How aifoundry-org/mhaq works
Project Type and Purpose
• This is an open‐source, Python-based quantization framework (“MHAQ”) built on PyTorch Lightning.
• It provides a modular environment for defining, integrating, and benchmarking custom quantization methods on vision (and other) models.
• Key use cases: model composition, quantization (QAT), training, logging, and visualization.
High-Level File Structure Analysis
– Top-level directories:
• config: YAML files defining quantization and training pipelines
• data: placeholder and pre-trained model files
• scripts: CLI or example scripts to launch specific quant workflows
• src: main framework code, subdivided by function (aux, callbacks, config loader, data, loggers, models, quantization, training)
• tests: performance tests for quant implementations
– Entry points: scripts/rniq_q_config.py, scripts/training/resnet_train.py
– Config/schema: src/config/config_loader.py + config_schema.py
– Training orchestration: src/training/trainer.py (wraps PyTorch Lightning Trainer)
– Quantization engine: src/quantization/quantizer.py delegates to different backends (rniq, dummy)
– Callbacks & loggers: src/callbacks/*.py and src/loggers/*.py for experiment control and metric tracking
Insights from README
• Framework leverages PyTorch Lightning for distributed training, checkpoints, metrics.
• Workflow: load YAML config → compose model → wrap in Quantizer → prepare datamodule → training fit → evaluate.
• Customization points: add new quant methods under src/quantization, extend callbacks, inject new data pipelines under src/data.
a. Main Components
1. User (CLI/API)
2. Config Loader
3. Model Composer
4. Quantizer (pluggable backends: RNIQ, Dummy)
5. Data Module (vision pipelines, e.g., CIFAR, ImageNet, DALI variants)
6. Trainer (PyTorch Lightning wrapper)
7. Callbacks & Loggers (early stopping, LR scheduler, WandB, visualizations)
8. Persistence (saved weights/models under data/models)
9. Tests & Scripts (validation and benchmarking in tests/ and scripts/)
b. Relationships & Data Flow
• User invokes a script → config loader reads YAML → config object
• Composer uses config → builds base model architecture
• Quantizer takes base model + config → produces quantized model (in-place or copy)
• Data Module reads config → sets up dataset transforms, loaders
• Trainer receives quantized model & data module → runs fit()
• During fit: callbacks adjust LR, temperature, noise ratio; loggers send metrics to console/WandB
• On complete: trainer checkpoints model → saved to data/models
• Tests import quantizer/trainer to run perf benchmarks
c. Architectural Patterns & Principles
• Modular plugin architecture (strategy pattern for quant backends, data pipelines, callbacks)
• Single responsibility: each subpackage (aux, quantization, data, models) focuses on a concern
• Inversion of control via central config object injected into all components
• Layered design: CLI → Configuration → Composition → Quantization → Training → Persistence
d. Key Technologies & Frameworks
• Python 3.x, PyTorch, PyTorch Lightning
• YAML for configs
• DALI (for high-performance data loading)
• WandB (external logging service)
• Standard Python packaging, pytest for tests