How rohanasthana/dinas works
Project Type & Purpose
• This is a research-oriented machine-learning framework (not a web app), focusing on neural architecture search (NAS) via graph diffusion.
• It orchestrates dataset loading, conditional diffusion-based generator models, training loops, metrics computation, and analysis/visualization.
• It also bundles several NAS benchmarks (NASBench101/201/301/NLP/HW), each accessed through dedicated APIs.
Core Components
A. Configuration Manager
– Location: configs/ + src/Settings.py
– Role: Load YAML config files (dataset settings, experiment settings, model/training hyperparameters) and expose them to the rest of the system.
B. Dataset Module
– Location: src/datasets/ + subfolders for each benchmark
– Role: AbstractDataset → specific dataset classes (nasbench101_dataset, guacamol_dataset, qm9_dataset, etc.)
– Interactions: Fetch raw data via API wrappers (e.g., src/datasets/nasbench/api.py) or external libraries.
C. Graph Diffusion Generator
– Location: src/diffusion_model.py, src/diffusion_model_discrete*.py, src/Generator.py
– Role: Implement the core diffusion process—discrete or continuous—conditioned on accuracy/latency.
– Dependencies: src/diffusion/layers.py, noise_schedule.py, distributions.py, extra_features.py, etc.
D. Model Definitions
– Location: src/models/transformer_*.py, src/models/layers.py
– Role: Define architecture of the transformer-based denoiser or regressor used during diffusion.
E. Training Orchestrator (CLI)
– Location: main_reg_free.py and src/main.py
– Role: Parse CLI arguments, invoke Settings, instantiate dataset + model + training loop, checkpoint saving.
F. Metrics & Evaluation
– Location: src/metrics/ + train_metrics.py
– Role: AbstractMetrics → specific metrics (molecular_metrics, nasbench201metrics, etc.). Computes quality of generated architectures.
G. Analysis & Visualization
– Location: src/analysis/ + visualization.py, spectre_utils.py, rdkit_functions*.py
– Role: Post-hoc analysis of generated samples, plotting, chemical graph operations (via RDKit), ORCA binding for quantum metrics.
H. External NASBenchX11 Submodule
– Location: src/nasbenchx11/… (a full subproject)
– Role: Alternate NAS API with ensemble surrogate models, own dataset loaders, encodings, predictors, and optimizers.
I. Utilities & Scripts
– Location: src/utils.py, src/job_script.sh, generate_samples.py
– Role: Helper functions, sample generation script, HPC job wrappers.
Component Interactions & Data Flow
– The user invokes the CLI (main_reg_free.py or main.py) with flags (e.g., --dataset nasbench101).
→ CLI loads config via Settings.
→ Settings selects a Dataset class and constructs it.
→ Dataset module either loads from local cache or queries the NASBench API subpackages.
→ CLI then constructs the Graph Diffusion Generator, injecting model hyperparameters from config.
→ The Training Loop kicks off:
• For each epoch/step: apply forward/backward passes through diffusion_model → compute loss → update weights.
• Periodically save checkpoints to disk.
• After training, or concurrently, generate new architectures via Generator.sample().
→ Generated architectures flow into Metrics module for evaluation (accuracy, latency, chemical properties, etc.).
→ Analysis module ingests metrics + samples to produce visualizations, graphs, or quantum-level metrics (via ORCA).
→ Final trained models and metrics are stored locally or uploaded to Google Drive (pre-trained models).
Architectural Patterns & Principles
– Modular separation by concern (Configs, Datasets, Models, Training, Metrics, Analysis).
– Factory pattern in Dataset selection (abstract base → concrete classes).
– Plug-in style for diffusion variants (continuous vs. discrete vs. reg-free).
– Clear layering:
• Top layer: CLI & Experiment scripts
• Middle layer: Orchestrator & Models
• Bottom layer: Data access & Utilities
Technologies & Frameworks
– Language: Python 3.x
– Config management: YAML + custom Settings loader
– ML libraries: PyTorch (implied by model code), RDKit for molecular graphs, ORCA for quantum metrics
– Environments: Conda (environment.yml)
– External APIs: NASBench101/201/301 libraries, GuacaMol benchmark, QM9 dataset, etc.
Final Checklist
– Have you named each component clearly?
– Are any key interactions (e.g., model → metrics → analysis) missing?