns
• Modular separation: orchestration (run.py) vs. model definitions vs. data handling
• Plugin‐style extensibility: new models or datasets can be added under models/ or by extending run.py
• Checkpoint abstraction to decouple training logic from storage
d. Key Technologies
• PyTorch: tensor ops, nn.Module, DataLoader, optim
• Brevitas: QuantTensor, QuantConv, QuantLinear
• Python CLI parsing (argparse) in run.py